2 * Copyright (C) 2011 Julien Desfossez <julien.desfossez@polymtl.ca>
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
10 #include "common/index/ctf-index.h"
17 #include <sys/socket.h>
18 #include <sys/types.h>
23 #include <bin/lttng-consumerd/health-consumerd.h>
24 #include <common/common.h>
25 #include <common/utils.h>
26 #include <common/time.h>
27 #include <common/compat/poll.h>
28 #include <common/compat/endian.h>
29 #include <common/index/index.h>
30 #include <common/kernel-ctl/kernel-ctl.h>
31 #include <common/sessiond-comm/relayd.h>
32 #include <common/sessiond-comm/sessiond-comm.h>
33 #include <common/kernel-consumer/kernel-consumer.h>
34 #include <common/relayd/relayd.h>
35 #include <common/ust-consumer/ust-consumer.h>
36 #include <common/consumer/consumer-timer.h>
37 #include <common/consumer/consumer.h>
38 #include <common/consumer/consumer-stream.h>
39 #include <common/consumer/consumer-testpoint.h>
40 #include <common/align.h>
41 #include <common/consumer/consumer-metadata-cache.h>
42 #include <common/trace-chunk.h>
43 #include <common/trace-chunk-registry.h>
44 #include <common/string-utils/format.h>
45 #include <common/dynamic-array.h>
47 struct lttng_consumer_global_data the_consumer_data
= {
50 .type
= LTTNG_CONSUMER_UNKNOWN
,
53 enum consumer_channel_action
{
56 CONSUMER_CHANNEL_QUIT
,
59 struct consumer_channel_msg
{
60 enum consumer_channel_action action
;
61 struct lttng_consumer_channel
*chan
; /* add */
62 uint64_t key
; /* del */
65 /* Flag used to temporarily pause data consumption from testpoints. */
66 int data_consumption_paused
;
69 * Flag to inform the polling thread to quit when all fd hung up. Updated by
70 * the consumer_thread_receive_fds when it notices that all fds has hung up.
71 * Also updated by the signal handler (consumer_should_exit()). Read by the
77 * Global hash table containing respectively metadata and data streams. The
78 * stream element in this ht should only be updated by the metadata poll thread
79 * for the metadata and the data poll thread for the data.
81 static struct lttng_ht
*metadata_ht
;
82 static struct lttng_ht
*data_ht
;
84 static const char *get_consumer_domain(void)
86 switch (the_consumer_data
.type
) {
87 case LTTNG_CONSUMER_KERNEL
:
88 return DEFAULT_KERNEL_TRACE_DIR
;
89 case LTTNG_CONSUMER64_UST
:
91 case LTTNG_CONSUMER32_UST
:
92 return DEFAULT_UST_TRACE_DIR
;
99 * Notify a thread lttng pipe to poll back again. This usually means that some
100 * global state has changed so we just send back the thread in a poll wait
103 static void notify_thread_lttng_pipe(struct lttng_pipe
*pipe
)
105 struct lttng_consumer_stream
*null_stream
= NULL
;
109 (void) lttng_pipe_write(pipe
, &null_stream
, sizeof(null_stream
));
112 static void notify_health_quit_pipe(int *pipe
)
116 ret
= lttng_write(pipe
[1], "4", 1);
118 PERROR("write consumer health quit");
122 static void notify_channel_pipe(struct lttng_consumer_local_data
*ctx
,
123 struct lttng_consumer_channel
*chan
,
125 enum consumer_channel_action action
)
127 struct consumer_channel_msg msg
;
130 memset(&msg
, 0, sizeof(msg
));
135 ret
= lttng_write(ctx
->consumer_channel_pipe
[1], &msg
, sizeof(msg
));
136 if (ret
< sizeof(msg
)) {
137 PERROR("notify_channel_pipe write error");
141 void notify_thread_del_channel(struct lttng_consumer_local_data
*ctx
,
144 notify_channel_pipe(ctx
, NULL
, key
, CONSUMER_CHANNEL_DEL
);
147 static int read_channel_pipe(struct lttng_consumer_local_data
*ctx
,
148 struct lttng_consumer_channel
**chan
,
150 enum consumer_channel_action
*action
)
152 struct consumer_channel_msg msg
;
155 ret
= lttng_read(ctx
->consumer_channel_pipe
[0], &msg
, sizeof(msg
));
156 if (ret
< sizeof(msg
)) {
160 *action
= msg
.action
;
168 * Cleanup the stream list of a channel. Those streams are not yet globally
171 static void clean_channel_stream_list(struct lttng_consumer_channel
*channel
)
173 struct lttng_consumer_stream
*stream
, *stmp
;
175 LTTNG_ASSERT(channel
);
177 /* Delete streams that might have been left in the stream list. */
178 cds_list_for_each_entry_safe(stream
, stmp
, &channel
->streams
.head
,
180 cds_list_del(&stream
->send_node
);
182 * Once a stream is added to this list, the buffers were created so we
183 * have a guarantee that this call will succeed. Setting the monitor
184 * mode to 0 so we don't lock nor try to delete the stream from the
188 consumer_stream_destroy(stream
, NULL
);
193 * Find a stream. The consumer_data.lock must be locked during this
196 static struct lttng_consumer_stream
*find_stream(uint64_t key
,
199 struct lttng_ht_iter iter
;
200 struct lttng_ht_node_u64
*node
;
201 struct lttng_consumer_stream
*stream
= NULL
;
205 /* -1ULL keys are lookup failures */
206 if (key
== (uint64_t) -1ULL) {
212 lttng_ht_lookup(ht
, &key
, &iter
);
213 node
= lttng_ht_iter_get_node_u64(&iter
);
215 stream
= caa_container_of(node
, struct lttng_consumer_stream
, node
);
223 static void steal_stream_key(uint64_t key
, struct lttng_ht
*ht
)
225 struct lttng_consumer_stream
*stream
;
228 stream
= find_stream(key
, ht
);
230 stream
->key
= (uint64_t) -1ULL;
232 * We don't want the lookup to match, but we still need
233 * to iterate on this stream when iterating over the hash table. Just
234 * change the node key.
236 stream
->node
.key
= (uint64_t) -1ULL;
242 * Return a channel object for the given key.
244 * RCU read side lock MUST be acquired before calling this function and
245 * protects the channel ptr.
247 struct lttng_consumer_channel
*consumer_find_channel(uint64_t key
)
249 struct lttng_ht_iter iter
;
250 struct lttng_ht_node_u64
*node
;
251 struct lttng_consumer_channel
*channel
= NULL
;
253 /* -1ULL keys are lookup failures */
254 if (key
== (uint64_t) -1ULL) {
258 lttng_ht_lookup(the_consumer_data
.channel_ht
, &key
, &iter
);
259 node
= lttng_ht_iter_get_node_u64(&iter
);
261 channel
= caa_container_of(node
, struct lttng_consumer_channel
, node
);
268 * There is a possibility that the consumer does not have enough time between
269 * the close of the channel on the session daemon and the cleanup in here thus
270 * once we have a channel add with an existing key, we know for sure that this
271 * channel will eventually get cleaned up by all streams being closed.
273 * This function just nullifies the already existing channel key.
275 static void steal_channel_key(uint64_t key
)
277 struct lttng_consumer_channel
*channel
;
280 channel
= consumer_find_channel(key
);
282 channel
->key
= (uint64_t) -1ULL;
284 * We don't want the lookup to match, but we still need to iterate on
285 * this channel when iterating over the hash table. Just change the
288 channel
->node
.key
= (uint64_t) -1ULL;
293 static void free_channel_rcu(struct rcu_head
*head
)
295 struct lttng_ht_node_u64
*node
=
296 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
297 struct lttng_consumer_channel
*channel
=
298 caa_container_of(node
, struct lttng_consumer_channel
, node
);
300 switch (the_consumer_data
.type
) {
301 case LTTNG_CONSUMER_KERNEL
:
303 case LTTNG_CONSUMER32_UST
:
304 case LTTNG_CONSUMER64_UST
:
305 lttng_ustconsumer_free_channel(channel
);
308 ERR("Unknown consumer_data type");
315 * RCU protected relayd socket pair free.
317 static void free_relayd_rcu(struct rcu_head
*head
)
319 struct lttng_ht_node_u64
*node
=
320 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
321 struct consumer_relayd_sock_pair
*relayd
=
322 caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
325 * Close all sockets. This is done in the call RCU since we don't want the
326 * socket fds to be reassigned thus potentially creating bad state of the
329 * We do not have to lock the control socket mutex here since at this stage
330 * there is no one referencing to this relayd object.
332 (void) relayd_close(&relayd
->control_sock
);
333 (void) relayd_close(&relayd
->data_sock
);
335 pthread_mutex_destroy(&relayd
->ctrl_sock_mutex
);
340 * Destroy and free relayd socket pair object.
342 void consumer_destroy_relayd(struct consumer_relayd_sock_pair
*relayd
)
345 struct lttng_ht_iter iter
;
347 if (relayd
== NULL
) {
351 DBG("Consumer destroy and close relayd socket pair");
353 iter
.iter
.node
= &relayd
->node
.node
;
354 ret
= lttng_ht_del(the_consumer_data
.relayd_ht
, &iter
);
356 /* We assume the relayd is being or is destroyed */
360 /* RCU free() call */
361 call_rcu(&relayd
->node
.head
, free_relayd_rcu
);
365 * Remove a channel from the global list protected by a mutex. This function is
366 * also responsible for freeing its data structures.
368 void consumer_del_channel(struct lttng_consumer_channel
*channel
)
370 struct lttng_ht_iter iter
;
372 DBG("Consumer delete channel key %" PRIu64
, channel
->key
);
374 pthread_mutex_lock(&the_consumer_data
.lock
);
375 pthread_mutex_lock(&channel
->lock
);
377 /* Destroy streams that might have been left in the stream list. */
378 clean_channel_stream_list(channel
);
380 if (channel
->live_timer_enabled
== 1) {
381 consumer_timer_live_stop(channel
);
383 if (channel
->monitor_timer_enabled
== 1) {
384 consumer_timer_monitor_stop(channel
);
387 switch (the_consumer_data
.type
) {
388 case LTTNG_CONSUMER_KERNEL
:
390 case LTTNG_CONSUMER32_UST
:
391 case LTTNG_CONSUMER64_UST
:
392 lttng_ustconsumer_del_channel(channel
);
395 ERR("Unknown consumer_data type");
400 lttng_trace_chunk_put(channel
->trace_chunk
);
401 channel
->trace_chunk
= NULL
;
403 if (channel
->is_published
) {
407 iter
.iter
.node
= &channel
->node
.node
;
408 ret
= lttng_ht_del(the_consumer_data
.channel_ht
, &iter
);
411 iter
.iter
.node
= &channel
->channels_by_session_id_ht_node
.node
;
412 ret
= lttng_ht_del(the_consumer_data
.channels_by_session_id_ht
,
418 channel
->is_deleted
= true;
419 call_rcu(&channel
->node
.head
, free_channel_rcu
);
421 pthread_mutex_unlock(&channel
->lock
);
422 pthread_mutex_unlock(&the_consumer_data
.lock
);
426 * Iterate over the relayd hash table and destroy each element. Finally,
427 * destroy the whole hash table.
429 static void cleanup_relayd_ht(void)
431 struct lttng_ht_iter iter
;
432 struct consumer_relayd_sock_pair
*relayd
;
436 cds_lfht_for_each_entry(the_consumer_data
.relayd_ht
->ht
, &iter
.iter
,
438 consumer_destroy_relayd(relayd
);
443 lttng_ht_destroy(the_consumer_data
.relayd_ht
);
447 * Update the end point status of all streams having the given network sequence
448 * index (relayd index).
450 * It's atomically set without having the stream mutex locked which is fine
451 * because we handle the write/read race with a pipe wakeup for each thread.
453 static void update_endpoint_status_by_netidx(uint64_t net_seq_idx
,
454 enum consumer_endpoint_status status
)
456 struct lttng_ht_iter iter
;
457 struct lttng_consumer_stream
*stream
;
459 DBG("Consumer set delete flag on stream by idx %" PRIu64
, net_seq_idx
);
463 /* Let's begin with metadata */
464 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
465 if (stream
->net_seq_idx
== net_seq_idx
) {
466 uatomic_set(&stream
->endpoint_status
, status
);
467 DBG("Delete flag set to metadata stream %d", stream
->wait_fd
);
471 /* Follow up by the data streams */
472 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
473 if (stream
->net_seq_idx
== net_seq_idx
) {
474 uatomic_set(&stream
->endpoint_status
, status
);
475 DBG("Delete flag set to data stream %d", stream
->wait_fd
);
482 * Cleanup a relayd object by flagging every associated streams for deletion,
483 * destroying the object meaning removing it from the relayd hash table,
484 * closing the sockets and freeing the memory in a RCU call.
486 * If a local data context is available, notify the threads that the streams'
487 * state have changed.
489 void lttng_consumer_cleanup_relayd(struct consumer_relayd_sock_pair
*relayd
)
493 LTTNG_ASSERT(relayd
);
495 DBG("Cleaning up relayd object ID %"PRIu64
, relayd
->net_seq_idx
);
497 /* Save the net sequence index before destroying the object */
498 netidx
= relayd
->net_seq_idx
;
501 * Delete the relayd from the relayd hash table, close the sockets and free
502 * the object in a RCU call.
504 consumer_destroy_relayd(relayd
);
506 /* Set inactive endpoint to all streams */
507 update_endpoint_status_by_netidx(netidx
, CONSUMER_ENDPOINT_INACTIVE
);
510 * With a local data context, notify the threads that the streams' state
511 * have changed. The write() action on the pipe acts as an "implicit"
512 * memory barrier ordering the updates of the end point status from the
513 * read of this status which happens AFTER receiving this notify.
515 notify_thread_lttng_pipe(relayd
->ctx
->consumer_data_pipe
);
516 notify_thread_lttng_pipe(relayd
->ctx
->consumer_metadata_pipe
);
520 * Flag a relayd socket pair for destruction. Destroy it if the refcount
523 * RCU read side lock MUST be aquired before calling this function.
525 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair
*relayd
)
527 LTTNG_ASSERT(relayd
);
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
);
641 lttng_ht_lookup(the_consumer_data
.relayd_ht
, &relayd
->net_seq_idx
,
643 node
= lttng_ht_iter_get_node_u64(&iter
);
647 lttng_ht_add_unique_u64(the_consumer_data
.relayd_ht
, &relayd
->node
);
654 * Allocate and return a consumer relayd socket.
656 static struct consumer_relayd_sock_pair
*consumer_allocate_relayd_sock_pair(
657 uint64_t net_seq_idx
)
659 struct consumer_relayd_sock_pair
*obj
= NULL
;
661 /* net sequence index of -1 is a failure */
662 if (net_seq_idx
== (uint64_t) -1ULL) {
666 obj
= zmalloc(sizeof(struct consumer_relayd_sock_pair
));
668 PERROR("zmalloc relayd sock");
672 obj
->net_seq_idx
= net_seq_idx
;
674 obj
->destroy_flag
= 0;
675 obj
->control_sock
.sock
.fd
= -1;
676 obj
->data_sock
.sock
.fd
= -1;
677 lttng_ht_node_init_u64(&obj
->node
, obj
->net_seq_idx
);
678 pthread_mutex_init(&obj
->ctrl_sock_mutex
, NULL
);
685 * Find a relayd socket pair in the global consumer data.
687 * Return the object if found else NULL.
688 * RCU read-side lock must be held across this call and while using the
691 struct consumer_relayd_sock_pair
*consumer_find_relayd(uint64_t key
)
693 struct lttng_ht_iter iter
;
694 struct lttng_ht_node_u64
*node
;
695 struct consumer_relayd_sock_pair
*relayd
= NULL
;
697 /* Negative keys are lookup failures */
698 if (key
== (uint64_t) -1ULL) {
702 lttng_ht_lookup(the_consumer_data
.relayd_ht
, &key
, &iter
);
703 node
= lttng_ht_iter_get_node_u64(&iter
);
705 relayd
= caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
713 * Find a relayd and send the stream
715 * Returns 0 on success, < 0 on error
717 int consumer_send_relayd_stream(struct lttng_consumer_stream
*stream
,
721 struct consumer_relayd_sock_pair
*relayd
;
723 LTTNG_ASSERT(stream
);
724 LTTNG_ASSERT(stream
->net_seq_idx
!= -1ULL);
727 /* The stream is not metadata. Get relayd reference if exists. */
729 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
730 if (relayd
!= NULL
) {
731 /* Add stream on the relayd */
732 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
733 ret
= relayd_add_stream(&relayd
->control_sock
, stream
->name
,
734 get_consumer_domain(), path
, &stream
->relayd_stream_id
,
735 stream
->chan
->tracefile_size
,
736 stream
->chan
->tracefile_count
,
737 stream
->trace_chunk
);
738 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
740 ERR("Relayd add stream failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
741 lttng_consumer_cleanup_relayd(relayd
);
745 uatomic_inc(&relayd
->refcount
);
746 stream
->sent_to_relayd
= 1;
748 ERR("Stream %" PRIu64
" relayd ID %" PRIu64
" unknown. Can't send it.",
749 stream
->key
, stream
->net_seq_idx
);
754 DBG("Stream %s with key %" PRIu64
" sent to relayd id %" PRIu64
,
755 stream
->name
, stream
->key
, stream
->net_seq_idx
);
763 * Find a relayd and send the streams sent message
765 * Returns 0 on success, < 0 on error
767 int consumer_send_relayd_streams_sent(uint64_t net_seq_idx
)
770 struct consumer_relayd_sock_pair
*relayd
;
772 LTTNG_ASSERT(net_seq_idx
!= -1ULL);
774 /* The stream is not metadata. Get relayd reference if exists. */
776 relayd
= consumer_find_relayd(net_seq_idx
);
777 if (relayd
!= NULL
) {
778 /* Add stream on the relayd */
779 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
780 ret
= relayd_streams_sent(&relayd
->control_sock
);
781 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
783 ERR("Relayd streams sent failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
784 lttng_consumer_cleanup_relayd(relayd
);
788 ERR("Relayd ID %" PRIu64
" unknown. Can't send streams_sent.",
795 DBG("All streams sent relayd id %" PRIu64
, net_seq_idx
);
803 * Find a relayd and close the stream
805 void close_relayd_stream(struct lttng_consumer_stream
*stream
)
807 struct consumer_relayd_sock_pair
*relayd
;
809 /* The stream is not metadata. Get relayd reference if exists. */
811 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
813 consumer_stream_relayd_close(stream
, relayd
);
819 * Handle stream for relayd transmission if the stream applies for network
820 * streaming where the net sequence index is set.
822 * Return destination file descriptor or negative value on error.
824 static int write_relayd_stream_header(struct lttng_consumer_stream
*stream
,
825 size_t data_size
, unsigned long padding
,
826 struct consumer_relayd_sock_pair
*relayd
)
829 struct lttcomm_relayd_data_hdr data_hdr
;
832 LTTNG_ASSERT(stream
);
833 LTTNG_ASSERT(relayd
);
835 /* Reset data header */
836 memset(&data_hdr
, 0, sizeof(data_hdr
));
838 if (stream
->metadata_flag
) {
839 /* Caller MUST acquire the relayd control socket lock */
840 ret
= relayd_send_metadata(&relayd
->control_sock
, data_size
);
845 /* Metadata are always sent on the control socket. */
846 outfd
= relayd
->control_sock
.sock
.fd
;
848 /* Set header with stream information */
849 data_hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
850 data_hdr
.data_size
= htobe32(data_size
);
851 data_hdr
.padding_size
= htobe32(padding
);
854 * Note that net_seq_num below is assigned with the *current* value of
855 * next_net_seq_num and only after that the next_net_seq_num will be
856 * increment. This is why when issuing a command on the relayd using
857 * this next value, 1 should always be substracted in order to compare
858 * the last seen sequence number on the relayd side to the last sent.
860 data_hdr
.net_seq_num
= htobe64(stream
->next_net_seq_num
);
861 /* Other fields are zeroed previously */
863 ret
= relayd_send_data_hdr(&relayd
->data_sock
, &data_hdr
,
869 ++stream
->next_net_seq_num
;
871 /* Set to go on data socket */
872 outfd
= relayd
->data_sock
.sock
.fd
;
880 * Write a character on the metadata poll pipe to wake the metadata thread.
881 * Returns 0 on success, -1 on error.
883 int consumer_metadata_wakeup_pipe(const struct lttng_consumer_channel
*channel
)
887 DBG("Waking up metadata poll thread (writing to pipe): channel name = '%s'",
889 if (channel
->monitor
&& channel
->metadata_stream
) {
890 const char dummy
= 'c';
891 const ssize_t write_ret
= lttng_write(
892 channel
->metadata_stream
->ust_metadata_poll_pipe
[1],
896 if (errno
== EWOULDBLOCK
) {
898 * This is fine, the metadata poll thread
899 * is having a hard time keeping-up, but
900 * it will eventually wake-up and consume
901 * the available data.
905 PERROR("Failed to write to UST metadata pipe while attempting to wake-up the metadata poll thread");
917 * Trigger a dump of the metadata content. Following/during the succesful
918 * completion of this call, the metadata poll thread will start receiving
919 * metadata packets to consume.
921 * The caller must hold the channel and stream locks.
924 int consumer_metadata_stream_dump(struct lttng_consumer_stream
*stream
)
928 ASSERT_LOCKED(stream
->chan
->lock
);
929 ASSERT_LOCKED(stream
->lock
);
930 LTTNG_ASSERT(stream
->metadata_flag
);
931 LTTNG_ASSERT(stream
->chan
->trace_chunk
);
933 switch (the_consumer_data
.type
) {
934 case LTTNG_CONSUMER_KERNEL
:
936 * Reset the position of what has been read from the
937 * metadata cache to 0 so we can dump it again.
939 ret
= kernctl_metadata_cache_dump(stream
->wait_fd
);
941 case LTTNG_CONSUMER32_UST
:
942 case LTTNG_CONSUMER64_UST
:
944 * Reset the position pushed from the metadata cache so it
945 * will write from the beginning on the next push.
947 stream
->ust_metadata_pushed
= 0;
948 ret
= consumer_metadata_wakeup_pipe(stream
->chan
);
951 ERR("Unknown consumer_data type");
955 ERR("Failed to dump the metadata cache");
961 int lttng_consumer_channel_set_trace_chunk(
962 struct lttng_consumer_channel
*channel
,
963 struct lttng_trace_chunk
*new_trace_chunk
)
965 pthread_mutex_lock(&channel
->lock
);
966 if (channel
->is_deleted
) {
968 * The channel has been logically deleted and should no longer
969 * be used. It has released its reference to its current trace
970 * chunk and should not acquire a new one.
972 * Return success as there is nothing for the caller to do.
978 * The acquisition of the reference cannot fail (barring
979 * a severe internal error) since a reference to the published
980 * chunk is already held by the caller.
982 if (new_trace_chunk
) {
983 const bool acquired_reference
= lttng_trace_chunk_get(
986 LTTNG_ASSERT(acquired_reference
);
989 lttng_trace_chunk_put(channel
->trace_chunk
);
990 channel
->trace_chunk
= new_trace_chunk
;
992 pthread_mutex_unlock(&channel
->lock
);
997 * Allocate and return a new lttng_consumer_channel object using the given key
998 * to initialize the hash table node.
1000 * On error, return NULL.
1002 struct lttng_consumer_channel
*consumer_allocate_channel(uint64_t key
,
1003 uint64_t session_id
,
1004 const uint64_t *chunk_id
,
1005 const char *pathname
,
1008 enum lttng_event_output output
,
1009 uint64_t tracefile_size
,
1010 uint64_t tracefile_count
,
1011 uint64_t session_id_per_pid
,
1012 unsigned int monitor
,
1013 unsigned int live_timer_interval
,
1014 bool is_in_live_session
,
1015 const char *root_shm_path
,
1016 const char *shm_path
)
1018 struct lttng_consumer_channel
*channel
= NULL
;
1019 struct lttng_trace_chunk
*trace_chunk
= NULL
;
1022 trace_chunk
= lttng_trace_chunk_registry_find_chunk(
1023 the_consumer_data
.chunk_registry
, session_id
,
1026 ERR("Failed to find trace chunk reference during creation of channel");
1031 channel
= zmalloc(sizeof(*channel
));
1032 if (channel
== NULL
) {
1033 PERROR("malloc struct lttng_consumer_channel");
1038 channel
->refcount
= 0;
1039 channel
->session_id
= session_id
;
1040 channel
->session_id_per_pid
= session_id_per_pid
;
1041 channel
->relayd_id
= relayd_id
;
1042 channel
->tracefile_size
= tracefile_size
;
1043 channel
->tracefile_count
= tracefile_count
;
1044 channel
->monitor
= monitor
;
1045 channel
->live_timer_interval
= live_timer_interval
;
1046 channel
->is_live
= is_in_live_session
;
1047 pthread_mutex_init(&channel
->lock
, NULL
);
1048 pthread_mutex_init(&channel
->timer_lock
, NULL
);
1051 case LTTNG_EVENT_SPLICE
:
1052 channel
->output
= CONSUMER_CHANNEL_SPLICE
;
1054 case LTTNG_EVENT_MMAP
:
1055 channel
->output
= CONSUMER_CHANNEL_MMAP
;
1065 * In monitor mode, the streams associated with the channel will be put in
1066 * a special list ONLY owned by this channel. So, the refcount is set to 1
1067 * here meaning that the channel itself has streams that are referenced.
1069 * On a channel deletion, once the channel is no longer visible, the
1070 * refcount is decremented and checked for a zero value to delete it. With
1071 * streams in no monitor mode, it will now be safe to destroy the channel.
1073 if (!channel
->monitor
) {
1074 channel
->refcount
= 1;
1077 strncpy(channel
->pathname
, pathname
, sizeof(channel
->pathname
));
1078 channel
->pathname
[sizeof(channel
->pathname
) - 1] = '\0';
1080 strncpy(channel
->name
, name
, sizeof(channel
->name
));
1081 channel
->name
[sizeof(channel
->name
) - 1] = '\0';
1083 if (root_shm_path
) {
1084 strncpy(channel
->root_shm_path
, root_shm_path
, sizeof(channel
->root_shm_path
));
1085 channel
->root_shm_path
[sizeof(channel
->root_shm_path
) - 1] = '\0';
1088 strncpy(channel
->shm_path
, shm_path
, sizeof(channel
->shm_path
));
1089 channel
->shm_path
[sizeof(channel
->shm_path
) - 1] = '\0';
1092 lttng_ht_node_init_u64(&channel
->node
, channel
->key
);
1093 lttng_ht_node_init_u64(&channel
->channels_by_session_id_ht_node
,
1094 channel
->session_id
);
1096 channel
->wait_fd
= -1;
1097 CDS_INIT_LIST_HEAD(&channel
->streams
.head
);
1100 int ret
= lttng_consumer_channel_set_trace_chunk(channel
,
1107 DBG("Allocated channel (key %" PRIu64
")", channel
->key
);
1110 lttng_trace_chunk_put(trace_chunk
);
1113 consumer_del_channel(channel
);
1119 * Add a channel to the global list protected by a mutex.
1121 * Always return 0 indicating success.
1123 int consumer_add_channel(struct lttng_consumer_channel
*channel
,
1124 struct lttng_consumer_local_data
*ctx
)
1126 pthread_mutex_lock(&the_consumer_data
.lock
);
1127 pthread_mutex_lock(&channel
->lock
);
1128 pthread_mutex_lock(&channel
->timer_lock
);
1131 * This gives us a guarantee that the channel we are about to add to the
1132 * channel hash table will be unique. See this function comment on the why
1133 * we need to steel the channel key at this stage.
1135 steal_channel_key(channel
->key
);
1138 lttng_ht_add_unique_u64(the_consumer_data
.channel_ht
, &channel
->node
);
1139 lttng_ht_add_u64(the_consumer_data
.channels_by_session_id_ht
,
1140 &channel
->channels_by_session_id_ht_node
);
1142 channel
->is_published
= true;
1144 pthread_mutex_unlock(&channel
->timer_lock
);
1145 pthread_mutex_unlock(&channel
->lock
);
1146 pthread_mutex_unlock(&the_consumer_data
.lock
);
1148 if (channel
->wait_fd
!= -1 && channel
->type
== CONSUMER_CHANNEL_TYPE_DATA
) {
1149 notify_channel_pipe(ctx
, channel
, -1, CONSUMER_CHANNEL_ADD
);
1156 * Allocate the pollfd structure and the local view of the out fds to avoid
1157 * doing a lookup in the linked list and concurrency issues when writing is
1158 * needed. Called with consumer_data.lock held.
1160 * Returns the number of fds in the structures.
1162 static int update_poll_array(struct lttng_consumer_local_data
*ctx
,
1163 struct pollfd
**pollfd
, struct lttng_consumer_stream
**local_stream
,
1164 struct lttng_ht
*ht
, int *nb_inactive_fd
)
1167 struct lttng_ht_iter iter
;
1168 struct lttng_consumer_stream
*stream
;
1172 LTTNG_ASSERT(pollfd
);
1173 LTTNG_ASSERT(local_stream
);
1175 DBG("Updating poll fd array");
1176 *nb_inactive_fd
= 0;
1178 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1180 * Only active streams with an active end point can be added to the
1181 * poll set and local stream storage of the thread.
1183 * There is a potential race here for endpoint_status to be updated
1184 * just after the check. However, this is OK since the stream(s) will
1185 * be deleted once the thread is notified that the end point state has
1186 * changed where this function will be called back again.
1188 * We track the number of inactive FDs because they still need to be
1189 * closed by the polling thread after a wakeup on the data_pipe or
1192 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_INACTIVE
) {
1193 (*nb_inactive_fd
)++;
1197 * This clobbers way too much the debug output. Uncomment that if you
1198 * need it for debugging purposes.
1200 (*pollfd
)[i
].fd
= stream
->wait_fd
;
1201 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1202 local_stream
[i
] = stream
;
1208 * Insert the consumer_data_pipe at the end of the array and don't
1209 * increment i so nb_fd is the number of real FD.
1211 (*pollfd
)[i
].fd
= lttng_pipe_get_readfd(ctx
->consumer_data_pipe
);
1212 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1214 (*pollfd
)[i
+ 1].fd
= lttng_pipe_get_readfd(ctx
->consumer_wakeup_pipe
);
1215 (*pollfd
)[i
+ 1].events
= POLLIN
| POLLPRI
;
1220 * Poll on the should_quit pipe and the command socket return -1 on
1221 * error, 1 if should exit, 0 if data is available on the command socket
1223 int lttng_consumer_poll_socket(struct pollfd
*consumer_sockpoll
)
1228 num_rdy
= poll(consumer_sockpoll
, 2, -1);
1229 if (num_rdy
== -1) {
1231 * Restart interrupted system call.
1233 if (errno
== EINTR
) {
1236 PERROR("Poll error");
1239 if (consumer_sockpoll
[0].revents
& (POLLIN
| POLLPRI
)) {
1240 DBG("consumer_should_quit wake up");
1247 * Set the error socket.
1249 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data
*ctx
,
1252 ctx
->consumer_error_socket
= sock
;
1256 * Set the command socket path.
1258 void lttng_consumer_set_command_sock_path(
1259 struct lttng_consumer_local_data
*ctx
, char *sock
)
1261 ctx
->consumer_command_sock_path
= sock
;
1265 * Send return code to the session daemon.
1266 * If the socket is not defined, we return 0, it is not a fatal error
1268 int lttng_consumer_send_error(struct lttng_consumer_local_data
*ctx
, int cmd
)
1270 if (ctx
->consumer_error_socket
> 0) {
1271 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
1272 sizeof(enum lttcomm_sessiond_command
));
1279 * Close all the tracefiles and stream fds and MUST be called when all
1280 * instances are destroyed i.e. when all threads were joined and are ended.
1282 void lttng_consumer_cleanup(void)
1284 struct lttng_ht_iter iter
;
1285 struct lttng_consumer_channel
*channel
;
1286 unsigned int trace_chunks_left
;
1290 cds_lfht_for_each_entry(the_consumer_data
.channel_ht
->ht
, &iter
.iter
,
1291 channel
, node
.node
) {
1292 consumer_del_channel(channel
);
1297 lttng_ht_destroy(the_consumer_data
.channel_ht
);
1298 lttng_ht_destroy(the_consumer_data
.channels_by_session_id_ht
);
1300 cleanup_relayd_ht();
1302 lttng_ht_destroy(the_consumer_data
.stream_per_chan_id_ht
);
1305 * This HT contains streams that are freed by either the metadata thread or
1306 * the data thread so we do *nothing* on the hash table and simply destroy
1309 lttng_ht_destroy(the_consumer_data
.stream_list_ht
);
1312 * Trace chunks in the registry may still exist if the session
1313 * daemon has encountered an internal error and could not
1314 * tear down its sessions and/or trace chunks properly.
1316 * Release the session daemon's implicit reference to any remaining
1317 * trace chunk and print an error if any trace chunk was found. Note
1318 * that there are _no_ legitimate cases for trace chunks to be left,
1319 * it is a leak. However, it can happen following a crash of the
1320 * session daemon and not emptying the registry would cause an assertion
1323 trace_chunks_left
= lttng_trace_chunk_registry_put_each_chunk(
1324 the_consumer_data
.chunk_registry
);
1325 if (trace_chunks_left
) {
1326 ERR("%u trace chunks are leaked by lttng-consumerd. "
1327 "This can be caused by an internal error of the session daemon.",
1330 /* Run all callbacks freeing each chunk. */
1332 lttng_trace_chunk_registry_destroy(the_consumer_data
.chunk_registry
);
1336 * Called from signal handler.
1338 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
1342 CMM_STORE_SHARED(consumer_quit
, 1);
1343 ret
= lttng_write(ctx
->consumer_should_quit
[1], "4", 1);
1345 PERROR("write consumer quit");
1348 DBG("Consumer flag that it should quit");
1353 * Flush pending writes to trace output disk file.
1356 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1360 int outfd
= stream
->out_fd
;
1363 * This does a blocking write-and-wait on any page that belongs to the
1364 * subbuffer prior to the one we just wrote.
1365 * Don't care about error values, as these are just hints and ways to
1366 * limit the amount of page cache used.
1368 if (orig_offset
< stream
->max_sb_size
) {
1371 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1372 stream
->max_sb_size
,
1373 SYNC_FILE_RANGE_WAIT_BEFORE
1374 | SYNC_FILE_RANGE_WRITE
1375 | SYNC_FILE_RANGE_WAIT_AFTER
);
1377 * Give hints to the kernel about how we access the file:
1378 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1381 * We need to call fadvise again after the file grows because the
1382 * kernel does not seem to apply fadvise to non-existing parts of the
1385 * Call fadvise _after_ having waited for the page writeback to
1386 * complete because the dirty page writeback semantic is not well
1387 * defined. So it can be expected to lead to lower throughput in
1390 ret
= posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1391 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1392 if (ret
&& ret
!= -ENOSYS
) {
1394 PERROR("posix_fadvise on fd %i", outfd
);
1399 * Initialise the necessary environnement :
1400 * - create a new context
1401 * - create the poll_pipe
1402 * - create the should_quit pipe (for signal handler)
1403 * - create the thread pipe (for splice)
1405 * Takes a function pointer as argument, this function is called when data is
1406 * available on a buffer. This function is responsible to do the
1407 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1408 * buffer configuration and then kernctl_put_next_subbuf at the end.
1410 * Returns a pointer to the new context or NULL on error.
1412 struct lttng_consumer_local_data
*lttng_consumer_create(
1413 enum lttng_consumer_type type
,
1414 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1415 struct lttng_consumer_local_data
*ctx
, bool locked_by_caller
),
1416 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1417 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1418 int (*update_stream
)(uint64_t stream_key
, uint32_t state
))
1421 struct lttng_consumer_local_data
*ctx
;
1423 LTTNG_ASSERT(the_consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1424 the_consumer_data
.type
== type
);
1425 the_consumer_data
.type
= type
;
1427 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
1429 PERROR("allocating context");
1433 ctx
->consumer_error_socket
= -1;
1434 ctx
->consumer_metadata_socket
= -1;
1435 pthread_mutex_init(&ctx
->metadata_socket_lock
, NULL
);
1436 /* assign the callbacks */
1437 ctx
->on_buffer_ready
= buffer_ready
;
1438 ctx
->on_recv_channel
= recv_channel
;
1439 ctx
->on_recv_stream
= recv_stream
;
1440 ctx
->on_update_stream
= update_stream
;
1442 ctx
->consumer_data_pipe
= lttng_pipe_open(0);
1443 if (!ctx
->consumer_data_pipe
) {
1444 goto error_poll_pipe
;
1447 ctx
->consumer_wakeup_pipe
= lttng_pipe_open(0);
1448 if (!ctx
->consumer_wakeup_pipe
) {
1449 goto error_wakeup_pipe
;
1452 ret
= pipe(ctx
->consumer_should_quit
);
1454 PERROR("Error creating recv pipe");
1455 goto error_quit_pipe
;
1458 ret
= pipe(ctx
->consumer_channel_pipe
);
1460 PERROR("Error creating channel pipe");
1461 goto error_channel_pipe
;
1464 ctx
->consumer_metadata_pipe
= lttng_pipe_open(0);
1465 if (!ctx
->consumer_metadata_pipe
) {
1466 goto error_metadata_pipe
;
1469 ctx
->channel_monitor_pipe
= -1;
1473 error_metadata_pipe
:
1474 utils_close_pipe(ctx
->consumer_channel_pipe
);
1476 utils_close_pipe(ctx
->consumer_should_quit
);
1478 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1480 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1488 * Iterate over all streams of the hashtable and free them properly.
1490 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1492 struct lttng_ht_iter iter
;
1493 struct lttng_consumer_stream
*stream
;
1500 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1502 * Ignore return value since we are currently cleaning up so any error
1505 (void) consumer_del_stream(stream
, ht
);
1509 lttng_ht_destroy(ht
);
1513 * Iterate over all streams of the metadata hashtable and free them
1516 static void destroy_metadata_stream_ht(struct lttng_ht
*ht
)
1518 struct lttng_ht_iter iter
;
1519 struct lttng_consumer_stream
*stream
;
1526 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1528 * Ignore return value since we are currently cleaning up so any error
1531 (void) consumer_del_metadata_stream(stream
, ht
);
1535 lttng_ht_destroy(ht
);
1539 * Close all fds associated with the instance and free the context.
1541 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1545 DBG("Consumer destroying it. Closing everything.");
1551 destroy_data_stream_ht(data_ht
);
1552 destroy_metadata_stream_ht(metadata_ht
);
1554 ret
= close(ctx
->consumer_error_socket
);
1558 ret
= close(ctx
->consumer_metadata_socket
);
1562 utils_close_pipe(ctx
->consumer_channel_pipe
);
1563 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1564 lttng_pipe_destroy(ctx
->consumer_metadata_pipe
);
1565 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1566 utils_close_pipe(ctx
->consumer_should_quit
);
1568 unlink(ctx
->consumer_command_sock_path
);
1573 * Write the metadata stream id on the specified file descriptor.
1575 static int write_relayd_metadata_id(int fd
,
1576 struct lttng_consumer_stream
*stream
,
1577 unsigned long padding
)
1580 struct lttcomm_relayd_metadata_payload hdr
;
1582 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1583 hdr
.padding_size
= htobe32(padding
);
1584 ret
= lttng_write(fd
, (void *) &hdr
, sizeof(hdr
));
1585 if (ret
< sizeof(hdr
)) {
1587 * This error means that the fd's end is closed so ignore the PERROR
1588 * not to clubber the error output since this can happen in a normal
1591 if (errno
!= EPIPE
) {
1592 PERROR("write metadata stream id");
1594 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1596 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1597 * handle writting the missing part so report that as an error and
1598 * don't lie to the caller.
1603 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1604 stream
->relayd_stream_id
, padding
);
1611 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1612 * core function for writing trace buffers to either the local filesystem or
1615 * It must be called with the stream and the channel lock held.
1617 * Careful review MUST be put if any changes occur!
1619 * Returns the number of bytes written
1621 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1622 struct lttng_consumer_stream
*stream
,
1623 const struct lttng_buffer_view
*buffer
,
1624 unsigned long padding
)
1627 off_t orig_offset
= stream
->out_fd_offset
;
1628 /* Default is on the disk */
1629 int outfd
= stream
->out_fd
;
1630 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1631 unsigned int relayd_hang_up
= 0;
1632 const size_t subbuf_content_size
= buffer
->size
- padding
;
1635 /* RCU lock for the relayd pointer */
1637 LTTNG_ASSERT(stream
->net_seq_idx
!= (uint64_t) -1ULL ||
1638 stream
->trace_chunk
);
1640 /* Flag that the current stream if set for network streaming. */
1641 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1642 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1643 if (relayd
== NULL
) {
1649 /* Handle stream on the relayd if the output is on the network */
1651 unsigned long netlen
= subbuf_content_size
;
1654 * Lock the control socket for the complete duration of the function
1655 * since from this point on we will use the socket.
1657 if (stream
->metadata_flag
) {
1658 /* Metadata requires the control socket. */
1659 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1660 if (stream
->reset_metadata_flag
) {
1661 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1662 stream
->relayd_stream_id
,
1663 stream
->metadata_version
);
1668 stream
->reset_metadata_flag
= 0;
1670 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1673 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1678 /* Use the returned socket. */
1681 /* Write metadata stream id before payload */
1682 if (stream
->metadata_flag
) {
1683 ret
= write_relayd_metadata_id(outfd
, stream
, padding
);
1690 write_len
= subbuf_content_size
;
1692 /* No streaming; we have to write the full padding. */
1693 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1694 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1696 ERR("Reset metadata file");
1699 stream
->reset_metadata_flag
= 0;
1703 * Check if we need to change the tracefile before writing the packet.
1705 if (stream
->chan
->tracefile_size
> 0 &&
1706 (stream
->tracefile_size_current
+ buffer
->size
) >
1707 stream
->chan
->tracefile_size
) {
1708 ret
= consumer_stream_rotate_output_files(stream
);
1712 outfd
= stream
->out_fd
;
1715 stream
->tracefile_size_current
+= buffer
->size
;
1716 write_len
= buffer
->size
;
1720 * This call guarantee that len or less is returned. It's impossible to
1721 * receive a ret value that is bigger than len.
1723 ret
= lttng_write(outfd
, buffer
->data
, write_len
);
1724 DBG("Consumer mmap write() ret %zd (len %zu)", ret
, write_len
);
1725 if (ret
< 0 || ((size_t) ret
!= write_len
)) {
1727 * Report error to caller if nothing was written else at least send the
1735 /* Socket operation failed. We consider the relayd dead */
1736 if (errno
== EPIPE
) {
1738 * This is possible if the fd is closed on the other side
1739 * (outfd) or any write problem. It can be verbose a bit for a
1740 * normal execution if for instance the relayd is stopped
1741 * abruptly. This can happen so set this to a DBG statement.
1743 DBG("Consumer mmap write detected relayd hang up");
1745 /* Unhandled error, print it and stop function right now. */
1746 PERROR("Error in write mmap (ret %zd != write_len %zu)", ret
,
1751 stream
->output_written
+= ret
;
1753 /* This call is useless on a socket so better save a syscall. */
1755 /* This won't block, but will start writeout asynchronously */
1756 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, write_len
,
1757 SYNC_FILE_RANGE_WRITE
);
1758 stream
->out_fd_offset
+= write_len
;
1759 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1764 * This is a special case that the relayd has closed its socket. Let's
1765 * cleanup the relayd object and all associated streams.
1767 if (relayd
&& relayd_hang_up
) {
1768 ERR("Relayd hangup. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
1769 lttng_consumer_cleanup_relayd(relayd
);
1773 /* Unlock only if ctrl socket used */
1774 if (relayd
&& stream
->metadata_flag
) {
1775 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1783 * Splice the data from the ring buffer to the tracefile.
1785 * It must be called with the stream lock held.
1787 * Returns the number of bytes spliced.
1789 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1790 struct lttng_consumer_local_data
*ctx
,
1791 struct lttng_consumer_stream
*stream
, unsigned long len
,
1792 unsigned long padding
)
1794 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1796 off_t orig_offset
= stream
->out_fd_offset
;
1797 int fd
= stream
->wait_fd
;
1798 /* Default is on the disk */
1799 int outfd
= stream
->out_fd
;
1800 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1802 unsigned int relayd_hang_up
= 0;
1804 switch (the_consumer_data
.type
) {
1805 case LTTNG_CONSUMER_KERNEL
:
1807 case LTTNG_CONSUMER32_UST
:
1808 case LTTNG_CONSUMER64_UST
:
1809 /* Not supported for user space tracing */
1812 ERR("Unknown consumer_data type");
1816 /* RCU lock for the relayd pointer */
1819 /* Flag that the current stream if set for network streaming. */
1820 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1821 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1822 if (relayd
== NULL
) {
1827 splice_pipe
= stream
->splice_pipe
;
1829 /* Write metadata stream id before payload */
1831 unsigned long total_len
= len
;
1833 if (stream
->metadata_flag
) {
1835 * Lock the control socket for the complete duration of the function
1836 * since from this point on we will use the socket.
1838 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1840 if (stream
->reset_metadata_flag
) {
1841 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1842 stream
->relayd_stream_id
,
1843 stream
->metadata_version
);
1848 stream
->reset_metadata_flag
= 0;
1850 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
,
1858 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1861 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1867 /* Use the returned socket. */
1870 /* No streaming, we have to set the len with the full padding */
1873 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1874 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1876 ERR("Reset metadata file");
1879 stream
->reset_metadata_flag
= 0;
1882 * Check if we need to change the tracefile before writing the packet.
1884 if (stream
->chan
->tracefile_size
> 0 &&
1885 (stream
->tracefile_size_current
+ len
) >
1886 stream
->chan
->tracefile_size
) {
1887 ret
= consumer_stream_rotate_output_files(stream
);
1892 outfd
= stream
->out_fd
;
1895 stream
->tracefile_size_current
+= len
;
1899 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1900 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1901 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1902 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1903 DBG("splice chan to pipe, ret %zd", ret_splice
);
1904 if (ret_splice
< 0) {
1907 PERROR("Error in relay splice");
1911 /* Handle stream on the relayd if the output is on the network */
1912 if (relayd
&& stream
->metadata_flag
) {
1913 size_t metadata_payload_size
=
1914 sizeof(struct lttcomm_relayd_metadata_payload
);
1916 /* Update counter to fit the spliced data */
1917 ret_splice
+= metadata_payload_size
;
1918 len
+= metadata_payload_size
;
1920 * We do this so the return value can match the len passed as
1921 * argument to this function.
1923 written
-= metadata_payload_size
;
1926 /* Splice data out */
1927 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1928 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1929 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
1931 if (ret_splice
< 0) {
1936 } else if (ret_splice
> len
) {
1938 * We don't expect this code path to be executed but you never know
1939 * so this is an extra protection agains a buggy splice().
1942 written
+= ret_splice
;
1943 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice
,
1947 /* All good, update current len and continue. */
1951 /* This call is useless on a socket so better save a syscall. */
1953 /* This won't block, but will start writeout asynchronously */
1954 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1955 SYNC_FILE_RANGE_WRITE
);
1956 stream
->out_fd_offset
+= ret_splice
;
1958 stream
->output_written
+= ret_splice
;
1959 written
+= ret_splice
;
1962 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1968 * This is a special case that the relayd has closed its socket. Let's
1969 * cleanup the relayd object and all associated streams.
1971 if (relayd
&& relayd_hang_up
) {
1972 ERR("Relayd hangup. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
1973 lttng_consumer_cleanup_relayd(relayd
);
1974 /* Skip splice error so the consumer does not fail */
1979 /* send the appropriate error description to sessiond */
1982 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1985 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1988 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
1993 if (relayd
&& stream
->metadata_flag
) {
1994 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
2002 * Sample the snapshot positions for a specific fd
2004 * Returns 0 on success, < 0 on error
2006 int lttng_consumer_sample_snapshot_positions(struct lttng_consumer_stream
*stream
)
2008 switch (the_consumer_data
.type
) {
2009 case LTTNG_CONSUMER_KERNEL
:
2010 return lttng_kconsumer_sample_snapshot_positions(stream
);
2011 case LTTNG_CONSUMER32_UST
:
2012 case LTTNG_CONSUMER64_UST
:
2013 return lttng_ustconsumer_sample_snapshot_positions(stream
);
2015 ERR("Unknown consumer_data type");
2021 * Take a snapshot for a specific fd
2023 * Returns 0 on success, < 0 on error
2025 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
2027 switch (the_consumer_data
.type
) {
2028 case LTTNG_CONSUMER_KERNEL
:
2029 return lttng_kconsumer_take_snapshot(stream
);
2030 case LTTNG_CONSUMER32_UST
:
2031 case LTTNG_CONSUMER64_UST
:
2032 return lttng_ustconsumer_take_snapshot(stream
);
2034 ERR("Unknown consumer_data type");
2041 * Get the produced position
2043 * Returns 0 on success, < 0 on error
2045 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
2048 switch (the_consumer_data
.type
) {
2049 case LTTNG_CONSUMER_KERNEL
:
2050 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
2051 case LTTNG_CONSUMER32_UST
:
2052 case LTTNG_CONSUMER64_UST
:
2053 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
2055 ERR("Unknown consumer_data type");
2062 * Get the consumed position (free-running counter position in bytes).
2064 * Returns 0 on success, < 0 on error
2066 int lttng_consumer_get_consumed_snapshot(struct lttng_consumer_stream
*stream
,
2069 switch (the_consumer_data
.type
) {
2070 case LTTNG_CONSUMER_KERNEL
:
2071 return lttng_kconsumer_get_consumed_snapshot(stream
, pos
);
2072 case LTTNG_CONSUMER32_UST
:
2073 case LTTNG_CONSUMER64_UST
:
2074 return lttng_ustconsumer_get_consumed_snapshot(stream
, pos
);
2076 ERR("Unknown consumer_data type");
2082 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
2083 int sock
, struct pollfd
*consumer_sockpoll
)
2085 switch (the_consumer_data
.type
) {
2086 case LTTNG_CONSUMER_KERNEL
:
2087 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2088 case LTTNG_CONSUMER32_UST
:
2089 case LTTNG_CONSUMER64_UST
:
2090 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2092 ERR("Unknown consumer_data type");
2099 void lttng_consumer_close_all_metadata(void)
2101 switch (the_consumer_data
.type
) {
2102 case LTTNG_CONSUMER_KERNEL
:
2104 * The Kernel consumer has a different metadata scheme so we don't
2105 * close anything because the stream will be closed by the session
2109 case LTTNG_CONSUMER32_UST
:
2110 case LTTNG_CONSUMER64_UST
:
2112 * Close all metadata streams. The metadata hash table is passed and
2113 * this call iterates over it by closing all wakeup fd. This is safe
2114 * because at this point we are sure that the metadata producer is
2115 * either dead or blocked.
2117 lttng_ustconsumer_close_all_metadata(metadata_ht
);
2120 ERR("Unknown consumer_data type");
2126 * Clean up a metadata stream and free its memory.
2128 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
2129 struct lttng_ht
*ht
)
2131 struct lttng_consumer_channel
*channel
= NULL
;
2132 bool free_channel
= false;
2134 LTTNG_ASSERT(stream
);
2136 * This call should NEVER receive regular stream. It must always be
2137 * metadata stream and this is crucial for data structure synchronization.
2139 LTTNG_ASSERT(stream
->metadata_flag
);
2141 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
2143 pthread_mutex_lock(&the_consumer_data
.lock
);
2145 * Note that this assumes that a stream's channel is never changed and
2146 * that the stream's lock doesn't need to be taken to sample its
2149 channel
= stream
->chan
;
2150 pthread_mutex_lock(&channel
->lock
);
2151 pthread_mutex_lock(&stream
->lock
);
2152 if (channel
->metadata_cache
) {
2153 /* Only applicable to userspace consumers. */
2154 pthread_mutex_lock(&channel
->metadata_cache
->lock
);
2157 /* Remove any reference to that stream. */
2158 consumer_stream_delete(stream
, ht
);
2160 /* Close down everything including the relayd if one. */
2161 consumer_stream_close(stream
);
2162 /* Destroy tracer buffers of the stream. */
2163 consumer_stream_destroy_buffers(stream
);
2165 /* Atomically decrement channel refcount since other threads can use it. */
2166 if (!uatomic_sub_return(&channel
->refcount
, 1)
2167 && !uatomic_read(&channel
->nb_init_stream_left
)) {
2168 /* Go for channel deletion! */
2169 free_channel
= true;
2171 stream
->chan
= NULL
;
2174 * Nullify the stream reference so it is not used after deletion. The
2175 * channel lock MUST be acquired before being able to check for a NULL
2178 channel
->metadata_stream
= NULL
;
2180 if (channel
->metadata_cache
) {
2181 pthread_mutex_unlock(&channel
->metadata_cache
->lock
);
2183 pthread_mutex_unlock(&stream
->lock
);
2184 pthread_mutex_unlock(&channel
->lock
);
2185 pthread_mutex_unlock(&the_consumer_data
.lock
);
2188 consumer_del_channel(channel
);
2191 lttng_trace_chunk_put(stream
->trace_chunk
);
2192 stream
->trace_chunk
= NULL
;
2193 consumer_stream_free(stream
);
2197 * Action done with the metadata stream when adding it to the consumer internal
2198 * data structures to handle it.
2200 void consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
2202 struct lttng_ht
*ht
= metadata_ht
;
2203 struct lttng_ht_iter iter
;
2204 struct lttng_ht_node_u64
*node
;
2206 LTTNG_ASSERT(stream
);
2209 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2211 pthread_mutex_lock(&the_consumer_data
.lock
);
2212 pthread_mutex_lock(&stream
->chan
->lock
);
2213 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2214 pthread_mutex_lock(&stream
->lock
);
2217 * From here, refcounts are updated so be _careful_ when returning an error
2224 * Lookup the stream just to make sure it does not exist in our internal
2225 * state. This should NEVER happen.
2227 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2228 node
= lttng_ht_iter_get_node_u64(&iter
);
2229 LTTNG_ASSERT(!node
);
2232 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2233 * in terms of destroying the associated channel, because the action that
2234 * causes the count to become 0 also causes a stream to be added. The
2235 * channel deletion will thus be triggered by the following removal of this
2238 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2239 /* Increment refcount before decrementing nb_init_stream_left */
2241 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2244 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2246 lttng_ht_add_u64(the_consumer_data
.stream_per_chan_id_ht
,
2247 &stream
->node_channel_id
);
2250 * Add stream to the stream_list_ht of the consumer data. No need to steal
2251 * the key since the HT does not use it and we allow to add redundant keys
2254 lttng_ht_add_u64(the_consumer_data
.stream_list_ht
,
2255 &stream
->node_session_id
);
2259 pthread_mutex_unlock(&stream
->lock
);
2260 pthread_mutex_unlock(&stream
->chan
->lock
);
2261 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2262 pthread_mutex_unlock(&the_consumer_data
.lock
);
2266 * Delete data stream that are flagged for deletion (endpoint_status).
2268 static void validate_endpoint_status_data_stream(void)
2270 struct lttng_ht_iter iter
;
2271 struct lttng_consumer_stream
*stream
;
2273 DBG("Consumer delete flagged data stream");
2276 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2277 /* Validate delete flag of the stream */
2278 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2281 /* Delete it right now */
2282 consumer_del_stream(stream
, data_ht
);
2288 * Delete metadata stream that are flagged for deletion (endpoint_status).
2290 static void validate_endpoint_status_metadata_stream(
2291 struct lttng_poll_event
*pollset
)
2293 struct lttng_ht_iter iter
;
2294 struct lttng_consumer_stream
*stream
;
2296 DBG("Consumer delete flagged metadata stream");
2298 LTTNG_ASSERT(pollset
);
2301 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2302 /* Validate delete flag of the stream */
2303 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2307 * Remove from pollset so the metadata thread can continue without
2308 * blocking on a deleted stream.
2310 lttng_poll_del(pollset
, stream
->wait_fd
);
2312 /* Delete it right now */
2313 consumer_del_metadata_stream(stream
, metadata_ht
);
2319 * Thread polls on metadata file descriptor and write them on disk or on the
2322 void *consumer_thread_metadata_poll(void *data
)
2324 int ret
, i
, pollfd
, err
= -1;
2325 uint32_t revents
, nb_fd
;
2326 struct lttng_consumer_stream
*stream
= NULL
;
2327 struct lttng_ht_iter iter
;
2328 struct lttng_ht_node_u64
*node
;
2329 struct lttng_poll_event events
;
2330 struct lttng_consumer_local_data
*ctx
= data
;
2333 rcu_register_thread();
2335 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2337 if (testpoint(consumerd_thread_metadata
)) {
2338 goto error_testpoint
;
2341 health_code_update();
2343 DBG("Thread metadata poll started");
2345 /* Size is set to 1 for the consumer_metadata pipe */
2346 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2348 ERR("Poll set creation failed");
2352 ret
= lttng_poll_add(&events
,
2353 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2359 DBG("Metadata main loop started");
2363 health_code_update();
2364 health_poll_entry();
2365 DBG("Metadata poll wait");
2366 ret
= lttng_poll_wait(&events
, -1);
2367 DBG("Metadata poll return from wait with %d fd(s)",
2368 LTTNG_POLL_GETNB(&events
));
2370 DBG("Metadata event caught in thread");
2372 if (errno
== EINTR
) {
2373 ERR("Poll EINTR caught");
2376 if (LTTNG_POLL_GETNB(&events
) == 0) {
2377 err
= 0; /* All is OK */
2384 /* From here, the event is a metadata wait fd */
2385 for (i
= 0; i
< nb_fd
; i
++) {
2386 health_code_update();
2388 revents
= LTTNG_POLL_GETEV(&events
, i
);
2389 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2391 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2392 if (revents
& LPOLLIN
) {
2395 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2396 &stream
, sizeof(stream
));
2397 if (pipe_len
< sizeof(stream
)) {
2399 PERROR("read metadata stream");
2402 * Remove the pipe from the poll set and continue the loop
2403 * since their might be data to consume.
2405 lttng_poll_del(&events
,
2406 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2407 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2411 /* A NULL stream means that the state has changed. */
2412 if (stream
== NULL
) {
2413 /* Check for deleted streams. */
2414 validate_endpoint_status_metadata_stream(&events
);
2418 DBG("Adding metadata stream %d to poll set",
2421 /* Add metadata stream to the global poll events list */
2422 lttng_poll_add(&events
, stream
->wait_fd
,
2423 LPOLLIN
| LPOLLPRI
| LPOLLHUP
);
2424 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2425 DBG("Metadata thread pipe hung up");
2427 * Remove the pipe from the poll set and continue the loop
2428 * since their might be data to consume.
2430 lttng_poll_del(&events
,
2431 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2432 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2435 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2439 /* Handle other stream */
2445 uint64_t tmp_id
= (uint64_t) pollfd
;
2447 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2449 node
= lttng_ht_iter_get_node_u64(&iter
);
2452 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2455 if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2456 /* Get the data out of the metadata file descriptor */
2457 DBG("Metadata available on fd %d", pollfd
);
2458 LTTNG_ASSERT(stream
->wait_fd
== pollfd
);
2461 health_code_update();
2463 len
= ctx
->on_buffer_ready(stream
, ctx
, false);
2465 * We don't check the return value here since if we get
2466 * a negative len, it means an error occurred thus we
2467 * simply remove it from the poll set and free the
2472 /* It's ok to have an unavailable sub-buffer */
2473 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2474 /* Clean up stream from consumer and free it. */
2475 lttng_poll_del(&events
, stream
->wait_fd
);
2476 consumer_del_metadata_stream(stream
, metadata_ht
);
2478 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2479 DBG("Metadata fd %d is hup|err.", pollfd
);
2480 if (!stream
->hangup_flush_done
&&
2481 (the_consumer_data
.type
== LTTNG_CONSUMER32_UST
||
2482 the_consumer_data
.type
==
2483 LTTNG_CONSUMER64_UST
)) {
2484 DBG("Attempting to flush and consume the UST buffers");
2485 lttng_ustconsumer_on_stream_hangup(stream
);
2487 /* We just flushed the stream now read it. */
2489 health_code_update();
2491 len
= ctx
->on_buffer_ready(stream
, ctx
, false);
2493 * We don't check the return value here since if we get
2494 * a negative len, it means an error occurred thus we
2495 * simply remove it from the poll set and free the
2501 lttng_poll_del(&events
, stream
->wait_fd
);
2503 * This call update the channel states, closes file descriptors
2504 * and securely free the stream.
2506 consumer_del_metadata_stream(stream
, metadata_ht
);
2508 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2512 /* Release RCU lock for the stream looked up */
2520 DBG("Metadata poll thread exiting");
2522 lttng_poll_clean(&events
);
2527 ERR("Health error occurred in %s", __func__
);
2529 health_unregister(health_consumerd
);
2530 rcu_unregister_thread();
2535 * This thread polls the fds in the set to consume the data and write
2536 * it to tracefile if necessary.
2538 void *consumer_thread_data_poll(void *data
)
2540 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2541 struct pollfd
*pollfd
= NULL
;
2542 /* local view of the streams */
2543 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2544 /* local view of consumer_data.fds_count */
2546 /* 2 for the consumer_data_pipe and wake up pipe */
2547 const int nb_pipes_fd
= 2;
2548 /* Number of FDs with CONSUMER_ENDPOINT_INACTIVE but still open. */
2549 int nb_inactive_fd
= 0;
2550 struct lttng_consumer_local_data
*ctx
= data
;
2553 rcu_register_thread();
2555 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2557 if (testpoint(consumerd_thread_data
)) {
2558 goto error_testpoint
;
2561 health_code_update();
2563 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
*));
2564 if (local_stream
== NULL
) {
2565 PERROR("local_stream malloc");
2570 health_code_update();
2576 * the fds set has been updated, we need to update our
2577 * local array as well
2579 pthread_mutex_lock(&the_consumer_data
.lock
);
2580 if (the_consumer_data
.need_update
) {
2585 local_stream
= NULL
;
2587 /* Allocate for all fds */
2588 pollfd
= zmalloc((the_consumer_data
.stream_count
+
2590 sizeof(struct pollfd
));
2591 if (pollfd
== NULL
) {
2592 PERROR("pollfd malloc");
2593 pthread_mutex_unlock(&the_consumer_data
.lock
);
2597 local_stream
= zmalloc((the_consumer_data
.stream_count
+
2599 sizeof(struct lttng_consumer_stream
*));
2600 if (local_stream
== NULL
) {
2601 PERROR("local_stream malloc");
2602 pthread_mutex_unlock(&the_consumer_data
.lock
);
2605 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2606 data_ht
, &nb_inactive_fd
);
2608 ERR("Error in allocating pollfd or local_outfds");
2609 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2610 pthread_mutex_unlock(&the_consumer_data
.lock
);
2614 the_consumer_data
.need_update
= 0;
2616 pthread_mutex_unlock(&the_consumer_data
.lock
);
2618 /* No FDs and consumer_quit, consumer_cleanup the thread */
2619 if (nb_fd
== 0 && nb_inactive_fd
== 0 &&
2620 CMM_LOAD_SHARED(consumer_quit
) == 1) {
2621 err
= 0; /* All is OK */
2624 /* poll on the array of fds */
2626 DBG("polling on %d fd", nb_fd
+ nb_pipes_fd
);
2627 if (testpoint(consumerd_thread_data_poll
)) {
2630 health_poll_entry();
2631 num_rdy
= poll(pollfd
, nb_fd
+ nb_pipes_fd
, -1);
2633 DBG("poll num_rdy : %d", num_rdy
);
2634 if (num_rdy
== -1) {
2636 * Restart interrupted system call.
2638 if (errno
== EINTR
) {
2641 PERROR("Poll error");
2642 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2644 } else if (num_rdy
== 0) {
2645 DBG("Polling thread timed out");
2649 if (caa_unlikely(data_consumption_paused
)) {
2650 DBG("Data consumption paused, sleeping...");
2656 * If the consumer_data_pipe triggered poll go directly to the
2657 * beginning of the loop to update the array. We want to prioritize
2658 * array update over low-priority reads.
2660 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2661 ssize_t pipe_readlen
;
2663 DBG("consumer_data_pipe wake up");
2664 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2665 &new_stream
, sizeof(new_stream
));
2666 if (pipe_readlen
< sizeof(new_stream
)) {
2667 PERROR("Consumer data pipe");
2668 /* Continue so we can at least handle the current stream(s). */
2673 * If the stream is NULL, just ignore it. It's also possible that
2674 * the sessiond poll thread changed the consumer_quit state and is
2675 * waking us up to test it.
2677 if (new_stream
== NULL
) {
2678 validate_endpoint_status_data_stream();
2682 /* Continue to update the local streams and handle prio ones */
2686 /* Handle wakeup pipe. */
2687 if (pollfd
[nb_fd
+ 1].revents
& (POLLIN
| POLLPRI
)) {
2689 ssize_t pipe_readlen
;
2691 pipe_readlen
= lttng_pipe_read(ctx
->consumer_wakeup_pipe
, &dummy
,
2693 if (pipe_readlen
< 0) {
2694 PERROR("Consumer data wakeup pipe");
2696 /* We've been awakened to handle stream(s). */
2697 ctx
->has_wakeup
= 0;
2700 /* Take care of high priority channels first. */
2701 for (i
= 0; i
< nb_fd
; i
++) {
2702 health_code_update();
2704 if (local_stream
[i
] == NULL
) {
2707 if (pollfd
[i
].revents
& POLLPRI
) {
2708 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2710 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
, false);
2711 /* it's ok to have an unavailable sub-buffer */
2712 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2713 /* Clean the stream and free it. */
2714 consumer_del_stream(local_stream
[i
], data_ht
);
2715 local_stream
[i
] = NULL
;
2716 } else if (len
> 0) {
2717 local_stream
[i
]->data_read
= 1;
2723 * If we read high prio channel in this loop, try again
2724 * for more high prio data.
2730 /* Take care of low priority channels. */
2731 for (i
= 0; i
< nb_fd
; i
++) {
2732 health_code_update();
2734 if (local_stream
[i
] == NULL
) {
2737 if ((pollfd
[i
].revents
& POLLIN
) ||
2738 local_stream
[i
]->hangup_flush_done
||
2739 local_stream
[i
]->has_data
) {
2740 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2741 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
, false);
2742 /* it's ok to have an unavailable sub-buffer */
2743 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2744 /* Clean the stream and free it. */
2745 consumer_del_stream(local_stream
[i
], data_ht
);
2746 local_stream
[i
] = NULL
;
2747 } else if (len
> 0) {
2748 local_stream
[i
]->data_read
= 1;
2753 /* Handle hangup and errors */
2754 for (i
= 0; i
< nb_fd
; i
++) {
2755 health_code_update();
2757 if (local_stream
[i
] == NULL
) {
2760 if (!local_stream
[i
]->hangup_flush_done
2761 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2762 && (the_consumer_data
.type
== LTTNG_CONSUMER32_UST
2763 || the_consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2764 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2766 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2767 /* Attempt read again, for the data we just flushed. */
2768 local_stream
[i
]->data_read
= 1;
2771 * If the poll flag is HUP/ERR/NVAL and we have
2772 * read no data in this pass, we can remove the
2773 * stream from its hash table.
2775 if ((pollfd
[i
].revents
& POLLHUP
)) {
2776 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2777 if (!local_stream
[i
]->data_read
) {
2778 consumer_del_stream(local_stream
[i
], data_ht
);
2779 local_stream
[i
] = NULL
;
2782 } else if (pollfd
[i
].revents
& POLLERR
) {
2783 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2784 if (!local_stream
[i
]->data_read
) {
2785 consumer_del_stream(local_stream
[i
], data_ht
);
2786 local_stream
[i
] = NULL
;
2789 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2790 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2791 if (!local_stream
[i
]->data_read
) {
2792 consumer_del_stream(local_stream
[i
], data_ht
);
2793 local_stream
[i
] = NULL
;
2797 if (local_stream
[i
] != NULL
) {
2798 local_stream
[i
]->data_read
= 0;
2805 DBG("polling thread exiting");
2810 * Close the write side of the pipe so epoll_wait() in
2811 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2812 * read side of the pipe. If we close them both, epoll_wait strangely does
2813 * not return and could create a endless wait period if the pipe is the
2814 * only tracked fd in the poll set. The thread will take care of closing
2817 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2822 ERR("Health error occurred in %s", __func__
);
2824 health_unregister(health_consumerd
);
2826 rcu_unregister_thread();
2831 * Close wake-up end of each stream belonging to the channel. This will
2832 * allow the poll() on the stream read-side to detect when the
2833 * write-side (application) finally closes them.
2836 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2838 struct lttng_ht
*ht
;
2839 struct lttng_consumer_stream
*stream
;
2840 struct lttng_ht_iter iter
;
2842 ht
= the_consumer_data
.stream_per_chan_id_ht
;
2845 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2846 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2847 ht
->match_fct
, &channel
->key
,
2848 &iter
.iter
, stream
, node_channel_id
.node
) {
2850 * Protect against teardown with mutex.
2852 pthread_mutex_lock(&stream
->lock
);
2853 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2856 switch (the_consumer_data
.type
) {
2857 case LTTNG_CONSUMER_KERNEL
:
2859 case LTTNG_CONSUMER32_UST
:
2860 case LTTNG_CONSUMER64_UST
:
2861 if (stream
->metadata_flag
) {
2862 /* Safe and protected by the stream lock. */
2863 lttng_ustconsumer_close_metadata(stream
->chan
);
2866 * Note: a mutex is taken internally within
2867 * liblttng-ust-ctl to protect timer wakeup_fd
2868 * use from concurrent close.
2870 lttng_ustconsumer_close_stream_wakeup(stream
);
2874 ERR("Unknown consumer_data type");
2878 pthread_mutex_unlock(&stream
->lock
);
2883 static void destroy_channel_ht(struct lttng_ht
*ht
)
2885 struct lttng_ht_iter iter
;
2886 struct lttng_consumer_channel
*channel
;
2894 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2895 ret
= lttng_ht_del(ht
, &iter
);
2896 LTTNG_ASSERT(ret
!= 0);
2900 lttng_ht_destroy(ht
);
2904 * This thread polls the channel fds to detect when they are being
2905 * closed. It closes all related streams if the channel is detected as
2906 * closed. It is currently only used as a shim layer for UST because the
2907 * consumerd needs to keep the per-stream wakeup end of pipes open for
2910 void *consumer_thread_channel_poll(void *data
)
2912 int ret
, i
, pollfd
, err
= -1;
2913 uint32_t revents
, nb_fd
;
2914 struct lttng_consumer_channel
*chan
= NULL
;
2915 struct lttng_ht_iter iter
;
2916 struct lttng_ht_node_u64
*node
;
2917 struct lttng_poll_event events
;
2918 struct lttng_consumer_local_data
*ctx
= data
;
2919 struct lttng_ht
*channel_ht
;
2921 rcu_register_thread();
2923 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
2925 if (testpoint(consumerd_thread_channel
)) {
2926 goto error_testpoint
;
2929 health_code_update();
2931 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2933 /* ENOMEM at this point. Better to bail out. */
2937 DBG("Thread channel poll started");
2939 /* Size is set to 1 for the consumer_channel pipe */
2940 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2942 ERR("Poll set creation failed");
2946 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2952 DBG("Channel main loop started");
2956 health_code_update();
2957 DBG("Channel poll wait");
2958 health_poll_entry();
2959 ret
= lttng_poll_wait(&events
, -1);
2960 DBG("Channel poll return from wait with %d fd(s)",
2961 LTTNG_POLL_GETNB(&events
));
2963 DBG("Channel event caught in thread");
2965 if (errno
== EINTR
) {
2966 ERR("Poll EINTR caught");
2969 if (LTTNG_POLL_GETNB(&events
) == 0) {
2970 err
= 0; /* All is OK */
2977 /* From here, the event is a channel wait fd */
2978 for (i
= 0; i
< nb_fd
; i
++) {
2979 health_code_update();
2981 revents
= LTTNG_POLL_GETEV(&events
, i
);
2982 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2984 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2985 if (revents
& LPOLLIN
) {
2986 enum consumer_channel_action action
;
2989 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
2992 ERR("Error reading channel pipe");
2994 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2999 case CONSUMER_CHANNEL_ADD
:
3000 DBG("Adding channel %d to poll set",
3003 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
3006 lttng_ht_add_unique_u64(channel_ht
,
3007 &chan
->wait_fd_node
);
3009 /* Add channel to the global poll events list */
3010 lttng_poll_add(&events
, chan
->wait_fd
,
3011 LPOLLERR
| LPOLLHUP
);
3013 case CONSUMER_CHANNEL_DEL
:
3016 * This command should never be called if the channel
3017 * has streams monitored by either the data or metadata
3018 * thread. The consumer only notify this thread with a
3019 * channel del. command if it receives a destroy
3020 * channel command from the session daemon that send it
3021 * if a command prior to the GET_CHANNEL failed.
3025 chan
= consumer_find_channel(key
);
3028 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
3031 lttng_poll_del(&events
, chan
->wait_fd
);
3032 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
3033 ret
= lttng_ht_del(channel_ht
, &iter
);
3034 LTTNG_ASSERT(ret
== 0);
3036 switch (the_consumer_data
.type
) {
3037 case LTTNG_CONSUMER_KERNEL
:
3039 case LTTNG_CONSUMER32_UST
:
3040 case LTTNG_CONSUMER64_UST
:
3041 health_code_update();
3042 /* Destroy streams that might have been left in the stream list. */
3043 clean_channel_stream_list(chan
);
3046 ERR("Unknown consumer_data type");
3051 * Release our own refcount. Force channel deletion even if
3052 * streams were not initialized.
3054 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
3055 consumer_del_channel(chan
);
3060 case CONSUMER_CHANNEL_QUIT
:
3062 * Remove the pipe from the poll set and continue the loop
3063 * since their might be data to consume.
3065 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
3068 ERR("Unknown action");
3071 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
3072 DBG("Channel thread pipe hung up");
3074 * Remove the pipe from the poll set and continue the loop
3075 * since their might be data to consume.
3077 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
3080 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3084 /* Handle other stream */
3090 uint64_t tmp_id
= (uint64_t) pollfd
;
3092 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
3094 node
= lttng_ht_iter_get_node_u64(&iter
);
3097 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
3100 /* Check for error event */
3101 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
3102 DBG("Channel fd %d is hup|err.", pollfd
);
3104 lttng_poll_del(&events
, chan
->wait_fd
);
3105 ret
= lttng_ht_del(channel_ht
, &iter
);
3106 LTTNG_ASSERT(ret
== 0);
3109 * This will close the wait fd for each stream associated to
3110 * this channel AND monitored by the data/metadata thread thus
3111 * will be clean by the right thread.
3113 consumer_close_channel_streams(chan
);
3115 /* Release our own refcount */
3116 if (!uatomic_sub_return(&chan
->refcount
, 1)
3117 && !uatomic_read(&chan
->nb_init_stream_left
)) {
3118 consumer_del_channel(chan
);
3121 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3126 /* Release RCU lock for the channel looked up */
3134 lttng_poll_clean(&events
);
3136 destroy_channel_ht(channel_ht
);
3139 DBG("Channel poll thread exiting");
3142 ERR("Health error occurred in %s", __func__
);
3144 health_unregister(health_consumerd
);
3145 rcu_unregister_thread();
3149 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
3150 struct pollfd
*sockpoll
, int client_socket
)
3155 LTTNG_ASSERT(sockpoll
);
3157 ret
= lttng_consumer_poll_socket(sockpoll
);
3161 DBG("Metadata connection on client_socket");
3163 /* Blocking call, waiting for transmission */
3164 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
3165 if (ctx
->consumer_metadata_socket
< 0) {
3166 WARN("On accept metadata");
3177 * This thread listens on the consumerd socket and receives the file
3178 * descriptors from the session daemon.
3180 void *consumer_thread_sessiond_poll(void *data
)
3182 int sock
= -1, client_socket
, ret
, err
= -1;
3184 * structure to poll for incoming data on communication socket avoids
3185 * making blocking sockets.
3187 struct pollfd consumer_sockpoll
[2];
3188 struct lttng_consumer_local_data
*ctx
= data
;
3190 rcu_register_thread();
3192 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
3194 if (testpoint(consumerd_thread_sessiond
)) {
3195 goto error_testpoint
;
3198 health_code_update();
3200 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
3201 unlink(ctx
->consumer_command_sock_path
);
3202 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
3203 if (client_socket
< 0) {
3204 ERR("Cannot create command socket");
3208 ret
= lttcomm_listen_unix_sock(client_socket
);
3213 DBG("Sending ready command to lttng-sessiond");
3214 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3215 /* return < 0 on error, but == 0 is not fatal */
3217 ERR("Error sending ready command to lttng-sessiond");
3221 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3222 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3223 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3224 consumer_sockpoll
[1].fd
= client_socket
;
3225 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3227 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3235 DBG("Connection on client_socket");
3237 /* Blocking call, waiting for transmission */
3238 sock
= lttcomm_accept_unix_sock(client_socket
);
3245 * Setup metadata socket which is the second socket connection on the
3246 * command unix socket.
3248 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3257 /* This socket is not useful anymore. */
3258 ret
= close(client_socket
);
3260 PERROR("close client_socket");
3264 /* update the polling structure to poll on the established socket */
3265 consumer_sockpoll
[1].fd
= sock
;
3266 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3269 health_code_update();
3271 health_poll_entry();
3272 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3281 DBG("Incoming command on sock");
3282 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3285 * This could simply be a session daemon quitting. Don't output
3288 DBG("Communication interrupted on command socket");
3292 if (CMM_LOAD_SHARED(consumer_quit
)) {
3293 DBG("consumer_thread_receive_fds received quit from signal");
3294 err
= 0; /* All is OK */
3297 DBG("Received command on sock");
3303 DBG("Consumer thread sessiond poll exiting");
3306 * Close metadata streams since the producer is the session daemon which
3309 * NOTE: for now, this only applies to the UST tracer.
3311 lttng_consumer_close_all_metadata();
3314 * when all fds have hung up, the polling thread
3317 CMM_STORE_SHARED(consumer_quit
, 1);
3320 * Notify the data poll thread to poll back again and test the
3321 * consumer_quit state that we just set so to quit gracefully.
3323 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3325 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3327 notify_health_quit_pipe(health_quit_pipe
);
3329 /* Cleaning up possibly open sockets. */
3333 PERROR("close sock sessiond poll");
3336 if (client_socket
>= 0) {
3337 ret
= close(client_socket
);
3339 PERROR("close client_socket sessiond poll");
3346 ERR("Health error occurred in %s", __func__
);
3348 health_unregister(health_consumerd
);
3350 rcu_unregister_thread();
3354 static int post_consume(struct lttng_consumer_stream
*stream
,
3355 const struct stream_subbuffer
*subbuffer
,
3356 struct lttng_consumer_local_data
*ctx
)
3360 const size_t count
= lttng_dynamic_array_get_count(
3361 &stream
->read_subbuffer_ops
.post_consume_cbs
);
3363 for (i
= 0; i
< count
; i
++) {
3364 const post_consume_cb op
= *(post_consume_cb
*) lttng_dynamic_array_get_element(
3365 &stream
->read_subbuffer_ops
.post_consume_cbs
,
3368 ret
= op(stream
, subbuffer
, ctx
);
3377 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3378 struct lttng_consumer_local_data
*ctx
,
3379 bool locked_by_caller
)
3381 ssize_t ret
, written_bytes
= 0;
3383 struct stream_subbuffer subbuffer
= {};
3384 enum get_next_subbuffer_status get_next_status
;
3386 if (!locked_by_caller
) {
3387 stream
->read_subbuffer_ops
.lock(stream
);
3390 if (stream
->read_subbuffer_ops
.on_wake_up
) {
3391 ret
= stream
->read_subbuffer_ops
.on_wake_up(stream
);
3398 * If the stream was flagged to be ready for rotation before we extract
3399 * the next packet, rotate it now.
3401 if (stream
->rotate_ready
) {
3402 DBG("Rotate stream before consuming data");
3403 ret
= lttng_consumer_rotate_stream(ctx
, stream
);
3405 ERR("Stream rotation error before consuming data");
3410 get_next_status
= stream
->read_subbuffer_ops
.get_next_subbuffer(
3411 stream
, &subbuffer
);
3412 switch (get_next_status
) {
3413 case GET_NEXT_SUBBUFFER_STATUS_OK
:
3415 case GET_NEXT_SUBBUFFER_STATUS_NO_DATA
:
3419 case GET_NEXT_SUBBUFFER_STATUS_ERROR
:
3426 ret
= stream
->read_subbuffer_ops
.pre_consume_subbuffer(
3427 stream
, &subbuffer
);
3429 goto error_put_subbuf
;
3432 written_bytes
= stream
->read_subbuffer_ops
.consume_subbuffer(
3433 ctx
, stream
, &subbuffer
);
3434 if (written_bytes
<= 0) {
3435 ERR("Error consuming subbuffer: (%zd)", written_bytes
);
3436 ret
= (int) written_bytes
;
3437 goto error_put_subbuf
;
3440 ret
= stream
->read_subbuffer_ops
.put_next_subbuffer(stream
, &subbuffer
);
3445 ret
= post_consume(stream
, &subbuffer
, ctx
);
3451 * After extracting the packet, we check if the stream is now ready to
3452 * be rotated and perform the action immediately.
3454 * Don't overwrite `ret` as callers expect the number of bytes
3455 * consumed to be returned on success.
3457 rotation_ret
= lttng_consumer_stream_is_rotate_ready(stream
);
3458 if (rotation_ret
== 1) {
3459 rotation_ret
= lttng_consumer_rotate_stream(ctx
, stream
);
3460 if (rotation_ret
< 0) {
3462 ERR("Stream rotation error after consuming data");
3466 } else if (rotation_ret
< 0) {
3468 ERR("Failed to check if stream was ready to rotate after consuming data");
3473 if (stream
->read_subbuffer_ops
.on_sleep
) {
3474 stream
->read_subbuffer_ops
.on_sleep(stream
, ctx
);
3477 ret
= written_bytes
;
3479 if (!locked_by_caller
) {
3480 stream
->read_subbuffer_ops
.unlock(stream
);
3485 (void) stream
->read_subbuffer_ops
.put_next_subbuffer(stream
, &subbuffer
);
3489 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3491 switch (the_consumer_data
.type
) {
3492 case LTTNG_CONSUMER_KERNEL
:
3493 return lttng_kconsumer_on_recv_stream(stream
);
3494 case LTTNG_CONSUMER32_UST
:
3495 case LTTNG_CONSUMER64_UST
:
3496 return lttng_ustconsumer_on_recv_stream(stream
);
3498 ERR("Unknown consumer_data type");
3505 * Allocate and set consumer data hash tables.
3507 int lttng_consumer_init(void)
3509 the_consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3510 if (!the_consumer_data
.channel_ht
) {
3514 the_consumer_data
.channels_by_session_id_ht
=
3515 lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3516 if (!the_consumer_data
.channels_by_session_id_ht
) {
3520 the_consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3521 if (!the_consumer_data
.relayd_ht
) {
3525 the_consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3526 if (!the_consumer_data
.stream_list_ht
) {
3530 the_consumer_data
.stream_per_chan_id_ht
=
3531 lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3532 if (!the_consumer_data
.stream_per_chan_id_ht
) {
3536 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3541 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3546 the_consumer_data
.chunk_registry
= lttng_trace_chunk_registry_create();
3547 if (!the_consumer_data
.chunk_registry
) {
3558 * Process the ADD_RELAYD command receive by a consumer.
3560 * This will create a relayd socket pair and add it to the relayd hash table.
3561 * The caller MUST acquire a RCU read side lock before calling it.
3563 void consumer_add_relayd_socket(uint64_t net_seq_idx
, int sock_type
,
3564 struct lttng_consumer_local_data
*ctx
, int sock
,
3565 struct pollfd
*consumer_sockpoll
,
3566 struct lttcomm_relayd_sock
*relayd_sock
, uint64_t sessiond_id
,
3567 uint64_t relayd_session_id
)
3569 int fd
= -1, ret
= -1, relayd_created
= 0;
3570 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3571 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3574 LTTNG_ASSERT(relayd_sock
);
3576 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", net_seq_idx
);
3578 /* Get relayd reference if exists. */
3579 relayd
= consumer_find_relayd(net_seq_idx
);
3580 if (relayd
== NULL
) {
3581 LTTNG_ASSERT(sock_type
== LTTNG_STREAM_CONTROL
);
3582 /* Not found. Allocate one. */
3583 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3584 if (relayd
== NULL
) {
3585 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3588 relayd
->sessiond_session_id
= sessiond_id
;
3593 * This code path MUST continue to the consumer send status message to
3594 * we can notify the session daemon and continue our work without
3595 * killing everything.
3599 * relayd key should never be found for control socket.
3601 LTTNG_ASSERT(sock_type
!= LTTNG_STREAM_CONTROL
);
3604 /* First send a status message before receiving the fds. */
3605 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3607 /* Somehow, the session daemon is not responding anymore. */
3608 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3609 goto error_nosignal
;
3612 /* Poll on consumer socket. */
3613 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3615 /* Needing to exit in the middle of a command: error. */
3616 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3617 goto error_nosignal
;
3620 /* Get relayd socket from session daemon */
3621 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3622 if (ret
!= sizeof(fd
)) {
3623 fd
= -1; /* Just in case it gets set with an invalid value. */
3626 * Failing to receive FDs might indicate a major problem such as
3627 * reaching a fd limit during the receive where the kernel returns a
3628 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3629 * don't take any chances and stop everything.
3631 * XXX: Feature request #558 will fix that and avoid this possible
3632 * issue when reaching the fd limit.
3634 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3635 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3639 /* Copy socket information and received FD */
3640 switch (sock_type
) {
3641 case LTTNG_STREAM_CONTROL
:
3642 /* Copy received lttcomm socket */
3643 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3644 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3645 /* Handle create_sock error. */
3647 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3651 * Close the socket created internally by
3652 * lttcomm_create_sock, so we can replace it by the one
3653 * received from sessiond.
3655 if (close(relayd
->control_sock
.sock
.fd
)) {
3659 /* Assign new file descriptor */
3660 relayd
->control_sock
.sock
.fd
= fd
;
3661 /* Assign version values. */
3662 relayd
->control_sock
.major
= relayd_sock
->major
;
3663 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3665 relayd
->relayd_session_id
= relayd_session_id
;
3668 case LTTNG_STREAM_DATA
:
3669 /* Copy received lttcomm socket */
3670 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3671 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3672 /* Handle create_sock error. */
3674 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3678 * Close the socket created internally by
3679 * lttcomm_create_sock, so we can replace it by the one
3680 * received from sessiond.
3682 if (close(relayd
->data_sock
.sock
.fd
)) {
3686 /* Assign new file descriptor */
3687 relayd
->data_sock
.sock
.fd
= fd
;
3688 /* Assign version values. */
3689 relayd
->data_sock
.major
= relayd_sock
->major
;
3690 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3693 ERR("Unknown relayd socket type (%d)", sock_type
);
3694 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3698 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3699 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3700 relayd
->net_seq_idx
, fd
);
3702 * We gave the ownership of the fd to the relayd structure. Set the
3703 * fd to -1 so we don't call close() on it in the error path below.
3707 /* We successfully added the socket. Send status back. */
3708 ret
= consumer_send_status_msg(sock
, ret_code
);
3710 /* Somehow, the session daemon is not responding anymore. */
3711 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3712 goto error_nosignal
;
3716 * Add relayd socket pair to consumer data hashtable. If object already
3717 * exists or on error, the function gracefully returns.
3726 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3727 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3731 /* Close received socket if valid. */
3734 PERROR("close received socket");
3738 if (relayd_created
) {
3744 * Search for a relayd associated to the session id and return the reference.
3746 * A rcu read side lock MUST be acquire before calling this function and locked
3747 * until the relayd object is no longer necessary.
3749 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3751 struct lttng_ht_iter iter
;
3752 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3754 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3755 cds_lfht_for_each_entry(the_consumer_data
.relayd_ht
->ht
, &iter
.iter
,
3756 relayd
, node
.node
) {
3758 * Check by sessiond id which is unique here where the relayd session
3759 * id might not be when having multiple relayd.
3761 if (relayd
->sessiond_session_id
== id
) {
3762 /* Found the relayd. There can be only one per id. */
3774 * Check if for a given session id there is still data needed to be extract
3777 * Return 1 if data is pending or else 0 meaning ready to be read.
3779 int consumer_data_pending(uint64_t id
)
3782 struct lttng_ht_iter iter
;
3783 struct lttng_ht
*ht
;
3784 struct lttng_consumer_stream
*stream
;
3785 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3786 int (*data_pending
)(struct lttng_consumer_stream
*);
3788 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3791 pthread_mutex_lock(&the_consumer_data
.lock
);
3793 switch (the_consumer_data
.type
) {
3794 case LTTNG_CONSUMER_KERNEL
:
3795 data_pending
= lttng_kconsumer_data_pending
;
3797 case LTTNG_CONSUMER32_UST
:
3798 case LTTNG_CONSUMER64_UST
:
3799 data_pending
= lttng_ustconsumer_data_pending
;
3802 ERR("Unknown consumer data type");
3806 /* Ease our life a bit */
3807 ht
= the_consumer_data
.stream_list_ht
;
3809 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3810 ht
->hash_fct(&id
, lttng_ht_seed
),
3812 &iter
.iter
, stream
, node_session_id
.node
) {
3813 pthread_mutex_lock(&stream
->lock
);
3816 * A removed node from the hash table indicates that the stream has
3817 * been deleted thus having a guarantee that the buffers are closed
3818 * on the consumer side. However, data can still be transmitted
3819 * over the network so don't skip the relayd check.
3821 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3823 /* Check the stream if there is data in the buffers. */
3824 ret
= data_pending(stream
);
3826 pthread_mutex_unlock(&stream
->lock
);
3831 pthread_mutex_unlock(&stream
->lock
);
3834 relayd
= find_relayd_by_session_id(id
);
3836 unsigned int is_data_inflight
= 0;
3838 /* Send init command for data pending. */
3839 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3840 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3841 relayd
->relayd_session_id
);
3843 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3844 /* Communication error thus the relayd so no data pending. */
3845 goto data_not_pending
;
3848 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3849 ht
->hash_fct(&id
, lttng_ht_seed
),
3851 &iter
.iter
, stream
, node_session_id
.node
) {
3852 if (stream
->metadata_flag
) {
3853 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3854 stream
->relayd_stream_id
);
3856 ret
= relayd_data_pending(&relayd
->control_sock
,
3857 stream
->relayd_stream_id
,
3858 stream
->next_net_seq_num
- 1);
3862 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3864 } else if (ret
< 0) {
3865 ERR("Relayd data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
3866 lttng_consumer_cleanup_relayd(relayd
);
3867 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3868 goto data_not_pending
;
3872 /* Send end command for data pending. */
3873 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3874 relayd
->relayd_session_id
, &is_data_inflight
);
3875 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3877 ERR("Relayd end data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
3878 lttng_consumer_cleanup_relayd(relayd
);
3879 goto data_not_pending
;
3881 if (is_data_inflight
) {
3887 * Finding _no_ node in the hash table and no inflight data means that the
3888 * stream(s) have been removed thus data is guaranteed to be available for
3889 * analysis from the trace files.
3893 /* Data is available to be read by a viewer. */
3894 pthread_mutex_unlock(&the_consumer_data
.lock
);
3899 /* Data is still being extracted from buffers. */
3900 pthread_mutex_unlock(&the_consumer_data
.lock
);
3906 * Send a ret code status message to the sessiond daemon.
3908 * Return the sendmsg() return value.
3910 int consumer_send_status_msg(int sock
, int ret_code
)
3912 struct lttcomm_consumer_status_msg msg
;
3914 memset(&msg
, 0, sizeof(msg
));
3915 msg
.ret_code
= ret_code
;
3917 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3921 * Send a channel status message to the sessiond daemon.
3923 * Return the sendmsg() return value.
3925 int consumer_send_status_channel(int sock
,
3926 struct lttng_consumer_channel
*channel
)
3928 struct lttcomm_consumer_status_channel msg
;
3930 LTTNG_ASSERT(sock
>= 0);
3932 memset(&msg
, 0, sizeof(msg
));
3934 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3936 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3937 msg
.key
= channel
->key
;
3938 msg
.stream_count
= channel
->streams
.count
;
3941 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3944 unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos
,
3945 unsigned long produced_pos
, uint64_t nb_packets_per_stream
,
3946 uint64_t max_sb_size
)
3948 unsigned long start_pos
;
3950 if (!nb_packets_per_stream
) {
3951 return consumed_pos
; /* Grab everything */
3953 start_pos
= produced_pos
- lttng_offset_align_floor(produced_pos
, max_sb_size
);
3954 start_pos
-= max_sb_size
* nb_packets_per_stream
;
3955 if ((long) (start_pos
- consumed_pos
) < 0) {
3956 return consumed_pos
; /* Grab everything */
3961 /* Stream lock must be held by the caller. */
3962 static int sample_stream_positions(struct lttng_consumer_stream
*stream
,
3963 unsigned long *produced
, unsigned long *consumed
)
3967 ASSERT_LOCKED(stream
->lock
);
3969 ret
= lttng_consumer_sample_snapshot_positions(stream
);
3971 ERR("Failed to sample snapshot positions");
3975 ret
= lttng_consumer_get_produced_snapshot(stream
, produced
);
3977 ERR("Failed to sample produced position");
3981 ret
= lttng_consumer_get_consumed_snapshot(stream
, consumed
);
3983 ERR("Failed to sample consumed position");
3992 * Sample the rotate position for all the streams of a channel. If a stream
3993 * is already at the rotate position (produced == consumed), we flag it as
3994 * ready for rotation. The rotation of ready streams occurs after we have
3995 * replied to the session daemon that we have finished sampling the positions.
3996 * Must be called with RCU read-side lock held to ensure existence of channel.
3998 * Returns 0 on success, < 0 on error
4000 int lttng_consumer_rotate_channel(struct lttng_consumer_channel
*channel
,
4001 uint64_t key
, uint64_t relayd_id
, uint32_t metadata
,
4002 struct lttng_consumer_local_data
*ctx
)
4005 struct lttng_consumer_stream
*stream
;
4006 struct lttng_ht_iter iter
;
4007 struct lttng_ht
*ht
= the_consumer_data
.stream_per_chan_id_ht
;
4008 struct lttng_dynamic_array stream_rotation_positions
;
4009 uint64_t next_chunk_id
, stream_count
= 0;
4010 enum lttng_trace_chunk_status chunk_status
;
4011 const bool is_local_trace
= relayd_id
== -1ULL;
4012 struct consumer_relayd_sock_pair
*relayd
= NULL
;
4013 bool rotating_to_new_chunk
= true;
4014 /* Array of `struct lttng_consumer_stream *` */
4015 struct lttng_dynamic_pointer_array streams_packet_to_open
;
4018 DBG("Consumer sample rotate position for channel %" PRIu64
, key
);
4020 lttng_dynamic_array_init(&stream_rotation_positions
,
4021 sizeof(struct relayd_stream_rotation_position
), NULL
);
4022 lttng_dynamic_pointer_array_init(&streams_packet_to_open
, NULL
);
4026 pthread_mutex_lock(&channel
->lock
);
4027 LTTNG_ASSERT(channel
->trace_chunk
);
4028 chunk_status
= lttng_trace_chunk_get_id(channel
->trace_chunk
,
4030 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4032 goto end_unlock_channel
;
4035 cds_lfht_for_each_entry_duplicate(ht
->ht
,
4036 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
4037 ht
->match_fct
, &channel
->key
, &iter
.iter
,
4038 stream
, node_channel_id
.node
) {
4039 unsigned long produced_pos
= 0, consumed_pos
= 0;
4041 health_code_update();
4044 * Lock stream because we are about to change its state.
4046 pthread_mutex_lock(&stream
->lock
);
4048 if (stream
->trace_chunk
== stream
->chan
->trace_chunk
) {
4049 rotating_to_new_chunk
= false;
4053 * Do not flush a packet when rotating from a NULL trace
4054 * chunk. The stream has no means to output data, and the prior
4055 * rotation which rotated to NULL performed that side-effect
4056 * already. No new data can be produced when a stream has no
4057 * associated trace chunk (e.g. a stop followed by a rotate).
4059 if (stream
->trace_chunk
) {
4062 if (stream
->metadata_flag
) {
4064 * Don't produce an empty metadata packet,
4065 * simply close the current one.
4067 * Metadata is regenerated on every trace chunk
4068 * switch; there is no concern that no data was
4071 flush_active
= true;
4074 * Only flush an empty packet if the "packet
4075 * open" could not be performed on transition
4076 * to a new trace chunk and no packets were
4077 * consumed within the chunk's lifetime.
4079 if (stream
->opened_packet_in_current_trace_chunk
) {
4080 flush_active
= true;
4083 * Stream could have been full at the
4084 * time of rotation, but then have had
4085 * no activity at all.
4087 * It is important to flush a packet
4088 * to prevent 0-length files from being
4089 * produced as most viewers choke on
4092 * Unfortunately viewers will not be
4093 * able to know that tracing was active
4094 * for this stream during this trace
4097 ret
= sample_stream_positions(stream
, &produced_pos
, &consumed_pos
);
4099 goto end_unlock_stream
;
4103 * Don't flush an empty packet if data
4104 * was produced; it will be consumed
4105 * before the rotation completes.
4107 flush_active
= produced_pos
!= consumed_pos
;
4108 if (!flush_active
) {
4109 const char *trace_chunk_name
;
4110 uint64_t trace_chunk_id
;
4112 chunk_status
= lttng_trace_chunk_get_name(
4113 stream
->trace_chunk
,
4116 if (chunk_status
== LTTNG_TRACE_CHUNK_STATUS_NONE
) {
4117 trace_chunk_name
= "none";
4121 * Consumer trace chunks are
4124 chunk_status
= lttng_trace_chunk_get_id(
4125 stream
->trace_chunk
,
4127 LTTNG_ASSERT(chunk_status
==
4128 LTTNG_TRACE_CHUNK_STATUS_OK
);
4130 DBG("Unable to open packet for stream during trace chunk's lifetime. "
4131 "Flushing an empty packet to prevent an empty file from being created: "
4132 "stream id = %" PRIu64
", trace chunk name = `%s`, trace chunk id = %" PRIu64
,
4133 stream
->key
, trace_chunk_name
, trace_chunk_id
);
4139 * Close the current packet before sampling the
4140 * ring buffer positions.
4142 ret
= consumer_stream_flush_buffer(stream
, flush_active
);
4144 ERR("Failed to flush stream %" PRIu64
" during channel rotation",
4146 goto end_unlock_stream
;
4150 ret
= lttng_consumer_take_snapshot(stream
);
4151 if (ret
< 0 && ret
!= -ENODATA
&& ret
!= -EAGAIN
) {
4152 ERR("Failed to sample snapshot position during channel rotation");
4153 goto end_unlock_stream
;
4156 ret
= lttng_consumer_get_produced_snapshot(stream
,
4159 ERR("Failed to sample produced position during channel rotation");
4160 goto end_unlock_stream
;
4163 ret
= lttng_consumer_get_consumed_snapshot(stream
,
4166 ERR("Failed to sample consumed position during channel rotation");
4167 goto end_unlock_stream
;
4171 * Align produced position on the start-of-packet boundary of the first
4172 * packet going into the next trace chunk.
4174 produced_pos
= lttng_align_floor(produced_pos
, stream
->max_sb_size
);
4175 if (consumed_pos
== produced_pos
) {
4176 DBG("Set rotate ready for stream %" PRIu64
" produced = %lu consumed = %lu",
4177 stream
->key
, produced_pos
, consumed_pos
);
4178 stream
->rotate_ready
= true;
4180 DBG("Different consumed and produced positions "
4181 "for stream %" PRIu64
" produced = %lu consumed = %lu",
4182 stream
->key
, produced_pos
, consumed_pos
);
4185 * The rotation position is based on the packet_seq_num of the
4186 * packet following the last packet that was consumed for this
4187 * stream, incremented by the offset between produced and
4188 * consumed positions. This rotation position is a lower bound
4189 * (inclusive) at which the next trace chunk starts. Since it
4190 * is a lower bound, it is OK if the packet_seq_num does not
4191 * correspond exactly to the same packet identified by the
4192 * consumed_pos, which can happen in overwrite mode.
4194 if (stream
->sequence_number_unavailable
) {
4196 * Rotation should never be performed on a session which
4197 * interacts with a pre-2.8 lttng-modules, which does
4198 * not implement packet sequence number.
4200 ERR("Failure to rotate stream %" PRIu64
": sequence number unavailable",
4203 goto end_unlock_stream
;
4205 stream
->rotate_position
= stream
->last_sequence_number
+ 1 +
4206 ((produced_pos
- consumed_pos
) / stream
->max_sb_size
);
4207 DBG("Set rotation position for stream %" PRIu64
" at position %" PRIu64
,
4208 stream
->key
, stream
->rotate_position
);
4210 if (!is_local_trace
) {
4212 * The relay daemon control protocol expects a rotation
4213 * position as "the sequence number of the first packet
4214 * _after_ the current trace chunk".
4216 const struct relayd_stream_rotation_position position
= {
4217 .stream_id
= stream
->relayd_stream_id
,
4218 .rotate_at_seq_num
= stream
->rotate_position
,
4221 ret
= lttng_dynamic_array_add_element(
4222 &stream_rotation_positions
,
4225 ERR("Failed to allocate stream rotation position");
4226 goto end_unlock_stream
;
4231 stream
->opened_packet_in_current_trace_chunk
= false;
4233 if (rotating_to_new_chunk
&& !stream
->metadata_flag
) {
4235 * Attempt to flush an empty packet as close to the
4236 * rotation point as possible. In the event where a
4237 * stream remains inactive after the rotation point,
4238 * this ensures that the new trace chunk has a
4239 * beginning timestamp set at the begining of the
4240 * trace chunk instead of only creating an empty
4241 * packet when the trace chunk is stopped.
4243 * This indicates to the viewers that the stream
4244 * was being recorded, but more importantly it
4245 * allows viewers to determine a useable trace
4248 * This presents a problem in the case where the
4249 * ring-buffer is completely full.
4251 * Consider the following scenario:
4252 * - The consumption of data is slow (slow network,
4254 * - The ring buffer is full,
4255 * - A rotation is initiated,
4256 * - The flush below does nothing (no space left to
4257 * open a new packet),
4258 * - The other streams rotate very soon, and new
4259 * data is produced in the new chunk,
4260 * - This stream completes its rotation long after the
4261 * rotation was initiated
4262 * - The session is stopped before any event can be
4263 * produced in this stream's buffers.
4265 * The resulting trace chunk will have a single packet
4266 * temporaly at the end of the trace chunk for this
4267 * stream making the stream intersection more narrow
4268 * than it should be.
4270 * To work-around this, an empty flush is performed
4271 * after the first consumption of a packet during a
4272 * rotation if open_packet fails. The idea is that
4273 * consuming a packet frees enough space to switch
4274 * packets in this scenario and allows the tracer to
4275 * "stamp" the beginning of the new trace chunk at the
4276 * earliest possible point.
4278 * The packet open is performed after the channel
4279 * rotation to ensure that no attempt to open a packet
4280 * is performed in a stream that has no active trace
4283 ret
= lttng_dynamic_pointer_array_add_pointer(
4284 &streams_packet_to_open
, stream
);
4286 PERROR("Failed to add a stream pointer to array of streams in which to open a packet");
4288 goto end_unlock_stream
;
4292 pthread_mutex_unlock(&stream
->lock
);
4296 if (!is_local_trace
) {
4297 relayd
= consumer_find_relayd(relayd_id
);
4299 ERR("Failed to find relayd %" PRIu64
, relayd_id
);
4301 goto end_unlock_channel
;
4304 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
4305 ret
= relayd_rotate_streams(&relayd
->control_sock
, stream_count
,
4306 rotating_to_new_chunk
? &next_chunk_id
: NULL
,
4307 (const struct relayd_stream_rotation_position
*)
4308 stream_rotation_positions
.buffer
4310 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
4312 ERR("Relayd rotate stream failed. Cleaning up relayd %" PRIu64
,
4313 relayd
->net_seq_idx
);
4314 lttng_consumer_cleanup_relayd(relayd
);
4315 goto end_unlock_channel
;
4319 for (stream_idx
= 0;
4320 stream_idx
< lttng_dynamic_pointer_array_get_count(
4321 &streams_packet_to_open
);
4323 enum consumer_stream_open_packet_status status
;
4325 stream
= lttng_dynamic_pointer_array_get_pointer(
4326 &streams_packet_to_open
, stream_idx
);
4328 pthread_mutex_lock(&stream
->lock
);
4329 status
= consumer_stream_open_packet(stream
);
4330 pthread_mutex_unlock(&stream
->lock
);
4332 case CONSUMER_STREAM_OPEN_PACKET_STATUS_OPENED
:
4333 DBG("Opened a packet after a rotation: stream id = %" PRIu64
4334 ", channel name = %s, session id = %" PRIu64
,
4335 stream
->key
, stream
->chan
->name
,
4336 stream
->chan
->session_id
);
4338 case CONSUMER_STREAM_OPEN_PACKET_STATUS_NO_SPACE
:
4340 * Can't open a packet as there is no space left
4341 * in the buffer. A new packet will be opened
4342 * once one has been consumed.
4344 DBG("No space left to open a packet after a rotation: stream id = %" PRIu64
4345 ", channel name = %s, session id = %" PRIu64
,
4346 stream
->key
, stream
->chan
->name
,
4347 stream
->chan
->session_id
);
4349 case CONSUMER_STREAM_OPEN_PACKET_STATUS_ERROR
:
4350 /* Logged by callee. */
4352 goto end_unlock_channel
;
4358 pthread_mutex_unlock(&channel
->lock
);
4363 pthread_mutex_unlock(&stream
->lock
);
4365 pthread_mutex_unlock(&channel
->lock
);
4368 lttng_dynamic_array_reset(&stream_rotation_positions
);
4369 lttng_dynamic_pointer_array_reset(&streams_packet_to_open
);
4374 int consumer_clear_buffer(struct lttng_consumer_stream
*stream
)
4377 unsigned long consumed_pos_before
, consumed_pos_after
;
4379 ret
= lttng_consumer_sample_snapshot_positions(stream
);
4381 ERR("Taking snapshot positions");
4385 ret
= lttng_consumer_get_consumed_snapshot(stream
, &consumed_pos_before
);
4387 ERR("Consumed snapshot position");
4391 switch (the_consumer_data
.type
) {
4392 case LTTNG_CONSUMER_KERNEL
:
4393 ret
= kernctl_buffer_clear(stream
->wait_fd
);
4395 ERR("Failed to clear kernel stream (ret = %d)", ret
);
4399 case LTTNG_CONSUMER32_UST
:
4400 case LTTNG_CONSUMER64_UST
:
4401 ret
= lttng_ustconsumer_clear_buffer(stream
);
4403 ERR("Failed to clear ust stream (ret = %d)", ret
);
4408 ERR("Unknown consumer_data type");
4412 ret
= lttng_consumer_sample_snapshot_positions(stream
);
4414 ERR("Taking snapshot positions");
4417 ret
= lttng_consumer_get_consumed_snapshot(stream
, &consumed_pos_after
);
4419 ERR("Consumed snapshot position");
4422 DBG("clear: before: %lu after: %lu", consumed_pos_before
, consumed_pos_after
);
4428 int consumer_clear_stream(struct lttng_consumer_stream
*stream
)
4432 ret
= consumer_stream_flush_buffer(stream
, 1);
4434 ERR("Failed to flush stream %" PRIu64
" during channel clear",
4436 ret
= LTTCOMM_CONSUMERD_FATAL
;
4440 ret
= consumer_clear_buffer(stream
);
4442 ERR("Failed to clear stream %" PRIu64
" during channel clear",
4444 ret
= LTTCOMM_CONSUMERD_FATAL
;
4448 ret
= LTTCOMM_CONSUMERD_SUCCESS
;
4454 int consumer_clear_unmonitored_channel(struct lttng_consumer_channel
*channel
)
4457 struct lttng_consumer_stream
*stream
;
4460 pthread_mutex_lock(&channel
->lock
);
4461 cds_list_for_each_entry(stream
, &channel
->streams
.head
, send_node
) {
4462 health_code_update();
4463 pthread_mutex_lock(&stream
->lock
);
4464 ret
= consumer_clear_stream(stream
);
4468 pthread_mutex_unlock(&stream
->lock
);
4470 pthread_mutex_unlock(&channel
->lock
);
4475 pthread_mutex_unlock(&stream
->lock
);
4476 pthread_mutex_unlock(&channel
->lock
);
4482 * Check if a stream is ready to be rotated after extracting it.
4484 * Return 1 if it is ready for rotation, 0 if it is not, a negative value on
4485 * error. Stream lock must be held.
4487 int lttng_consumer_stream_is_rotate_ready(struct lttng_consumer_stream
*stream
)
4489 DBG("Check is rotate ready for stream %" PRIu64
4490 " ready %u rotate_position %" PRIu64
4491 " last_sequence_number %" PRIu64
,
4492 stream
->key
, stream
->rotate_ready
,
4493 stream
->rotate_position
, stream
->last_sequence_number
);
4494 if (stream
->rotate_ready
) {
4499 * If packet seq num is unavailable, it means we are interacting
4500 * with a pre-2.8 lttng-modules which does not implement the
4501 * sequence number. Rotation should never be used by sessiond in this
4504 if (stream
->sequence_number_unavailable
) {
4505 ERR("Internal error: rotation used on stream %" PRIu64
4506 " with unavailable sequence number",
4511 if (stream
->rotate_position
== -1ULL ||
4512 stream
->last_sequence_number
== -1ULL) {
4517 * Rotate position not reached yet. The stream rotate position is
4518 * the position of the next packet belonging to the next trace chunk,
4519 * but consumerd considers rotation ready when reaching the last
4520 * packet of the current chunk, hence the "rotate_position - 1".
4523 DBG("Check is rotate ready for stream %" PRIu64
4524 " last_sequence_number %" PRIu64
4525 " rotate_position %" PRIu64
,
4526 stream
->key
, stream
->last_sequence_number
,
4527 stream
->rotate_position
);
4528 if (stream
->last_sequence_number
>= stream
->rotate_position
- 1) {
4536 * Reset the state for a stream after a rotation occurred.
4538 void lttng_consumer_reset_stream_rotate_state(struct lttng_consumer_stream
*stream
)
4540 DBG("lttng_consumer_reset_stream_rotate_state for stream %" PRIu64
,
4542 stream
->rotate_position
= -1ULL;
4543 stream
->rotate_ready
= false;
4547 * Perform the rotation a local stream file.
4550 int rotate_local_stream(struct lttng_consumer_local_data
*ctx
,
4551 struct lttng_consumer_stream
*stream
)
4555 DBG("Rotate local stream: stream key %" PRIu64
", channel key %" PRIu64
,
4558 stream
->tracefile_size_current
= 0;
4559 stream
->tracefile_count_current
= 0;
4561 if (stream
->out_fd
>= 0) {
4562 ret
= close(stream
->out_fd
);
4564 PERROR("Failed to close stream out_fd of channel \"%s\"",
4565 stream
->chan
->name
);
4567 stream
->out_fd
= -1;
4570 if (stream
->index_file
) {
4571 lttng_index_file_put(stream
->index_file
);
4572 stream
->index_file
= NULL
;
4575 if (!stream
->trace_chunk
) {
4579 ret
= consumer_stream_create_output_files(stream
, true);
4585 * Performs the stream rotation for the rotate session feature if needed.
4586 * It must be called with the channel and stream locks held.
4588 * Return 0 on success, a negative number of error.
4590 int lttng_consumer_rotate_stream(struct lttng_consumer_local_data
*ctx
,
4591 struct lttng_consumer_stream
*stream
)
4595 DBG("Consumer rotate stream %" PRIu64
, stream
->key
);
4598 * Update the stream's 'current' chunk to the session's (channel)
4599 * now-current chunk.
4601 lttng_trace_chunk_put(stream
->trace_chunk
);
4602 if (stream
->chan
->trace_chunk
== stream
->trace_chunk
) {
4604 * A channel can be rotated and not have a "next" chunk
4605 * to transition to. In that case, the channel's "current chunk"
4606 * has not been closed yet, but it has not been updated to
4607 * a "next" trace chunk either. Hence, the stream, like its
4608 * parent channel, becomes part of no chunk and can't output
4609 * anything until a new trace chunk is created.
4611 stream
->trace_chunk
= NULL
;
4612 } else if (stream
->chan
->trace_chunk
&&
4613 !lttng_trace_chunk_get(stream
->chan
->trace_chunk
)) {
4614 ERR("Failed to acquire a reference to channel's trace chunk during stream rotation");
4619 * Update the stream's trace chunk to its parent channel's
4620 * current trace chunk.
4622 stream
->trace_chunk
= stream
->chan
->trace_chunk
;
4625 if (stream
->net_seq_idx
== (uint64_t) -1ULL) {
4626 ret
= rotate_local_stream(ctx
, stream
);
4628 ERR("Failed to rotate stream, ret = %i", ret
);
4633 if (stream
->metadata_flag
&& stream
->trace_chunk
) {
4635 * If the stream has transitioned to a new trace
4636 * chunk, the metadata should be re-dumped to the
4639 * However, it is possible for a stream to transition to
4640 * a "no-chunk" state. This can happen if a rotation
4641 * occurs on an inactive session. In such cases, the metadata
4642 * regeneration will happen when the next trace chunk is
4645 ret
= consumer_metadata_stream_dump(stream
);
4650 lttng_consumer_reset_stream_rotate_state(stream
);
4659 * Rotate all the ready streams now.
4661 * This is especially important for low throughput streams that have already
4662 * been consumed, we cannot wait for their next packet to perform the
4664 * Need to be called with RCU read-side lock held to ensure existence of
4667 * Returns 0 on success, < 0 on error
4669 int lttng_consumer_rotate_ready_streams(struct lttng_consumer_channel
*channel
,
4670 uint64_t key
, struct lttng_consumer_local_data
*ctx
)
4673 struct lttng_consumer_stream
*stream
;
4674 struct lttng_ht_iter iter
;
4675 struct lttng_ht
*ht
= the_consumer_data
.stream_per_chan_id_ht
;
4679 DBG("Consumer rotate ready streams in channel %" PRIu64
, key
);
4681 cds_lfht_for_each_entry_duplicate(ht
->ht
,
4682 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
4683 ht
->match_fct
, &channel
->key
, &iter
.iter
,
4684 stream
, node_channel_id
.node
) {
4685 health_code_update();
4687 pthread_mutex_lock(&stream
->chan
->lock
);
4688 pthread_mutex_lock(&stream
->lock
);
4690 if (!stream
->rotate_ready
) {
4691 pthread_mutex_unlock(&stream
->lock
);
4692 pthread_mutex_unlock(&stream
->chan
->lock
);
4695 DBG("Consumer rotate ready stream %" PRIu64
, stream
->key
);
4697 ret
= lttng_consumer_rotate_stream(ctx
, stream
);
4698 pthread_mutex_unlock(&stream
->lock
);
4699 pthread_mutex_unlock(&stream
->chan
->lock
);
4712 enum lttcomm_return_code
lttng_consumer_init_command(
4713 struct lttng_consumer_local_data
*ctx
,
4714 const lttng_uuid sessiond_uuid
)
4716 enum lttcomm_return_code ret
;
4717 char uuid_str
[LTTNG_UUID_STR_LEN
];
4719 if (ctx
->sessiond_uuid
.is_set
) {
4720 ret
= LTTCOMM_CONSUMERD_ALREADY_SET
;
4724 ctx
->sessiond_uuid
.is_set
= true;
4725 memcpy(ctx
->sessiond_uuid
.value
, sessiond_uuid
, sizeof(lttng_uuid
));
4726 ret
= LTTCOMM_CONSUMERD_SUCCESS
;
4727 lttng_uuid_to_str(sessiond_uuid
, uuid_str
);
4728 DBG("Received session daemon UUID: %s", uuid_str
);
4733 enum lttcomm_return_code
lttng_consumer_create_trace_chunk(
4734 const uint64_t *relayd_id
, uint64_t session_id
,
4736 time_t chunk_creation_timestamp
,
4737 const char *chunk_override_name
,
4738 const struct lttng_credentials
*credentials
,
4739 struct lttng_directory_handle
*chunk_directory_handle
)
4742 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
4743 struct lttng_trace_chunk
*created_chunk
= NULL
, *published_chunk
= NULL
;
4744 enum lttng_trace_chunk_status chunk_status
;
4745 char relayd_id_buffer
[MAX_INT_DEC_LEN(*relayd_id
)];
4746 char creation_timestamp_buffer
[ISO8601_STR_LEN
];
4747 const char *relayd_id_str
= "(none)";
4748 const char *creation_timestamp_str
;
4749 struct lttng_ht_iter iter
;
4750 struct lttng_consumer_channel
*channel
;
4753 /* Only used for logging purposes. */
4754 ret
= snprintf(relayd_id_buffer
, sizeof(relayd_id_buffer
),
4755 "%" PRIu64
, *relayd_id
);
4756 if (ret
> 0 && ret
< sizeof(relayd_id_buffer
)) {
4757 relayd_id_str
= relayd_id_buffer
;
4759 relayd_id_str
= "(formatting error)";
4763 /* Local protocol error. */
4764 LTTNG_ASSERT(chunk_creation_timestamp
);
4765 ret
= time_to_iso8601_str(chunk_creation_timestamp
,
4766 creation_timestamp_buffer
,
4767 sizeof(creation_timestamp_buffer
));
4768 creation_timestamp_str
= !ret
? creation_timestamp_buffer
:
4769 "(formatting error)";
4771 DBG("Consumer create trace chunk command: relay_id = %s"
4772 ", session_id = %" PRIu64
", chunk_id = %" PRIu64
4773 ", chunk_override_name = %s"
4774 ", chunk_creation_timestamp = %s",
4775 relayd_id_str
, session_id
, chunk_id
,
4776 chunk_override_name
? : "(none)",
4777 creation_timestamp_str
);
4780 * The trace chunk registry, as used by the consumer daemon, implicitly
4781 * owns the trace chunks. This is only needed in the consumer since
4782 * the consumer has no notion of a session beyond session IDs being
4783 * used to identify other objects.
4785 * The lttng_trace_chunk_registry_publish() call below provides a
4786 * reference which is not released; it implicitly becomes the session
4787 * daemon's reference to the chunk in the consumer daemon.
4789 * The lifetime of trace chunks in the consumer daemon is managed by
4790 * the session daemon through the LTTNG_CONSUMER_CREATE_TRACE_CHUNK
4791 * and LTTNG_CONSUMER_DESTROY_TRACE_CHUNK commands.
4793 created_chunk
= lttng_trace_chunk_create(chunk_id
,
4794 chunk_creation_timestamp
, NULL
);
4795 if (!created_chunk
) {
4796 ERR("Failed to create trace chunk");
4797 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4801 if (chunk_override_name
) {
4802 chunk_status
= lttng_trace_chunk_override_name(created_chunk
,
4803 chunk_override_name
);
4804 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4805 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4810 if (chunk_directory_handle
) {
4811 chunk_status
= lttng_trace_chunk_set_credentials(created_chunk
,
4813 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4814 ERR("Failed to set trace chunk credentials");
4815 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4819 * The consumer daemon has no ownership of the chunk output
4822 chunk_status
= lttng_trace_chunk_set_as_user(created_chunk
,
4823 chunk_directory_handle
);
4824 chunk_directory_handle
= NULL
;
4825 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4826 ERR("Failed to set trace chunk's directory handle");
4827 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4832 published_chunk
= lttng_trace_chunk_registry_publish_chunk(
4833 the_consumer_data
.chunk_registry
, session_id
,
4835 lttng_trace_chunk_put(created_chunk
);
4836 created_chunk
= NULL
;
4837 if (!published_chunk
) {
4838 ERR("Failed to publish trace chunk");
4839 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4844 cds_lfht_for_each_entry_duplicate(
4845 the_consumer_data
.channels_by_session_id_ht
->ht
,
4846 the_consumer_data
.channels_by_session_id_ht
->hash_fct(
4847 &session_id
, lttng_ht_seed
),
4848 the_consumer_data
.channels_by_session_id_ht
->match_fct
,
4849 &session_id
, &iter
.iter
, channel
,
4850 channels_by_session_id_ht_node
.node
) {
4851 ret
= lttng_consumer_channel_set_trace_chunk(channel
,
4855 * Roll-back the creation of this chunk.
4857 * This is important since the session daemon will
4858 * assume that the creation of this chunk failed and
4859 * will never ask for it to be closed, resulting
4860 * in a leak and an inconsistent state for some
4863 enum lttcomm_return_code close_ret
;
4864 char path
[LTTNG_PATH_MAX
];
4866 DBG("Failed to set new trace chunk on existing channels, rolling back");
4867 close_ret
= lttng_consumer_close_trace_chunk(relayd_id
,
4868 session_id
, chunk_id
,
4869 chunk_creation_timestamp
, NULL
,
4871 if (close_ret
!= LTTCOMM_CONSUMERD_SUCCESS
) {
4872 ERR("Failed to roll-back the creation of new chunk: session_id = %" PRIu64
", chunk_id = %" PRIu64
,
4873 session_id
, chunk_id
);
4876 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4882 struct consumer_relayd_sock_pair
*relayd
;
4884 relayd
= consumer_find_relayd(*relayd_id
);
4886 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
4887 ret
= relayd_create_trace_chunk(
4888 &relayd
->control_sock
, published_chunk
);
4889 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
4891 ERR("Failed to find relay daemon socket: relayd_id = %" PRIu64
, *relayd_id
);
4894 if (!relayd
|| ret
) {
4895 enum lttcomm_return_code close_ret
;
4896 char path
[LTTNG_PATH_MAX
];
4898 close_ret
= lttng_consumer_close_trace_chunk(relayd_id
,
4901 chunk_creation_timestamp
,
4903 if (close_ret
!= LTTCOMM_CONSUMERD_SUCCESS
) {
4904 ERR("Failed to roll-back the creation of new chunk: session_id = %" PRIu64
", chunk_id = %" PRIu64
,
4909 ret_code
= LTTCOMM_CONSUMERD_CREATE_TRACE_CHUNK_FAILED
;
4916 /* Release the reference returned by the "publish" operation. */
4917 lttng_trace_chunk_put(published_chunk
);
4918 lttng_trace_chunk_put(created_chunk
);
4922 enum lttcomm_return_code
lttng_consumer_close_trace_chunk(
4923 const uint64_t *relayd_id
, uint64_t session_id
,
4924 uint64_t chunk_id
, time_t chunk_close_timestamp
,
4925 const enum lttng_trace_chunk_command_type
*close_command
,
4928 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
4929 struct lttng_trace_chunk
*chunk
;
4930 char relayd_id_buffer
[MAX_INT_DEC_LEN(*relayd_id
)];
4931 const char *relayd_id_str
= "(none)";
4932 const char *close_command_name
= "none";
4933 struct lttng_ht_iter iter
;
4934 struct lttng_consumer_channel
*channel
;
4935 enum lttng_trace_chunk_status chunk_status
;
4940 /* Only used for logging purposes. */
4941 ret
= snprintf(relayd_id_buffer
, sizeof(relayd_id_buffer
),
4942 "%" PRIu64
, *relayd_id
);
4943 if (ret
> 0 && ret
< sizeof(relayd_id_buffer
)) {
4944 relayd_id_str
= relayd_id_buffer
;
4946 relayd_id_str
= "(formatting error)";
4949 if (close_command
) {
4950 close_command_name
= lttng_trace_chunk_command_type_get_name(
4954 DBG("Consumer close trace chunk command: relayd_id = %s"
4955 ", session_id = %" PRIu64
", chunk_id = %" PRIu64
4956 ", close command = %s",
4957 relayd_id_str
, session_id
, chunk_id
,
4958 close_command_name
);
4960 chunk
= lttng_trace_chunk_registry_find_chunk(
4961 the_consumer_data
.chunk_registry
, session_id
, chunk_id
);
4963 ERR("Failed to find chunk: session_id = %" PRIu64
4964 ", chunk_id = %" PRIu64
,
4965 session_id
, chunk_id
);
4966 ret_code
= LTTCOMM_CONSUMERD_UNKNOWN_TRACE_CHUNK
;
4970 chunk_status
= lttng_trace_chunk_set_close_timestamp(chunk
,
4971 chunk_close_timestamp
);
4972 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4973 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
4977 if (close_command
) {
4978 chunk_status
= lttng_trace_chunk_set_close_command(
4979 chunk
, *close_command
);
4980 if (chunk_status
!= LTTNG_TRACE_CHUNK_STATUS_OK
) {
4981 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
4987 * chunk is now invalid to access as we no longer hold a reference to
4988 * it; it is only kept around to compare it (by address) to the
4989 * current chunk found in the session's channels.
4992 cds_lfht_for_each_entry(the_consumer_data
.channel_ht
->ht
, &iter
.iter
,
4993 channel
, node
.node
) {
4997 * Only change the channel's chunk to NULL if it still
4998 * references the chunk being closed. The channel may
4999 * reference a newer channel in the case of a session
5000 * rotation. When a session rotation occurs, the "next"
5001 * chunk is created before the "current" chunk is closed.
5003 if (channel
->trace_chunk
!= chunk
) {
5006 ret
= lttng_consumer_channel_set_trace_chunk(channel
, NULL
);
5009 * Attempt to close the chunk on as many channels as
5012 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
5018 struct consumer_relayd_sock_pair
*relayd
;
5020 relayd
= consumer_find_relayd(*relayd_id
);
5022 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
5023 ret
= relayd_close_trace_chunk(
5024 &relayd
->control_sock
, chunk
,
5026 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
5028 ERR("Failed to find relay daemon socket: relayd_id = %" PRIu64
,
5032 if (!relayd
|| ret
) {
5033 ret_code
= LTTCOMM_CONSUMERD_CLOSE_TRACE_CHUNK_FAILED
;
5041 * Release the reference returned by the "find" operation and
5042 * the session daemon's implicit reference to the chunk.
5044 lttng_trace_chunk_put(chunk
);
5045 lttng_trace_chunk_put(chunk
);
5050 enum lttcomm_return_code
lttng_consumer_trace_chunk_exists(
5051 const uint64_t *relayd_id
, uint64_t session_id
,
5055 enum lttcomm_return_code ret_code
;
5056 char relayd_id_buffer
[MAX_INT_DEC_LEN(*relayd_id
)];
5057 const char *relayd_id_str
= "(none)";
5058 const bool is_local_trace
= !relayd_id
;
5059 struct consumer_relayd_sock_pair
*relayd
= NULL
;
5060 bool chunk_exists_local
, chunk_exists_remote
;
5063 /* Only used for logging purposes. */
5064 ret
= snprintf(relayd_id_buffer
, sizeof(relayd_id_buffer
),
5065 "%" PRIu64
, *relayd_id
);
5066 if (ret
> 0 && ret
< sizeof(relayd_id_buffer
)) {
5067 relayd_id_str
= relayd_id_buffer
;
5069 relayd_id_str
= "(formatting error)";
5073 DBG("Consumer trace chunk exists command: relayd_id = %s"
5074 ", chunk_id = %" PRIu64
, relayd_id_str
,
5076 ret
= lttng_trace_chunk_registry_chunk_exists(
5077 the_consumer_data
.chunk_registry
, session_id
, chunk_id
,
5078 &chunk_exists_local
);
5080 /* Internal error. */
5081 ERR("Failed to query the existence of a trace chunk");
5082 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
5085 DBG("Trace chunk %s locally",
5086 chunk_exists_local
? "exists" : "does not exist");
5087 if (chunk_exists_local
) {
5088 ret_code
= LTTCOMM_CONSUMERD_TRACE_CHUNK_EXISTS_LOCAL
;
5090 } else if (is_local_trace
) {
5091 ret_code
= LTTCOMM_CONSUMERD_UNKNOWN_TRACE_CHUNK
;
5096 relayd
= consumer_find_relayd(*relayd_id
);
5098 ERR("Failed to find relayd %" PRIu64
, *relayd_id
);
5099 ret_code
= LTTCOMM_CONSUMERD_INVALID_PARAMETERS
;
5100 goto end_rcu_unlock
;
5102 DBG("Looking up existence of trace chunk on relay daemon");
5103 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
5104 ret
= relayd_trace_chunk_exists(&relayd
->control_sock
, chunk_id
,
5105 &chunk_exists_remote
);
5106 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
5108 ERR("Failed to look-up the existence of trace chunk on relay daemon");
5109 ret_code
= LTTCOMM_CONSUMERD_RELAYD_FAIL
;
5110 goto end_rcu_unlock
;
5113 ret_code
= chunk_exists_remote
?
5114 LTTCOMM_CONSUMERD_TRACE_CHUNK_EXISTS_REMOTE
:
5115 LTTCOMM_CONSUMERD_UNKNOWN_TRACE_CHUNK
;
5116 DBG("Trace chunk %s on relay daemon",
5117 chunk_exists_remote
? "exists" : "does not exist");
5126 int consumer_clear_monitored_channel(struct lttng_consumer_channel
*channel
)
5128 struct lttng_ht
*ht
;
5129 struct lttng_consumer_stream
*stream
;
5130 struct lttng_ht_iter iter
;
5133 ht
= the_consumer_data
.stream_per_chan_id_ht
;
5136 cds_lfht_for_each_entry_duplicate(ht
->ht
,
5137 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
5138 ht
->match_fct
, &channel
->key
,
5139 &iter
.iter
, stream
, node_channel_id
.node
) {
5141 * Protect against teardown with mutex.
5143 pthread_mutex_lock(&stream
->lock
);
5144 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
5147 ret
= consumer_clear_stream(stream
);
5152 pthread_mutex_unlock(&stream
->lock
);
5155 return LTTCOMM_CONSUMERD_SUCCESS
;
5158 pthread_mutex_unlock(&stream
->lock
);
5163 int lttng_consumer_clear_channel(struct lttng_consumer_channel
*channel
)
5167 DBG("Consumer clear channel %" PRIu64
, channel
->key
);
5169 if (channel
->type
== CONSUMER_CHANNEL_TYPE_METADATA
) {
5171 * Nothing to do for the metadata channel/stream.
5172 * Snapshot mechanism already take care of the metadata
5173 * handling/generation, and monitored channels only need to
5174 * have their data stream cleared..
5176 ret
= LTTCOMM_CONSUMERD_SUCCESS
;
5180 if (!channel
->monitor
) {
5181 ret
= consumer_clear_unmonitored_channel(channel
);
5183 ret
= consumer_clear_monitored_channel(channel
);
5189 enum lttcomm_return_code
lttng_consumer_open_channel_packets(
5190 struct lttng_consumer_channel
*channel
)
5192 struct lttng_consumer_stream
*stream
;
5193 enum lttcomm_return_code ret
= LTTCOMM_CONSUMERD_SUCCESS
;
5195 if (channel
->metadata_stream
) {
5196 ERR("Open channel packets command attempted on a metadata channel");
5197 ret
= LTTCOMM_CONSUMERD_INVALID_PARAMETERS
;
5202 cds_list_for_each_entry(stream
, &channel
->streams
.head
, send_node
) {
5203 enum consumer_stream_open_packet_status status
;
5205 pthread_mutex_lock(&stream
->lock
);
5206 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
5210 status
= consumer_stream_open_packet(stream
);
5212 case CONSUMER_STREAM_OPEN_PACKET_STATUS_OPENED
:
5213 DBG("Opened a packet in \"open channel packets\" command: stream id = %" PRIu64
5214 ", channel name = %s, session id = %" PRIu64
,
5215 stream
->key
, stream
->chan
->name
,
5216 stream
->chan
->session_id
);
5217 stream
->opened_packet_in_current_trace_chunk
= true;
5219 case CONSUMER_STREAM_OPEN_PACKET_STATUS_NO_SPACE
:
5220 DBG("No space left to open a packet in \"open channel packets\" command: stream id = %" PRIu64
5221 ", channel name = %s, session id = %" PRIu64
,
5222 stream
->key
, stream
->chan
->name
,
5223 stream
->chan
->session_id
);
5225 case CONSUMER_STREAM_OPEN_PACKET_STATUS_ERROR
:
5227 * Only unexpected internal errors can lead to this
5228 * failing. Report an unknown error.
5230 ERR("Failed to flush empty buffer in \"open channel packets\" command: stream id = %" PRIu64
5231 ", channel id = %" PRIu64
5232 ", channel name = %s"
5233 ", session id = %" PRIu64
,
5234 stream
->key
, channel
->key
,
5235 channel
->name
, channel
->session_id
);
5236 ret
= LTTCOMM_CONSUMERD_UNKNOWN_ERROR
;
5243 pthread_mutex_unlock(&stream
->lock
);
5252 pthread_mutex_unlock(&stream
->lock
);
5253 goto end_rcu_unlock
;
5256 void lttng_consumer_sigbus_handle(void *addr
)
5258 lttng_ustconsumer_sigbus_handle(addr
);