2 * Copyright (C) 2011 EfficiOS Inc.
3 * Copyright (C) 2011 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
4 * Copyright (C) 2012 David Goulet <dgoulet@efficios.com>
6 * SPDX-License-Identifier: GPL-2.0-only
18 #include <sys/socket.h>
19 #include <sys/types.h>
22 #include <bin/lttng-consumerd/health-consumerd.hpp>
23 #include <common/align.hpp>
24 #include <common/common.hpp>
25 #include <common/compat/endian.hpp>
26 #include <common/compat/poll.hpp>
27 #include <common/consumer/consumer-metadata-cache.hpp>
28 #include <common/consumer/consumer-stream.hpp>
29 #include <common/consumer/consumer-testpoint.hpp>
30 #include <common/consumer/consumer-timer.hpp>
31 #include <common/consumer/consumer.hpp>
32 #include <common/dynamic-array.hpp>
33 #include <common/index/ctf-index.hpp>
34 #include <common/index/index.hpp>
35 #include <common/kernel-consumer/kernel-consumer.hpp>
36 #include <common/kernel-ctl/kernel-ctl.hpp>
37 #include <common/relayd/relayd.hpp>
38 #include <common/sessiond-comm/relayd.hpp>
39 #include <common/sessiond-comm/sessiond-comm.hpp>
40 #include <common/string-utils/format.hpp>
41 #include <common/time.hpp>
42 #include <common/trace-chunk-registry.hpp>
43 #include <common/trace-chunk.hpp>
44 #include <common/ust-consumer/ust-consumer.hpp>
45 #include <common/utils.hpp>
47 lttng_consumer_global_data the_consumer_data
;
49 enum consumer_channel_action
{
52 CONSUMER_CHANNEL_QUIT
,
56 struct consumer_channel_msg
{
57 enum consumer_channel_action action
;
58 struct lttng_consumer_channel
*chan
; /* add */
59 uint64_t key
; /* del */
63 * Global hash table containing respectively metadata and data streams. The
64 * stream element in this ht should only be updated by the metadata poll thread
65 * for the metadata and the data poll thread for the data.
67 struct lttng_ht
*metadata_ht
;
68 struct lttng_ht
*data_ht
;
71 /* Flag used to temporarily pause data consumption from testpoints. */
72 int data_consumption_paused
;
75 * Flag to inform the polling thread to quit when all fd hung up. Updated by
76 * the consumer_thread_receive_fds when it notices that all fds has hung up.
77 * Also updated by the signal handler (consumer_should_exit()). Read by the
82 static const char *get_consumer_domain(void)
84 switch (the_consumer_data
.type
) {
85 case LTTNG_CONSUMER_KERNEL
:
86 return DEFAULT_KERNEL_TRACE_DIR
;
87 case LTTNG_CONSUMER64_UST
:
89 case LTTNG_CONSUMER32_UST
:
90 return DEFAULT_UST_TRACE_DIR
;
97 * Notify a thread lttng pipe to poll back again. This usually means that some
98 * global state has changed so we just send back the thread in a poll wait
101 static void notify_thread_lttng_pipe(struct lttng_pipe
*pipe
)
103 struct lttng_consumer_stream
*null_stream
= NULL
;
107 (void) lttng_pipe_write(pipe
, &null_stream
, sizeof(null_stream
));
110 static void notify_health_quit_pipe(int *pipe
)
114 ret
= lttng_write(pipe
[1], "4", 1);
116 PERROR("write consumer health quit");
120 static void notify_channel_pipe(struct lttng_consumer_local_data
*ctx
,
121 struct lttng_consumer_channel
*chan
,
123 enum consumer_channel_action action
)
125 struct consumer_channel_msg msg
;
128 memset(&msg
, 0, sizeof(msg
));
133 ret
= lttng_write(ctx
->consumer_channel_pipe
[1], &msg
, sizeof(msg
));
134 if (ret
< sizeof(msg
)) {
135 PERROR("notify_channel_pipe write error");
139 void notify_thread_del_channel(struct lttng_consumer_local_data
*ctx
,
142 notify_channel_pipe(ctx
, NULL
, key
, CONSUMER_CHANNEL_DEL
);
145 static int read_channel_pipe(struct lttng_consumer_local_data
*ctx
,
146 struct lttng_consumer_channel
**chan
,
148 enum consumer_channel_action
*action
)
150 struct consumer_channel_msg msg
;
153 ret
= lttng_read(ctx
->consumer_channel_pipe
[0], &msg
, sizeof(msg
));
154 if (ret
< sizeof(msg
)) {
158 *action
= msg
.action
;
166 * Cleanup the stream list of a channel. Those streams are not yet globally
169 static void clean_channel_stream_list(struct lttng_consumer_channel
*channel
)
171 struct lttng_consumer_stream
*stream
, *stmp
;
173 LTTNG_ASSERT(channel
);
175 /* Delete streams that might have been left in the stream list. */
176 cds_list_for_each_entry_safe(stream
, stmp
, &channel
->streams
.head
,
179 * Once a stream is added to this list, the buffers were created so we
180 * have a guarantee that this call will succeed. Setting the monitor
181 * mode to 0 so we don't lock nor try to delete the stream from the
185 consumer_stream_destroy(stream
, NULL
);
190 * Find a stream. The consumer_data.lock must be locked during this
193 static struct lttng_consumer_stream
*find_stream(uint64_t key
,
196 struct lttng_ht_iter iter
;
197 struct lttng_ht_node_u64
*node
;
198 struct lttng_consumer_stream
*stream
= NULL
;
202 /* -1ULL keys are lookup failures */
203 if (key
== (uint64_t) -1ULL) {
209 lttng_ht_lookup(ht
, &key
, &iter
);
210 node
= lttng_ht_iter_get_node_u64(&iter
);
212 stream
= lttng::utils::container_of(node
, <tng_consumer_stream::node
);
220 static void steal_stream_key(uint64_t key
, struct lttng_ht
*ht
)
222 struct lttng_consumer_stream
*stream
;
225 stream
= find_stream(key
, ht
);
227 stream
->key
= (uint64_t) -1ULL;
229 * We don't want the lookup to match, but we still need
230 * to iterate on this stream when iterating over the hash table. Just
231 * change the node key.
233 stream
->node
.key
= (uint64_t) -1ULL;
239 * Return a channel object for the given key.
241 * RCU read side lock MUST be acquired before calling this function and
242 * protects the channel ptr.
244 struct lttng_consumer_channel
*consumer_find_channel(uint64_t key
)
246 struct lttng_ht_iter iter
;
247 struct lttng_ht_node_u64
*node
;
248 struct lttng_consumer_channel
*channel
= NULL
;
250 ASSERT_RCU_READ_LOCKED();
252 /* -1ULL keys are lookup failures */
253 if (key
== (uint64_t) -1ULL) {
257 lttng_ht_lookup(the_consumer_data
.channel_ht
, &key
, &iter
);
258 node
= lttng_ht_iter_get_node_u64(&iter
);
260 channel
= lttng::utils::container_of(node
, <tng_consumer_channel::node
);
267 * There is a possibility that the consumer does not have enough time between
268 * the close of the channel on the session daemon and the cleanup in here thus
269 * once we have a channel add with an existing key, we know for sure that this
270 * channel will eventually get cleaned up by all streams being closed.
272 * This function just nullifies the already existing channel key.
274 static void steal_channel_key(uint64_t key
)
276 struct lttng_consumer_channel
*channel
;
279 channel
= consumer_find_channel(key
);
281 channel
->key
= (uint64_t) -1ULL;
283 * We don't want the lookup to match, but we still need to iterate on
284 * this channel when iterating over the hash table. Just change the
287 channel
->node
.key
= (uint64_t) -1ULL;
292 static void free_channel_rcu(struct rcu_head
*head
)
294 struct lttng_ht_node_u64
*node
=
295 lttng::utils::container_of(head
, <tng_ht_node_u64::head
);
296 struct lttng_consumer_channel
*channel
=
297 lttng::utils::container_of(node
, <tng_consumer_channel::node
);
299 switch (the_consumer_data
.type
) {
300 case LTTNG_CONSUMER_KERNEL
:
302 case LTTNG_CONSUMER32_UST
:
303 case LTTNG_CONSUMER64_UST
:
304 lttng_ustconsumer_free_channel(channel
);
307 ERR("Unknown consumer_data type");
314 * RCU protected relayd socket pair free.
316 static void free_relayd_rcu(struct rcu_head
*head
)
318 struct lttng_ht_node_u64
*node
=
319 lttng::utils::container_of(head
, <tng_ht_node_u64::head
);
320 struct consumer_relayd_sock_pair
*relayd
=
321 lttng::utils::container_of(node
, &consumer_relayd_sock_pair::node
);
324 * Close all sockets. This is done in the call RCU since we don't want the
325 * socket fds to be reassigned thus potentially creating bad state of the
328 * We do not have to lock the control socket mutex here since at this stage
329 * there is no one referencing to this relayd object.
331 (void) relayd_close(&relayd
->control_sock
);
332 (void) relayd_close(&relayd
->data_sock
);
334 pthread_mutex_destroy(&relayd
->ctrl_sock_mutex
);
339 * Destroy and free relayd socket pair object.
341 void consumer_destroy_relayd(struct consumer_relayd_sock_pair
*relayd
)
344 struct lttng_ht_iter iter
;
346 if (relayd
== NULL
) {
350 DBG("Consumer destroy and close relayd socket pair");
352 iter
.iter
.node
= &relayd
->node
.node
;
353 ret
= lttng_ht_del(the_consumer_data
.relayd_ht
, &iter
);
355 /* We assume the relayd is being or is destroyed */
359 /* RCU free() call */
360 call_rcu(&relayd
->node
.head
, free_relayd_rcu
);
364 * Remove a channel from the global list protected by a mutex. This function is
365 * also responsible for freeing its data structures.
367 void consumer_del_channel(struct lttng_consumer_channel
*channel
)
369 struct lttng_ht_iter iter
;
371 DBG("Consumer delete channel key %" PRIu64
, channel
->key
);
373 pthread_mutex_lock(&the_consumer_data
.lock
);
374 pthread_mutex_lock(&channel
->lock
);
376 /* Destroy streams that might have been left in the stream list. */
377 clean_channel_stream_list(channel
);
379 if (channel
->live_timer_enabled
== 1) {
380 consumer_timer_live_stop(channel
);
382 if (channel
->monitor_timer_enabled
== 1) {
383 consumer_timer_monitor_stop(channel
);
386 switch (the_consumer_data
.type
) {
387 case LTTNG_CONSUMER_KERNEL
:
389 case LTTNG_CONSUMER32_UST
:
390 case LTTNG_CONSUMER64_UST
:
391 lttng_ustconsumer_del_channel(channel
);
394 ERR("Unknown consumer_data type");
399 lttng_trace_chunk_put(channel
->trace_chunk
);
400 channel
->trace_chunk
= NULL
;
402 if (channel
->is_published
) {
406 iter
.iter
.node
= &channel
->node
.node
;
407 ret
= lttng_ht_del(the_consumer_data
.channel_ht
, &iter
);
410 iter
.iter
.node
= &channel
->channels_by_session_id_ht_node
.node
;
411 ret
= lttng_ht_del(the_consumer_data
.channels_by_session_id_ht
,
417 channel
->is_deleted
= true;
418 call_rcu(&channel
->node
.head
, free_channel_rcu
);
420 pthread_mutex_unlock(&channel
->lock
);
421 pthread_mutex_unlock(&the_consumer_data
.lock
);
425 * Iterate over the relayd hash table and destroy each element. Finally,
426 * destroy the whole hash table.
428 static void cleanup_relayd_ht(void)
430 struct lttng_ht_iter iter
;
431 struct consumer_relayd_sock_pair
*relayd
;
435 cds_lfht_for_each_entry(the_consumer_data
.relayd_ht
->ht
, &iter
.iter
,
437 consumer_destroy_relayd(relayd
);
442 lttng_ht_destroy(the_consumer_data
.relayd_ht
);
446 * Update the end point status of all streams having the given network sequence
447 * index (relayd index).
449 * It's atomically set without having the stream mutex locked which is fine
450 * because we handle the write/read race with a pipe wakeup for each thread.
452 static void update_endpoint_status_by_netidx(uint64_t net_seq_idx
,
453 enum consumer_endpoint_status status
)
455 struct lttng_ht_iter iter
;
456 struct lttng_consumer_stream
*stream
;
458 DBG("Consumer set delete flag on stream by idx %" PRIu64
, net_seq_idx
);
462 /* Let's begin with metadata */
463 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
464 if (stream
->net_seq_idx
== net_seq_idx
) {
465 uatomic_set(&stream
->endpoint_status
, status
);
466 DBG("Delete flag set to metadata stream %d", stream
->wait_fd
);
470 /* Follow up by the data streams */
471 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
472 if (stream
->net_seq_idx
== net_seq_idx
) {
473 uatomic_set(&stream
->endpoint_status
, status
);
474 DBG("Delete flag set to data stream %d", stream
->wait_fd
);
481 * Cleanup a relayd object by flagging every associated streams for deletion,
482 * destroying the object meaning removing it from the relayd hash table,
483 * closing the sockets and freeing the memory in a RCU call.
485 * If a local data context is available, notify the threads that the streams'
486 * state have changed.
488 void lttng_consumer_cleanup_relayd(struct consumer_relayd_sock_pair
*relayd
)
492 LTTNG_ASSERT(relayd
);
494 DBG("Cleaning up relayd object ID %" PRIu64
, relayd
->net_seq_idx
);
496 /* Save the net sequence index before destroying the object */
497 netidx
= relayd
->net_seq_idx
;
500 * Delete the relayd from the relayd hash table, close the sockets and free
501 * the object in a RCU call.
503 consumer_destroy_relayd(relayd
);
505 /* Set inactive endpoint to all streams */
506 update_endpoint_status_by_netidx(netidx
, CONSUMER_ENDPOINT_INACTIVE
);
509 * With a local data context, notify the threads that the streams' state
510 * have changed. The write() action on the pipe acts as an "implicit"
511 * memory barrier ordering the updates of the end point status from the
512 * read of this status which happens AFTER receiving this notify.
514 notify_thread_lttng_pipe(relayd
->ctx
->consumer_data_pipe
);
515 notify_thread_lttng_pipe(relayd
->ctx
->consumer_metadata_pipe
);
519 * Flag a relayd socket pair for destruction. Destroy it if the refcount
522 * RCU read side lock MUST be aquired before calling this function.
524 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair
*relayd
)
526 LTTNG_ASSERT(relayd
);
527 ASSERT_RCU_READ_LOCKED();
529 /* Set destroy flag for this object */
530 uatomic_set(&relayd
->destroy_flag
, 1);
532 /* Destroy the relayd if refcount is 0 */
533 if (uatomic_read(&relayd
->refcount
) == 0) {
534 consumer_destroy_relayd(relayd
);
539 * Completly destroy stream from every visiable data structure and the given
542 * One this call returns, the stream object is not longer usable nor visible.
544 void consumer_del_stream(struct lttng_consumer_stream
*stream
,
547 consumer_stream_destroy(stream
, ht
);
551 * XXX naming of del vs destroy is all mixed up.
553 void consumer_del_stream_for_data(struct lttng_consumer_stream
*stream
)
555 consumer_stream_destroy(stream
, data_ht
);
558 void consumer_del_stream_for_metadata(struct lttng_consumer_stream
*stream
)
560 consumer_stream_destroy(stream
, metadata_ht
);
563 void consumer_stream_update_channel_attributes(
564 struct lttng_consumer_stream
*stream
,
565 struct lttng_consumer_channel
*channel
)
567 stream
->channel_read_only_attributes
.tracefile_size
=
568 channel
->tracefile_size
;
572 * Add a stream to the global list protected by a mutex.
574 void consumer_add_data_stream(struct lttng_consumer_stream
*stream
)
576 struct lttng_ht
*ht
= data_ht
;
578 LTTNG_ASSERT(stream
);
581 DBG3("Adding consumer stream %" PRIu64
, stream
->key
);
583 pthread_mutex_lock(&the_consumer_data
.lock
);
584 pthread_mutex_lock(&stream
->chan
->lock
);
585 pthread_mutex_lock(&stream
->chan
->timer_lock
);
586 pthread_mutex_lock(&stream
->lock
);
589 /* Steal stream identifier to avoid having streams with the same key */
590 steal_stream_key(stream
->key
, ht
);
592 lttng_ht_add_unique_u64(ht
, &stream
->node
);
594 lttng_ht_add_u64(the_consumer_data
.stream_per_chan_id_ht
,
595 &stream
->node_channel_id
);
598 * Add stream to the stream_list_ht of the consumer data. No need to steal
599 * the key since the HT does not use it and we allow to add redundant keys
602 lttng_ht_add_u64(the_consumer_data
.stream_list_ht
,
603 &stream
->node_session_id
);
606 * When nb_init_stream_left reaches 0, we don't need to trigger any action
607 * in terms of destroying the associated channel, because the action that
608 * causes the count to become 0 also causes a stream to be added. The
609 * channel deletion will thus be triggered by the following removal of this
612 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
613 /* Increment refcount before decrementing nb_init_stream_left */
615 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
618 /* Update consumer data once the node is inserted. */
619 the_consumer_data
.stream_count
++;
620 the_consumer_data
.need_update
= 1;
623 pthread_mutex_unlock(&stream
->lock
);
624 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
625 pthread_mutex_unlock(&stream
->chan
->lock
);
626 pthread_mutex_unlock(&the_consumer_data
.lock
);
630 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
631 * be acquired before calling this.
633 static int add_relayd(struct consumer_relayd_sock_pair
*relayd
)
636 struct lttng_ht_node_u64
*node
;
637 struct lttng_ht_iter iter
;
639 LTTNG_ASSERT(relayd
);
640 ASSERT_RCU_READ_LOCKED();
642 lttng_ht_lookup(the_consumer_data
.relayd_ht
, &relayd
->net_seq_idx
,
644 node
= lttng_ht_iter_get_node_u64(&iter
);
648 lttng_ht_add_unique_u64(the_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
<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 ASSERT_RCU_READ_LOCKED();
700 /* Negative keys are lookup failures */
701 if (key
== (uint64_t) -1ULL) {
705 lttng_ht_lookup(the_consumer_data
.relayd_ht
, &key
, &iter
);
706 node
= lttng_ht_iter_get_node_u64(&iter
);
708 relayd
= lttng::utils::container_of(node
, &consumer_relayd_sock_pair::node
);
716 * Find a relayd and send the stream
718 * Returns 0 on success, < 0 on error
720 int consumer_send_relayd_stream(struct lttng_consumer_stream
*stream
,
724 struct consumer_relayd_sock_pair
*relayd
;
726 LTTNG_ASSERT(stream
);
727 LTTNG_ASSERT(stream
->net_seq_idx
!= -1ULL);
730 /* The stream is not metadata. Get relayd reference if exists. */
732 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
733 if (relayd
!= NULL
) {
734 /* Add stream on the relayd */
735 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
736 ret
= relayd_add_stream(&relayd
->control_sock
, stream
->name
,
737 get_consumer_domain(), path
, &stream
->relayd_stream_id
,
738 stream
->chan
->tracefile_size
,
739 stream
->chan
->tracefile_count
,
740 stream
->trace_chunk
);
741 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
743 ERR("Relayd add stream failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
744 lttng_consumer_cleanup_relayd(relayd
);
748 uatomic_inc(&relayd
->refcount
);
749 stream
->sent_to_relayd
= 1;
751 ERR("Stream %" PRIu64
" relayd ID %" PRIu64
" unknown. Can't send it.",
752 stream
->key
, stream
->net_seq_idx
);
757 DBG("Stream %s with key %" PRIu64
" sent to relayd id %" PRIu64
,
758 stream
->name
, stream
->key
, stream
->net_seq_idx
);
766 * Find a relayd and send the streams sent message
768 * Returns 0 on success, < 0 on error
770 int consumer_send_relayd_streams_sent(uint64_t net_seq_idx
)
773 struct consumer_relayd_sock_pair
*relayd
;
775 LTTNG_ASSERT(net_seq_idx
!= -1ULL);
777 /* The stream is not metadata. Get relayd reference if exists. */
779 relayd
= consumer_find_relayd(net_seq_idx
);
780 if (relayd
!= NULL
) {
781 /* Add stream on the relayd */
782 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
783 ret
= relayd_streams_sent(&relayd
->control_sock
);
784 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
786 ERR("Relayd streams sent failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
787 lttng_consumer_cleanup_relayd(relayd
);
791 ERR("Relayd ID %" PRIu64
" unknown. Can't send streams_sent.",
798 DBG("All streams sent relayd id %" PRIu64
, net_seq_idx
);
806 * Find a relayd and close the stream
808 void close_relayd_stream(struct lttng_consumer_stream
*stream
)
810 struct consumer_relayd_sock_pair
*relayd
;
812 /* The stream is not metadata. Get relayd reference if exists. */
814 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
816 consumer_stream_relayd_close(stream
, relayd
);
822 * Handle stream for relayd transmission if the stream applies for network
823 * streaming where the net sequence index is set.
825 * Return destination file descriptor or negative value on error.
827 static int write_relayd_stream_header(struct lttng_consumer_stream
*stream
,
828 size_t data_size
, unsigned long padding
,
829 struct consumer_relayd_sock_pair
*relayd
)
832 struct lttcomm_relayd_data_hdr data_hdr
;
835 LTTNG_ASSERT(stream
);
836 LTTNG_ASSERT(relayd
);
838 /* Reset data header */
839 memset(&data_hdr
, 0, sizeof(data_hdr
));
841 if (stream
->metadata_flag
) {
842 /* Caller MUST acquire the relayd control socket lock */
843 ret
= relayd_send_metadata(&relayd
->control_sock
, data_size
);
848 /* Metadata are always sent on the control socket. */
849 outfd
= relayd
->control_sock
.sock
.fd
;
851 /* Set header with stream information */
852 data_hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
853 data_hdr
.data_size
= htobe32(data_size
);
854 data_hdr
.padding_size
= htobe32(padding
);
857 * Note that net_seq_num below is assigned with the *current* value of
858 * next_net_seq_num and only after that the next_net_seq_num will be
859 * increment. This is why when issuing a command on the relayd using
860 * this next value, 1 should always be substracted in order to compare
861 * the last seen sequence number on the relayd side to the last sent.
863 data_hdr
.net_seq_num
= htobe64(stream
->next_net_seq_num
);
864 /* Other fields are zeroed previously */
866 ret
= relayd_send_data_hdr(&relayd
->data_sock
, &data_hdr
,
872 ++stream
->next_net_seq_num
;
874 /* Set to go on data socket */
875 outfd
= relayd
->data_sock
.sock
.fd
;
883 * Write a character on the metadata poll pipe to wake the metadata thread.
884 * Returns 0 on success, -1 on error.
886 int consumer_metadata_wakeup_pipe(const struct lttng_consumer_channel
*channel
)
890 DBG("Waking up metadata poll thread (writing to pipe): channel name = '%s'",
892 if (channel
->monitor
&& channel
->metadata_stream
) {
893 const char dummy
= 'c';
894 const ssize_t write_ret
= lttng_write(
895 channel
->metadata_stream
->ust_metadata_poll_pipe
[1],
899 if (errno
== EWOULDBLOCK
) {
901 * This is fine, the metadata poll thread
902 * is having a hard time keeping-up, but
903 * it will eventually wake-up and consume
904 * the available data.
908 PERROR("Failed to write to UST metadata pipe while attempting to wake-up the metadata poll thread");
920 * Trigger a dump of the metadata content. Following/during the succesful
921 * completion of this call, the metadata poll thread will start receiving
922 * metadata packets to consume.
924 * The caller must hold the channel and stream locks.
927 int consumer_metadata_stream_dump(struct lttng_consumer_stream
*stream
)
931 ASSERT_LOCKED(stream
->chan
->lock
);
932 ASSERT_LOCKED(stream
->lock
);
933 LTTNG_ASSERT(stream
->metadata_flag
);
934 LTTNG_ASSERT(stream
->chan
->trace_chunk
);
936 switch (the_consumer_data
.type
) {
937 case LTTNG_CONSUMER_KERNEL
:
939 * Reset the position of what has been read from the
940 * metadata cache to 0 so we can dump it again.
942 ret
= kernctl_metadata_cache_dump(stream
->wait_fd
);
944 case LTTNG_CONSUMER32_UST
:
945 case LTTNG_CONSUMER64_UST
:
947 * Reset the position pushed from the metadata cache so it
948 * will write from the beginning on the next push.
950 stream
->ust_metadata_pushed
= 0;
951 ret
= consumer_metadata_wakeup_pipe(stream
->chan
);
954 ERR("Unknown consumer_data type");
958 ERR("Failed to dump the metadata cache");
964 int lttng_consumer_channel_set_trace_chunk(
965 struct lttng_consumer_channel
*channel
,
966 struct lttng_trace_chunk
*new_trace_chunk
)
968 pthread_mutex_lock(&channel
->lock
);
969 if (channel
->is_deleted
) {
971 * The channel has been logically deleted and should no longer
972 * be used. It has released its reference to its current trace
973 * chunk and should not acquire a new one.
975 * Return success as there is nothing for the caller to do.
981 * The acquisition of the reference cannot fail (barring
982 * a severe internal error) since a reference to the published
983 * chunk is already held by the caller.
985 if (new_trace_chunk
) {
986 const bool acquired_reference
= lttng_trace_chunk_get(
989 LTTNG_ASSERT(acquired_reference
);
992 lttng_trace_chunk_put(channel
->trace_chunk
);
993 channel
->trace_chunk
= new_trace_chunk
;
995 pthread_mutex_unlock(&channel
->lock
);
1000 * Allocate and return a new lttng_consumer_channel object using the given key
1001 * to initialize the hash table node.
1003 * On error, return NULL.
1005 struct lttng_consumer_channel
*consumer_allocate_channel(uint64_t key
,
1006 uint64_t session_id
,
1007 const uint64_t *chunk_id
,
1008 const char *pathname
,
1011 enum lttng_event_output output
,
1012 uint64_t tracefile_size
,
1013 uint64_t tracefile_count
,
1014 uint64_t session_id_per_pid
,
1015 unsigned int monitor
,
1016 unsigned int live_timer_interval
,
1017 bool is_in_live_session
,
1018 const char *root_shm_path
,
1019 const char *shm_path
)
1021 struct lttng_consumer_channel
*channel
= NULL
;
1022 struct lttng_trace_chunk
*trace_chunk
= NULL
;
1025 trace_chunk
= lttng_trace_chunk_registry_find_chunk(
1026 the_consumer_data
.chunk_registry
, session_id
,
1029 ERR("Failed to find trace chunk reference during creation of channel");
1034 channel
= zmalloc
<lttng_consumer_channel
>();
1035 if (channel
== NULL
) {
1036 PERROR("malloc struct lttng_consumer_channel");
1041 channel
->refcount
= 0;
1042 channel
->session_id
= session_id
;
1043 channel
->session_id_per_pid
= session_id_per_pid
;
1044 channel
->relayd_id
= relayd_id
;
1045 channel
->tracefile_size
= tracefile_size
;
1046 channel
->tracefile_count
= tracefile_count
;
1047 channel
->monitor
= monitor
;
1048 channel
->live_timer_interval
= live_timer_interval
;
1049 channel
->is_live
= is_in_live_session
;
1050 pthread_mutex_init(&channel
->lock
, NULL
);
1051 pthread_mutex_init(&channel
->timer_lock
, NULL
);
1054 case LTTNG_EVENT_SPLICE
:
1055 channel
->output
= CONSUMER_CHANNEL_SPLICE
;
1057 case LTTNG_EVENT_MMAP
:
1058 channel
->output
= CONSUMER_CHANNEL_MMAP
;
1068 * In monitor mode, the streams associated with the channel will be put in
1069 * a special list ONLY owned by this channel. So, the refcount is set to 1
1070 * here meaning that the channel itself has streams that are referenced.
1072 * On a channel deletion, once the channel is no longer visible, the
1073 * refcount is decremented and checked for a zero value to delete it. With
1074 * streams in no monitor mode, it will now be safe to destroy the channel.
1076 if (!channel
->monitor
) {
1077 channel
->refcount
= 1;
1080 strncpy(channel
->pathname
, pathname
, sizeof(channel
->pathname
));
1081 channel
->pathname
[sizeof(channel
->pathname
) - 1] = '\0';
1083 strncpy(channel
->name
, name
, sizeof(channel
->name
));
1084 channel
->name
[sizeof(channel
->name
) - 1] = '\0';
1086 if (root_shm_path
) {
1087 strncpy(channel
->root_shm_path
, root_shm_path
, sizeof(channel
->root_shm_path
));
1088 channel
->root_shm_path
[sizeof(channel
->root_shm_path
) - 1] = '\0';
1091 strncpy(channel
->shm_path
, shm_path
, sizeof(channel
->shm_path
));
1092 channel
->shm_path
[sizeof(channel
->shm_path
) - 1] = '\0';
1095 lttng_ht_node_init_u64(&channel
->node
, channel
->key
);
1096 lttng_ht_node_init_u64(&channel
->channels_by_session_id_ht_node
,
1097 channel
->session_id
);
1099 channel
->wait_fd
= -1;
1100 CDS_INIT_LIST_HEAD(&channel
->streams
.head
);
1103 int ret
= lttng_consumer_channel_set_trace_chunk(channel
,
1110 DBG("Allocated channel (key %" PRIu64
")", channel
->key
);
1113 lttng_trace_chunk_put(trace_chunk
);
1116 consumer_del_channel(channel
);
1122 * Add a channel to the global list protected by a mutex.
1124 * Always return 0 indicating success.
1126 int consumer_add_channel(struct lttng_consumer_channel
*channel
,
1127 struct lttng_consumer_local_data
*ctx
)
1129 pthread_mutex_lock(&the_consumer_data
.lock
);
1130 pthread_mutex_lock(&channel
->lock
);
1131 pthread_mutex_lock(&channel
->timer_lock
);
1134 * This gives us a guarantee that the channel we are about to add to the
1135 * channel hash table will be unique. See this function comment on the why
1136 * we need to steel the channel key at this stage.
1138 steal_channel_key(channel
->key
);
1141 lttng_ht_add_unique_u64(the_consumer_data
.channel_ht
, &channel
->node
);
1142 lttng_ht_add_u64(the_consumer_data
.channels_by_session_id_ht
,
1143 &channel
->channels_by_session_id_ht_node
);
1145 channel
->is_published
= true;
1147 pthread_mutex_unlock(&channel
->timer_lock
);
1148 pthread_mutex_unlock(&channel
->lock
);
1149 pthread_mutex_unlock(&the_consumer_data
.lock
);
1151 if (channel
->wait_fd
!= -1 && channel
->type
== CONSUMER_CHANNEL_TYPE_DATA
) {
1152 notify_channel_pipe(ctx
, channel
, -1, CONSUMER_CHANNEL_ADD
);
1159 * Allocate the pollfd structure and the local view of the out fds to avoid
1160 * doing a lookup in the linked list and concurrency issues when writing is
1161 * needed. Called with consumer_data.lock held.
1163 * Returns the number of fds in the structures.
1165 static int update_poll_array(struct lttng_consumer_local_data
*ctx
,
1166 struct pollfd
**pollfd
, struct lttng_consumer_stream
**local_stream
,
1167 struct lttng_ht
*ht
, int *nb_inactive_fd
)
1170 struct lttng_ht_iter iter
;
1171 struct lttng_consumer_stream
*stream
;
1175 LTTNG_ASSERT(pollfd
);
1176 LTTNG_ASSERT(local_stream
);
1178 DBG("Updating poll fd array");
1179 *nb_inactive_fd
= 0;
1181 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1183 * Only active streams with an active end point can be added to the
1184 * poll set and local stream storage of the thread.
1186 * There is a potential race here for endpoint_status to be updated
1187 * just after the check. However, this is OK since the stream(s) will
1188 * be deleted once the thread is notified that the end point state has
1189 * changed where this function will be called back again.
1191 * We track the number of inactive FDs because they still need to be
1192 * closed by the polling thread after a wakeup on the data_pipe or
1195 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_INACTIVE
) {
1196 (*nb_inactive_fd
)++;
1200 * This clobbers way too much the debug output. Uncomment that if you
1201 * need it for debugging purposes.
1203 (*pollfd
)[i
].fd
= stream
->wait_fd
;
1204 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1205 local_stream
[i
] = stream
;
1211 * Insert the consumer_data_pipe at the end of the array and don't
1212 * increment i so nb_fd is the number of real FD.
1214 (*pollfd
)[i
].fd
= lttng_pipe_get_readfd(ctx
->consumer_data_pipe
);
1215 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1217 (*pollfd
)[i
+ 1].fd
= lttng_pipe_get_readfd(ctx
->consumer_wakeup_pipe
);
1218 (*pollfd
)[i
+ 1].events
= POLLIN
| POLLPRI
;
1223 * Poll on the should_quit pipe and the command socket return -1 on
1224 * error, 1 if should exit, 0 if data is available on the command socket
1226 int lttng_consumer_poll_socket(struct pollfd
*consumer_sockpoll
)
1231 num_rdy
= poll(consumer_sockpoll
, 2, -1);
1232 if (num_rdy
== -1) {
1234 * Restart interrupted system call.
1236 if (errno
== EINTR
) {
1239 PERROR("Poll error");
1242 if (consumer_sockpoll
[0].revents
& (POLLIN
| POLLPRI
)) {
1243 DBG("consumer_should_quit wake up");
1250 * Set the error socket.
1252 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data
*ctx
,
1255 ctx
->consumer_error_socket
= sock
;
1259 * Set the command socket path.
1261 void lttng_consumer_set_command_sock_path(
1262 struct lttng_consumer_local_data
*ctx
, char *sock
)
1264 ctx
->consumer_command_sock_path
= sock
;
1268 * Send return code to the session daemon.
1269 * If the socket is not defined, we return 0, it is not a fatal error
1271 int lttng_consumer_send_error(struct lttng_consumer_local_data
*ctx
, int cmd
)
1273 if (ctx
->consumer_error_socket
> 0) {
1274 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
1275 sizeof(enum lttcomm_sessiond_command
));
1282 * Close all the tracefiles and stream fds and MUST be called when all
1283 * instances are destroyed i.e. when all threads were joined and are ended.
1285 void lttng_consumer_cleanup(void)
1287 struct lttng_ht_iter iter
;
1288 struct lttng_consumer_channel
*channel
;
1289 unsigned int trace_chunks_left
;
1293 cds_lfht_for_each_entry(the_consumer_data
.channel_ht
->ht
, &iter
.iter
,
1294 channel
, node
.node
) {
1295 consumer_del_channel(channel
);
1300 lttng_ht_destroy(the_consumer_data
.channel_ht
);
1301 lttng_ht_destroy(the_consumer_data
.channels_by_session_id_ht
);
1303 cleanup_relayd_ht();
1305 lttng_ht_destroy(the_consumer_data
.stream_per_chan_id_ht
);
1308 * This HT contains streams that are freed by either the metadata thread or
1309 * the data thread so we do *nothing* on the hash table and simply destroy
1312 lttng_ht_destroy(the_consumer_data
.stream_list_ht
);
1315 * Trace chunks in the registry may still exist if the session
1316 * daemon has encountered an internal error and could not
1317 * tear down its sessions and/or trace chunks properly.
1319 * Release the session daemon's implicit reference to any remaining
1320 * trace chunk and print an error if any trace chunk was found. Note
1321 * that there are _no_ legitimate cases for trace chunks to be left,
1322 * it is a leak. However, it can happen following a crash of the
1323 * session daemon and not emptying the registry would cause an assertion
1326 trace_chunks_left
= lttng_trace_chunk_registry_put_each_chunk(
1327 the_consumer_data
.chunk_registry
);
1328 if (trace_chunks_left
) {
1329 ERR("%u trace chunks are leaked by lttng-consumerd. "
1330 "This can be caused by an internal error of the session daemon.",
1333 /* Run all callbacks freeing each chunk. */
1335 lttng_trace_chunk_registry_destroy(the_consumer_data
.chunk_registry
);
1339 * Called from signal handler.
1341 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
1345 CMM_STORE_SHARED(consumer_quit
, 1);
1346 ret
= lttng_write(ctx
->consumer_should_quit
[1], "4", 1);
1348 PERROR("write consumer quit");
1351 DBG("Consumer flag that it should quit");
1356 * Flush pending writes to trace output disk file.
1359 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1363 int outfd
= stream
->out_fd
;
1366 * This does a blocking write-and-wait on any page that belongs to the
1367 * subbuffer prior to the one we just wrote.
1368 * Don't care about error values, as these are just hints and ways to
1369 * limit the amount of page cache used.
1371 if (orig_offset
< stream
->max_sb_size
) {
1374 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1375 stream
->max_sb_size
,
1376 SYNC_FILE_RANGE_WAIT_BEFORE
1377 | SYNC_FILE_RANGE_WRITE
1378 | SYNC_FILE_RANGE_WAIT_AFTER
);
1380 * Give hints to the kernel about how we access the file:
1381 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1384 * We need to call fadvise again after the file grows because the
1385 * kernel does not seem to apply fadvise to non-existing parts of the
1388 * Call fadvise _after_ having waited for the page writeback to
1389 * complete because the dirty page writeback semantic is not well
1390 * defined. So it can be expected to lead to lower throughput in
1393 ret
= posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1394 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1395 if (ret
&& ret
!= -ENOSYS
) {
1397 PERROR("posix_fadvise on fd %i", outfd
);
1402 * Initialise the necessary environnement :
1403 * - create a new context
1404 * - create the poll_pipe
1405 * - create the should_quit pipe (for signal handler)
1406 * - create the thread pipe (for splice)
1408 * Takes a function pointer as argument, this function is called when data is
1409 * available on a buffer. This function is responsible to do the
1410 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1411 * buffer configuration and then kernctl_put_next_subbuf at the end.
1413 * Returns a pointer to the new context or NULL on error.
1415 struct lttng_consumer_local_data
*lttng_consumer_create(
1416 enum lttng_consumer_type type
,
1417 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1418 struct lttng_consumer_local_data
*ctx
, bool locked_by_caller
),
1419 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1420 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1421 int (*update_stream
)(uint64_t stream_key
, uint32_t state
))
1424 struct lttng_consumer_local_data
*ctx
;
1426 LTTNG_ASSERT(the_consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1427 the_consumer_data
.type
== type
);
1428 the_consumer_data
.type
= type
;
1430 ctx
= zmalloc
<lttng_consumer_local_data
>();
1432 PERROR("allocating context");
1436 ctx
->consumer_error_socket
= -1;
1437 ctx
->consumer_metadata_socket
= -1;
1438 pthread_mutex_init(&ctx
->metadata_socket_lock
, NULL
);
1439 /* assign the callbacks */
1440 ctx
->on_buffer_ready
= buffer_ready
;
1441 ctx
->on_recv_channel
= recv_channel
;
1442 ctx
->on_recv_stream
= recv_stream
;
1443 ctx
->on_update_stream
= update_stream
;
1445 ctx
->consumer_data_pipe
= lttng_pipe_open(0);
1446 if (!ctx
->consumer_data_pipe
) {
1447 goto error_poll_pipe
;
1450 ctx
->consumer_wakeup_pipe
= lttng_pipe_open(0);
1451 if (!ctx
->consumer_wakeup_pipe
) {
1452 goto error_wakeup_pipe
;
1455 ret
= pipe(ctx
->consumer_should_quit
);
1457 PERROR("Error creating recv pipe");
1458 goto error_quit_pipe
;
1461 ret
= pipe(ctx
->consumer_channel_pipe
);
1463 PERROR("Error creating channel pipe");
1464 goto error_channel_pipe
;
1467 ctx
->consumer_metadata_pipe
= lttng_pipe_open(0);
1468 if (!ctx
->consumer_metadata_pipe
) {
1469 goto error_metadata_pipe
;
1472 ctx
->channel_monitor_pipe
= -1;
1476 error_metadata_pipe
:
1477 utils_close_pipe(ctx
->consumer_channel_pipe
);
1479 utils_close_pipe(ctx
->consumer_should_quit
);
1481 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1483 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1491 * Iterate over all streams of the hashtable and free them properly.
1493 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1495 struct lttng_ht_iter iter
;
1496 struct lttng_consumer_stream
*stream
;
1503 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1505 * Ignore return value since we are currently cleaning up so any error
1508 (void) consumer_del_stream(stream
, ht
);
1512 lttng_ht_destroy(ht
);
1516 * Iterate over all streams of the metadata hashtable and free them
1519 static void destroy_metadata_stream_ht(struct lttng_ht
*ht
)
1521 struct lttng_ht_iter iter
;
1522 struct lttng_consumer_stream
*stream
;
1529 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1531 * Ignore return value since we are currently cleaning up so any error
1534 (void) consumer_del_metadata_stream(stream
, ht
);
1538 lttng_ht_destroy(ht
);
1542 * Close all fds associated with the instance and free the context.
1544 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1548 DBG("Consumer destroying it. Closing everything.");
1554 destroy_data_stream_ht(data_ht
);
1555 destroy_metadata_stream_ht(metadata_ht
);
1557 ret
= close(ctx
->consumer_error_socket
);
1561 ret
= close(ctx
->consumer_metadata_socket
);
1565 utils_close_pipe(ctx
->consumer_channel_pipe
);
1566 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1567 lttng_pipe_destroy(ctx
->consumer_metadata_pipe
);
1568 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1569 utils_close_pipe(ctx
->consumer_should_quit
);
1571 unlink(ctx
->consumer_command_sock_path
);
1576 * Write the metadata stream id on the specified file descriptor.
1578 static int write_relayd_metadata_id(int fd
,
1579 struct lttng_consumer_stream
*stream
,
1580 unsigned long padding
)
1583 struct lttcomm_relayd_metadata_payload hdr
;
1585 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1586 hdr
.padding_size
= htobe32(padding
);
1587 ret
= lttng_write(fd
, (void *) &hdr
, sizeof(hdr
));
1588 if (ret
< sizeof(hdr
)) {
1590 * This error means that the fd's end is closed so ignore the PERROR
1591 * not to clubber the error output since this can happen in a normal
1594 if (errno
!= EPIPE
) {
1595 PERROR("write metadata stream id");
1597 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1599 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1600 * handle writting the missing part so report that as an error and
1601 * don't lie to the caller.
1606 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1607 stream
->relayd_stream_id
, padding
);
1614 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1615 * core function for writing trace buffers to either the local filesystem or
1618 * It must be called with the stream and the channel lock held.
1620 * Careful review MUST be put if any changes occur!
1622 * Returns the number of bytes written
1624 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1625 struct lttng_consumer_stream
*stream
,
1626 const struct lttng_buffer_view
*buffer
,
1627 unsigned long padding
)
1630 off_t orig_offset
= stream
->out_fd_offset
;
1631 /* Default is on the disk */
1632 int outfd
= stream
->out_fd
;
1633 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1634 unsigned int relayd_hang_up
= 0;
1635 const size_t subbuf_content_size
= buffer
->size
- padding
;
1638 /* RCU lock for the relayd pointer */
1640 LTTNG_ASSERT(stream
->net_seq_idx
!= (uint64_t) -1ULL ||
1641 stream
->trace_chunk
);
1643 /* Flag that the current stream if set for network streaming. */
1644 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1645 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1646 if (relayd
== NULL
) {
1652 /* Handle stream on the relayd if the output is on the network */
1654 unsigned long netlen
= subbuf_content_size
;
1657 * Lock the control socket for the complete duration of the function
1658 * since from this point on we will use the socket.
1660 if (stream
->metadata_flag
) {
1661 /* Metadata requires the control socket. */
1662 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1663 if (stream
->reset_metadata_flag
) {
1664 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1665 stream
->relayd_stream_id
,
1666 stream
->metadata_version
);
1671 stream
->reset_metadata_flag
= 0;
1673 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1676 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1681 /* Use the returned socket. */
1684 /* Write metadata stream id before payload */
1685 if (stream
->metadata_flag
) {
1686 ret
= write_relayd_metadata_id(outfd
, stream
, padding
);
1693 write_len
= subbuf_content_size
;
1695 /* No streaming; we have to write the full padding. */
1696 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1697 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1699 ERR("Reset metadata file");
1702 stream
->reset_metadata_flag
= 0;
1706 * Check if we need to change the tracefile before writing the packet.
1708 if (stream
->chan
->tracefile_size
> 0 &&
1709 (stream
->tracefile_size_current
+ buffer
->size
) >
1710 stream
->chan
->tracefile_size
) {
1711 ret
= consumer_stream_rotate_output_files(stream
);
1715 outfd
= stream
->out_fd
;
1718 stream
->tracefile_size_current
+= buffer
->size
;
1719 write_len
= buffer
->size
;
1723 * This call guarantee that len or less is returned. It's impossible to
1724 * receive a ret value that is bigger than len.
1726 ret
= lttng_write(outfd
, buffer
->data
, write_len
);
1727 DBG("Consumer mmap write() ret %zd (len %zu)", ret
, write_len
);
1728 if (ret
< 0 || ((size_t) ret
!= write_len
)) {
1730 * Report error to caller if nothing was written else at least send the
1738 /* Socket operation failed. We consider the relayd dead */
1739 if (errno
== EPIPE
) {
1741 * This is possible if the fd is closed on the other side
1742 * (outfd) or any write problem. It can be verbose a bit for a
1743 * normal execution if for instance the relayd is stopped
1744 * abruptly. This can happen so set this to a DBG statement.
1746 DBG("Consumer mmap write detected relayd hang up");
1748 /* Unhandled error, print it and stop function right now. */
1749 PERROR("Error in write mmap (ret %zd != write_len %zu)", ret
,
1754 stream
->output_written
+= ret
;
1756 /* This call is useless on a socket so better save a syscall. */
1758 /* This won't block, but will start writeout asynchronously */
1759 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, write_len
,
1760 SYNC_FILE_RANGE_WRITE
);
1761 stream
->out_fd_offset
+= write_len
;
1762 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1767 * This is a special case that the relayd has closed its socket. Let's
1768 * cleanup the relayd object and all associated streams.
1770 if (relayd
&& relayd_hang_up
) {
1771 ERR("Relayd hangup. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
1772 lttng_consumer_cleanup_relayd(relayd
);
1776 /* Unlock only if ctrl socket used */
1777 if (relayd
&& stream
->metadata_flag
) {
1778 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1786 * Splice the data from the ring buffer to the tracefile.
1788 * It must be called with the stream lock held.
1790 * Returns the number of bytes spliced.
1792 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1793 struct lttng_consumer_local_data
*ctx
,
1794 struct lttng_consumer_stream
*stream
, unsigned long len
,
1795 unsigned long padding
)
1797 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1799 off_t orig_offset
= stream
->out_fd_offset
;
1800 int fd
= stream
->wait_fd
;
1801 /* Default is on the disk */
1802 int outfd
= stream
->out_fd
;
1803 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1805 unsigned int relayd_hang_up
= 0;
1807 switch (the_consumer_data
.type
) {
1808 case LTTNG_CONSUMER_KERNEL
:
1810 case LTTNG_CONSUMER32_UST
:
1811 case LTTNG_CONSUMER64_UST
:
1812 /* Not supported for user space tracing */
1815 ERR("Unknown consumer_data type");
1819 /* RCU lock for the relayd pointer */
1822 /* Flag that the current stream if set for network streaming. */
1823 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1824 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1825 if (relayd
== NULL
) {
1830 splice_pipe
= stream
->splice_pipe
;
1832 /* Write metadata stream id before payload */
1834 unsigned long total_len
= len
;
1836 if (stream
->metadata_flag
) {
1838 * Lock the control socket for the complete duration of the function
1839 * since from this point on we will use the socket.
1841 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1843 if (stream
->reset_metadata_flag
) {
1844 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1845 stream
->relayd_stream_id
,
1846 stream
->metadata_version
);
1851 stream
->reset_metadata_flag
= 0;
1853 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
,
1861 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1864 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1870 /* Use the returned socket. */
1873 /* No streaming, we have to set the len with the full padding */
1876 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1877 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1879 ERR("Reset metadata file");
1882 stream
->reset_metadata_flag
= 0;
1885 * Check if we need to change the tracefile before writing the packet.
1887 if (stream
->chan
->tracefile_size
> 0 &&
1888 (stream
->tracefile_size_current
+ len
) >
1889 stream
->chan
->tracefile_size
) {
1890 ret
= consumer_stream_rotate_output_files(stream
);
1895 outfd
= stream
->out_fd
;
1898 stream
->tracefile_size_current
+= len
;
1902 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1903 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1904 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1905 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1906 DBG("splice chan to pipe, ret %zd", ret_splice
);
1907 if (ret_splice
< 0) {
1910 PERROR("Error in relay splice");
1914 /* Handle stream on the relayd if the output is on the network */
1915 if (relayd
&& stream
->metadata_flag
) {
1916 size_t metadata_payload_size
=
1917 sizeof(struct lttcomm_relayd_metadata_payload
);
1919 /* Update counter to fit the spliced data */
1920 ret_splice
+= metadata_payload_size
;
1921 len
+= metadata_payload_size
;
1923 * We do this so the return value can match the len passed as
1924 * argument to this function.
1926 written
-= metadata_payload_size
;
1929 /* Splice data out */
1930 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1931 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1932 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
1934 if (ret_splice
< 0) {
1939 } else if (ret_splice
> len
) {
1941 * We don't expect this code path to be executed but you never know
1942 * so this is an extra protection agains a buggy splice().
1945 written
+= ret_splice
;
1946 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice
,
1950 /* All good, update current len and continue. */
1954 /* This call is useless on a socket so better save a syscall. */
1956 /* This won't block, but will start writeout asynchronously */
1957 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1958 SYNC_FILE_RANGE_WRITE
);
1959 stream
->out_fd_offset
+= ret_splice
;
1961 stream
->output_written
+= ret_splice
;
1962 written
+= ret_splice
;
1965 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1971 * This is a special case that the relayd has closed its socket. Let's
1972 * cleanup the relayd object and all associated streams.
1974 if (relayd
&& relayd_hang_up
) {
1975 ERR("Relayd hangup. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
1976 lttng_consumer_cleanup_relayd(relayd
);
1977 /* Skip splice error so the consumer does not fail */
1982 /* send the appropriate error description to sessiond */
1985 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1988 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1991 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
1996 if (relayd
&& stream
->metadata_flag
) {
1997 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
2005 * Sample the snapshot positions for a specific fd
2007 * Returns 0 on success, < 0 on error
2009 int lttng_consumer_sample_snapshot_positions(struct lttng_consumer_stream
*stream
)
2011 switch (the_consumer_data
.type
) {
2012 case LTTNG_CONSUMER_KERNEL
:
2013 return lttng_kconsumer_sample_snapshot_positions(stream
);
2014 case LTTNG_CONSUMER32_UST
:
2015 case LTTNG_CONSUMER64_UST
:
2016 return lttng_ustconsumer_sample_snapshot_positions(stream
);
2018 ERR("Unknown consumer_data type");
2024 * Take a snapshot for a specific fd
2026 * Returns 0 on success, < 0 on error
2028 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
2030 switch (the_consumer_data
.type
) {
2031 case LTTNG_CONSUMER_KERNEL
:
2032 return lttng_kconsumer_take_snapshot(stream
);
2033 case LTTNG_CONSUMER32_UST
:
2034 case LTTNG_CONSUMER64_UST
:
2035 return lttng_ustconsumer_take_snapshot(stream
);
2037 ERR("Unknown consumer_data type");
2044 * Get the produced position
2046 * Returns 0 on success, < 0 on error
2048 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
2051 switch (the_consumer_data
.type
) {
2052 case LTTNG_CONSUMER_KERNEL
:
2053 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
2054 case LTTNG_CONSUMER32_UST
:
2055 case LTTNG_CONSUMER64_UST
:
2056 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
2058 ERR("Unknown consumer_data type");
2065 * Get the consumed position (free-running counter position in bytes).
2067 * Returns 0 on success, < 0 on error
2069 int lttng_consumer_get_consumed_snapshot(struct lttng_consumer_stream
*stream
,
2072 switch (the_consumer_data
.type
) {
2073 case LTTNG_CONSUMER_KERNEL
:
2074 return lttng_kconsumer_get_consumed_snapshot(stream
, pos
);
2075 case LTTNG_CONSUMER32_UST
:
2076 case LTTNG_CONSUMER64_UST
:
2077 return lttng_ustconsumer_get_consumed_snapshot(stream
, pos
);
2079 ERR("Unknown consumer_data type");
2085 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
2086 int sock
, struct pollfd
*consumer_sockpoll
)
2088 switch (the_consumer_data
.type
) {
2089 case LTTNG_CONSUMER_KERNEL
:
2090 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2091 case LTTNG_CONSUMER32_UST
:
2092 case LTTNG_CONSUMER64_UST
:
2093 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2095 ERR("Unknown consumer_data type");
2102 void lttng_consumer_close_all_metadata(void)
2104 switch (the_consumer_data
.type
) {
2105 case LTTNG_CONSUMER_KERNEL
:
2107 * The Kernel consumer has a different metadata scheme so we don't
2108 * close anything because the stream will be closed by the session
2112 case LTTNG_CONSUMER32_UST
:
2113 case LTTNG_CONSUMER64_UST
:
2115 * Close all metadata streams. The metadata hash table is passed and
2116 * this call iterates over it by closing all wakeup fd. This is safe
2117 * because at this point we are sure that the metadata producer is
2118 * either dead or blocked.
2120 lttng_ustconsumer_close_all_metadata(metadata_ht
);
2123 ERR("Unknown consumer_data type");
2129 * Clean up a metadata stream and free its memory.
2131 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
2132 struct lttng_ht
*ht
)
2134 struct lttng_consumer_channel
*channel
= NULL
;
2135 bool free_channel
= false;
2137 LTTNG_ASSERT(stream
);
2139 * This call should NEVER receive regular stream. It must always be
2140 * metadata stream and this is crucial for data structure synchronization.
2142 LTTNG_ASSERT(stream
->metadata_flag
);
2144 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
2146 pthread_mutex_lock(&the_consumer_data
.lock
);
2148 * Note that this assumes that a stream's channel is never changed and
2149 * that the stream's lock doesn't need to be taken to sample its
2152 channel
= stream
->chan
;
2153 pthread_mutex_lock(&channel
->lock
);
2154 pthread_mutex_lock(&stream
->lock
);
2155 if (channel
->metadata_cache
) {
2156 /* Only applicable to userspace consumers. */
2157 pthread_mutex_lock(&channel
->metadata_cache
->lock
);
2160 /* Remove any reference to that stream. */
2161 consumer_stream_delete(stream
, ht
);
2163 /* Close down everything including the relayd if one. */
2164 consumer_stream_close(stream
);
2165 /* Destroy tracer buffers of the stream. */
2166 consumer_stream_destroy_buffers(stream
);
2168 /* Atomically decrement channel refcount since other threads can use it. */
2169 if (!uatomic_sub_return(&channel
->refcount
, 1)
2170 && !uatomic_read(&channel
->nb_init_stream_left
)) {
2171 /* Go for channel deletion! */
2172 free_channel
= true;
2174 stream
->chan
= NULL
;
2177 * Nullify the stream reference so it is not used after deletion. The
2178 * channel lock MUST be acquired before being able to check for a NULL
2181 channel
->metadata_stream
= NULL
;
2183 if (channel
->metadata_cache
) {
2184 pthread_mutex_unlock(&channel
->metadata_cache
->lock
);
2186 pthread_mutex_unlock(&stream
->lock
);
2187 pthread_mutex_unlock(&channel
->lock
);
2188 pthread_mutex_unlock(&the_consumer_data
.lock
);
2191 consumer_del_channel(channel
);
2194 lttng_trace_chunk_put(stream
->trace_chunk
);
2195 stream
->trace_chunk
= NULL
;
2196 consumer_stream_free(stream
);
2200 * Action done with the metadata stream when adding it to the consumer internal
2201 * data structures to handle it.
2203 void consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
2205 struct lttng_ht
*ht
= metadata_ht
;
2206 struct lttng_ht_iter iter
;
2207 struct lttng_ht_node_u64
*node
;
2209 LTTNG_ASSERT(stream
);
2212 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2214 pthread_mutex_lock(&the_consumer_data
.lock
);
2215 pthread_mutex_lock(&stream
->chan
->lock
);
2216 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2217 pthread_mutex_lock(&stream
->lock
);
2220 * From here, refcounts are updated so be _careful_ when returning an error
2227 * Lookup the stream just to make sure it does not exist in our internal
2228 * state. This should NEVER happen.
2230 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2231 node
= lttng_ht_iter_get_node_u64(&iter
);
2232 LTTNG_ASSERT(!node
);
2235 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2236 * in terms of destroying the associated channel, because the action that
2237 * causes the count to become 0 also causes a stream to be added. The
2238 * channel deletion will thus be triggered by the following removal of this
2241 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2242 /* Increment refcount before decrementing nb_init_stream_left */
2244 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2247 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2249 lttng_ht_add_u64(the_consumer_data
.stream_per_chan_id_ht
,
2250 &stream
->node_channel_id
);
2253 * Add stream to the stream_list_ht of the consumer data. No need to steal
2254 * the key since the HT does not use it and we allow to add redundant keys
2257 lttng_ht_add_u64(the_consumer_data
.stream_list_ht
,
2258 &stream
->node_session_id
);
2262 pthread_mutex_unlock(&stream
->lock
);
2263 pthread_mutex_unlock(&stream
->chan
->lock
);
2264 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2265 pthread_mutex_unlock(&the_consumer_data
.lock
);
2269 * Delete data stream that are flagged for deletion (endpoint_status).
2271 static void validate_endpoint_status_data_stream(void)
2273 struct lttng_ht_iter iter
;
2274 struct lttng_consumer_stream
*stream
;
2276 DBG("Consumer delete flagged data stream");
2279 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2280 /* Validate delete flag of the stream */
2281 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2284 /* Delete it right now */
2285 consumer_del_stream(stream
, data_ht
);
2291 * Delete metadata stream that are flagged for deletion (endpoint_status).
2293 static void validate_endpoint_status_metadata_stream(
2294 struct lttng_poll_event
*pollset
)
2296 struct lttng_ht_iter iter
;
2297 struct lttng_consumer_stream
*stream
;
2299 DBG("Consumer delete flagged metadata stream");
2301 LTTNG_ASSERT(pollset
);
2304 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2305 /* Validate delete flag of the stream */
2306 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2310 * Remove from pollset so the metadata thread can continue without
2311 * blocking on a deleted stream.
2313 lttng_poll_del(pollset
, stream
->wait_fd
);
2315 /* Delete it right now */
2316 consumer_del_metadata_stream(stream
, metadata_ht
);
2322 * Thread polls on metadata file descriptor and write them on disk or on the
2325 void *consumer_thread_metadata_poll(void *data
)
2327 int ret
, i
, pollfd
, err
= -1;
2328 uint32_t revents
, nb_fd
;
2329 struct lttng_consumer_stream
*stream
= NULL
;
2330 struct lttng_ht_iter iter
;
2331 struct lttng_ht_node_u64
*node
;
2332 struct lttng_poll_event events
;
2333 struct lttng_consumer_local_data
*ctx
= (lttng_consumer_local_data
*) data
;
2336 rcu_register_thread();
2338 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2340 if (testpoint(consumerd_thread_metadata
)) {
2341 goto error_testpoint
;
2344 health_code_update();
2346 DBG("Thread metadata poll started");
2348 /* Size is set to 1 for the consumer_metadata pipe */
2349 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2351 ERR("Poll set creation failed");
2355 ret
= lttng_poll_add(&events
,
2356 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2362 DBG("Metadata main loop started");
2366 health_code_update();
2367 health_poll_entry();
2368 DBG("Metadata poll wait");
2369 ret
= lttng_poll_wait(&events
, -1);
2370 DBG("Metadata poll return from wait with %d fd(s)",
2371 LTTNG_POLL_GETNB(&events
));
2373 DBG("Metadata event caught in thread");
2375 if (errno
== EINTR
) {
2376 ERR("Poll EINTR caught");
2379 if (LTTNG_POLL_GETNB(&events
) == 0) {
2380 err
= 0; /* All is OK */
2387 /* From here, the event is a metadata wait fd */
2388 for (i
= 0; i
< nb_fd
; i
++) {
2389 health_code_update();
2391 revents
= LTTNG_POLL_GETEV(&events
, i
);
2392 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2394 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2395 if (revents
& LPOLLIN
) {
2398 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2399 &stream
, sizeof(stream
));
2400 if (pipe_len
< sizeof(stream
)) {
2402 PERROR("read metadata stream");
2405 * Remove the pipe from the poll set and continue the loop
2406 * since their might be data to consume.
2408 lttng_poll_del(&events
,
2409 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2410 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2414 /* A NULL stream means that the state has changed. */
2415 if (stream
== NULL
) {
2416 /* Check for deleted streams. */
2417 validate_endpoint_status_metadata_stream(&events
);
2421 DBG("Adding metadata stream %d to poll set",
2424 /* Add metadata stream to the global poll events list */
2425 lttng_poll_add(&events
, stream
->wait_fd
,
2426 LPOLLIN
| LPOLLPRI
| LPOLLHUP
);
2427 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2428 DBG("Metadata thread pipe hung up");
2430 * Remove the pipe from the poll set and continue the loop
2431 * since their might be data to consume.
2433 lttng_poll_del(&events
,
2434 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2435 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2438 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2442 /* Handle other stream */
2448 uint64_t tmp_id
= (uint64_t) pollfd
;
2450 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2452 node
= lttng_ht_iter_get_node_u64(&iter
);
2455 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2458 if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2459 /* Get the data out of the metadata file descriptor */
2460 DBG("Metadata available on fd %d", pollfd
);
2461 LTTNG_ASSERT(stream
->wait_fd
== pollfd
);
2464 health_code_update();
2466 len
= ctx
->on_buffer_ready(stream
, ctx
, false);
2468 * We don't check the return value here since if we get
2469 * a negative len, it means an error occurred thus we
2470 * simply remove it from the poll set and free the
2475 /* It's ok to have an unavailable sub-buffer */
2476 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2477 /* Clean up stream from consumer and free it. */
2478 lttng_poll_del(&events
, stream
->wait_fd
);
2479 consumer_del_metadata_stream(stream
, metadata_ht
);
2481 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2482 DBG("Metadata fd %d is hup|err.", pollfd
);
2483 if (!stream
->hangup_flush_done
&&
2484 (the_consumer_data
.type
== LTTNG_CONSUMER32_UST
||
2485 the_consumer_data
.type
==
2486 LTTNG_CONSUMER64_UST
)) {
2487 DBG("Attempting to flush and consume the UST buffers");
2488 lttng_ustconsumer_on_stream_hangup(stream
);
2490 /* We just flushed the stream now read it. */
2492 health_code_update();
2494 len
= ctx
->on_buffer_ready(stream
, ctx
, false);
2496 * We don't check the return value here since if we get
2497 * a negative len, it means an error occurred thus we
2498 * simply remove it from the poll set and free the
2504 lttng_poll_del(&events
, stream
->wait_fd
);
2506 * This call update the channel states, closes file descriptors
2507 * and securely free the stream.
2509 consumer_del_metadata_stream(stream
, metadata_ht
);
2511 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2515 /* Release RCU lock for the stream looked up */
2523 DBG("Metadata poll thread exiting");
2525 lttng_poll_clean(&events
);
2530 ERR("Health error occurred in %s", __func__
);
2532 health_unregister(health_consumerd
);
2533 rcu_unregister_thread();
2538 * This thread polls the fds in the set to consume the data and write
2539 * it to tracefile if necessary.
2541 void *consumer_thread_data_poll(void *data
)
2543 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2544 struct pollfd
*pollfd
= NULL
;
2545 /* local view of the streams */
2546 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2547 /* local view of consumer_data.fds_count */
2549 /* 2 for the consumer_data_pipe and wake up pipe */
2550 const int nb_pipes_fd
= 2;
2551 /* Number of FDs with CONSUMER_ENDPOINT_INACTIVE but still open. */
2552 int nb_inactive_fd
= 0;
2553 struct lttng_consumer_local_data
*ctx
= (lttng_consumer_local_data
*) data
;
2556 rcu_register_thread();
2558 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2560 if (testpoint(consumerd_thread_data
)) {
2561 goto error_testpoint
;
2564 health_code_update();
2566 local_stream
= zmalloc
<lttng_consumer_stream
*>();
2567 if (local_stream
== NULL
) {
2568 PERROR("local_stream malloc");
2573 health_code_update();
2579 * the fds set has been updated, we need to update our
2580 * local array as well
2582 pthread_mutex_lock(&the_consumer_data
.lock
);
2583 if (the_consumer_data
.need_update
) {
2588 local_stream
= NULL
;
2590 /* Allocate for all fds */
2591 pollfd
= calloc
<struct pollfd
>(the_consumer_data
.stream_count
+ nb_pipes_fd
);
2592 if (pollfd
== NULL
) {
2593 PERROR("pollfd malloc");
2594 pthread_mutex_unlock(&the_consumer_data
.lock
);
2598 local_stream
= calloc
<lttng_consumer_stream
*>(the_consumer_data
.stream_count
+ nb_pipes_fd
);
2599 if (local_stream
== NULL
) {
2600 PERROR("local_stream malloc");
2601 pthread_mutex_unlock(&the_consumer_data
.lock
);
2604 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2605 data_ht
, &nb_inactive_fd
);
2607 ERR("Error in allocating pollfd or local_outfds");
2608 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2609 pthread_mutex_unlock(&the_consumer_data
.lock
);
2613 the_consumer_data
.need_update
= 0;
2615 pthread_mutex_unlock(&the_consumer_data
.lock
);
2617 /* No FDs and consumer_quit, consumer_cleanup the thread */
2618 if (nb_fd
== 0 && nb_inactive_fd
== 0 &&
2619 CMM_LOAD_SHARED(consumer_quit
) == 1) {
2620 err
= 0; /* All is OK */
2623 /* poll on the array of fds */
2625 DBG("polling on %d fd", nb_fd
+ nb_pipes_fd
);
2626 if (testpoint(consumerd_thread_data_poll
)) {
2629 health_poll_entry();
2630 num_rdy
= poll(pollfd
, nb_fd
+ nb_pipes_fd
, -1);
2632 DBG("poll num_rdy : %d", num_rdy
);
2633 if (num_rdy
== -1) {
2635 * Restart interrupted system call.
2637 if (errno
== EINTR
) {
2640 PERROR("Poll error");
2641 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2643 } else if (num_rdy
== 0) {
2644 DBG("Polling thread timed out");
2648 if (caa_unlikely(data_consumption_paused
)) {
2649 DBG("Data consumption paused, sleeping...");
2655 * If the consumer_data_pipe triggered poll go directly to the
2656 * beginning of the loop to update the array. We want to prioritize
2657 * array update over low-priority reads.
2659 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2660 ssize_t pipe_readlen
;
2662 DBG("consumer_data_pipe wake up");
2663 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2664 &new_stream
, sizeof(new_stream
));
2665 if (pipe_readlen
< sizeof(new_stream
)) {
2666 PERROR("Consumer data pipe");
2667 /* Continue so we can at least handle the current stream(s). */
2672 * If the stream is NULL, just ignore it. It's also possible that
2673 * the sessiond poll thread changed the consumer_quit state and is
2674 * waking us up to test it.
2676 if (new_stream
== NULL
) {
2677 validate_endpoint_status_data_stream();
2681 /* Continue to update the local streams and handle prio ones */
2685 /* Handle wakeup pipe. */
2686 if (pollfd
[nb_fd
+ 1].revents
& (POLLIN
| POLLPRI
)) {
2688 ssize_t pipe_readlen
;
2690 pipe_readlen
= lttng_pipe_read(ctx
->consumer_wakeup_pipe
, &dummy
,
2692 if (pipe_readlen
< 0) {
2693 PERROR("Consumer data wakeup pipe");
2695 /* We've been awakened to handle stream(s). */
2696 ctx
->has_wakeup
= 0;
2699 /* Take care of high priority channels first. */
2700 for (i
= 0; i
< nb_fd
; i
++) {
2701 health_code_update();
2703 if (local_stream
[i
] == NULL
) {
2706 if (pollfd
[i
].revents
& POLLPRI
) {
2707 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2709 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
, false);
2710 /* it's ok to have an unavailable sub-buffer */
2711 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2712 /* Clean the stream and free it. */
2713 consumer_del_stream(local_stream
[i
], data_ht
);
2714 local_stream
[i
] = NULL
;
2715 } else if (len
> 0) {
2716 local_stream
[i
]->has_data_left_to_be_read_before_teardown
= 1;
2722 * If we read high prio channel in this loop, try again
2723 * for more high prio data.
2729 /* Take care of low priority channels. */
2730 for (i
= 0; i
< nb_fd
; i
++) {
2731 health_code_update();
2733 if (local_stream
[i
] == NULL
) {
2736 if ((pollfd
[i
].revents
& POLLIN
) ||
2737 local_stream
[i
]->hangup_flush_done
||
2738 local_stream
[i
]->has_data
) {
2739 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2740 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
, false);
2741 /* it's ok to have an unavailable sub-buffer */
2742 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2743 /* Clean the stream and free it. */
2744 consumer_del_stream(local_stream
[i
], data_ht
);
2745 local_stream
[i
] = NULL
;
2746 } else if (len
> 0) {
2747 local_stream
[i
]->has_data_left_to_be_read_before_teardown
= 1;
2752 /* Handle hangup and errors */
2753 for (i
= 0; i
< nb_fd
; i
++) {
2754 health_code_update();
2756 if (local_stream
[i
] == NULL
) {
2759 if (!local_stream
[i
]->hangup_flush_done
2760 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2761 && (the_consumer_data
.type
== LTTNG_CONSUMER32_UST
2762 || the_consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2763 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2765 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2766 /* Attempt read again, for the data we just flushed. */
2767 local_stream
[i
]->has_data_left_to_be_read_before_teardown
= 1;
2770 * When a stream's pipe dies (hup/err/nval), an "inactive producer" flush is
2771 * performed. This type of flush ensures that a new packet is produced no
2772 * matter the consumed/produced positions are.
2774 * This, in turn, causes the next pass to see that data available for the
2775 * stream. When we come back here, we can be assured that all available
2776 * data has been consumed and we can finally destroy the stream.
2778 * If the poll flag is HUP/ERR/NVAL and we have
2779 * read no data in this pass, we can remove the
2780 * stream from its hash table.
2782 if ((pollfd
[i
].revents
& POLLHUP
)) {
2783 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2784 if (!local_stream
[i
]->has_data_left_to_be_read_before_teardown
) {
2785 consumer_del_stream(local_stream
[i
], data_ht
);
2786 local_stream
[i
] = NULL
;
2789 } else if (pollfd
[i
].revents
& POLLERR
) {
2790 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2791 if (!local_stream
[i
]->has_data_left_to_be_read_before_teardown
) {
2792 consumer_del_stream(local_stream
[i
], data_ht
);
2793 local_stream
[i
] = NULL
;
2796 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2797 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2798 if (!local_stream
[i
]->has_data_left_to_be_read_before_teardown
) {
2799 consumer_del_stream(local_stream
[i
], data_ht
);
2800 local_stream
[i
] = NULL
;
2804 if (local_stream
[i
] != NULL
) {
2805 local_stream
[i
]->has_data_left_to_be_read_before_teardown
= 0;
2812 DBG("polling thread exiting");
2817 * Close the write side of the pipe so epoll_wait() in
2818 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2819 * read side of the pipe. If we close them both, epoll_wait strangely does
2820 * not return and could create a endless wait period if the pipe is the
2821 * only tracked fd in the poll set. The thread will take care of closing
2824 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2829 ERR("Health error occurred in %s", __func__
);
2831 health_unregister(health_consumerd
);
2833 rcu_unregister_thread();
2838 * Close wake-up end of each stream belonging to the channel. This will
2839 * allow the poll() on the stream read-side to detect when the
2840 * write-side (application) finally closes them.
2843 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2845 struct lttng_ht
*ht
;
2846 struct lttng_consumer_stream
*stream
;
2847 struct lttng_ht_iter iter
;
2849 ht
= the_consumer_data
.stream_per_chan_id_ht
;
2852 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2853 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2854 ht
->match_fct
, &channel
->key
,
2855 &iter
.iter
, stream
, node_channel_id
.node
) {
2857 * Protect against teardown with mutex.
2859 pthread_mutex_lock(&stream
->lock
);
2860 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2863 switch (the_consumer_data
.type
) {
2864 case LTTNG_CONSUMER_KERNEL
:
2866 case LTTNG_CONSUMER32_UST
:
2867 case LTTNG_CONSUMER64_UST
:
2868 if (stream
->metadata_flag
) {
2869 /* Safe and protected by the stream lock. */
2870 lttng_ustconsumer_close_metadata(stream
->chan
);
2873 * Note: a mutex is taken internally within
2874 * liblttng-ust-ctl to protect timer wakeup_fd
2875 * use from concurrent close.
2877 lttng_ustconsumer_close_stream_wakeup(stream
);
2881 ERR("Unknown consumer_data type");
2885 pthread_mutex_unlock(&stream
->lock
);
2890 static void destroy_channel_ht(struct lttng_ht
*ht
)
2892 struct lttng_ht_iter iter
;
2893 struct lttng_consumer_channel
*channel
;
2901 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2902 ret
= lttng_ht_del(ht
, &iter
);
2903 LTTNG_ASSERT(ret
!= 0);
2907 lttng_ht_destroy(ht
);
2911 * This thread polls the channel fds to detect when they are being
2912 * closed. It closes all related streams if the channel is detected as
2913 * closed. It is currently only used as a shim layer for UST because the
2914 * consumerd needs to keep the per-stream wakeup end of pipes open for
2917 void *consumer_thread_channel_poll(void *data
)
2919 int ret
, i
, pollfd
, err
= -1;
2920 uint32_t revents
, nb_fd
;
2921 struct lttng_consumer_channel
*chan
= NULL
;
2922 struct lttng_ht_iter iter
;
2923 struct lttng_ht_node_u64
*node
;
2924 struct lttng_poll_event events
;
2925 struct lttng_consumer_local_data
*ctx
= (lttng_consumer_local_data
*) data
;
2926 struct lttng_ht
*channel_ht
;
2928 rcu_register_thread();
2930 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
2932 if (testpoint(consumerd_thread_channel
)) {
2933 goto error_testpoint
;
2936 health_code_update();
2938 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2940 /* ENOMEM at this point. Better to bail out. */
2944 DBG("Thread channel poll started");
2946 /* Size is set to 1 for the consumer_channel pipe */
2947 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2949 ERR("Poll set creation failed");
2953 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2959 DBG("Channel main loop started");
2963 health_code_update();
2964 DBG("Channel poll wait");
2965 health_poll_entry();
2966 ret
= lttng_poll_wait(&events
, -1);
2967 DBG("Channel poll return from wait with %d fd(s)",
2968 LTTNG_POLL_GETNB(&events
));
2970 DBG("Channel event caught in thread");
2972 if (errno
== EINTR
) {
2973 ERR("Poll EINTR caught");
2976 if (LTTNG_POLL_GETNB(&events
) == 0) {
2977 err
= 0; /* All is OK */
2984 /* From here, the event is a channel wait fd */
2985 for (i
= 0; i
< nb_fd
; i
++) {
2986 health_code_update();
2988 revents
= LTTNG_POLL_GETEV(&events
, i
);
2989 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2991 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2992 if (revents
& LPOLLIN
) {
2993 enum consumer_channel_action action
;
2996 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
2999 ERR("Error reading channel pipe");
3001 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
3006 case CONSUMER_CHANNEL_ADD
:
3007 DBG("Adding channel %d to poll set",
3010 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
3013 lttng_ht_add_unique_u64(channel_ht
,
3014 &chan
->wait_fd_node
);
3016 /* Add channel to the global poll events list */
3017 lttng_poll_add(&events
, chan
->wait_fd
,
3018 LPOLLERR
| LPOLLHUP
);
3020 case CONSUMER_CHANNEL_DEL
:
3023 * This command should never be called if the channel
3024 * has streams monitored by either the data or metadata
3025 * thread. The consumer only notify this thread with a
3026 * channel del. command if it receives a destroy
3027 * channel command from the session daemon that send it
3028 * if a command prior to the GET_CHANNEL failed.
3032 chan
= consumer_find_channel(key
);
3035 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
3038 lttng_poll_del(&events
, chan
->wait_fd
);
3039 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
3040 ret
= lttng_ht_del(channel_ht
, &iter
);
3041 LTTNG_ASSERT(ret
== 0);
3043 switch (the_consumer_data
.type
) {
3044 case LTTNG_CONSUMER_KERNEL
:
3046 case LTTNG_CONSUMER32_UST
:
3047 case LTTNG_CONSUMER64_UST
:
3048 health_code_update();
3049 /* Destroy streams that might have been left in the stream list. */
3050 clean_channel_stream_list(chan
);
3053 ERR("Unknown consumer_data type");
3058 * Release our own refcount. Force channel deletion even if
3059 * streams were not initialized.
3061 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
3062 consumer_del_channel(chan
);
3067 case CONSUMER_CHANNEL_QUIT
:
3069 * Remove the pipe from the poll set and continue the loop
3070 * since their might be data to consume.
3072 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
3075 ERR("Unknown action");
3078 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
3079 DBG("Channel thread pipe hung up");
3081 * Remove the pipe from the poll set and continue the loop
3082 * since their might be data to consume.
3084 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
3087 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3091 /* Handle other stream */
3097 uint64_t tmp_id
= (uint64_t) pollfd
;
3099 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
3101 node
= lttng_ht_iter_get_node_u64(&iter
);
3104 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
3107 /* Check for error event */
3108 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
3109 DBG("Channel fd %d is hup|err.", pollfd
);
3111 lttng_poll_del(&events
, chan
->wait_fd
);
3112 ret
= lttng_ht_del(channel_ht
, &iter
);
3113 LTTNG_ASSERT(ret
== 0);
3116 * This will close the wait fd for each stream associated to
3117 * this channel AND monitored by the data/metadata thread thus
3118 * will be clean by the right thread.
3120 consumer_close_channel_streams(chan
);
3122 /* Release our own refcount */
3123 if (!uatomic_sub_return(&chan
->refcount
, 1)
3124 && !uatomic_read(&chan
->nb_init_stream_left
)) {
3125 consumer_del_channel(chan
);
3128 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3133 /* Release RCU lock for the channel looked up */
3141 lttng_poll_clean(&events
);
3143 destroy_channel_ht(channel_ht
);
3146 DBG("Channel poll thread exiting");
3149 ERR("Health error occurred in %s", __func__
);
3151 health_unregister(health_consumerd
);
3152 rcu_unregister_thread();
3156 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
3157 struct pollfd
*sockpoll
, int client_socket
)
3162 LTTNG_ASSERT(sockpoll
);
3164 ret
= lttng_consumer_poll_socket(sockpoll
);
3168 DBG("Metadata connection on client_socket");
3170 /* Blocking call, waiting for transmission */
3171 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
3172 if (ctx
->consumer_metadata_socket
< 0) {
3173 WARN("On accept metadata");
3184 * This thread listens on the consumerd socket and receives the file
3185 * descriptors from the session daemon.
3187 void *consumer_thread_sessiond_poll(void *data
)
3189 int sock
= -1, client_socket
, ret
, err
= -1;
3191 * structure to poll for incoming data on communication socket avoids
3192 * making blocking sockets.
3194 struct pollfd consumer_sockpoll
[2];
3195 struct lttng_consumer_local_data
*ctx
= (lttng_consumer_local_data
*) data
;
3197 rcu_register_thread();
3199 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
3201 if (testpoint(consumerd_thread_sessiond
)) {
3202 goto error_testpoint
;
3205 health_code_update();
3207 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
3208 unlink(ctx
->consumer_command_sock_path
);
3209 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
3210 if (client_socket
< 0) {
3211 ERR("Cannot create command socket");
3215 ret
= lttcomm_listen_unix_sock(client_socket
);
3220 DBG("Sending ready command to lttng-sessiond");
3221 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3222 /* return < 0 on error, but == 0 is not fatal */
3224 ERR("Error sending ready command to lttng-sessiond");
3228 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3229 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3230 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3231 consumer_sockpoll
[1].fd
= client_socket
;
3232 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3234 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3242 DBG("Connection on client_socket");
3244 /* Blocking call, waiting for transmission */
3245 sock
= lttcomm_accept_unix_sock(client_socket
);
3252 * Setup metadata socket which is the second socket connection on the
3253 * command unix socket.
3255 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3264 /* This socket is not useful anymore. */
3265 ret
= close(client_socket
);
3267 PERROR("close client_socket");
3271 /* update the polling structure to poll on the established socket */
3272 consumer_sockpoll
[1].fd
= sock
;
3273 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3276 health_code_update();
3278 health_poll_entry();
3279 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3288 DBG("Incoming command on sock");
3289 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3292 * This could simply be a session daemon quitting. Don't output
3295 DBG("Communication interrupted on command socket");
3299 if (CMM_LOAD_SHARED(consumer_quit
)) {
3300 DBG("consumer_thread_receive_fds received quit from signal");
3301 err
= 0; /* All is OK */
3304 DBG("Received command on sock");
3310 DBG("Consumer thread sessiond poll exiting");
3313 * Close metadata streams since the producer is the session daemon which
3316 * NOTE: for now, this only applies to the UST tracer.
3318 lttng_consumer_close_all_metadata();
3321 * when all fds have hung up, the polling thread
3324 CMM_STORE_SHARED(consumer_quit
, 1);
3327 * Notify the data poll thread to poll back again and test the
3328 * consumer_quit state that we just set so to quit gracefully.
3330 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3332 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3334 notify_health_quit_pipe(health_quit_pipe
);
3336 /* Cleaning up possibly open sockets. */
3340 PERROR("close sock sessiond poll");
3343 if (client_socket
>= 0) {
3344 ret
= close(client_socket
);
3346 PERROR("close client_socket sessiond poll");
3353 ERR("Health error occurred in %s", __func__
);
3355 health_unregister(health_consumerd
);
3357 rcu_unregister_thread();
3361 static int post_consume(struct lttng_consumer_stream
*stream
,
3362 const struct stream_subbuffer
*subbuffer
,
3363 struct lttng_consumer_local_data
*ctx
)
3367 const size_t count
= lttng_dynamic_array_get_count(
3368 &stream
->read_subbuffer_ops
.post_consume_cbs
);
3370 for (i
= 0; i
< count
; i
++) {
3371 const post_consume_cb op
= *(post_consume_cb
*) lttng_dynamic_array_get_element(
3372 &stream
->read_subbuffer_ops
.post_consume_cbs
,
3375 ret
= op(stream
, subbuffer
, ctx
);
3384 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3385 struct lttng_consumer_local_data
*ctx
,
3386 bool locked_by_caller
)
3388 ssize_t ret
, written_bytes
= 0;
3390 struct stream_subbuffer subbuffer
= {};
3391 enum get_next_subbuffer_status get_next_status
;
3393 if (!locked_by_caller
) {
3394 stream
->read_subbuffer_ops
.lock(stream
);
3396 stream
->read_subbuffer_ops
.assert_locked(stream
);
3399 if (stream
->read_subbuffer_ops
.on_wake_up
) {
3400 ret
= stream
->read_subbuffer_ops
.on_wake_up(stream
);
3407 * If the stream was flagged to be ready for rotation before we extract
3408 * the next packet, rotate it now.
3410 if (stream
->rotate_ready
) {
3411 DBG("Rotate stream before consuming data");
3412 ret
= lttng_consumer_rotate_stream(stream
);
3414 ERR("Stream rotation error before consuming data");
3419 get_next_status
= stream
->read_subbuffer_ops
.get_next_subbuffer(
3420 stream
, &subbuffer
);
3421 switch (get_next_status
) {
3422 case GET_NEXT_SUBBUFFER_STATUS_OK
:
3424 case GET_NEXT_SUBBUFFER_STATUS_NO_DATA
:
3428 case GET_NEXT_SUBBUFFER_STATUS_ERROR
:
3435 ret
= stream
->read_subbuffer_ops
.pre_consume_subbuffer(
3436 stream
, &subbuffer
);
3438 goto error_put_subbuf
;
3441 written_bytes
= stream
->read_subbuffer_ops
.consume_subbuffer(
3442 ctx
, stream
, &subbuffer
);
3443 if (written_bytes
<= 0) {
3444 ERR("Error consuming subbuffer: (%zd)", written_bytes
);
3445 ret
= (int) written_bytes
;
3446 goto error_put_subbuf
;
3449 ret
= stream
->read_subbuffer_ops
.put_next_subbuffer(stream
, &subbuffer
);
3454 ret
= post_consume(stream
, &subbuffer
, ctx
);
3460 * After extracting the packet, we check if the stream is now ready to
3461 * be rotated and perform the action immediately.
3463 * Don't overwrite `ret` as callers expect the number of bytes
3464 * consumed to be returned on success.
3466 rotation_ret
= lttng_consumer_stream_is_rotate_ready(stream
);
3467 if (rotation_ret
== 1) {
3468 rotation_ret
= lttng_consumer_rotate_stream(stream
);
3469 if (rotation_ret
< 0) {
3471 ERR("Stream rotation error after consuming data");
3475 } else if (rotation_ret
< 0) {
3477 ERR("Failed to check if stream was ready to rotate after consuming data");
3482 if (stream
->read_subbuffer_ops
.on_sleep
) {
3483 stream
->read_subbuffer_ops
.on_sleep(stream
, ctx
);
3486 ret
= written_bytes
;
3488 if (!locked_by_caller
) {
3489 stream
->read_subbuffer_ops
.unlock(stream
);
3494 (void) stream
->read_subbuffer_ops
.put_next_subbuffer(stream
, &subbuffer
);
3498 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3500 switch (the_consumer_data
.type
) {
3501 case LTTNG_CONSUMER_KERNEL
:
3502 return lttng_kconsumer_on_recv_stream(stream
);
3503 case LTTNG_CONSUMER32_UST
:
3504 case LTTNG_CONSUMER64_UST
:
3505 return lttng_ustconsumer_on_recv_stream(stream
);
3507 ERR("Unknown consumer_data type");
3514 * Allocate and set consumer data hash tables.
3516 int lttng_consumer_init(void)
3518 the_consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3519 if (!the_consumer_data
.channel_ht
) {
3523 the_consumer_data
.channels_by_session_id_ht
=
3524 lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3525 if (!the_consumer_data
.channels_by_session_id_ht
) {
3529 the_consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3530 if (!the_consumer_data
.relayd_ht
) {
3534 the_consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3535 if (!the_consumer_data
.stream_list_ht
) {
3539 the_consumer_data
.stream_per_chan_id_ht
=
3540 lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3541 if (!the_consumer_data
.stream_per_chan_id_ht
) {
3545 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3550 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3555 the_consumer_data
.chunk_registry
= lttng_trace_chunk_registry_create();
3556 if (!the_consumer_data
.chunk_registry
) {
3567 * Process the ADD_RELAYD command receive by a consumer.
3569 * This will create a relayd socket pair and add it to the relayd hash table.
3570 * The caller MUST acquire a RCU read side lock before calling it.
3572 void consumer_add_relayd_socket(uint64_t net_seq_idx
,
3574 struct lttng_consumer_local_data
*ctx
,
3576 struct pollfd
*consumer_sockpoll
,
3577 uint64_t sessiond_id
,
3578 uint64_t relayd_session_id
,
3579 uint32_t relayd_version_major
,
3580 uint32_t relayd_version_minor
,
3581 enum lttcomm_sock_proto relayd_socket_protocol
)
3583 int fd
= -1, ret
= -1, relayd_created
= 0;
3584 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3585 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3588 LTTNG_ASSERT(sock
>= 0);
3589 ASSERT_RCU_READ_LOCKED();
3591 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", net_seq_idx
);
3593 /* Get relayd reference if exists. */
3594 relayd
= consumer_find_relayd(net_seq_idx
);
3595 if (relayd
== NULL
) {
3596 LTTNG_ASSERT(sock_type
== LTTNG_STREAM_CONTROL
);
3597 /* Not found. Allocate one. */
3598 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3599 if (relayd
== NULL
) {
3600 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3603 relayd
->sessiond_session_id
= sessiond_id
;
3608 * This code path MUST continue to the consumer send status message to
3609 * we can notify the session daemon and continue our work without
3610 * killing everything.
3614 * relayd key should never be found for control socket.
3616 LTTNG_ASSERT(sock_type
!= LTTNG_STREAM_CONTROL
);
3619 /* First send a status message before receiving the fds. */
3620 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3622 /* Somehow, the session daemon is not responding anymore. */
3623 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3624 goto error_nosignal
;
3627 /* Poll on consumer socket. */
3628 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3630 /* Needing to exit in the middle of a command: error. */
3631 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3632 goto error_nosignal
;
3635 /* Get relayd socket from session daemon */
3636 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3637 if (ret
!= sizeof(fd
)) {
3638 fd
= -1; /* Just in case it gets set with an invalid value. */
3641 * Failing to receive FDs might indicate a major problem such as
3642 * reaching a fd limit during the receive where the kernel returns a
3643 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3644 * don't take any chances and stop everything.
3646 * XXX: Feature request #558 will fix that and avoid this possible
3647 * issue when reaching the fd limit.
3649 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3650 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3654 /* Copy socket information and received FD */
3655 switch (sock_type
) {
3656 case LTTNG_STREAM_CONTROL
:
3657 /* Copy received lttcomm socket */
3658 ret
= lttcomm_populate_sock_from_open_socket(
3659 &relayd
->control_sock
.sock
, fd
,
3660 relayd_socket_protocol
);
3662 /* Assign version values. */
3663 relayd
->control_sock
.major
= relayd_version_major
;
3664 relayd
->control_sock
.minor
= relayd_version_minor
;
3666 relayd
->relayd_session_id
= relayd_session_id
;
3669 case LTTNG_STREAM_DATA
:
3670 /* Copy received lttcomm socket */
3671 ret
= lttcomm_populate_sock_from_open_socket(
3672 &relayd
->data_sock
.sock
, fd
,
3673 relayd_socket_protocol
);
3674 /* Assign version values. */
3675 relayd
->data_sock
.major
= relayd_version_major
;
3676 relayd
->data_sock
.minor
= relayd_version_minor
;
3679 ERR("Unknown relayd socket type (%d)", sock_type
);
3680 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3685 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3689 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3690 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3691 relayd
->net_seq_idx
, fd
);
3693 * We gave the ownership of the fd to the relayd structure. Set the
3694 * fd to -1 so we don't call close() on it in the error path below.
3698 /* We successfully added the socket. Send status back. */
3699 ret
= consumer_send_status_msg(sock
, ret_code
);
3701 /* Somehow, the session daemon is not responding anymore. */
3702 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3703 goto error_nosignal
;
3707 * Add relayd socket pair to consumer data hashtable. If object already
3708 * exists or on error, the function gracefully returns.
3717 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3718 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3722 /* Close received socket if valid. */
3725 PERROR("close received socket");
3729 if (relayd_created
) {
3735 * Search for a relayd associated to the session id and return the reference.
3737 * A rcu read side lock MUST be acquire before calling this function and locked
3738 * until the relayd object is no longer necessary.
3740 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3742 struct lttng_ht_iter iter
;
3743 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3745 ASSERT_RCU_READ_LOCKED();
3747 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3748 cds_lfht_for_each_entry(the_consumer_data
.relayd_ht
->ht
, &iter
.iter
,
3749 relayd
, node
.node
) {
3751 * Check by sessiond id which is unique here where the relayd session
3752 * id might not be when having multiple relayd.
3754 if (relayd
->sessiond_session_id
== id
) {
3755 /* Found the relayd. There can be only one per id. */
3767 * Check if for a given session id there is still data needed to be extract
3770 * Return 1 if data is pending or else 0 meaning ready to be read.
3772 int consumer_data_pending(uint64_t id
)
3775 struct lttng_ht_iter iter
;
3776 struct lttng_ht
*ht
;
3777 struct lttng_consumer_stream
*stream
;
3778 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3779 int (*data_pending
)(struct lttng_consumer_stream
*);
3781 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3784 pthread_mutex_lock(&the_consumer_data
.lock
);
3786 switch (the_consumer_data
.type
) {
3787 case LTTNG_CONSUMER_KERNEL
:
3788 data_pending
= lttng_kconsumer_data_pending
;
3790 case LTTNG_CONSUMER32_UST
:
3791 case LTTNG_CONSUMER64_UST
:
3792 data_pending
= lttng_ustconsumer_data_pending
;
3795 ERR("Unknown consumer data type");
3799 /* Ease our life a bit */
3800 ht
= the_consumer_data
.stream_list_ht
;
3802 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3803 ht
->hash_fct(&id
, lttng_ht_seed
),
3805 &iter
.iter
, stream
, node_session_id
.node
) {
3806 pthread_mutex_lock(&stream
->lock
);
3809 * A removed node from the hash table indicates that the stream has
3810 * been deleted thus having a guarantee that the buffers are closed
3811 * on the consumer side. However, data can still be transmitted
3812 * over the network so don't skip the relayd check.
3814 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3816 /* Check the stream if there is data in the buffers. */
3817 ret
= data_pending(stream
);
3819 pthread_mutex_unlock(&stream
->lock
);
3824 pthread_mutex_unlock(&stream
->lock
);
3827 relayd
= find_relayd_by_session_id(id
);
3829 unsigned int is_data_inflight
= 0;
3831 /* Send init command for data pending. */
3832 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3833 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3834 relayd
->relayd_session_id
);
3836 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3837 /* Communication error thus the relayd so no data pending. */
3838 goto data_not_pending
;
3841 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3842 ht
->hash_fct(&id
, lttng_ht_seed
),
3844 &iter
.iter
, stream
, node_session_id
.node
) {
3845 if (stream
->metadata_flag
) {
3846 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3847 stream
->relayd_stream_id
);
3849 ret
= relayd_data_pending(&relayd
->control_sock
,
3850 stream
->relayd_stream_id
,
3851 stream
->next_net_seq_num
- 1);
3855 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3857 } else if (ret
< 0) {
3858 ERR("Relayd data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
3859 lttng_consumer_cleanup_relayd(relayd
);
3860 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3861 goto data_not_pending
;
3865 /* Send end command for data pending. */
3866 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3867 relayd
->relayd_session_id
, &is_data_inflight
);
3868 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3870 ERR("Relayd end data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
3871 lttng_consumer_cleanup_relayd(relayd
);
3872 goto data_not_pending
;
3874 if (is_data_inflight
) {
3880 * Finding _no_ node in the hash table and no inflight data means that the
3881 * stream(s) have been removed thus data is guaranteed to be available for
3882 * analysis from the trace files.
3886 /* Data is available to be read by a viewer. */
3887 pthread_mutex_unlock(&the_consumer_data
.lock
);
3892 /* Data is still being extracted from buffers. */
3893 pthread_mutex_unlock(&the_consumer_data
.lock
);
3899 * Send a ret code status message to the sessiond daemon.
3901 * Return the sendmsg() return value.
3903 int consumer_send_status_msg(int sock
, int ret_code
)
3905 struct lttcomm_consumer_status_msg msg
;
3907 memset(&msg
, 0, sizeof(msg
));
3908 msg
.ret_code
= (lttcomm_return_code
) ret_code
;
3910 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3914 * Send a channel status message to the sessiond daemon.
3916 * Return the sendmsg() return value.
3918 int consumer_send_status_channel(int sock
,
3919 struct lttng_consumer_channel
*channel
)
3921 struct lttcomm_consumer_status_channel msg
;
3923 LTTNG_ASSERT(sock
>= 0);
3925 memset(&msg
, 0, sizeof(msg
));
3927 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3929 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3930 msg
.key
= channel
->key
;
3931 msg
.stream_count
= channel
->streams
.count
;
3934 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3937 unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos
,
3938 unsigned long produced_pos
, uint64_t nb_packets_per_stream
,
3939 uint64_t max_sb_size
)
3941 unsigned long start_pos
;
3943 if (!nb_packets_per_stream
) {
3944 return consumed_pos
; /* Grab everything */
3946 start_pos
= produced_pos
- lttng_offset_align_floor(produced_pos
, max_sb_size
);
3947 start_pos
-= max_sb_size
* nb_packets_per_stream
;
3948 if ((long) (start_pos
- consumed_pos
) < 0) {
3949 return consumed_pos
; /* Grab everything */
3954 /* Stream lock must be held by the caller. */
3955 static int sample_stream_positions(struct lttng_consumer_stream
*stream
,
3956 unsigned long *produced
, unsigned long *consumed
)
3960 ASSERT_LOCKED(stream
->lock
);
3962 ret
= lttng_consumer_sample_snapshot_positions(stream
);
3964 ERR("Failed to sample snapshot positions");
3968 ret
= lttng_consumer_get_produced_snapshot(stream
, produced
);
3970 ERR("Failed to sample produced position");
3974 ret
= lttng_consumer_get_consumed_snapshot(stream
, consumed
);
3976 ERR("Failed to sample consumed position");
3985 * Sample the rotate position for all the streams of a channel. If a stream
3986 * is already at the rotate position (produced == consumed), we flag it as
3987 * ready for rotation. The rotation of ready streams occurs after we have
3988 * replied to the session daemon that we have finished sampling the positions.
3989 * Must be called with RCU read-side lock held to ensure existence of channel.
3991 * Returns 0 on success, < 0 on error
3993 int lttng_consumer_rotate_channel(struct lttng_consumer_channel
*channel
,
3994 uint64_t key
, uint64_t relayd_id
)
3997 struct lttng_consumer_stream
*stream
;
3998 struct lttng_ht_iter iter
;
3999 struct lttng_ht
*ht
= the_consumer_data
.stream_per_chan_id_ht
;
4000 struct lttng_dynamic_array stream_rotation_positions
;
4001 uint64_t next_chunk_id
, stream_count
= 0;
4002 enum lttng_trace_chunk_status chunk_status
;
4003 const bool is_local_trace
= relayd_id
== -1ULL;
4004 struct consumer_relayd_sock_pair
*relayd
= NULL
;
4005 bool rotating_to_new_chunk
= true;
4006 /* Array of `struct lttng_consumer_stream *` */
4007 struct lttng_dynamic_pointer_array streams_packet_to_open
;
4010 ASSERT_RCU_READ_LOCKED();
4012 DBG("Consumer sample rotate position for channel %" PRIu64
, key
);
4014 lttng_dynamic_array_init(&stream_rotation_positions
,
4015 sizeof(struct relayd_stream_rotation_position
), NULL
);
4016 lttng_dynamic_pointer_array_init(&streams_packet_to_open
, NULL
);
4020 pthread_mutex_lock(&channel
->lock
);
4021 LTTNG_ASSERT(channel
->trace_chunk
);
4022 chunk_status
= lttng_trace_chunk_get_id(channel
->trace_chunk
,
4024 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4026 goto end_unlock_channel
;
4029 cds_lfht_for_each_entry_duplicate(ht
->ht
,
4030 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
4031 ht
->match_fct
, &channel
->key
, &iter
.iter
,
4032 stream
, node_channel_id
.node
) {
4033 unsigned long produced_pos
= 0, consumed_pos
= 0;
4035 health_code_update();
4038 * Lock stream because we are about to change its state.
4040 pthread_mutex_lock(&stream
->lock
);
4042 if (stream
->trace_chunk
== stream
->chan
->trace_chunk
) {
4043 rotating_to_new_chunk
= false;
4047 * Do not flush a packet when rotating from a NULL trace
4048 * chunk. The stream has no means to output data, and the prior
4049 * rotation which rotated to NULL performed that side-effect
4050 * already. No new data can be produced when a stream has no
4051 * associated trace chunk (e.g. a stop followed by a rotate).
4053 if (stream
->trace_chunk
) {
4056 if (stream
->metadata_flag
) {
4058 * Don't produce an empty metadata packet,
4059 * simply close the current one.
4061 * Metadata is regenerated on every trace chunk
4062 * switch; there is no concern that no data was
4065 flush_active
= true;
4068 * Only flush an empty packet if the "packet
4069 * open" could not be performed on transition
4070 * to a new trace chunk and no packets were
4071 * consumed within the chunk's lifetime.
4073 if (stream
->opened_packet_in_current_trace_chunk
) {
4074 flush_active
= true;
4077 * Stream could have been full at the
4078 * time of rotation, but then have had
4079 * no activity at all.
4081 * It is important to flush a packet
4082 * to prevent 0-length files from being
4083 * produced as most viewers choke on
4086 * Unfortunately viewers will not be
4087 * able to know that tracing was active
4088 * for this stream during this trace
4091 ret
= sample_stream_positions(stream
, &produced_pos
, &consumed_pos
);
4093 goto end_unlock_stream
;
4097 * Don't flush an empty packet if data
4098 * was produced; it will be consumed
4099 * before the rotation completes.
4101 flush_active
= produced_pos
!= consumed_pos
;
4102 if (!flush_active
) {
4103 const char *trace_chunk_name
;
4104 uint64_t trace_chunk_id
;
4106 chunk_status
= lttng_trace_chunk_get_name(
4107 stream
->trace_chunk
,
4110 if (chunk_status
== LTTNG_TRACE_CHUNK_STATUS_NONE
) {
4111 trace_chunk_name
= "none";
4115 * Consumer trace chunks are
4118 chunk_status
= lttng_trace_chunk_get_id(
4119 stream
->trace_chunk
,
4121 LTTNG_ASSERT(chunk_status
==
4122 LTTNG_TRACE_CHUNK_STATUS_OK
);
4124 DBG("Unable to open packet for stream during trace chunk's lifetime. "
4125 "Flushing an empty packet to prevent an empty file from being created: "
4126 "stream id = %" PRIu64
", trace chunk name = `%s`, trace chunk id = %" PRIu64
,
4127 stream
->key
, trace_chunk_name
, trace_chunk_id
);
4133 * Close the current packet before sampling the
4134 * ring buffer positions.
4136 ret
= consumer_stream_flush_buffer(stream
, flush_active
);
4138 ERR("Failed to flush stream %" PRIu64
" during channel rotation",
4140 goto end_unlock_stream
;
4144 ret
= lttng_consumer_take_snapshot(stream
);
4145 if (ret
< 0 && ret
!= -ENODATA
&& ret
!= -EAGAIN
) {
4146 ERR("Failed to sample snapshot position during channel rotation");
4147 goto end_unlock_stream
;
4150 ret
= lttng_consumer_get_produced_snapshot(stream
,
4153 ERR("Failed to sample produced position during channel rotation");
4154 goto end_unlock_stream
;
4157 ret
= lttng_consumer_get_consumed_snapshot(stream
,
4160 ERR("Failed to sample consumed position during channel rotation");
4161 goto end_unlock_stream
;
4165 * Align produced position on the start-of-packet boundary of the first
4166 * packet going into the next trace chunk.
4168 produced_pos
= lttng_align_floor(produced_pos
, stream
->max_sb_size
);
4169 if (consumed_pos
== produced_pos
) {
4170 DBG("Set rotate ready for stream %" PRIu64
" produced = %lu consumed = %lu",
4171 stream
->key
, produced_pos
, consumed_pos
);
4172 stream
->rotate_ready
= true;
4174 DBG("Different consumed and produced positions "
4175 "for stream %" PRIu64
" produced = %lu consumed = %lu",
4176 stream
->key
, produced_pos
, consumed_pos
);
4179 * The rotation position is based on the packet_seq_num of the
4180 * packet following the last packet that was consumed for this
4181 * stream, incremented by the offset between produced and
4182 * consumed positions. This rotation position is a lower bound
4183 * (inclusive) at which the next trace chunk starts. Since it
4184 * is a lower bound, it is OK if the packet_seq_num does not
4185 * correspond exactly to the same packet identified by the
4186 * consumed_pos, which can happen in overwrite mode.
4188 if (stream
->sequence_number_unavailable
) {
4190 * Rotation should never be performed on a session which
4191 * interacts with a pre-2.8 lttng-modules, which does
4192 * not implement packet sequence number.
4194 ERR("Failure to rotate stream %" PRIu64
": sequence number unavailable",
4197 goto end_unlock_stream
;
4199 stream
->rotate_position
= stream
->last_sequence_number
+ 1 +
4200 ((produced_pos
- consumed_pos
) / stream
->max_sb_size
);
4201 DBG("Set rotation position for stream %" PRIu64
" at position %" PRIu64
,
4202 stream
->key
, stream
->rotate_position
);
4204 if (!is_local_trace
) {
4206 * The relay daemon control protocol expects a rotation
4207 * position as "the sequence number of the first packet
4208 * _after_ the current trace chunk".
4210 const struct relayd_stream_rotation_position position
= {
4211 .stream_id
= stream
->relayd_stream_id
,
4212 .rotate_at_seq_num
= stream
->rotate_position
,
4215 ret
= lttng_dynamic_array_add_element(
4216 &stream_rotation_positions
,
4219 ERR("Failed to allocate stream rotation position");
4220 goto end_unlock_stream
;
4225 stream
->opened_packet_in_current_trace_chunk
= false;
4227 if (rotating_to_new_chunk
&& !stream
->metadata_flag
) {
4229 * Attempt to flush an empty packet as close to the
4230 * rotation point as possible. In the event where a
4231 * stream remains inactive after the rotation point,
4232 * this ensures that the new trace chunk has a
4233 * beginning timestamp set at the begining of the
4234 * trace chunk instead of only creating an empty
4235 * packet when the trace chunk is stopped.
4237 * This indicates to the viewers that the stream
4238 * was being recorded, but more importantly it
4239 * allows viewers to determine a useable trace
4242 * This presents a problem in the case where the
4243 * ring-buffer is completely full.
4245 * Consider the following scenario:
4246 * - The consumption of data is slow (slow network,
4248 * - The ring buffer is full,
4249 * - A rotation is initiated,
4250 * - The flush below does nothing (no space left to
4251 * open a new packet),
4252 * - The other streams rotate very soon, and new
4253 * data is produced in the new chunk,
4254 * - This stream completes its rotation long after the
4255 * rotation was initiated
4256 * - The session is stopped before any event can be
4257 * produced in this stream's buffers.
4259 * The resulting trace chunk will have a single packet
4260 * temporaly at the end of the trace chunk for this
4261 * stream making the stream intersection more narrow
4262 * than it should be.
4264 * To work-around this, an empty flush is performed
4265 * after the first consumption of a packet during a
4266 * rotation if open_packet fails. The idea is that
4267 * consuming a packet frees enough space to switch
4268 * packets in this scenario and allows the tracer to
4269 * "stamp" the beginning of the new trace chunk at the
4270 * earliest possible point.
4272 * The packet open is performed after the channel
4273 * rotation to ensure that no attempt to open a packet
4274 * is performed in a stream that has no active trace
4277 ret
= lttng_dynamic_pointer_array_add_pointer(
4278 &streams_packet_to_open
, stream
);
4280 PERROR("Failed to add a stream pointer to array of streams in which to open a packet");
4282 goto end_unlock_stream
;
4286 pthread_mutex_unlock(&stream
->lock
);
4290 if (!is_local_trace
) {
4291 relayd
= consumer_find_relayd(relayd_id
);
4293 ERR("Failed to find relayd %" PRIu64
, relayd_id
);
4295 goto end_unlock_channel
;
4298 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
4299 ret
= relayd_rotate_streams(&relayd
->control_sock
, stream_count
,
4300 rotating_to_new_chunk
? &next_chunk_id
: NULL
,
4301 (const struct relayd_stream_rotation_position
*)
4302 stream_rotation_positions
.buffer
4304 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
4306 ERR("Relayd rotate stream failed. Cleaning up relayd %" PRIu64
,
4307 relayd
->net_seq_idx
);
4308 lttng_consumer_cleanup_relayd(relayd
);
4309 goto end_unlock_channel
;
4313 for (stream_idx
= 0;
4314 stream_idx
< lttng_dynamic_pointer_array_get_count(
4315 &streams_packet_to_open
);
4317 enum consumer_stream_open_packet_status status
;
4319 stream
= (lttng_consumer_stream
*) lttng_dynamic_pointer_array_get_pointer(
4320 &streams_packet_to_open
, stream_idx
);
4322 pthread_mutex_lock(&stream
->lock
);
4323 status
= consumer_stream_open_packet(stream
);
4324 pthread_mutex_unlock(&stream
->lock
);
4326 case CONSUMER_STREAM_OPEN_PACKET_STATUS_OPENED
:
4327 DBG("Opened a packet after a rotation: stream id = %" PRIu64
4328 ", channel name = %s, session id = %" PRIu64
,
4329 stream
->key
, stream
->chan
->name
,
4330 stream
->chan
->session_id
);
4332 case CONSUMER_STREAM_OPEN_PACKET_STATUS_NO_SPACE
:
4334 * Can't open a packet as there is no space left
4335 * in the buffer. A new packet will be opened
4336 * once one has been consumed.
4338 DBG("No space left to open a packet after a rotation: stream id = %" PRIu64
4339 ", channel name = %s, session id = %" PRIu64
,
4340 stream
->key
, stream
->chan
->name
,
4341 stream
->chan
->session_id
);
4343 case CONSUMER_STREAM_OPEN_PACKET_STATUS_ERROR
:
4344 /* Logged by callee. */
4346 goto end_unlock_channel
;
4352 pthread_mutex_unlock(&channel
->lock
);
4357 pthread_mutex_unlock(&stream
->lock
);
4359 pthread_mutex_unlock(&channel
->lock
);
4362 lttng_dynamic_array_reset(&stream_rotation_positions
);
4363 lttng_dynamic_pointer_array_reset(&streams_packet_to_open
);
4368 int consumer_clear_buffer(struct lttng_consumer_stream
*stream
)
4371 unsigned long consumed_pos_before
, consumed_pos_after
;
4373 ret
= lttng_consumer_sample_snapshot_positions(stream
);
4375 ERR("Taking snapshot positions");
4379 ret
= lttng_consumer_get_consumed_snapshot(stream
, &consumed_pos_before
);
4381 ERR("Consumed snapshot position");
4385 switch (the_consumer_data
.type
) {
4386 case LTTNG_CONSUMER_KERNEL
:
4387 ret
= kernctl_buffer_clear(stream
->wait_fd
);
4389 ERR("Failed to clear kernel stream (ret = %d)", ret
);
4393 case LTTNG_CONSUMER32_UST
:
4394 case LTTNG_CONSUMER64_UST
:
4395 ret
= lttng_ustconsumer_clear_buffer(stream
);
4397 ERR("Failed to clear ust stream (ret = %d)", ret
);
4402 ERR("Unknown consumer_data type");
4406 ret
= lttng_consumer_sample_snapshot_positions(stream
);
4408 ERR("Taking snapshot positions");
4411 ret
= lttng_consumer_get_consumed_snapshot(stream
, &consumed_pos_after
);
4413 ERR("Consumed snapshot position");
4416 DBG("clear: before: %lu after: %lu", consumed_pos_before
, consumed_pos_after
);
4422 int consumer_clear_stream(struct lttng_consumer_stream
*stream
)
4426 ret
= consumer_stream_flush_buffer(stream
, 1);
4428 ERR("Failed to flush stream %" PRIu64
" during channel clear",
4430 ret
= LTTCOMM_CONSUMERD_FATAL
;
4434 ret
= consumer_clear_buffer(stream
);
4436 ERR("Failed to clear stream %" PRIu64
" during channel clear",
4438 ret
= LTTCOMM_CONSUMERD_FATAL
;
4442 ret
= LTTCOMM_CONSUMERD_SUCCESS
;
4448 int consumer_clear_unmonitored_channel(struct lttng_consumer_channel
*channel
)
4451 struct lttng_consumer_stream
*stream
;
4454 pthread_mutex_lock(&channel
->lock
);
4455 cds_list_for_each_entry(stream
, &channel
->streams
.head
, send_node
) {
4456 health_code_update();
4457 pthread_mutex_lock(&stream
->lock
);
4458 ret
= consumer_clear_stream(stream
);
4462 pthread_mutex_unlock(&stream
->lock
);
4464 pthread_mutex_unlock(&channel
->lock
);
4469 pthread_mutex_unlock(&stream
->lock
);
4470 pthread_mutex_unlock(&channel
->lock
);
4476 * Check if a stream is ready to be rotated after extracting it.
4478 * Return 1 if it is ready for rotation, 0 if it is not, a negative value on
4479 * error. Stream lock must be held.
4481 int lttng_consumer_stream_is_rotate_ready(struct lttng_consumer_stream
*stream
)
4483 DBG("Check is rotate ready for stream %" PRIu64
4484 " ready %u rotate_position %" PRIu64
4485 " last_sequence_number %" PRIu64
,
4486 stream
->key
, stream
->rotate_ready
,
4487 stream
->rotate_position
, stream
->last_sequence_number
);
4488 if (stream
->rotate_ready
) {
4493 * If packet seq num is unavailable, it means we are interacting
4494 * with a pre-2.8 lttng-modules which does not implement the
4495 * sequence number. Rotation should never be used by sessiond in this
4498 if (stream
->sequence_number_unavailable
) {
4499 ERR("Internal error: rotation used on stream %" PRIu64
4500 " with unavailable sequence number",
4505 if (stream
->rotate_position
== -1ULL ||
4506 stream
->last_sequence_number
== -1ULL) {
4511 * Rotate position not reached yet. The stream rotate position is
4512 * the position of the next packet belonging to the next trace chunk,
4513 * but consumerd considers rotation ready when reaching the last
4514 * packet of the current chunk, hence the "rotate_position - 1".
4517 DBG("Check is rotate ready for stream %" PRIu64
4518 " last_sequence_number %" PRIu64
4519 " rotate_position %" PRIu64
,
4520 stream
->key
, stream
->last_sequence_number
,
4521 stream
->rotate_position
);
4522 if (stream
->last_sequence_number
>= stream
->rotate_position
- 1) {
4530 * Reset the state for a stream after a rotation occurred.
4532 void lttng_consumer_reset_stream_rotate_state(struct lttng_consumer_stream
*stream
)
4534 DBG("lttng_consumer_reset_stream_rotate_state for stream %" PRIu64
,
4536 stream
->rotate_position
= -1ULL;
4537 stream
->rotate_ready
= false;
4541 * Perform the rotation a local stream file.
4544 int rotate_local_stream(struct lttng_consumer_stream
*stream
)
4548 DBG("Rotate local stream: stream key %" PRIu64
", channel key %" PRIu64
,
4551 stream
->tracefile_size_current
= 0;
4552 stream
->tracefile_count_current
= 0;
4554 if (stream
->out_fd
>= 0) {
4555 ret
= close(stream
->out_fd
);
4557 PERROR("Failed to close stream out_fd of channel \"%s\"",
4558 stream
->chan
->name
);
4560 stream
->out_fd
= -1;
4563 if (stream
->index_file
) {
4564 lttng_index_file_put(stream
->index_file
);
4565 stream
->index_file
= NULL
;
4568 if (!stream
->trace_chunk
) {
4572 ret
= consumer_stream_create_output_files(stream
, true);
4578 * Performs the stream rotation for the rotate session feature if needed.
4579 * It must be called with the channel and stream locks held.
4581 * Return 0 on success, a negative number of error.
4583 int lttng_consumer_rotate_stream(struct lttng_consumer_stream
*stream
)
4587 DBG("Consumer rotate stream %" PRIu64
, stream
->key
);
4590 * Update the stream's 'current' chunk to the session's (channel)
4591 * now-current chunk.
4593 lttng_trace_chunk_put(stream
->trace_chunk
);
4594 if (stream
->chan
->trace_chunk
== stream
->trace_chunk
) {
4596 * A channel can be rotated and not have a "next" chunk
4597 * to transition to. In that case, the channel's "current chunk"
4598 * has not been closed yet, but it has not been updated to
4599 * a "next" trace chunk either. Hence, the stream, like its
4600 * parent channel, becomes part of no chunk and can't output
4601 * anything until a new trace chunk is created.
4603 stream
->trace_chunk
= NULL
;
4604 } else if (stream
->chan
->trace_chunk
&&
4605 !lttng_trace_chunk_get(stream
->chan
->trace_chunk
)) {
4606 ERR("Failed to acquire a reference to channel's trace chunk during stream rotation");
4611 * Update the stream's trace chunk to its parent channel's
4612 * current trace chunk.
4614 stream
->trace_chunk
= stream
->chan
->trace_chunk
;
4617 if (stream
->net_seq_idx
== (uint64_t) -1ULL) {
4618 ret
= rotate_local_stream(stream
);
4620 ERR("Failed to rotate stream, ret = %i", ret
);
4625 if (stream
->metadata_flag
&& stream
->trace_chunk
) {
4627 * If the stream has transitioned to a new trace
4628 * chunk, the metadata should be re-dumped to the
4631 * However, it is possible for a stream to transition to
4632 * a "no-chunk" state. This can happen if a rotation
4633 * occurs on an inactive session. In such cases, the metadata
4634 * regeneration will happen when the next trace chunk is
4637 ret
= consumer_metadata_stream_dump(stream
);
4642 lttng_consumer_reset_stream_rotate_state(stream
);
4651 * Rotate all the ready streams now.
4653 * This is especially important for low throughput streams that have already
4654 * been consumed, we cannot wait for their next packet to perform the
4656 * Need to be called with RCU read-side lock held to ensure existence of
4659 * Returns 0 on success, < 0 on error
4661 int lttng_consumer_rotate_ready_streams(struct lttng_consumer_channel
*channel
,
4665 struct lttng_consumer_stream
*stream
;
4666 struct lttng_ht_iter iter
;
4667 struct lttng_ht
*ht
= the_consumer_data
.stream_per_chan_id_ht
;
4669 ASSERT_RCU_READ_LOCKED();
4673 DBG("Consumer rotate ready streams in channel %" PRIu64
, key
);
4675 cds_lfht_for_each_entry_duplicate(ht
->ht
,
4676 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
4677 ht
->match_fct
, &channel
->key
, &iter
.iter
,
4678 stream
, node_channel_id
.node
) {
4679 health_code_update();
4681 pthread_mutex_lock(&stream
->chan
->lock
);
4682 pthread_mutex_lock(&stream
->lock
);
4684 if (!stream
->rotate_ready
) {
4685 pthread_mutex_unlock(&stream
->lock
);
4686 pthread_mutex_unlock(&stream
->chan
->lock
);
4689 DBG("Consumer rotate ready stream %" PRIu64
, stream
->key
);
4691 ret
= lttng_consumer_rotate_stream(stream
);
4692 pthread_mutex_unlock(&stream
->lock
);
4693 pthread_mutex_unlock(&stream
->chan
->lock
);
4706 enum lttcomm_return_code
lttng_consumer_init_command(
4707 struct lttng_consumer_local_data
*ctx
,
4708 const lttng_uuid
& sessiond_uuid
)
4710 enum lttcomm_return_code ret
;
4711 char uuid_str
[LTTNG_UUID_STR_LEN
];
4713 if (ctx
->sessiond_uuid
.is_set
) {
4714 ret
= LTTCOMM_CONSUMERD_ALREADY_SET
;
4718 ctx
->sessiond_uuid
.is_set
= true;
4719 ctx
->sessiond_uuid
.value
= sessiond_uuid
;
4720 ret
= LTTCOMM_CONSUMERD_SUCCESS
;
4721 lttng_uuid_to_str(sessiond_uuid
, uuid_str
);
4722 DBG("Received session daemon UUID: %s", uuid_str
);
4727 enum lttcomm_return_code
lttng_consumer_create_trace_chunk(
4728 const uint64_t *relayd_id
, uint64_t session_id
,
4730 time_t chunk_creation_timestamp
,
4731 const char *chunk_override_name
,
4732 const struct lttng_credentials
*credentials
,
4733 struct lttng_directory_handle
*chunk_directory_handle
)
4736 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
4737 struct lttng_trace_chunk
*created_chunk
= NULL
, *published_chunk
= NULL
;
4738 enum lttng_trace_chunk_status chunk_status
;
4739 char relayd_id_buffer
[MAX_INT_DEC_LEN(*relayd_id
)];
4740 char creation_timestamp_buffer
[ISO8601_STR_LEN
];
4741 const char *relayd_id_str
= "(none)";
4742 const char *creation_timestamp_str
;
4743 struct lttng_ht_iter iter
;
4744 struct lttng_consumer_channel
*channel
;
4747 /* Only used for logging purposes. */
4748 ret
= snprintf(relayd_id_buffer
, sizeof(relayd_id_buffer
),
4749 "%" PRIu64
, *relayd_id
);
4750 if (ret
> 0 && ret
< sizeof(relayd_id_buffer
)) {
4751 relayd_id_str
= relayd_id_buffer
;
4753 relayd_id_str
= "(formatting error)";
4757 /* Local protocol error. */
4758 LTTNG_ASSERT(chunk_creation_timestamp
);
4759 ret
= time_to_iso8601_str(chunk_creation_timestamp
,
4760 creation_timestamp_buffer
,
4761 sizeof(creation_timestamp_buffer
));
4762 creation_timestamp_str
= !ret
? creation_timestamp_buffer
:
4763 "(formatting error)";
4765 DBG("Consumer create trace chunk command: relay_id = %s"
4766 ", session_id = %" PRIu64
", chunk_id = %" PRIu64
4767 ", chunk_override_name = %s"
4768 ", chunk_creation_timestamp = %s",
4769 relayd_id_str
, session_id
, chunk_id
,
4770 chunk_override_name
? : "(none)",
4771 creation_timestamp_str
);
4774 * The trace chunk registry, as used by the consumer daemon, implicitly
4775 * owns the trace chunks. This is only needed in the consumer since
4776 * the consumer has no notion of a session beyond session IDs being
4777 * used to identify other objects.
4779 * The lttng_trace_chunk_registry_publish() call below provides a
4780 * reference which is not released; it implicitly becomes the session
4781 * daemon's reference to the chunk in the consumer daemon.
4783 * The lifetime of trace chunks in the consumer daemon is managed by
4784 * the session daemon through the LTTNG_CONSUMER_CREATE_TRACE_CHUNK
4785 * and LTTNG_CONSUMER_DESTROY_TRACE_CHUNK commands.
4787 created_chunk
= lttng_trace_chunk_create(chunk_id
,
4788 chunk_creation_timestamp
, NULL
);
4789 if (!created_chunk
) {
4790 ERR("Failed to create trace chunk");
4791 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4795 if (chunk_override_name
) {
4796 chunk_status
= lttng_trace_chunk_override_name(created_chunk
,
4797 chunk_override_name
);
4798 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4799 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4804 if (chunk_directory_handle
) {
4805 chunk_status
= lttng_trace_chunk_set_credentials(created_chunk
,
4807 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4808 ERR("Failed to set trace chunk credentials");
4809 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4813 * The consumer daemon has no ownership of the chunk output
4816 chunk_status
= lttng_trace_chunk_set_as_user(created_chunk
,
4817 chunk_directory_handle
);
4818 chunk_directory_handle
= NULL
;
4819 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4820 ERR("Failed to set trace chunk's directory handle");
4821 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4826 published_chunk
= lttng_trace_chunk_registry_publish_chunk(
4827 the_consumer_data
.chunk_registry
, session_id
,
4829 lttng_trace_chunk_put(created_chunk
);
4830 created_chunk
= NULL
;
4831 if (!published_chunk
) {
4832 ERR("Failed to publish trace chunk");
4833 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4838 cds_lfht_for_each_entry_duplicate(
4839 the_consumer_data
.channels_by_session_id_ht
->ht
,
4840 the_consumer_data
.channels_by_session_id_ht
->hash_fct(
4841 &session_id
, lttng_ht_seed
),
4842 the_consumer_data
.channels_by_session_id_ht
->match_fct
,
4843 &session_id
, &iter
.iter
, channel
,
4844 channels_by_session_id_ht_node
.node
) {
4845 ret
= lttng_consumer_channel_set_trace_chunk(channel
,
4849 * Roll-back the creation of this chunk.
4851 * This is important since the session daemon will
4852 * assume that the creation of this chunk failed and
4853 * will never ask for it to be closed, resulting
4854 * in a leak and an inconsistent state for some
4857 enum lttcomm_return_code close_ret
;
4858 char path
[LTTNG_PATH_MAX
];
4860 DBG("Failed to set new trace chunk on existing channels, rolling back");
4861 close_ret
= lttng_consumer_close_trace_chunk(relayd_id
,
4862 session_id
, chunk_id
,
4863 chunk_creation_timestamp
, NULL
,
4865 if (close_ret
!= LTTCOMM_CONSUMERD_SUCCESS
) {
4866 ERR("Failed to roll-back the creation of new chunk: session_id = %" PRIu64
", chunk_id = %" PRIu64
,
4867 session_id
, chunk_id
);
4870 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4876 struct consumer_relayd_sock_pair
*relayd
;
4878 relayd
= consumer_find_relayd(*relayd_id
);
4880 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
4881 ret
= relayd_create_trace_chunk(
4882 &relayd
->control_sock
, published_chunk
);
4883 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
4885 ERR("Failed to find relay daemon socket: relayd_id = %" PRIu64
, *relayd_id
);
4888 if (!relayd
|| ret
) {
4889 enum lttcomm_return_code close_ret
;
4890 char path
[LTTNG_PATH_MAX
];
4892 close_ret
= lttng_consumer_close_trace_chunk(relayd_id
,
4895 chunk_creation_timestamp
,
4897 if (close_ret
!= LTTCOMM_CONSUMERD_SUCCESS
) {
4898 ERR("Failed to roll-back the creation of new chunk: session_id = %" PRIu64
", chunk_id = %" PRIu64
,
4903 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4910 /* Release the reference returned by the "publish" operation. */
4911 lttng_trace_chunk_put(published_chunk
);
4912 lttng_trace_chunk_put(created_chunk
);
4916 enum lttcomm_return_code
lttng_consumer_close_trace_chunk(
4917 const uint64_t *relayd_id
, uint64_t session_id
,
4918 uint64_t chunk_id
, time_t chunk_close_timestamp
,
4919 const enum lttng_trace_chunk_command_type
*close_command
,
4922 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
4923 struct lttng_trace_chunk
*chunk
;
4924 char relayd_id_buffer
[MAX_INT_DEC_LEN(*relayd_id
)];
4925 const char *relayd_id_str
= "(none)";
4926 const char *close_command_name
= "none";
4927 struct lttng_ht_iter iter
;
4928 struct lttng_consumer_channel
*channel
;
4929 enum lttng_trace_chunk_status chunk_status
;
4934 /* Only used for logging purposes. */
4935 ret
= snprintf(relayd_id_buffer
, sizeof(relayd_id_buffer
),
4936 "%" PRIu64
, *relayd_id
);
4937 if (ret
> 0 && ret
< sizeof(relayd_id_buffer
)) {
4938 relayd_id_str
= relayd_id_buffer
;
4940 relayd_id_str
= "(formatting error)";
4943 if (close_command
) {
4944 close_command_name
= lttng_trace_chunk_command_type_get_name(
4948 DBG("Consumer close trace chunk command: relayd_id = %s"
4949 ", session_id = %" PRIu64
", chunk_id = %" PRIu64
4950 ", close command = %s",
4951 relayd_id_str
, session_id
, chunk_id
,
4952 close_command_name
);
4954 chunk
= lttng_trace_chunk_registry_find_chunk(
4955 the_consumer_data
.chunk_registry
, session_id
, chunk_id
);
4957 ERR("Failed to find chunk: session_id = %" PRIu64
4958 ", chunk_id = %" PRIu64
,
4959 session_id
, chunk_id
);
4960 ret_code
= LTTCOMM_CONSUMERD_UNKNOWN_TRACE_CHUNK
;
4964 chunk_status
= lttng_trace_chunk_set_close_timestamp(chunk
,
4965 chunk_close_timestamp
);
4966 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4967 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
4971 if (close_command
) {
4972 chunk_status
= lttng_trace_chunk_set_close_command(
4973 chunk
, *close_command
);
4974 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4975 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
4981 * chunk is now invalid to access as we no longer hold a reference to
4982 * it; it is only kept around to compare it (by address) to the
4983 * current chunk found in the session's channels.
4986 cds_lfht_for_each_entry(the_consumer_data
.channel_ht
->ht
, &iter
.iter
,
4987 channel
, node
.node
) {
4991 * Only change the channel's chunk to NULL if it still
4992 * references the chunk being closed. The channel may
4993 * reference a newer channel in the case of a session
4994 * rotation. When a session rotation occurs, the "next"
4995 * chunk is created before the "current" chunk is closed.
4997 if (channel
->trace_chunk
!= chunk
) {
5000 ret
= lttng_consumer_channel_set_trace_chunk(channel
, NULL
);
5003 * Attempt to close the chunk on as many channels as
5006 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
5012 struct consumer_relayd_sock_pair
*relayd
;
5014 relayd
= consumer_find_relayd(*relayd_id
);
5016 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
5017 ret
= relayd_close_trace_chunk(
5018 &relayd
->control_sock
, chunk
,
5020 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
5022 ERR("Failed to find relay daemon socket: relayd_id = %" PRIu64
,
5026 if (!relayd
|| ret
) {
5027 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
5035 * Release the reference returned by the "find" operation and
5036 * the session daemon's implicit reference to the chunk.
5038 lttng_trace_chunk_put(chunk
);
5039 lttng_trace_chunk_put(chunk
);
5044 enum lttcomm_return_code
lttng_consumer_trace_chunk_exists(
5045 const uint64_t *relayd_id
, uint64_t session_id
,
5049 enum lttcomm_return_code ret_code
;
5050 char relayd_id_buffer
[MAX_INT_DEC_LEN(*relayd_id
)];
5051 const char *relayd_id_str
= "(none)";
5052 const bool is_local_trace
= !relayd_id
;
5053 struct consumer_relayd_sock_pair
*relayd
= NULL
;
5054 bool chunk_exists_local
, chunk_exists_remote
;
5057 /* Only used for logging purposes. */
5058 ret
= snprintf(relayd_id_buffer
, sizeof(relayd_id_buffer
),
5059 "%" PRIu64
, *relayd_id
);
5060 if (ret
> 0 && ret
< sizeof(relayd_id_buffer
)) {
5061 relayd_id_str
= relayd_id_buffer
;
5063 relayd_id_str
= "(formatting error)";
5067 DBG("Consumer trace chunk exists command: relayd_id = %s"
5068 ", chunk_id = %" PRIu64
, relayd_id_str
,
5070 ret
= lttng_trace_chunk_registry_chunk_exists(
5071 the_consumer_data
.chunk_registry
, session_id
, chunk_id
,
5072 &chunk_exists_local
);
5074 /* Internal error. */
5075 ERR("Failed to query the existence of a trace chunk");
5076 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
5079 DBG("Trace chunk %s locally",
5080 chunk_exists_local
? "exists" : "does not exist");
5081 if (chunk_exists_local
) {
5082 ret_code
= LTTCOMM_CONSUMERD_TRACE_CHUNK_EXISTS_LOCAL
;
5084 } else if (is_local_trace
) {
5085 ret_code
= LTTCOMM_CONSUMERD_UNKNOWN_TRACE_CHUNK
;
5090 relayd
= consumer_find_relayd(*relayd_id
);
5092 ERR("Failed to find relayd %" PRIu64
, *relayd_id
);
5093 ret_code
= LTTCOMM_CONSUMERD_INVALID_PARAMETERS
;
5094 goto end_rcu_unlock
;
5096 DBG("Looking up existence of trace chunk on relay daemon");
5097 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
5098 ret
= relayd_trace_chunk_exists(&relayd
->control_sock
, chunk_id
,
5099 &chunk_exists_remote
);
5100 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
5102 ERR("Failed to look-up the existence of trace chunk on relay daemon");
5103 ret_code
= LTTCOMM_CONSUMERD_RELAYD_FAIL
;
5104 goto end_rcu_unlock
;
5107 ret_code
= chunk_exists_remote
?
5108 LTTCOMM_CONSUMERD_TRACE_CHUNK_EXISTS_REMOTE
:
5109 LTTCOMM_CONSUMERD_UNKNOWN_TRACE_CHUNK
;
5110 DBG("Trace chunk %s on relay daemon",
5111 chunk_exists_remote
? "exists" : "does not exist");
5120 int consumer_clear_monitored_channel(struct lttng_consumer_channel
*channel
)
5122 struct lttng_ht
*ht
;
5123 struct lttng_consumer_stream
*stream
;
5124 struct lttng_ht_iter iter
;
5127 ht
= the_consumer_data
.stream_per_chan_id_ht
;
5130 cds_lfht_for_each_entry_duplicate(ht
->ht
,
5131 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
5132 ht
->match_fct
, &channel
->key
,
5133 &iter
.iter
, stream
, node_channel_id
.node
) {
5135 * Protect against teardown with mutex.
5137 pthread_mutex_lock(&stream
->lock
);
5138 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
5141 ret
= consumer_clear_stream(stream
);
5146 pthread_mutex_unlock(&stream
->lock
);
5149 return LTTCOMM_CONSUMERD_SUCCESS
;
5152 pthread_mutex_unlock(&stream
->lock
);
5157 int lttng_consumer_clear_channel(struct lttng_consumer_channel
*channel
)
5161 DBG("Consumer clear channel %" PRIu64
, channel
->key
);
5163 if (channel
->type
== CONSUMER_CHANNEL_TYPE_METADATA
) {
5165 * Nothing to do for the metadata channel/stream.
5166 * Snapshot mechanism already take care of the metadata
5167 * handling/generation, and monitored channels only need to
5168 * have their data stream cleared..
5170 ret
= LTTCOMM_CONSUMERD_SUCCESS
;
5174 if (!channel
->monitor
) {
5175 ret
= consumer_clear_unmonitored_channel(channel
);
5177 ret
= consumer_clear_monitored_channel(channel
);
5183 enum lttcomm_return_code
lttng_consumer_open_channel_packets(
5184 struct lttng_consumer_channel
*channel
)
5186 struct lttng_consumer_stream
*stream
;
5187 enum lttcomm_return_code ret
= LTTCOMM_CONSUMERD_SUCCESS
;
5189 if (channel
->metadata_stream
) {
5190 ERR("Open channel packets command attempted on a metadata channel");
5191 ret
= LTTCOMM_CONSUMERD_INVALID_PARAMETERS
;
5196 cds_list_for_each_entry(stream
, &channel
->streams
.head
, send_node
) {
5197 enum consumer_stream_open_packet_status status
;
5199 pthread_mutex_lock(&stream
->lock
);
5200 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
5204 status
= consumer_stream_open_packet(stream
);
5206 case CONSUMER_STREAM_OPEN_PACKET_STATUS_OPENED
:
5207 DBG("Opened a packet in \"open channel packets\" command: stream id = %" PRIu64
5208 ", channel name = %s, session id = %" PRIu64
,
5209 stream
->key
, stream
->chan
->name
,
5210 stream
->chan
->session_id
);
5211 stream
->opened_packet_in_current_trace_chunk
= true;
5213 case CONSUMER_STREAM_OPEN_PACKET_STATUS_NO_SPACE
:
5214 DBG("No space left to open a packet in \"open channel packets\" command: stream id = %" PRIu64
5215 ", channel name = %s, session id = %" PRIu64
,
5216 stream
->key
, stream
->chan
->name
,
5217 stream
->chan
->session_id
);
5219 case CONSUMER_STREAM_OPEN_PACKET_STATUS_ERROR
:
5221 * Only unexpected internal errors can lead to this
5222 * failing. Report an unknown error.
5224 ERR("Failed to flush empty buffer in \"open channel packets\" command: stream id = %" PRIu64
5225 ", channel id = %" PRIu64
5226 ", channel name = %s"
5227 ", session id = %" PRIu64
,
5228 stream
->key
, channel
->key
,
5229 channel
->name
, channel
->session_id
);
5230 ret
= LTTCOMM_CONSUMERD_UNKNOWN_ERROR
;
5237 pthread_mutex_unlock(&stream
->lock
);
5246 pthread_mutex_unlock(&stream
->lock
);
5247 goto end_rcu_unlock
;
5250 void lttng_consumer_sigbus_handle(void *addr
)
5252 lttng_ustconsumer_sigbus_handle(addr
);