2 * Copyright (C) 2011 - Julien Desfossez <julien.desfossez@polymtl.ca>
3 * Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
4 * 2012 - David Goulet <dgoulet@efficios.com>
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License, version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
15 * You should have received a copy of the GNU General Public License along
16 * with this program; if not, write to the Free Software Foundation, Inc.,
17 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
27 #include <sys/socket.h>
28 #include <sys/types.h>
33 #include <bin/lttng-consumerd/health-consumerd.h>
34 #include <common/common.h>
35 #include <common/utils.h>
36 #include <common/compat/poll.h>
37 #include <common/compat/endian.h>
38 #include <common/index/index.h>
39 #include <common/kernel-ctl/kernel-ctl.h>
40 #include <common/sessiond-comm/relayd.h>
41 #include <common/sessiond-comm/sessiond-comm.h>
42 #include <common/kernel-consumer/kernel-consumer.h>
43 #include <common/relayd/relayd.h>
44 #include <common/ust-consumer/ust-consumer.h>
45 #include <common/consumer/consumer-timer.h>
46 #include <common/consumer/consumer.h>
47 #include <common/consumer/consumer-stream.h>
48 #include <common/consumer/consumer-testpoint.h>
49 #include <common/align.h>
50 #include <common/consumer/consumer-metadata-cache.h>
52 struct lttng_consumer_global_data consumer_data
= {
55 .type
= LTTNG_CONSUMER_UNKNOWN
,
58 enum consumer_channel_action
{
61 CONSUMER_CHANNEL_QUIT
,
64 struct consumer_channel_msg
{
65 enum consumer_channel_action action
;
66 struct lttng_consumer_channel
*chan
; /* add */
67 uint64_t key
; /* del */
70 int data_consumption_paused
;
73 * Flag to inform the polling thread to quit when all fd hung up. Updated by
74 * the consumer_thread_receive_fds when it notices that all fds has hung up.
75 * Also updated by the signal handler (consumer_should_exit()). Read by the
78 volatile int consumer_quit
;
81 * Global hash table containing respectively metadata and data streams. The
82 * stream element in this ht should only be updated by the metadata poll thread
83 * for the metadata and the data poll thread for the data.
85 static struct lttng_ht
*metadata_ht
;
86 static struct lttng_ht
*data_ht
;
89 * Notify a thread lttng pipe to poll back again. This usually means that some
90 * global state has changed so we just send back the thread in a poll wait
93 static void notify_thread_lttng_pipe(struct lttng_pipe
*pipe
)
95 struct lttng_consumer_stream
*null_stream
= NULL
;
99 (void) lttng_pipe_write(pipe
, &null_stream
, sizeof(null_stream
));
102 static void notify_health_quit_pipe(int *pipe
)
106 ret
= lttng_write(pipe
[1], "4", 1);
108 PERROR("write consumer health quit");
112 static void notify_channel_pipe(struct lttng_consumer_local_data
*ctx
,
113 struct lttng_consumer_channel
*chan
,
115 enum consumer_channel_action action
)
117 struct consumer_channel_msg msg
;
120 memset(&msg
, 0, sizeof(msg
));
125 ret
= lttng_write(ctx
->consumer_channel_pipe
[1], &msg
, sizeof(msg
));
126 if (ret
< sizeof(msg
)) {
127 PERROR("notify_channel_pipe write error");
131 void notify_thread_del_channel(struct lttng_consumer_local_data
*ctx
,
134 notify_channel_pipe(ctx
, NULL
, key
, CONSUMER_CHANNEL_DEL
);
137 static int read_channel_pipe(struct lttng_consumer_local_data
*ctx
,
138 struct lttng_consumer_channel
**chan
,
140 enum consumer_channel_action
*action
)
142 struct consumer_channel_msg msg
;
145 ret
= lttng_read(ctx
->consumer_channel_pipe
[0], &msg
, sizeof(msg
));
146 if (ret
< sizeof(msg
)) {
150 *action
= msg
.action
;
158 * Cleanup the stream list of a channel. Those streams are not yet globally
161 static void clean_channel_stream_list(struct lttng_consumer_channel
*channel
)
163 struct lttng_consumer_stream
*stream
, *stmp
;
167 /* Delete streams that might have been left in the stream list. */
168 cds_list_for_each_entry_safe(stream
, stmp
, &channel
->streams
.head
,
170 cds_list_del(&stream
->send_node
);
172 * Once a stream is added to this list, the buffers were created so we
173 * have a guarantee that this call will succeed. Setting the monitor
174 * mode to 0 so we don't lock nor try to delete the stream from the
178 consumer_stream_destroy(stream
, NULL
);
183 * Find a stream. The consumer_data.lock must be locked during this
186 static struct lttng_consumer_stream
*find_stream(uint64_t key
,
189 struct lttng_ht_iter iter
;
190 struct lttng_ht_node_u64
*node
;
191 struct lttng_consumer_stream
*stream
= NULL
;
195 /* -1ULL keys are lookup failures */
196 if (key
== (uint64_t) -1ULL) {
202 lttng_ht_lookup(ht
, &key
, &iter
);
203 node
= lttng_ht_iter_get_node_u64(&iter
);
205 stream
= caa_container_of(node
, struct lttng_consumer_stream
, node
);
213 static void steal_stream_key(uint64_t key
, struct lttng_ht
*ht
)
215 struct lttng_consumer_stream
*stream
;
218 stream
= find_stream(key
, ht
);
220 stream
->key
= (uint64_t) -1ULL;
222 * We don't want the lookup to match, but we still need
223 * to iterate on this stream when iterating over the hash table. Just
224 * change the node key.
226 stream
->node
.key
= (uint64_t) -1ULL;
232 * Return a channel object for the given key.
234 * RCU read side lock MUST be acquired before calling this function and
235 * protects the channel ptr.
237 struct lttng_consumer_channel
*consumer_find_channel(uint64_t key
)
239 struct lttng_ht_iter iter
;
240 struct lttng_ht_node_u64
*node
;
241 struct lttng_consumer_channel
*channel
= NULL
;
243 /* -1ULL keys are lookup failures */
244 if (key
== (uint64_t) -1ULL) {
248 lttng_ht_lookup(consumer_data
.channel_ht
, &key
, &iter
);
249 node
= lttng_ht_iter_get_node_u64(&iter
);
251 channel
= caa_container_of(node
, struct lttng_consumer_channel
, node
);
258 * There is a possibility that the consumer does not have enough time between
259 * the close of the channel on the session daemon and the cleanup in here thus
260 * once we have a channel add with an existing key, we know for sure that this
261 * channel will eventually get cleaned up by all streams being closed.
263 * This function just nullifies the already existing channel key.
265 static void steal_channel_key(uint64_t key
)
267 struct lttng_consumer_channel
*channel
;
270 channel
= consumer_find_channel(key
);
272 channel
->key
= (uint64_t) -1ULL;
274 * We don't want the lookup to match, but we still need to iterate on
275 * this channel when iterating over the hash table. Just change the
278 channel
->node
.key
= (uint64_t) -1ULL;
283 static void free_channel_rcu(struct rcu_head
*head
)
285 struct lttng_ht_node_u64
*node
=
286 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
287 struct lttng_consumer_channel
*channel
=
288 caa_container_of(node
, struct lttng_consumer_channel
, node
);
290 switch (consumer_data
.type
) {
291 case LTTNG_CONSUMER_KERNEL
:
293 case LTTNG_CONSUMER32_UST
:
294 case LTTNG_CONSUMER64_UST
:
295 lttng_ustconsumer_free_channel(channel
);
298 ERR("Unknown consumer_data type");
305 * RCU protected relayd socket pair free.
307 static void free_relayd_rcu(struct rcu_head
*head
)
309 struct lttng_ht_node_u64
*node
=
310 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
311 struct consumer_relayd_sock_pair
*relayd
=
312 caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
315 * Close all sockets. This is done in the call RCU since we don't want the
316 * socket fds to be reassigned thus potentially creating bad state of the
319 * We do not have to lock the control socket mutex here since at this stage
320 * there is no one referencing to this relayd object.
322 (void) relayd_close(&relayd
->control_sock
);
323 (void) relayd_close(&relayd
->data_sock
);
329 * Destroy and free relayd socket pair object.
331 void consumer_destroy_relayd(struct consumer_relayd_sock_pair
*relayd
)
334 struct lttng_ht_iter iter
;
336 if (relayd
== NULL
) {
340 DBG("Consumer destroy and close relayd socket pair");
342 iter
.iter
.node
= &relayd
->node
.node
;
343 ret
= lttng_ht_del(consumer_data
.relayd_ht
, &iter
);
345 /* We assume the relayd is being or is destroyed */
349 /* RCU free() call */
350 call_rcu(&relayd
->node
.head
, free_relayd_rcu
);
354 * Remove a channel from the global list protected by a mutex. This function is
355 * also responsible for freeing its data structures.
357 void consumer_del_channel(struct lttng_consumer_channel
*channel
)
360 struct lttng_ht_iter iter
;
362 DBG("Consumer delete channel key %" PRIu64
, channel
->key
);
364 pthread_mutex_lock(&consumer_data
.lock
);
365 pthread_mutex_lock(&channel
->lock
);
367 /* Destroy streams that might have been left in the stream list. */
368 clean_channel_stream_list(channel
);
370 if (channel
->live_timer_enabled
== 1) {
371 consumer_timer_live_stop(channel
);
373 if (channel
->monitor_timer_enabled
== 1) {
374 consumer_timer_monitor_stop(channel
);
377 switch (consumer_data
.type
) {
378 case LTTNG_CONSUMER_KERNEL
:
380 case LTTNG_CONSUMER32_UST
:
381 case LTTNG_CONSUMER64_UST
:
382 lttng_ustconsumer_del_channel(channel
);
385 ERR("Unknown consumer_data type");
391 iter
.iter
.node
= &channel
->node
.node
;
392 ret
= lttng_ht_del(consumer_data
.channel_ht
, &iter
);
396 call_rcu(&channel
->node
.head
, free_channel_rcu
);
398 pthread_mutex_unlock(&channel
->lock
);
399 pthread_mutex_unlock(&consumer_data
.lock
);
403 * Iterate over the relayd hash table and destroy each element. Finally,
404 * destroy the whole hash table.
406 static void cleanup_relayd_ht(void)
408 struct lttng_ht_iter iter
;
409 struct consumer_relayd_sock_pair
*relayd
;
413 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
415 consumer_destroy_relayd(relayd
);
420 lttng_ht_destroy(consumer_data
.relayd_ht
);
424 * Update the end point status of all streams having the given network sequence
425 * index (relayd index).
427 * It's atomically set without having the stream mutex locked which is fine
428 * because we handle the write/read race with a pipe wakeup for each thread.
430 static void update_endpoint_status_by_netidx(uint64_t net_seq_idx
,
431 enum consumer_endpoint_status status
)
433 struct lttng_ht_iter iter
;
434 struct lttng_consumer_stream
*stream
;
436 DBG("Consumer set delete flag on stream by idx %" PRIu64
, net_seq_idx
);
440 /* Let's begin with metadata */
441 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
442 if (stream
->net_seq_idx
== net_seq_idx
) {
443 uatomic_set(&stream
->endpoint_status
, status
);
444 DBG("Delete flag set to metadata stream %d", stream
->wait_fd
);
448 /* Follow up by the data streams */
449 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
450 if (stream
->net_seq_idx
== net_seq_idx
) {
451 uatomic_set(&stream
->endpoint_status
, status
);
452 DBG("Delete flag set to data stream %d", stream
->wait_fd
);
459 * Cleanup a relayd object by flagging every associated streams for deletion,
460 * destroying the object meaning removing it from the relayd hash table,
461 * closing the sockets and freeing the memory in a RCU call.
463 * If a local data context is available, notify the threads that the streams'
464 * state have changed.
466 static void cleanup_relayd(struct consumer_relayd_sock_pair
*relayd
,
467 struct lttng_consumer_local_data
*ctx
)
473 DBG("Cleaning up relayd sockets");
475 /* Save the net sequence index before destroying the object */
476 netidx
= relayd
->net_seq_idx
;
479 * Delete the relayd from the relayd hash table, close the sockets and free
480 * the object in a RCU call.
482 consumer_destroy_relayd(relayd
);
484 /* Set inactive endpoint to all streams */
485 update_endpoint_status_by_netidx(netidx
, CONSUMER_ENDPOINT_INACTIVE
);
488 * With a local data context, notify the threads that the streams' state
489 * have changed. The write() action on the pipe acts as an "implicit"
490 * memory barrier ordering the updates of the end point status from the
491 * read of this status which happens AFTER receiving this notify.
494 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
495 notify_thread_lttng_pipe(ctx
->consumer_metadata_pipe
);
500 * Flag a relayd socket pair for destruction. Destroy it if the refcount
503 * RCU read side lock MUST be aquired before calling this function.
505 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair
*relayd
)
509 /* Set destroy flag for this object */
510 uatomic_set(&relayd
->destroy_flag
, 1);
512 /* Destroy the relayd if refcount is 0 */
513 if (uatomic_read(&relayd
->refcount
) == 0) {
514 consumer_destroy_relayd(relayd
);
519 * Completly destroy stream from every visiable data structure and the given
522 * One this call returns, the stream object is not longer usable nor visible.
524 void consumer_del_stream(struct lttng_consumer_stream
*stream
,
527 consumer_stream_destroy(stream
, ht
);
531 * XXX naming of del vs destroy is all mixed up.
533 void consumer_del_stream_for_data(struct lttng_consumer_stream
*stream
)
535 consumer_stream_destroy(stream
, data_ht
);
538 void consumer_del_stream_for_metadata(struct lttng_consumer_stream
*stream
)
540 consumer_stream_destroy(stream
, metadata_ht
);
543 struct lttng_consumer_stream
*consumer_allocate_stream(uint64_t channel_key
,
545 enum lttng_consumer_stream_state state
,
546 const char *channel_name
,
553 enum consumer_channel_type type
,
554 unsigned int monitor
)
557 struct lttng_consumer_stream
*stream
;
559 stream
= zmalloc(sizeof(*stream
));
560 if (stream
== NULL
) {
561 PERROR("malloc struct lttng_consumer_stream");
568 stream
->key
= stream_key
;
570 stream
->out_fd_offset
= 0;
571 stream
->output_written
= 0;
572 stream
->state
= state
;
575 stream
->net_seq_idx
= relayd_id
;
576 stream
->session_id
= session_id
;
577 stream
->monitor
= monitor
;
578 stream
->endpoint_status
= CONSUMER_ENDPOINT_ACTIVE
;
579 stream
->index_file
= NULL
;
580 stream
->last_sequence_number
= -1ULL;
581 pthread_mutex_init(&stream
->lock
, NULL
);
582 pthread_mutex_init(&stream
->metadata_timer_lock
, NULL
);
584 /* If channel is the metadata, flag this stream as metadata. */
585 if (type
== CONSUMER_CHANNEL_TYPE_METADATA
) {
586 stream
->metadata_flag
= 1;
587 /* Metadata is flat out. */
588 strncpy(stream
->name
, DEFAULT_METADATA_NAME
, sizeof(stream
->name
));
589 /* Live rendez-vous point. */
590 pthread_cond_init(&stream
->metadata_rdv
, NULL
);
591 pthread_mutex_init(&stream
->metadata_rdv_lock
, NULL
);
593 /* Format stream name to <channel_name>_<cpu_number> */
594 ret
= snprintf(stream
->name
, sizeof(stream
->name
), "%s_%d",
597 PERROR("snprintf stream name");
602 /* Key is always the wait_fd for streams. */
603 lttng_ht_node_init_u64(&stream
->node
, stream
->key
);
605 /* Init node per channel id key */
606 lttng_ht_node_init_u64(&stream
->node_channel_id
, channel_key
);
608 /* Init session id node with the stream session id */
609 lttng_ht_node_init_u64(&stream
->node_session_id
, stream
->session_id
);
611 DBG3("Allocated stream %s (key %" PRIu64
", chan_key %" PRIu64
612 " relayd_id %" PRIu64
", session_id %" PRIu64
,
613 stream
->name
, stream
->key
, channel_key
,
614 stream
->net_seq_idx
, stream
->session_id
);
630 * Add a stream to the global list protected by a mutex.
632 int consumer_add_data_stream(struct lttng_consumer_stream
*stream
)
634 struct lttng_ht
*ht
= data_ht
;
640 DBG3("Adding consumer stream %" PRIu64
, stream
->key
);
642 pthread_mutex_lock(&consumer_data
.lock
);
643 pthread_mutex_lock(&stream
->chan
->lock
);
644 pthread_mutex_lock(&stream
->chan
->timer_lock
);
645 pthread_mutex_lock(&stream
->lock
);
648 /* Steal stream identifier to avoid having streams with the same key */
649 steal_stream_key(stream
->key
, ht
);
651 lttng_ht_add_unique_u64(ht
, &stream
->node
);
653 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
654 &stream
->node_channel_id
);
657 * Add stream to the stream_list_ht of the consumer data. No need to steal
658 * the key since the HT does not use it and we allow to add redundant keys
661 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
664 * When nb_init_stream_left reaches 0, we don't need to trigger any action
665 * in terms of destroying the associated channel, because the action that
666 * causes the count to become 0 also causes a stream to be added. The
667 * channel deletion will thus be triggered by the following removal of this
670 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
671 /* Increment refcount before decrementing nb_init_stream_left */
673 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
676 /* Update consumer data once the node is inserted. */
677 consumer_data
.stream_count
++;
678 consumer_data
.need_update
= 1;
681 pthread_mutex_unlock(&stream
->lock
);
682 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
683 pthread_mutex_unlock(&stream
->chan
->lock
);
684 pthread_mutex_unlock(&consumer_data
.lock
);
689 void consumer_del_data_stream(struct lttng_consumer_stream
*stream
)
691 consumer_del_stream(stream
, data_ht
);
695 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
696 * be acquired before calling this.
698 static int add_relayd(struct consumer_relayd_sock_pair
*relayd
)
701 struct lttng_ht_node_u64
*node
;
702 struct lttng_ht_iter iter
;
706 lttng_ht_lookup(consumer_data
.relayd_ht
,
707 &relayd
->net_seq_idx
, &iter
);
708 node
= lttng_ht_iter_get_node_u64(&iter
);
712 lttng_ht_add_unique_u64(consumer_data
.relayd_ht
, &relayd
->node
);
719 * Allocate and return a consumer relayd socket.
721 static struct consumer_relayd_sock_pair
*consumer_allocate_relayd_sock_pair(
722 uint64_t net_seq_idx
)
724 struct consumer_relayd_sock_pair
*obj
= NULL
;
726 /* net sequence index of -1 is a failure */
727 if (net_seq_idx
== (uint64_t) -1ULL) {
731 obj
= zmalloc(sizeof(struct consumer_relayd_sock_pair
));
733 PERROR("zmalloc relayd sock");
737 obj
->net_seq_idx
= net_seq_idx
;
739 obj
->destroy_flag
= 0;
740 obj
->control_sock
.sock
.fd
= -1;
741 obj
->data_sock
.sock
.fd
= -1;
742 lttng_ht_node_init_u64(&obj
->node
, obj
->net_seq_idx
);
743 pthread_mutex_init(&obj
->ctrl_sock_mutex
, NULL
);
750 * Find a relayd socket pair in the global consumer data.
752 * Return the object if found else NULL.
753 * RCU read-side lock must be held across this call and while using the
756 struct consumer_relayd_sock_pair
*consumer_find_relayd(uint64_t key
)
758 struct lttng_ht_iter iter
;
759 struct lttng_ht_node_u64
*node
;
760 struct consumer_relayd_sock_pair
*relayd
= NULL
;
762 /* Negative keys are lookup failures */
763 if (key
== (uint64_t) -1ULL) {
767 lttng_ht_lookup(consumer_data
.relayd_ht
, &key
,
769 node
= lttng_ht_iter_get_node_u64(&iter
);
771 relayd
= caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
779 * Find a relayd and send the stream
781 * Returns 0 on success, < 0 on error
783 int consumer_send_relayd_stream(struct lttng_consumer_stream
*stream
,
787 struct consumer_relayd_sock_pair
*relayd
;
790 assert(stream
->net_seq_idx
!= -1ULL);
793 /* The stream is not metadata. Get relayd reference if exists. */
795 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
796 if (relayd
!= NULL
) {
797 /* Add stream on the relayd */
798 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
799 ret
= relayd_add_stream(&relayd
->control_sock
, stream
->name
,
800 path
, &stream
->relayd_stream_id
,
801 stream
->chan
->tracefile_size
, stream
->chan
->tracefile_count
);
802 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
807 uatomic_inc(&relayd
->refcount
);
808 stream
->sent_to_relayd
= 1;
810 ERR("Stream %" PRIu64
" relayd ID %" PRIu64
" unknown. Can't send it.",
811 stream
->key
, stream
->net_seq_idx
);
816 DBG("Stream %s with key %" PRIu64
" sent to relayd id %" PRIu64
,
817 stream
->name
, stream
->key
, stream
->net_seq_idx
);
825 * Find a relayd and send the streams sent message
827 * Returns 0 on success, < 0 on error
829 int consumer_send_relayd_streams_sent(uint64_t net_seq_idx
)
832 struct consumer_relayd_sock_pair
*relayd
;
834 assert(net_seq_idx
!= -1ULL);
836 /* The stream is not metadata. Get relayd reference if exists. */
838 relayd
= consumer_find_relayd(net_seq_idx
);
839 if (relayd
!= NULL
) {
840 /* Add stream on the relayd */
841 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
842 ret
= relayd_streams_sent(&relayd
->control_sock
);
843 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
848 ERR("Relayd ID %" PRIu64
" unknown. Can't send streams_sent.",
855 DBG("All streams sent relayd id %" PRIu64
, net_seq_idx
);
863 * Find a relayd and close the stream
865 void close_relayd_stream(struct lttng_consumer_stream
*stream
)
867 struct consumer_relayd_sock_pair
*relayd
;
869 /* The stream is not metadata. Get relayd reference if exists. */
871 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
873 consumer_stream_relayd_close(stream
, relayd
);
879 * Handle stream for relayd transmission if the stream applies for network
880 * streaming where the net sequence index is set.
882 * Return destination file descriptor or negative value on error.
884 static int write_relayd_stream_header(struct lttng_consumer_stream
*stream
,
885 size_t data_size
, unsigned long padding
,
886 struct consumer_relayd_sock_pair
*relayd
)
889 struct lttcomm_relayd_data_hdr data_hdr
;
895 /* Reset data header */
896 memset(&data_hdr
, 0, sizeof(data_hdr
));
898 if (stream
->metadata_flag
) {
899 /* Caller MUST acquire the relayd control socket lock */
900 ret
= relayd_send_metadata(&relayd
->control_sock
, data_size
);
905 /* Metadata are always sent on the control socket. */
906 outfd
= relayd
->control_sock
.sock
.fd
;
908 /* Set header with stream information */
909 data_hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
910 data_hdr
.data_size
= htobe32(data_size
);
911 data_hdr
.padding_size
= htobe32(padding
);
913 * Note that net_seq_num below is assigned with the *current* value of
914 * next_net_seq_num and only after that the next_net_seq_num will be
915 * increment. This is why when issuing a command on the relayd using
916 * this next value, 1 should always be substracted in order to compare
917 * the last seen sequence number on the relayd side to the last sent.
919 data_hdr
.net_seq_num
= htobe64(stream
->next_net_seq_num
);
920 /* Other fields are zeroed previously */
922 ret
= relayd_send_data_hdr(&relayd
->data_sock
, &data_hdr
,
928 ++stream
->next_net_seq_num
;
930 /* Set to go on data socket */
931 outfd
= relayd
->data_sock
.sock
.fd
;
939 * Allocate and return a new lttng_consumer_channel object using the given key
940 * to initialize the hash table node.
942 * On error, return NULL.
944 struct lttng_consumer_channel
*consumer_allocate_channel(uint64_t key
,
946 const char *pathname
,
951 enum lttng_event_output output
,
952 uint64_t tracefile_size
,
953 uint64_t tracefile_count
,
954 uint64_t session_id_per_pid
,
955 unsigned int monitor
,
956 unsigned int live_timer_interval
,
957 const char *root_shm_path
,
958 const char *shm_path
)
960 struct lttng_consumer_channel
*channel
;
962 channel
= zmalloc(sizeof(*channel
));
963 if (channel
== NULL
) {
964 PERROR("malloc struct lttng_consumer_channel");
969 channel
->refcount
= 0;
970 channel
->session_id
= session_id
;
971 channel
->session_id_per_pid
= session_id_per_pid
;
974 channel
->relayd_id
= relayd_id
;
975 channel
->tracefile_size
= tracefile_size
;
976 channel
->tracefile_count
= tracefile_count
;
977 channel
->monitor
= monitor
;
978 channel
->live_timer_interval
= live_timer_interval
;
979 pthread_mutex_init(&channel
->lock
, NULL
);
980 pthread_mutex_init(&channel
->timer_lock
, NULL
);
983 case LTTNG_EVENT_SPLICE
:
984 channel
->output
= CONSUMER_CHANNEL_SPLICE
;
986 case LTTNG_EVENT_MMAP
:
987 channel
->output
= CONSUMER_CHANNEL_MMAP
;
997 * In monitor mode, the streams associated with the channel will be put in
998 * a special list ONLY owned by this channel. So, the refcount is set to 1
999 * here meaning that the channel itself has streams that are referenced.
1001 * On a channel deletion, once the channel is no longer visible, the
1002 * refcount is decremented and checked for a zero value to delete it. With
1003 * streams in no monitor mode, it will now be safe to destroy the channel.
1005 if (!channel
->monitor
) {
1006 channel
->refcount
= 1;
1009 strncpy(channel
->pathname
, pathname
, sizeof(channel
->pathname
));
1010 channel
->pathname
[sizeof(channel
->pathname
) - 1] = '\0';
1012 strncpy(channel
->name
, name
, sizeof(channel
->name
));
1013 channel
->name
[sizeof(channel
->name
) - 1] = '\0';
1015 if (root_shm_path
) {
1016 strncpy(channel
->root_shm_path
, root_shm_path
, sizeof(channel
->root_shm_path
));
1017 channel
->root_shm_path
[sizeof(channel
->root_shm_path
) - 1] = '\0';
1020 strncpy(channel
->shm_path
, shm_path
, sizeof(channel
->shm_path
));
1021 channel
->shm_path
[sizeof(channel
->shm_path
) - 1] = '\0';
1024 lttng_ht_node_init_u64(&channel
->node
, channel
->key
);
1026 channel
->wait_fd
= -1;
1028 CDS_INIT_LIST_HEAD(&channel
->streams
.head
);
1030 DBG("Allocated channel (key %" PRIu64
")", channel
->key
);
1037 * Add a channel to the global list protected by a mutex.
1039 * Always return 0 indicating success.
1041 int consumer_add_channel(struct lttng_consumer_channel
*channel
,
1042 struct lttng_consumer_local_data
*ctx
)
1044 pthread_mutex_lock(&consumer_data
.lock
);
1045 pthread_mutex_lock(&channel
->lock
);
1046 pthread_mutex_lock(&channel
->timer_lock
);
1049 * This gives us a guarantee that the channel we are about to add to the
1050 * channel hash table will be unique. See this function comment on the why
1051 * we need to steel the channel key at this stage.
1053 steal_channel_key(channel
->key
);
1056 lttng_ht_add_unique_u64(consumer_data
.channel_ht
, &channel
->node
);
1059 pthread_mutex_unlock(&channel
->timer_lock
);
1060 pthread_mutex_unlock(&channel
->lock
);
1061 pthread_mutex_unlock(&consumer_data
.lock
);
1063 if (channel
->wait_fd
!= -1 && channel
->type
== CONSUMER_CHANNEL_TYPE_DATA
) {
1064 notify_channel_pipe(ctx
, channel
, -1, CONSUMER_CHANNEL_ADD
);
1071 * Allocate the pollfd structure and the local view of the out fds to avoid
1072 * doing a lookup in the linked list and concurrency issues when writing is
1073 * needed. Called with consumer_data.lock held.
1075 * Returns the number of fds in the structures.
1077 static int update_poll_array(struct lttng_consumer_local_data
*ctx
,
1078 struct pollfd
**pollfd
, struct lttng_consumer_stream
**local_stream
,
1079 struct lttng_ht
*ht
)
1082 struct lttng_ht_iter iter
;
1083 struct lttng_consumer_stream
*stream
;
1088 assert(local_stream
);
1090 DBG("Updating poll fd array");
1092 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1094 * Only active streams with an active end point can be added to the
1095 * poll set and local stream storage of the thread.
1097 * There is a potential race here for endpoint_status to be updated
1098 * just after the check. However, this is OK since the stream(s) will
1099 * be deleted once the thread is notified that the end point state has
1100 * changed where this function will be called back again.
1102 if (stream
->state
!= LTTNG_CONSUMER_ACTIVE_STREAM
||
1103 stream
->endpoint_status
== CONSUMER_ENDPOINT_INACTIVE
) {
1107 * This clobbers way too much the debug output. Uncomment that if you
1108 * need it for debugging purposes.
1110 * DBG("Active FD %d", stream->wait_fd);
1112 (*pollfd
)[i
].fd
= stream
->wait_fd
;
1113 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1114 local_stream
[i
] = stream
;
1120 * Insert the consumer_data_pipe at the end of the array and don't
1121 * increment i so nb_fd is the number of real FD.
1123 (*pollfd
)[i
].fd
= lttng_pipe_get_readfd(ctx
->consumer_data_pipe
);
1124 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1126 (*pollfd
)[i
+ 1].fd
= lttng_pipe_get_readfd(ctx
->consumer_wakeup_pipe
);
1127 (*pollfd
)[i
+ 1].events
= POLLIN
| POLLPRI
;
1132 * Poll on the should_quit pipe and the command socket return -1 on
1133 * error, 1 if should exit, 0 if data is available on the command socket
1135 int lttng_consumer_poll_socket(struct pollfd
*consumer_sockpoll
)
1140 num_rdy
= poll(consumer_sockpoll
, 2, -1);
1141 if (num_rdy
== -1) {
1143 * Restart interrupted system call.
1145 if (errno
== EINTR
) {
1148 PERROR("Poll error");
1151 if (consumer_sockpoll
[0].revents
& (POLLIN
| POLLPRI
)) {
1152 DBG("consumer_should_quit wake up");
1159 * Set the error socket.
1161 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data
*ctx
,
1164 ctx
->consumer_error_socket
= sock
;
1168 * Set the command socket path.
1170 void lttng_consumer_set_command_sock_path(
1171 struct lttng_consumer_local_data
*ctx
, char *sock
)
1173 ctx
->consumer_command_sock_path
= sock
;
1177 * Send return code to the session daemon.
1178 * If the socket is not defined, we return 0, it is not a fatal error
1180 int lttng_consumer_send_error(struct lttng_consumer_local_data
*ctx
, int cmd
)
1182 if (ctx
->consumer_error_socket
> 0) {
1183 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
1184 sizeof(enum lttcomm_sessiond_command
));
1191 * Close all the tracefiles and stream fds and MUST be called when all
1192 * instances are destroyed i.e. when all threads were joined and are ended.
1194 void lttng_consumer_cleanup(void)
1196 struct lttng_ht_iter iter
;
1197 struct lttng_consumer_channel
*channel
;
1201 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
, channel
,
1203 consumer_del_channel(channel
);
1208 lttng_ht_destroy(consumer_data
.channel_ht
);
1210 cleanup_relayd_ht();
1212 lttng_ht_destroy(consumer_data
.stream_per_chan_id_ht
);
1215 * This HT contains streams that are freed by either the metadata thread or
1216 * the data thread so we do *nothing* on the hash table and simply destroy
1219 lttng_ht_destroy(consumer_data
.stream_list_ht
);
1223 * Called from signal handler.
1225 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
1230 ret
= lttng_write(ctx
->consumer_should_quit
[1], "4", 1);
1232 PERROR("write consumer quit");
1235 DBG("Consumer flag that it should quit");
1240 * Flush pending writes to trace output disk file.
1243 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1247 int outfd
= stream
->out_fd
;
1250 * This does a blocking write-and-wait on any page that belongs to the
1251 * subbuffer prior to the one we just wrote.
1252 * Don't care about error values, as these are just hints and ways to
1253 * limit the amount of page cache used.
1255 if (orig_offset
< stream
->max_sb_size
) {
1258 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1259 stream
->max_sb_size
,
1260 SYNC_FILE_RANGE_WAIT_BEFORE
1261 | SYNC_FILE_RANGE_WRITE
1262 | SYNC_FILE_RANGE_WAIT_AFTER
);
1264 * Give hints to the kernel about how we access the file:
1265 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1268 * We need to call fadvise again after the file grows because the
1269 * kernel does not seem to apply fadvise to non-existing parts of the
1272 * Call fadvise _after_ having waited for the page writeback to
1273 * complete because the dirty page writeback semantic is not well
1274 * defined. So it can be expected to lead to lower throughput in
1277 ret
= posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1278 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1279 if (ret
&& ret
!= -ENOSYS
) {
1281 PERROR("posix_fadvise on fd %i", outfd
);
1286 * Initialise the necessary environnement :
1287 * - create a new context
1288 * - create the poll_pipe
1289 * - create the should_quit pipe (for signal handler)
1290 * - create the thread pipe (for splice)
1292 * Takes a function pointer as argument, this function is called when data is
1293 * available on a buffer. This function is responsible to do the
1294 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1295 * buffer configuration and then kernctl_put_next_subbuf at the end.
1297 * Returns a pointer to the new context or NULL on error.
1299 struct lttng_consumer_local_data
*lttng_consumer_create(
1300 enum lttng_consumer_type type
,
1301 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1302 struct lttng_consumer_local_data
*ctx
),
1303 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1304 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1305 int (*update_stream
)(uint64_t stream_key
, uint32_t state
))
1308 struct lttng_consumer_local_data
*ctx
;
1310 assert(consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1311 consumer_data
.type
== type
);
1312 consumer_data
.type
= type
;
1314 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
1316 PERROR("allocating context");
1320 ctx
->consumer_error_socket
= -1;
1321 ctx
->consumer_metadata_socket
= -1;
1322 pthread_mutex_init(&ctx
->metadata_socket_lock
, NULL
);
1323 /* assign the callbacks */
1324 ctx
->on_buffer_ready
= buffer_ready
;
1325 ctx
->on_recv_channel
= recv_channel
;
1326 ctx
->on_recv_stream
= recv_stream
;
1327 ctx
->on_update_stream
= update_stream
;
1329 ctx
->consumer_data_pipe
= lttng_pipe_open(0);
1330 if (!ctx
->consumer_data_pipe
) {
1331 goto error_poll_pipe
;
1334 ctx
->consumer_wakeup_pipe
= lttng_pipe_open(0);
1335 if (!ctx
->consumer_wakeup_pipe
) {
1336 goto error_wakeup_pipe
;
1339 ret
= pipe(ctx
->consumer_should_quit
);
1341 PERROR("Error creating recv pipe");
1342 goto error_quit_pipe
;
1345 ret
= pipe(ctx
->consumer_channel_pipe
);
1347 PERROR("Error creating channel pipe");
1348 goto error_channel_pipe
;
1351 ctx
->consumer_metadata_pipe
= lttng_pipe_open(0);
1352 if (!ctx
->consumer_metadata_pipe
) {
1353 goto error_metadata_pipe
;
1356 ctx
->channel_monitor_pipe
= -1;
1360 error_metadata_pipe
:
1361 utils_close_pipe(ctx
->consumer_channel_pipe
);
1363 utils_close_pipe(ctx
->consumer_should_quit
);
1365 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1367 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1375 * Iterate over all streams of the hashtable and free them properly.
1377 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1379 struct lttng_ht_iter iter
;
1380 struct lttng_consumer_stream
*stream
;
1387 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1389 * Ignore return value since we are currently cleaning up so any error
1392 (void) consumer_del_stream(stream
, ht
);
1396 lttng_ht_destroy(ht
);
1400 * Iterate over all streams of the metadata hashtable and free them
1403 static void destroy_metadata_stream_ht(struct lttng_ht
*ht
)
1405 struct lttng_ht_iter iter
;
1406 struct lttng_consumer_stream
*stream
;
1413 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1415 * Ignore return value since we are currently cleaning up so any error
1418 (void) consumer_del_metadata_stream(stream
, ht
);
1422 lttng_ht_destroy(ht
);
1426 * Close all fds associated with the instance and free the context.
1428 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1432 DBG("Consumer destroying it. Closing everything.");
1438 destroy_data_stream_ht(data_ht
);
1439 destroy_metadata_stream_ht(metadata_ht
);
1441 ret
= close(ctx
->consumer_error_socket
);
1445 ret
= close(ctx
->consumer_metadata_socket
);
1449 utils_close_pipe(ctx
->consumer_channel_pipe
);
1450 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1451 lttng_pipe_destroy(ctx
->consumer_metadata_pipe
);
1452 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1453 utils_close_pipe(ctx
->consumer_should_quit
);
1455 unlink(ctx
->consumer_command_sock_path
);
1460 * Write the metadata stream id on the specified file descriptor.
1462 static int write_relayd_metadata_id(int fd
,
1463 struct lttng_consumer_stream
*stream
,
1464 struct consumer_relayd_sock_pair
*relayd
, unsigned long padding
)
1467 struct lttcomm_relayd_metadata_payload hdr
;
1469 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1470 hdr
.padding_size
= htobe32(padding
);
1471 ret
= lttng_write(fd
, (void *) &hdr
, sizeof(hdr
));
1472 if (ret
< sizeof(hdr
)) {
1474 * This error means that the fd's end is closed so ignore the PERROR
1475 * not to clubber the error output since this can happen in a normal
1478 if (errno
!= EPIPE
) {
1479 PERROR("write metadata stream id");
1481 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1483 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1484 * handle writting the missing part so report that as an error and
1485 * don't lie to the caller.
1490 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1491 stream
->relayd_stream_id
, padding
);
1498 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1499 * core function for writing trace buffers to either the local filesystem or
1502 * It must be called with the stream lock held.
1504 * Careful review MUST be put if any changes occur!
1506 * Returns the number of bytes written
1508 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1509 struct lttng_consumer_local_data
*ctx
,
1510 struct lttng_consumer_stream
*stream
, unsigned long len
,
1511 unsigned long padding
,
1512 struct ctf_packet_index
*index
)
1514 unsigned long mmap_offset
;
1517 off_t orig_offset
= stream
->out_fd_offset
;
1518 /* Default is on the disk */
1519 int outfd
= stream
->out_fd
;
1520 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1521 unsigned int relayd_hang_up
= 0;
1523 /* RCU lock for the relayd pointer */
1526 /* Flag that the current stream if set for network streaming. */
1527 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1528 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1529 if (relayd
== NULL
) {
1535 /* get the offset inside the fd to mmap */
1536 switch (consumer_data
.type
) {
1537 case LTTNG_CONSUMER_KERNEL
:
1538 mmap_base
= stream
->mmap_base
;
1539 ret
= kernctl_get_mmap_read_offset(stream
->wait_fd
, &mmap_offset
);
1541 PERROR("tracer ctl get_mmap_read_offset");
1545 case LTTNG_CONSUMER32_UST
:
1546 case LTTNG_CONSUMER64_UST
:
1547 mmap_base
= lttng_ustctl_get_mmap_base(stream
);
1549 ERR("read mmap get mmap base for stream %s", stream
->name
);
1553 ret
= lttng_ustctl_get_mmap_read_offset(stream
, &mmap_offset
);
1555 PERROR("tracer ctl get_mmap_read_offset");
1561 ERR("Unknown consumer_data type");
1565 /* Handle stream on the relayd if the output is on the network */
1567 unsigned long netlen
= len
;
1570 * Lock the control socket for the complete duration of the function
1571 * since from this point on we will use the socket.
1573 if (stream
->metadata_flag
) {
1574 /* Metadata requires the control socket. */
1575 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1576 if (stream
->reset_metadata_flag
) {
1577 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1578 stream
->relayd_stream_id
,
1579 stream
->metadata_version
);
1584 stream
->reset_metadata_flag
= 0;
1586 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1589 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1594 /* Use the returned socket. */
1597 /* Write metadata stream id before payload */
1598 if (stream
->metadata_flag
) {
1599 ret
= write_relayd_metadata_id(outfd
, stream
, relayd
, padding
);
1606 /* No streaming, we have to set the len with the full padding */
1609 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1610 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1612 ERR("Reset metadata file");
1615 stream
->reset_metadata_flag
= 0;
1619 * Check if we need to change the tracefile before writing the packet.
1621 if (stream
->chan
->tracefile_size
> 0 &&
1622 (stream
->tracefile_size_current
+ len
) >
1623 stream
->chan
->tracefile_size
) {
1624 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1625 stream
->name
, stream
->chan
->tracefile_size
,
1626 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1627 stream
->out_fd
, &(stream
->tracefile_count_current
),
1630 ERR("Rotating output file");
1633 outfd
= stream
->out_fd
;
1635 if (stream
->index_file
) {
1636 lttng_index_file_put(stream
->index_file
);
1637 stream
->index_file
= lttng_index_file_create(stream
->chan
->pathname
,
1638 stream
->name
, stream
->uid
, stream
->gid
,
1639 stream
->chan
->tracefile_size
,
1640 stream
->tracefile_count_current
,
1641 CTF_INDEX_MAJOR
, CTF_INDEX_MINOR
);
1642 if (!stream
->index_file
) {
1647 /* Reset current size because we just perform a rotation. */
1648 stream
->tracefile_size_current
= 0;
1649 stream
->out_fd_offset
= 0;
1652 stream
->tracefile_size_current
+= len
;
1654 index
->offset
= htobe64(stream
->out_fd_offset
);
1659 * This call guarantee that len or less is returned. It's impossible to
1660 * receive a ret value that is bigger than len.
1662 ret
= lttng_write(outfd
, mmap_base
+ mmap_offset
, len
);
1663 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, len
);
1664 if (ret
< 0 || ((size_t) ret
!= len
)) {
1666 * Report error to caller if nothing was written else at least send the
1674 /* Socket operation failed. We consider the relayd dead */
1675 if (errno
== EPIPE
|| errno
== EINVAL
|| errno
== EBADF
) {
1677 * This is possible if the fd is closed on the other side
1678 * (outfd) or any write problem. It can be verbose a bit for a
1679 * normal execution if for instance the relayd is stopped
1680 * abruptly. This can happen so set this to a DBG statement.
1682 DBG("Consumer mmap write detected relayd hang up");
1684 /* Unhandled error, print it and stop function right now. */
1685 PERROR("Error in write mmap (ret %zd != len %lu)", ret
, len
);
1689 stream
->output_written
+= ret
;
1691 /* This call is useless on a socket so better save a syscall. */
1693 /* This won't block, but will start writeout asynchronously */
1694 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, len
,
1695 SYNC_FILE_RANGE_WRITE
);
1696 stream
->out_fd_offset
+= len
;
1697 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1702 * This is a special case that the relayd has closed its socket. Let's
1703 * cleanup the relayd object and all associated streams.
1705 if (relayd
&& relayd_hang_up
) {
1706 cleanup_relayd(relayd
, ctx
);
1710 /* Unlock only if ctrl socket used */
1711 if (relayd
&& stream
->metadata_flag
) {
1712 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1720 * Splice the data from the ring buffer to the tracefile.
1722 * It must be called with the stream lock held.
1724 * Returns the number of bytes spliced.
1726 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1727 struct lttng_consumer_local_data
*ctx
,
1728 struct lttng_consumer_stream
*stream
, unsigned long len
,
1729 unsigned long padding
,
1730 struct ctf_packet_index
*index
)
1732 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1734 off_t orig_offset
= stream
->out_fd_offset
;
1735 int fd
= stream
->wait_fd
;
1736 /* Default is on the disk */
1737 int outfd
= stream
->out_fd
;
1738 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1740 unsigned int relayd_hang_up
= 0;
1742 switch (consumer_data
.type
) {
1743 case LTTNG_CONSUMER_KERNEL
:
1745 case LTTNG_CONSUMER32_UST
:
1746 case LTTNG_CONSUMER64_UST
:
1747 /* Not supported for user space tracing */
1750 ERR("Unknown consumer_data type");
1754 /* RCU lock for the relayd pointer */
1757 /* Flag that the current stream if set for network streaming. */
1758 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1759 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1760 if (relayd
== NULL
) {
1765 splice_pipe
= stream
->splice_pipe
;
1767 /* Write metadata stream id before payload */
1769 unsigned long total_len
= len
;
1771 if (stream
->metadata_flag
) {
1773 * Lock the control socket for the complete duration of the function
1774 * since from this point on we will use the socket.
1776 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1778 if (stream
->reset_metadata_flag
) {
1779 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1780 stream
->relayd_stream_id
,
1781 stream
->metadata_version
);
1786 stream
->reset_metadata_flag
= 0;
1788 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
, relayd
,
1796 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1799 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1805 /* Use the returned socket. */
1808 /* No streaming, we have to set the len with the full padding */
1811 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1812 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1814 ERR("Reset metadata file");
1817 stream
->reset_metadata_flag
= 0;
1820 * Check if we need to change the tracefile before writing the packet.
1822 if (stream
->chan
->tracefile_size
> 0 &&
1823 (stream
->tracefile_size_current
+ len
) >
1824 stream
->chan
->tracefile_size
) {
1825 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1826 stream
->name
, stream
->chan
->tracefile_size
,
1827 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1828 stream
->out_fd
, &(stream
->tracefile_count_current
),
1832 ERR("Rotating output file");
1835 outfd
= stream
->out_fd
;
1837 if (stream
->index_file
) {
1838 lttng_index_file_put(stream
->index_file
);
1839 stream
->index_file
= lttng_index_file_create(stream
->chan
->pathname
,
1840 stream
->name
, stream
->uid
, stream
->gid
,
1841 stream
->chan
->tracefile_size
,
1842 stream
->tracefile_count_current
,
1843 CTF_INDEX_MAJOR
, CTF_INDEX_MINOR
);
1844 if (!stream
->index_file
) {
1849 /* Reset current size because we just perform a rotation. */
1850 stream
->tracefile_size_current
= 0;
1851 stream
->out_fd_offset
= 0;
1854 stream
->tracefile_size_current
+= len
;
1855 index
->offset
= htobe64(stream
->out_fd_offset
);
1859 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1860 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1861 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1862 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1863 DBG("splice chan to pipe, ret %zd", ret_splice
);
1864 if (ret_splice
< 0) {
1867 PERROR("Error in relay splice");
1871 /* Handle stream on the relayd if the output is on the network */
1872 if (relayd
&& stream
->metadata_flag
) {
1873 size_t metadata_payload_size
=
1874 sizeof(struct lttcomm_relayd_metadata_payload
);
1876 /* Update counter to fit the spliced data */
1877 ret_splice
+= metadata_payload_size
;
1878 len
+= metadata_payload_size
;
1880 * We do this so the return value can match the len passed as
1881 * argument to this function.
1883 written
-= metadata_payload_size
;
1886 /* Splice data out */
1887 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1888 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1889 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
1891 if (ret_splice
< 0) {
1896 } else if (ret_splice
> len
) {
1898 * We don't expect this code path to be executed but you never know
1899 * so this is an extra protection agains a buggy splice().
1902 written
+= ret_splice
;
1903 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice
,
1907 /* All good, update current len and continue. */
1911 /* This call is useless on a socket so better save a syscall. */
1913 /* This won't block, but will start writeout asynchronously */
1914 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1915 SYNC_FILE_RANGE_WRITE
);
1916 stream
->out_fd_offset
+= ret_splice
;
1918 stream
->output_written
+= ret_splice
;
1919 written
+= ret_splice
;
1922 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1928 * This is a special case that the relayd has closed its socket. Let's
1929 * cleanup the relayd object and all associated streams.
1931 if (relayd
&& relayd_hang_up
) {
1932 cleanup_relayd(relayd
, ctx
);
1933 /* Skip splice error so the consumer does not fail */
1938 /* send the appropriate error description to sessiond */
1941 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1944 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1947 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
1952 if (relayd
&& stream
->metadata_flag
) {
1953 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1961 * Take a snapshot for a specific fd
1963 * Returns 0 on success, < 0 on error
1965 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
1967 switch (consumer_data
.type
) {
1968 case LTTNG_CONSUMER_KERNEL
:
1969 return lttng_kconsumer_take_snapshot(stream
);
1970 case LTTNG_CONSUMER32_UST
:
1971 case LTTNG_CONSUMER64_UST
:
1972 return lttng_ustconsumer_take_snapshot(stream
);
1974 ERR("Unknown consumer_data type");
1981 * Get the produced position
1983 * Returns 0 on success, < 0 on error
1985 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
1988 switch (consumer_data
.type
) {
1989 case LTTNG_CONSUMER_KERNEL
:
1990 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
1991 case LTTNG_CONSUMER32_UST
:
1992 case LTTNG_CONSUMER64_UST
:
1993 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
1995 ERR("Unknown consumer_data type");
2001 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
2002 int sock
, struct pollfd
*consumer_sockpoll
)
2004 switch (consumer_data
.type
) {
2005 case LTTNG_CONSUMER_KERNEL
:
2006 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2007 case LTTNG_CONSUMER32_UST
:
2008 case LTTNG_CONSUMER64_UST
:
2009 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2011 ERR("Unknown consumer_data type");
2017 void lttng_consumer_close_all_metadata(void)
2019 switch (consumer_data
.type
) {
2020 case LTTNG_CONSUMER_KERNEL
:
2022 * The Kernel consumer has a different metadata scheme so we don't
2023 * close anything because the stream will be closed by the session
2027 case LTTNG_CONSUMER32_UST
:
2028 case LTTNG_CONSUMER64_UST
:
2030 * Close all metadata streams. The metadata hash table is passed and
2031 * this call iterates over it by closing all wakeup fd. This is safe
2032 * because at this point we are sure that the metadata producer is
2033 * either dead or blocked.
2035 lttng_ustconsumer_close_all_metadata(metadata_ht
);
2038 ERR("Unknown consumer_data type");
2044 * Clean up a metadata stream and free its memory.
2046 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
2047 struct lttng_ht
*ht
)
2049 struct lttng_consumer_channel
*free_chan
= NULL
;
2053 * This call should NEVER receive regular stream. It must always be
2054 * metadata stream and this is crucial for data structure synchronization.
2056 assert(stream
->metadata_flag
);
2058 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
2060 pthread_mutex_lock(&consumer_data
.lock
);
2061 pthread_mutex_lock(&stream
->chan
->lock
);
2062 pthread_mutex_lock(&stream
->lock
);
2063 if (stream
->chan
->metadata_cache
) {
2064 /* Only applicable to userspace consumers. */
2065 pthread_mutex_lock(&stream
->chan
->metadata_cache
->lock
);
2068 /* Remove any reference to that stream. */
2069 consumer_stream_delete(stream
, ht
);
2071 /* Close down everything including the relayd if one. */
2072 consumer_stream_close(stream
);
2073 /* Destroy tracer buffers of the stream. */
2074 consumer_stream_destroy_buffers(stream
);
2076 /* Atomically decrement channel refcount since other threads can use it. */
2077 if (!uatomic_sub_return(&stream
->chan
->refcount
, 1)
2078 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
2079 /* Go for channel deletion! */
2080 free_chan
= stream
->chan
;
2084 * Nullify the stream reference so it is not used after deletion. The
2085 * channel lock MUST be acquired before being able to check for a NULL
2088 stream
->chan
->metadata_stream
= NULL
;
2090 if (stream
->chan
->metadata_cache
) {
2091 pthread_mutex_unlock(&stream
->chan
->metadata_cache
->lock
);
2093 pthread_mutex_unlock(&stream
->lock
);
2094 pthread_mutex_unlock(&stream
->chan
->lock
);
2095 pthread_mutex_unlock(&consumer_data
.lock
);
2098 consumer_del_channel(free_chan
);
2101 consumer_stream_free(stream
);
2105 * Action done with the metadata stream when adding it to the consumer internal
2106 * data structures to handle it.
2108 int consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
2110 struct lttng_ht
*ht
= metadata_ht
;
2112 struct lttng_ht_iter iter
;
2113 struct lttng_ht_node_u64
*node
;
2118 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2120 pthread_mutex_lock(&consumer_data
.lock
);
2121 pthread_mutex_lock(&stream
->chan
->lock
);
2122 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2123 pthread_mutex_lock(&stream
->lock
);
2126 * From here, refcounts are updated so be _careful_ when returning an error
2133 * Lookup the stream just to make sure it does not exist in our internal
2134 * state. This should NEVER happen.
2136 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2137 node
= lttng_ht_iter_get_node_u64(&iter
);
2141 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2142 * in terms of destroying the associated channel, because the action that
2143 * causes the count to become 0 also causes a stream to be added. The
2144 * channel deletion will thus be triggered by the following removal of this
2147 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2148 /* Increment refcount before decrementing nb_init_stream_left */
2150 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2153 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2155 lttng_ht_add_unique_u64(consumer_data
.stream_per_chan_id_ht
,
2156 &stream
->node_channel_id
);
2159 * Add stream to the stream_list_ht of the consumer data. No need to steal
2160 * the key since the HT does not use it and we allow to add redundant keys
2163 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2167 pthread_mutex_unlock(&stream
->lock
);
2168 pthread_mutex_unlock(&stream
->chan
->lock
);
2169 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2170 pthread_mutex_unlock(&consumer_data
.lock
);
2175 * Delete data stream that are flagged for deletion (endpoint_status).
2177 static void validate_endpoint_status_data_stream(void)
2179 struct lttng_ht_iter iter
;
2180 struct lttng_consumer_stream
*stream
;
2182 DBG("Consumer delete flagged data stream");
2185 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2186 /* Validate delete flag of the stream */
2187 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2190 /* Delete it right now */
2191 consumer_del_stream(stream
, data_ht
);
2197 * Delete metadata stream that are flagged for deletion (endpoint_status).
2199 static void validate_endpoint_status_metadata_stream(
2200 struct lttng_poll_event
*pollset
)
2202 struct lttng_ht_iter iter
;
2203 struct lttng_consumer_stream
*stream
;
2205 DBG("Consumer delete flagged metadata stream");
2210 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2211 /* Validate delete flag of the stream */
2212 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2216 * Remove from pollset so the metadata thread can continue without
2217 * blocking on a deleted stream.
2219 lttng_poll_del(pollset
, stream
->wait_fd
);
2221 /* Delete it right now */
2222 consumer_del_metadata_stream(stream
, metadata_ht
);
2228 * Thread polls on metadata file descriptor and write them on disk or on the
2231 void *consumer_thread_metadata_poll(void *data
)
2233 int ret
, i
, pollfd
, err
= -1;
2234 uint32_t revents
, nb_fd
;
2235 struct lttng_consumer_stream
*stream
= NULL
;
2236 struct lttng_ht_iter iter
;
2237 struct lttng_ht_node_u64
*node
;
2238 struct lttng_poll_event events
;
2239 struct lttng_consumer_local_data
*ctx
= data
;
2242 rcu_register_thread();
2244 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2246 if (testpoint(consumerd_thread_metadata
)) {
2247 goto error_testpoint
;
2250 health_code_update();
2252 DBG("Thread metadata poll started");
2254 /* Size is set to 1 for the consumer_metadata pipe */
2255 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2257 ERR("Poll set creation failed");
2261 ret
= lttng_poll_add(&events
,
2262 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2268 DBG("Metadata main loop started");
2272 health_code_update();
2273 health_poll_entry();
2274 DBG("Metadata poll wait");
2275 ret
= lttng_poll_wait(&events
, -1);
2276 DBG("Metadata poll return from wait with %d fd(s)",
2277 LTTNG_POLL_GETNB(&events
));
2279 DBG("Metadata event caught in thread");
2281 if (errno
== EINTR
) {
2282 ERR("Poll EINTR caught");
2285 if (LTTNG_POLL_GETNB(&events
) == 0) {
2286 err
= 0; /* All is OK */
2293 /* From here, the event is a metadata wait fd */
2294 for (i
= 0; i
< nb_fd
; i
++) {
2295 health_code_update();
2297 revents
= LTTNG_POLL_GETEV(&events
, i
);
2298 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2301 /* No activity for this FD (poll implementation). */
2305 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2306 if (revents
& LPOLLIN
) {
2309 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2310 &stream
, sizeof(stream
));
2311 if (pipe_len
< sizeof(stream
)) {
2313 PERROR("read metadata stream");
2316 * Remove the pipe from the poll set and continue the loop
2317 * since their might be data to consume.
2319 lttng_poll_del(&events
,
2320 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2321 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2325 /* A NULL stream means that the state has changed. */
2326 if (stream
== NULL
) {
2327 /* Check for deleted streams. */
2328 validate_endpoint_status_metadata_stream(&events
);
2332 DBG("Adding metadata stream %d to poll set",
2335 /* Add metadata stream to the global poll events list */
2336 lttng_poll_add(&events
, stream
->wait_fd
,
2337 LPOLLIN
| LPOLLPRI
| LPOLLHUP
);
2338 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2339 DBG("Metadata thread pipe hung up");
2341 * Remove the pipe from the poll set and continue the loop
2342 * since their might be data to consume.
2344 lttng_poll_del(&events
,
2345 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2346 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2349 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2353 /* Handle other stream */
2359 uint64_t tmp_id
= (uint64_t) pollfd
;
2361 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2363 node
= lttng_ht_iter_get_node_u64(&iter
);
2366 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2369 if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2370 /* Get the data out of the metadata file descriptor */
2371 DBG("Metadata available on fd %d", pollfd
);
2372 assert(stream
->wait_fd
== pollfd
);
2375 health_code_update();
2377 len
= ctx
->on_buffer_ready(stream
, ctx
);
2379 * We don't check the return value here since if we get
2380 * a negative len, it means an error occurred thus we
2381 * simply remove it from the poll set and free the
2386 /* It's ok to have an unavailable sub-buffer */
2387 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2388 /* Clean up stream from consumer and free it. */
2389 lttng_poll_del(&events
, stream
->wait_fd
);
2390 consumer_del_metadata_stream(stream
, metadata_ht
);
2392 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2393 DBG("Metadata fd %d is hup|err.", pollfd
);
2394 if (!stream
->hangup_flush_done
2395 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2396 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2397 DBG("Attempting to flush and consume the UST buffers");
2398 lttng_ustconsumer_on_stream_hangup(stream
);
2400 /* We just flushed the stream now read it. */
2402 health_code_update();
2404 len
= ctx
->on_buffer_ready(stream
, ctx
);
2406 * We don't check the return value here since if we get
2407 * a negative len, it means an error occurred thus we
2408 * simply remove it from the poll set and free the
2414 lttng_poll_del(&events
, stream
->wait_fd
);
2416 * This call update the channel states, closes file descriptors
2417 * and securely free the stream.
2419 consumer_del_metadata_stream(stream
, metadata_ht
);
2421 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2425 /* Release RCU lock for the stream looked up */
2433 DBG("Metadata poll thread exiting");
2435 lttng_poll_clean(&events
);
2440 ERR("Health error occurred in %s", __func__
);
2442 health_unregister(health_consumerd
);
2443 rcu_unregister_thread();
2448 * This thread polls the fds in the set to consume the data and write
2449 * it to tracefile if necessary.
2451 void *consumer_thread_data_poll(void *data
)
2453 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2454 struct pollfd
*pollfd
= NULL
;
2455 /* local view of the streams */
2456 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2457 /* local view of consumer_data.fds_count */
2459 struct lttng_consumer_local_data
*ctx
= data
;
2462 rcu_register_thread();
2464 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2466 if (testpoint(consumerd_thread_data
)) {
2467 goto error_testpoint
;
2470 health_code_update();
2472 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
*));
2473 if (local_stream
== NULL
) {
2474 PERROR("local_stream malloc");
2479 health_code_update();
2485 * the fds set has been updated, we need to update our
2486 * local array as well
2488 pthread_mutex_lock(&consumer_data
.lock
);
2489 if (consumer_data
.need_update
) {
2494 local_stream
= NULL
;
2497 * Allocate for all fds +1 for the consumer_data_pipe and +1 for
2500 pollfd
= zmalloc((consumer_data
.stream_count
+ 2) * sizeof(struct pollfd
));
2501 if (pollfd
== NULL
) {
2502 PERROR("pollfd malloc");
2503 pthread_mutex_unlock(&consumer_data
.lock
);
2507 local_stream
= zmalloc((consumer_data
.stream_count
+ 2) *
2508 sizeof(struct lttng_consumer_stream
*));
2509 if (local_stream
== NULL
) {
2510 PERROR("local_stream malloc");
2511 pthread_mutex_unlock(&consumer_data
.lock
);
2514 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2517 ERR("Error in allocating pollfd or local_outfds");
2518 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2519 pthread_mutex_unlock(&consumer_data
.lock
);
2523 consumer_data
.need_update
= 0;
2525 pthread_mutex_unlock(&consumer_data
.lock
);
2527 /* No FDs and consumer_quit, consumer_cleanup the thread */
2528 if (nb_fd
== 0 && consumer_quit
== 1) {
2529 err
= 0; /* All is OK */
2532 /* poll on the array of fds */
2534 DBG("polling on %d fd", nb_fd
+ 2);
2535 if (testpoint(consumerd_thread_data_poll
)) {
2538 health_poll_entry();
2539 num_rdy
= poll(pollfd
, nb_fd
+ 2, -1);
2541 DBG("poll num_rdy : %d", num_rdy
);
2542 if (num_rdy
== -1) {
2544 * Restart interrupted system call.
2546 if (errno
== EINTR
) {
2549 PERROR("Poll error");
2550 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2552 } else if (num_rdy
== 0) {
2553 DBG("Polling thread timed out");
2558 * If the consumer_data_pipe triggered poll go directly to the
2559 * beginning of the loop to update the array. We want to prioritize
2560 * array update over low-priority reads.
2562 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2563 ssize_t pipe_readlen
;
2565 DBG("consumer_data_pipe wake up");
2566 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2567 &new_stream
, sizeof(new_stream
));
2568 if (pipe_readlen
< sizeof(new_stream
)) {
2569 PERROR("Consumer data pipe");
2570 /* Continue so we can at least handle the current stream(s). */
2575 * If the stream is NULL, just ignore it. It's also possible that
2576 * the sessiond poll thread changed the consumer_quit state and is
2577 * waking us up to test it.
2579 if (new_stream
== NULL
) {
2580 validate_endpoint_status_data_stream();
2584 /* Continue to update the local streams and handle prio ones */
2588 /* Handle wakeup pipe. */
2589 if (pollfd
[nb_fd
+ 1].revents
& (POLLIN
| POLLPRI
)) {
2591 ssize_t pipe_readlen
;
2593 pipe_readlen
= lttng_pipe_read(ctx
->consumer_wakeup_pipe
, &dummy
,
2595 if (pipe_readlen
< 0) {
2596 PERROR("Consumer data wakeup pipe");
2598 /* We've been awakened to handle stream(s). */
2599 ctx
->has_wakeup
= 0;
2602 /* Take care of high priority channels first. */
2603 for (i
= 0; i
< nb_fd
; i
++) {
2604 health_code_update();
2606 if (local_stream
[i
] == NULL
) {
2609 if (pollfd
[i
].revents
& POLLPRI
) {
2610 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2612 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2613 /* it's ok to have an unavailable sub-buffer */
2614 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2615 /* Clean the stream and free it. */
2616 consumer_del_stream(local_stream
[i
], data_ht
);
2617 local_stream
[i
] = NULL
;
2618 } else if (len
> 0) {
2619 local_stream
[i
]->data_read
= 1;
2625 * If we read high prio channel in this loop, try again
2626 * for more high prio data.
2632 /* Take care of low priority channels. */
2633 for (i
= 0; i
< nb_fd
; i
++) {
2634 health_code_update();
2636 if (local_stream
[i
] == NULL
) {
2639 if ((pollfd
[i
].revents
& POLLIN
) ||
2640 local_stream
[i
]->hangup_flush_done
||
2641 local_stream
[i
]->has_data
) {
2642 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2643 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2644 /* it's ok to have an unavailable sub-buffer */
2645 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2646 /* Clean the stream and free it. */
2647 consumer_del_stream(local_stream
[i
], data_ht
);
2648 local_stream
[i
] = NULL
;
2649 } else if (len
> 0) {
2650 local_stream
[i
]->data_read
= 1;
2655 /* Handle hangup and errors */
2656 for (i
= 0; i
< nb_fd
; i
++) {
2657 health_code_update();
2659 if (local_stream
[i
] == NULL
) {
2662 if (!local_stream
[i
]->hangup_flush_done
2663 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2664 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2665 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2666 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2668 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2669 /* Attempt read again, for the data we just flushed. */
2670 local_stream
[i
]->data_read
= 1;
2673 * If the poll flag is HUP/ERR/NVAL and we have
2674 * read no data in this pass, we can remove the
2675 * stream from its hash table.
2677 if ((pollfd
[i
].revents
& POLLHUP
)) {
2678 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2679 if (!local_stream
[i
]->data_read
) {
2680 consumer_del_stream(local_stream
[i
], data_ht
);
2681 local_stream
[i
] = NULL
;
2684 } else if (pollfd
[i
].revents
& POLLERR
) {
2685 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2686 if (!local_stream
[i
]->data_read
) {
2687 consumer_del_stream(local_stream
[i
], data_ht
);
2688 local_stream
[i
] = NULL
;
2691 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2692 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2693 if (!local_stream
[i
]->data_read
) {
2694 consumer_del_stream(local_stream
[i
], data_ht
);
2695 local_stream
[i
] = NULL
;
2699 if (local_stream
[i
] != NULL
) {
2700 local_stream
[i
]->data_read
= 0;
2707 DBG("polling thread exiting");
2712 * Close the write side of the pipe so epoll_wait() in
2713 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2714 * read side of the pipe. If we close them both, epoll_wait strangely does
2715 * not return and could create a endless wait period if the pipe is the
2716 * only tracked fd in the poll set. The thread will take care of closing
2719 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2724 ERR("Health error occurred in %s", __func__
);
2726 health_unregister(health_consumerd
);
2728 rcu_unregister_thread();
2733 * Close wake-up end of each stream belonging to the channel. This will
2734 * allow the poll() on the stream read-side to detect when the
2735 * write-side (application) finally closes them.
2738 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2740 struct lttng_ht
*ht
;
2741 struct lttng_consumer_stream
*stream
;
2742 struct lttng_ht_iter iter
;
2744 ht
= consumer_data
.stream_per_chan_id_ht
;
2747 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2748 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2749 ht
->match_fct
, &channel
->key
,
2750 &iter
.iter
, stream
, node_channel_id
.node
) {
2752 * Protect against teardown with mutex.
2754 pthread_mutex_lock(&stream
->lock
);
2755 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2758 switch (consumer_data
.type
) {
2759 case LTTNG_CONSUMER_KERNEL
:
2761 case LTTNG_CONSUMER32_UST
:
2762 case LTTNG_CONSUMER64_UST
:
2763 if (stream
->metadata_flag
) {
2764 /* Safe and protected by the stream lock. */
2765 lttng_ustconsumer_close_metadata(stream
->chan
);
2768 * Note: a mutex is taken internally within
2769 * liblttng-ust-ctl to protect timer wakeup_fd
2770 * use from concurrent close.
2772 lttng_ustconsumer_close_stream_wakeup(stream
);
2776 ERR("Unknown consumer_data type");
2780 pthread_mutex_unlock(&stream
->lock
);
2785 static void destroy_channel_ht(struct lttng_ht
*ht
)
2787 struct lttng_ht_iter iter
;
2788 struct lttng_consumer_channel
*channel
;
2796 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2797 ret
= lttng_ht_del(ht
, &iter
);
2802 lttng_ht_destroy(ht
);
2806 * This thread polls the channel fds to detect when they are being
2807 * closed. It closes all related streams if the channel is detected as
2808 * closed. It is currently only used as a shim layer for UST because the
2809 * consumerd needs to keep the per-stream wakeup end of pipes open for
2812 void *consumer_thread_channel_poll(void *data
)
2814 int ret
, i
, pollfd
, err
= -1;
2815 uint32_t revents
, nb_fd
;
2816 struct lttng_consumer_channel
*chan
= NULL
;
2817 struct lttng_ht_iter iter
;
2818 struct lttng_ht_node_u64
*node
;
2819 struct lttng_poll_event events
;
2820 struct lttng_consumer_local_data
*ctx
= data
;
2821 struct lttng_ht
*channel_ht
;
2823 rcu_register_thread();
2825 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
2827 if (testpoint(consumerd_thread_channel
)) {
2828 goto error_testpoint
;
2831 health_code_update();
2833 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2835 /* ENOMEM at this point. Better to bail out. */
2839 DBG("Thread channel poll started");
2841 /* Size is set to 1 for the consumer_channel pipe */
2842 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2844 ERR("Poll set creation failed");
2848 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2854 DBG("Channel main loop started");
2858 health_code_update();
2859 DBG("Channel poll wait");
2860 health_poll_entry();
2861 ret
= lttng_poll_wait(&events
, -1);
2862 DBG("Channel poll return from wait with %d fd(s)",
2863 LTTNG_POLL_GETNB(&events
));
2865 DBG("Channel event caught in thread");
2867 if (errno
== EINTR
) {
2868 ERR("Poll EINTR caught");
2871 if (LTTNG_POLL_GETNB(&events
) == 0) {
2872 err
= 0; /* All is OK */
2879 /* From here, the event is a channel wait fd */
2880 for (i
= 0; i
< nb_fd
; i
++) {
2881 health_code_update();
2883 revents
= LTTNG_POLL_GETEV(&events
, i
);
2884 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2887 /* No activity for this FD (poll implementation). */
2891 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2892 if (revents
& LPOLLIN
) {
2893 enum consumer_channel_action action
;
2896 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
2899 ERR("Error reading channel pipe");
2901 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2906 case CONSUMER_CHANNEL_ADD
:
2907 DBG("Adding channel %d to poll set",
2910 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
2913 lttng_ht_add_unique_u64(channel_ht
,
2914 &chan
->wait_fd_node
);
2916 /* Add channel to the global poll events list */
2917 lttng_poll_add(&events
, chan
->wait_fd
,
2918 LPOLLERR
| LPOLLHUP
);
2920 case CONSUMER_CHANNEL_DEL
:
2923 * This command should never be called if the channel
2924 * has streams monitored by either the data or metadata
2925 * thread. The consumer only notify this thread with a
2926 * channel del. command if it receives a destroy
2927 * channel command from the session daemon that send it
2928 * if a command prior to the GET_CHANNEL failed.
2932 chan
= consumer_find_channel(key
);
2935 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
2938 lttng_poll_del(&events
, chan
->wait_fd
);
2939 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
2940 ret
= lttng_ht_del(channel_ht
, &iter
);
2943 switch (consumer_data
.type
) {
2944 case LTTNG_CONSUMER_KERNEL
:
2946 case LTTNG_CONSUMER32_UST
:
2947 case LTTNG_CONSUMER64_UST
:
2948 health_code_update();
2949 /* Destroy streams that might have been left in the stream list. */
2950 clean_channel_stream_list(chan
);
2953 ERR("Unknown consumer_data type");
2958 * Release our own refcount. Force channel deletion even if
2959 * streams were not initialized.
2961 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
2962 consumer_del_channel(chan
);
2967 case CONSUMER_CHANNEL_QUIT
:
2969 * Remove the pipe from the poll set and continue the loop
2970 * since their might be data to consume.
2972 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2975 ERR("Unknown action");
2978 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2979 DBG("Channel thread pipe hung up");
2981 * Remove the pipe from the poll set and continue the loop
2982 * since their might be data to consume.
2984 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2987 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2991 /* Handle other stream */
2997 uint64_t tmp_id
= (uint64_t) pollfd
;
2999 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
3001 node
= lttng_ht_iter_get_node_u64(&iter
);
3004 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
3007 /* Check for error event */
3008 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
3009 DBG("Channel fd %d is hup|err.", pollfd
);
3011 lttng_poll_del(&events
, chan
->wait_fd
);
3012 ret
= lttng_ht_del(channel_ht
, &iter
);
3016 * This will close the wait fd for each stream associated to
3017 * this channel AND monitored by the data/metadata thread thus
3018 * will be clean by the right thread.
3020 consumer_close_channel_streams(chan
);
3022 /* Release our own refcount */
3023 if (!uatomic_sub_return(&chan
->refcount
, 1)
3024 && !uatomic_read(&chan
->nb_init_stream_left
)) {
3025 consumer_del_channel(chan
);
3028 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3033 /* Release RCU lock for the channel looked up */
3041 lttng_poll_clean(&events
);
3043 destroy_channel_ht(channel_ht
);
3046 DBG("Channel poll thread exiting");
3049 ERR("Health error occurred in %s", __func__
);
3051 health_unregister(health_consumerd
);
3052 rcu_unregister_thread();
3056 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
3057 struct pollfd
*sockpoll
, int client_socket
)
3064 ret
= lttng_consumer_poll_socket(sockpoll
);
3068 DBG("Metadata connection on client_socket");
3070 /* Blocking call, waiting for transmission */
3071 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
3072 if (ctx
->consumer_metadata_socket
< 0) {
3073 WARN("On accept metadata");
3084 * This thread listens on the consumerd socket and receives the file
3085 * descriptors from the session daemon.
3087 void *consumer_thread_sessiond_poll(void *data
)
3089 int sock
= -1, client_socket
, ret
, err
= -1;
3091 * structure to poll for incoming data on communication socket avoids
3092 * making blocking sockets.
3094 struct pollfd consumer_sockpoll
[2];
3095 struct lttng_consumer_local_data
*ctx
= data
;
3097 rcu_register_thread();
3099 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
3101 if (testpoint(consumerd_thread_sessiond
)) {
3102 goto error_testpoint
;
3105 health_code_update();
3107 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
3108 unlink(ctx
->consumer_command_sock_path
);
3109 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
3110 if (client_socket
< 0) {
3111 ERR("Cannot create command socket");
3115 ret
= lttcomm_listen_unix_sock(client_socket
);
3120 DBG("Sending ready command to lttng-sessiond");
3121 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3122 /* return < 0 on error, but == 0 is not fatal */
3124 ERR("Error sending ready command to lttng-sessiond");
3128 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3129 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3130 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3131 consumer_sockpoll
[1].fd
= client_socket
;
3132 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3134 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3142 DBG("Connection on client_socket");
3144 /* Blocking call, waiting for transmission */
3145 sock
= lttcomm_accept_unix_sock(client_socket
);
3152 * Setup metadata socket which is the second socket connection on the
3153 * command unix socket.
3155 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3164 /* This socket is not useful anymore. */
3165 ret
= close(client_socket
);
3167 PERROR("close client_socket");
3171 /* update the polling structure to poll on the established socket */
3172 consumer_sockpoll
[1].fd
= sock
;
3173 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3176 health_code_update();
3178 health_poll_entry();
3179 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3188 DBG("Incoming command on sock");
3189 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3192 * This could simply be a session daemon quitting. Don't output
3195 DBG("Communication interrupted on command socket");
3199 if (consumer_quit
) {
3200 DBG("consumer_thread_receive_fds received quit from signal");
3201 err
= 0; /* All is OK */
3204 DBG("received command on sock");
3210 DBG("Consumer thread sessiond poll exiting");
3213 * Close metadata streams since the producer is the session daemon which
3216 * NOTE: for now, this only applies to the UST tracer.
3218 lttng_consumer_close_all_metadata();
3221 * when all fds have hung up, the polling thread
3227 * Notify the data poll thread to poll back again and test the
3228 * consumer_quit state that we just set so to quit gracefully.
3230 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3232 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3234 notify_health_quit_pipe(health_quit_pipe
);
3236 /* Cleaning up possibly open sockets. */
3240 PERROR("close sock sessiond poll");
3243 if (client_socket
>= 0) {
3244 ret
= close(client_socket
);
3246 PERROR("close client_socket sessiond poll");
3253 ERR("Health error occurred in %s", __func__
);
3255 health_unregister(health_consumerd
);
3257 rcu_unregister_thread();
3261 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3262 struct lttng_consumer_local_data
*ctx
)
3266 pthread_mutex_lock(&stream
->lock
);
3267 if (stream
->metadata_flag
) {
3268 pthread_mutex_lock(&stream
->metadata_rdv_lock
);
3271 switch (consumer_data
.type
) {
3272 case LTTNG_CONSUMER_KERNEL
:
3273 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
3275 case LTTNG_CONSUMER32_UST
:
3276 case LTTNG_CONSUMER64_UST
:
3277 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
3280 ERR("Unknown consumer_data type");
3286 if (stream
->metadata_flag
) {
3287 pthread_cond_broadcast(&stream
->metadata_rdv
);
3288 pthread_mutex_unlock(&stream
->metadata_rdv_lock
);
3290 pthread_mutex_unlock(&stream
->lock
);
3294 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3296 switch (consumer_data
.type
) {
3297 case LTTNG_CONSUMER_KERNEL
:
3298 return lttng_kconsumer_on_recv_stream(stream
);
3299 case LTTNG_CONSUMER32_UST
:
3300 case LTTNG_CONSUMER64_UST
:
3301 return lttng_ustconsumer_on_recv_stream(stream
);
3303 ERR("Unknown consumer_data type");
3310 * Allocate and set consumer data hash tables.
3312 int lttng_consumer_init(void)
3314 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3315 if (!consumer_data
.channel_ht
) {
3319 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3320 if (!consumer_data
.relayd_ht
) {
3324 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3325 if (!consumer_data
.stream_list_ht
) {
3329 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3330 if (!consumer_data
.stream_per_chan_id_ht
) {
3334 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3339 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3351 * Process the ADD_RELAYD command receive by a consumer.
3353 * This will create a relayd socket pair and add it to the relayd hash table.
3354 * The caller MUST acquire a RCU read side lock before calling it.
3356 int consumer_add_relayd_socket(uint64_t net_seq_idx
, int sock_type
,
3357 struct lttng_consumer_local_data
*ctx
, int sock
,
3358 struct pollfd
*consumer_sockpoll
,
3359 struct lttcomm_relayd_sock
*relayd_sock
, uint64_t sessiond_id
,
3360 uint64_t relayd_session_id
)
3362 int fd
= -1, ret
= -1, relayd_created
= 0;
3363 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3364 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3367 assert(relayd_sock
);
3369 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", net_seq_idx
);
3371 /* Get relayd reference if exists. */
3372 relayd
= consumer_find_relayd(net_seq_idx
);
3373 if (relayd
== NULL
) {
3374 assert(sock_type
== LTTNG_STREAM_CONTROL
);
3375 /* Not found. Allocate one. */
3376 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3377 if (relayd
== NULL
) {
3379 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3382 relayd
->sessiond_session_id
= sessiond_id
;
3387 * This code path MUST continue to the consumer send status message to
3388 * we can notify the session daemon and continue our work without
3389 * killing everything.
3393 * relayd key should never be found for control socket.
3395 assert(sock_type
!= LTTNG_STREAM_CONTROL
);
3398 /* First send a status message before receiving the fds. */
3399 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3401 /* Somehow, the session daemon is not responding anymore. */
3402 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3403 goto error_nosignal
;
3406 /* Poll on consumer socket. */
3407 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3409 /* Needing to exit in the middle of a command: error. */
3410 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3412 goto error_nosignal
;
3415 /* Get relayd socket from session daemon */
3416 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3417 if (ret
!= sizeof(fd
)) {
3419 fd
= -1; /* Just in case it gets set with an invalid value. */
3422 * Failing to receive FDs might indicate a major problem such as
3423 * reaching a fd limit during the receive where the kernel returns a
3424 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3425 * don't take any chances and stop everything.
3427 * XXX: Feature request #558 will fix that and avoid this possible
3428 * issue when reaching the fd limit.
3430 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3431 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3435 /* Copy socket information and received FD */
3436 switch (sock_type
) {
3437 case LTTNG_STREAM_CONTROL
:
3438 /* Copy received lttcomm socket */
3439 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3440 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3441 /* Handle create_sock error. */
3443 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3447 * Close the socket created internally by
3448 * lttcomm_create_sock, so we can replace it by the one
3449 * received from sessiond.
3451 if (close(relayd
->control_sock
.sock
.fd
)) {
3455 /* Assign new file descriptor */
3456 relayd
->control_sock
.sock
.fd
= fd
;
3457 fd
= -1; /* For error path */
3458 /* Assign version values. */
3459 relayd
->control_sock
.major
= relayd_sock
->major
;
3460 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3462 relayd
->relayd_session_id
= relayd_session_id
;
3465 case LTTNG_STREAM_DATA
:
3466 /* Copy received lttcomm socket */
3467 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3468 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3469 /* Handle create_sock error. */
3471 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3475 * Close the socket created internally by
3476 * lttcomm_create_sock, so we can replace it by the one
3477 * received from sessiond.
3479 if (close(relayd
->data_sock
.sock
.fd
)) {
3483 /* Assign new file descriptor */
3484 relayd
->data_sock
.sock
.fd
= fd
;
3485 fd
= -1; /* for eventual error paths */
3486 /* Assign version values. */
3487 relayd
->data_sock
.major
= relayd_sock
->major
;
3488 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3491 ERR("Unknown relayd socket type (%d)", sock_type
);
3493 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3497 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3498 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3499 relayd
->net_seq_idx
, fd
);
3501 /* We successfully added the socket. Send status back. */
3502 ret
= consumer_send_status_msg(sock
, ret_code
);
3504 /* Somehow, the session daemon is not responding anymore. */
3505 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3506 goto error_nosignal
;
3510 * Add relayd socket pair to consumer data hashtable. If object already
3511 * exists or on error, the function gracefully returns.
3519 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3520 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3524 /* Close received socket if valid. */
3527 PERROR("close received socket");
3531 if (relayd_created
) {
3539 * Try to lock the stream mutex.
3541 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
3543 static int stream_try_lock(struct lttng_consumer_stream
*stream
)
3550 * Try to lock the stream mutex. On failure, we know that the stream is
3551 * being used else where hence there is data still being extracted.
3553 ret
= pthread_mutex_trylock(&stream
->lock
);
3555 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
3567 * Search for a relayd associated to the session id and return the reference.
3569 * A rcu read side lock MUST be acquire before calling this function and locked
3570 * until the relayd object is no longer necessary.
3572 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3574 struct lttng_ht_iter iter
;
3575 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3577 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3578 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3581 * Check by sessiond id which is unique here where the relayd session
3582 * id might not be when having multiple relayd.
3584 if (relayd
->sessiond_session_id
== id
) {
3585 /* Found the relayd. There can be only one per id. */
3597 * Check if for a given session id there is still data needed to be extract
3600 * Return 1 if data is pending or else 0 meaning ready to be read.
3602 int consumer_data_pending(uint64_t id
)
3605 struct lttng_ht_iter iter
;
3606 struct lttng_ht
*ht
;
3607 struct lttng_consumer_stream
*stream
;
3608 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3609 int (*data_pending
)(struct lttng_consumer_stream
*);
3611 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3614 pthread_mutex_lock(&consumer_data
.lock
);
3616 switch (consumer_data
.type
) {
3617 case LTTNG_CONSUMER_KERNEL
:
3618 data_pending
= lttng_kconsumer_data_pending
;
3620 case LTTNG_CONSUMER32_UST
:
3621 case LTTNG_CONSUMER64_UST
:
3622 data_pending
= lttng_ustconsumer_data_pending
;
3625 ERR("Unknown consumer data type");
3629 /* Ease our life a bit */
3630 ht
= consumer_data
.stream_list_ht
;
3632 relayd
= find_relayd_by_session_id(id
);
3634 /* Send init command for data pending. */
3635 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3636 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3637 relayd
->relayd_session_id
);
3638 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3640 /* Communication error thus the relayd so no data pending. */
3641 goto data_not_pending
;
3645 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3646 ht
->hash_fct(&id
, lttng_ht_seed
),
3648 &iter
.iter
, stream
, node_session_id
.node
) {
3649 /* If this call fails, the stream is being used hence data pending. */
3650 ret
= stream_try_lock(stream
);
3656 * A removed node from the hash table indicates that the stream has
3657 * been deleted thus having a guarantee that the buffers are closed
3658 * on the consumer side. However, data can still be transmitted
3659 * over the network so don't skip the relayd check.
3661 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3663 /* Check the stream if there is data in the buffers. */
3664 ret
= data_pending(stream
);
3666 pthread_mutex_unlock(&stream
->lock
);
3673 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3674 if (stream
->metadata_flag
) {
3675 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3676 stream
->relayd_stream_id
);
3678 ret
= relayd_data_pending(&relayd
->control_sock
,
3679 stream
->relayd_stream_id
,
3680 stream
->next_net_seq_num
- 1);
3682 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3684 pthread_mutex_unlock(&stream
->lock
);
3688 pthread_mutex_unlock(&stream
->lock
);
3692 unsigned int is_data_inflight
= 0;
3694 /* Send init command for data pending. */
3695 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3696 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3697 relayd
->relayd_session_id
, &is_data_inflight
);
3698 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3700 goto data_not_pending
;
3702 if (is_data_inflight
) {
3708 * Finding _no_ node in the hash table and no inflight data means that the
3709 * stream(s) have been removed thus data is guaranteed to be available for
3710 * analysis from the trace files.
3714 /* Data is available to be read by a viewer. */
3715 pthread_mutex_unlock(&consumer_data
.lock
);
3720 /* Data is still being extracted from buffers. */
3721 pthread_mutex_unlock(&consumer_data
.lock
);
3727 * Send a ret code status message to the sessiond daemon.
3729 * Return the sendmsg() return value.
3731 int consumer_send_status_msg(int sock
, int ret_code
)
3733 struct lttcomm_consumer_status_msg msg
;
3735 memset(&msg
, 0, sizeof(msg
));
3736 msg
.ret_code
= ret_code
;
3738 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3742 * Send a channel status message to the sessiond daemon.
3744 * Return the sendmsg() return value.
3746 int consumer_send_status_channel(int sock
,
3747 struct lttng_consumer_channel
*channel
)
3749 struct lttcomm_consumer_status_channel msg
;
3753 memset(&msg
, 0, sizeof(msg
));
3755 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3757 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3758 msg
.key
= channel
->key
;
3759 msg
.stream_count
= channel
->streams
.count
;
3762 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3765 unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos
,
3766 unsigned long produced_pos
, uint64_t nb_packets_per_stream
,
3767 uint64_t max_sb_size
)
3769 unsigned long start_pos
;
3771 if (!nb_packets_per_stream
) {
3772 return consumed_pos
; /* Grab everything */
3774 start_pos
= produced_pos
- offset_align_floor(produced_pos
, max_sb_size
);
3775 start_pos
-= max_sb_size
* nb_packets_per_stream
;
3776 if ((long) (start_pos
- consumed_pos
) < 0) {
3777 return consumed_pos
; /* Grab everything */