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-timer.h>
48 #include "consumer-stream.h"
49 #include "consumer-testpoint.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 */
71 * Flag to inform the polling thread to quit when all fd hung up. Updated by
72 * the consumer_thread_receive_fds when it notices that all fds has hung up.
73 * Also updated by the signal handler (consumer_should_exit()). Read by the
76 volatile int consumer_quit
;
79 * Global hash table containing respectively metadata and data streams. The
80 * stream element in this ht should only be updated by the metadata poll thread
81 * for the metadata and the data poll thread for the data.
83 static struct lttng_ht
*metadata_ht
;
84 static struct lttng_ht
*data_ht
;
87 * Notify a thread lttng pipe to poll back again. This usually means that some
88 * global state has changed so we just send back the thread in a poll wait
91 static void notify_thread_lttng_pipe(struct lttng_pipe
*pipe
)
93 struct lttng_consumer_stream
*null_stream
= NULL
;
97 (void) lttng_pipe_write(pipe
, &null_stream
, sizeof(null_stream
));
100 static void notify_health_quit_pipe(int *pipe
)
104 ret
= lttng_write(pipe
[1], "4", 1);
106 PERROR("write consumer health quit");
110 static void notify_channel_pipe(struct lttng_consumer_local_data
*ctx
,
111 struct lttng_consumer_channel
*chan
,
113 enum consumer_channel_action action
)
115 struct consumer_channel_msg msg
;
118 memset(&msg
, 0, sizeof(msg
));
123 ret
= lttng_write(ctx
->consumer_channel_pipe
[1], &msg
, sizeof(msg
));
124 if (ret
< sizeof(msg
)) {
125 PERROR("notify_channel_pipe write error");
129 void notify_thread_del_channel(struct lttng_consumer_local_data
*ctx
,
132 notify_channel_pipe(ctx
, NULL
, key
, CONSUMER_CHANNEL_DEL
);
135 static int read_channel_pipe(struct lttng_consumer_local_data
*ctx
,
136 struct lttng_consumer_channel
**chan
,
138 enum consumer_channel_action
*action
)
140 struct consumer_channel_msg msg
;
143 ret
= lttng_read(ctx
->consumer_channel_pipe
[0], &msg
, sizeof(msg
));
144 if (ret
< sizeof(msg
)) {
148 *action
= msg
.action
;
156 * Cleanup the stream list of a channel. Those streams are not yet globally
159 static void clean_channel_stream_list(struct lttng_consumer_channel
*channel
)
161 struct lttng_consumer_stream
*stream
, *stmp
;
165 /* Delete streams that might have been left in the stream list. */
166 cds_list_for_each_entry_safe(stream
, stmp
, &channel
->streams
.head
,
168 cds_list_del(&stream
->send_node
);
170 * Once a stream is added to this list, the buffers were created so we
171 * have a guarantee that this call will succeed. Setting the monitor
172 * mode to 0 so we don't lock nor try to delete the stream from the
176 consumer_stream_destroy(stream
, NULL
);
181 * Find a stream. The consumer_data.lock must be locked during this
184 static struct lttng_consumer_stream
*find_stream(uint64_t key
,
187 struct lttng_ht_iter iter
;
188 struct lttng_ht_node_u64
*node
;
189 struct lttng_consumer_stream
*stream
= NULL
;
193 /* -1ULL keys are lookup failures */
194 if (key
== (uint64_t) -1ULL) {
200 lttng_ht_lookup(ht
, &key
, &iter
);
201 node
= lttng_ht_iter_get_node_u64(&iter
);
203 stream
= caa_container_of(node
, struct lttng_consumer_stream
, node
);
211 static void steal_stream_key(uint64_t key
, struct lttng_ht
*ht
)
213 struct lttng_consumer_stream
*stream
;
216 stream
= find_stream(key
, ht
);
218 stream
->key
= (uint64_t) -1ULL;
220 * We don't want the lookup to match, but we still need
221 * to iterate on this stream when iterating over the hash table. Just
222 * change the node key.
224 stream
->node
.key
= (uint64_t) -1ULL;
230 * Return a channel object for the given key.
232 * RCU read side lock MUST be acquired before calling this function and
233 * protects the channel ptr.
235 struct lttng_consumer_channel
*consumer_find_channel(uint64_t key
)
237 struct lttng_ht_iter iter
;
238 struct lttng_ht_node_u64
*node
;
239 struct lttng_consumer_channel
*channel
= NULL
;
241 /* -1ULL keys are lookup failures */
242 if (key
== (uint64_t) -1ULL) {
246 lttng_ht_lookup(consumer_data
.channel_ht
, &key
, &iter
);
247 node
= lttng_ht_iter_get_node_u64(&iter
);
249 channel
= caa_container_of(node
, struct lttng_consumer_channel
, node
);
256 * There is a possibility that the consumer does not have enough time between
257 * the close of the channel on the session daemon and the cleanup in here thus
258 * once we have a channel add with an existing key, we know for sure that this
259 * channel will eventually get cleaned up by all streams being closed.
261 * This function just nullifies the already existing channel key.
263 static void steal_channel_key(uint64_t key
)
265 struct lttng_consumer_channel
*channel
;
268 channel
= consumer_find_channel(key
);
270 channel
->key
= (uint64_t) -1ULL;
272 * We don't want the lookup to match, but we still need to iterate on
273 * this channel when iterating over the hash table. Just change the
276 channel
->node
.key
= (uint64_t) -1ULL;
281 static void free_channel_rcu(struct rcu_head
*head
)
283 struct lttng_ht_node_u64
*node
=
284 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
285 struct lttng_consumer_channel
*channel
=
286 caa_container_of(node
, struct lttng_consumer_channel
, node
);
292 * RCU protected relayd socket pair free.
294 static void free_relayd_rcu(struct rcu_head
*head
)
296 struct lttng_ht_node_u64
*node
=
297 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
298 struct consumer_relayd_sock_pair
*relayd
=
299 caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
302 * Close all sockets. This is done in the call RCU since we don't want the
303 * socket fds to be reassigned thus potentially creating bad state of the
306 * We do not have to lock the control socket mutex here since at this stage
307 * there is no one referencing to this relayd object.
309 (void) relayd_close(&relayd
->control_sock
);
310 (void) relayd_close(&relayd
->data_sock
);
316 * Destroy and free relayd socket pair object.
318 void consumer_destroy_relayd(struct consumer_relayd_sock_pair
*relayd
)
321 struct lttng_ht_iter iter
;
323 if (relayd
== NULL
) {
327 DBG("Consumer destroy and close relayd socket pair");
329 iter
.iter
.node
= &relayd
->node
.node
;
330 ret
= lttng_ht_del(consumer_data
.relayd_ht
, &iter
);
332 /* We assume the relayd is being or is destroyed */
336 /* RCU free() call */
337 call_rcu(&relayd
->node
.head
, free_relayd_rcu
);
341 * Remove a channel from the global list protected by a mutex. This function is
342 * also responsible for freeing its data structures.
344 void consumer_del_channel(struct lttng_consumer_channel
*channel
)
347 struct lttng_ht_iter iter
;
349 DBG("Consumer delete channel key %" PRIu64
, channel
->key
);
351 pthread_mutex_lock(&consumer_data
.lock
);
352 pthread_mutex_lock(&channel
->lock
);
354 /* Destroy streams that might have been left in the stream list. */
355 clean_channel_stream_list(channel
);
357 if (channel
->live_timer_enabled
== 1) {
358 consumer_timer_live_stop(channel
);
361 switch (consumer_data
.type
) {
362 case LTTNG_CONSUMER_KERNEL
:
364 case LTTNG_CONSUMER32_UST
:
365 case LTTNG_CONSUMER64_UST
:
366 lttng_ustconsumer_del_channel(channel
);
369 ERR("Unknown consumer_data type");
375 iter
.iter
.node
= &channel
->node
.node
;
376 ret
= lttng_ht_del(consumer_data
.channel_ht
, &iter
);
380 call_rcu(&channel
->node
.head
, free_channel_rcu
);
382 pthread_mutex_unlock(&channel
->lock
);
383 pthread_mutex_unlock(&consumer_data
.lock
);
387 * Iterate over the relayd hash table and destroy each element. Finally,
388 * destroy the whole hash table.
390 static void cleanup_relayd_ht(void)
392 struct lttng_ht_iter iter
;
393 struct consumer_relayd_sock_pair
*relayd
;
397 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
399 consumer_destroy_relayd(relayd
);
404 lttng_ht_destroy(consumer_data
.relayd_ht
);
408 * Update the end point status of all streams having the given network sequence
409 * index (relayd index).
411 * It's atomically set without having the stream mutex locked which is fine
412 * because we handle the write/read race with a pipe wakeup for each thread.
414 static void update_endpoint_status_by_netidx(uint64_t net_seq_idx
,
415 enum consumer_endpoint_status status
)
417 struct lttng_ht_iter iter
;
418 struct lttng_consumer_stream
*stream
;
420 DBG("Consumer set delete flag on stream by idx %" PRIu64
, net_seq_idx
);
424 /* Let's begin with metadata */
425 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
426 if (stream
->net_seq_idx
== net_seq_idx
) {
427 uatomic_set(&stream
->endpoint_status
, status
);
428 DBG("Delete flag set to metadata stream %d", stream
->wait_fd
);
432 /* Follow up by the data streams */
433 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
434 if (stream
->net_seq_idx
== net_seq_idx
) {
435 uatomic_set(&stream
->endpoint_status
, status
);
436 DBG("Delete flag set to data stream %d", stream
->wait_fd
);
443 * Cleanup a relayd object by flagging every associated streams for deletion,
444 * destroying the object meaning removing it from the relayd hash table,
445 * closing the sockets and freeing the memory in a RCU call.
447 * If a local data context is available, notify the threads that the streams'
448 * state have changed.
450 static void cleanup_relayd(struct consumer_relayd_sock_pair
*relayd
,
451 struct lttng_consumer_local_data
*ctx
)
457 DBG("Cleaning up relayd sockets");
459 /* Save the net sequence index before destroying the object */
460 netidx
= relayd
->net_seq_idx
;
463 * Delete the relayd from the relayd hash table, close the sockets and free
464 * the object in a RCU call.
466 consumer_destroy_relayd(relayd
);
468 /* Set inactive endpoint to all streams */
469 update_endpoint_status_by_netidx(netidx
, CONSUMER_ENDPOINT_INACTIVE
);
472 * With a local data context, notify the threads that the streams' state
473 * have changed. The write() action on the pipe acts as an "implicit"
474 * memory barrier ordering the updates of the end point status from the
475 * read of this status which happens AFTER receiving this notify.
478 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
479 notify_thread_lttng_pipe(ctx
->consumer_metadata_pipe
);
484 * Flag a relayd socket pair for destruction. Destroy it if the refcount
487 * RCU read side lock MUST be aquired before calling this function.
489 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair
*relayd
)
493 /* Set destroy flag for this object */
494 uatomic_set(&relayd
->destroy_flag
, 1);
496 /* Destroy the relayd if refcount is 0 */
497 if (uatomic_read(&relayd
->refcount
) == 0) {
498 consumer_destroy_relayd(relayd
);
503 * Completly destroy stream from every visiable data structure and the given
506 * One this call returns, the stream object is not longer usable nor visible.
508 void consumer_del_stream(struct lttng_consumer_stream
*stream
,
511 consumer_stream_destroy(stream
, ht
);
515 * XXX naming of del vs destroy is all mixed up.
517 void consumer_del_stream_for_data(struct lttng_consumer_stream
*stream
)
519 consumer_stream_destroy(stream
, data_ht
);
522 void consumer_del_stream_for_metadata(struct lttng_consumer_stream
*stream
)
524 consumer_stream_destroy(stream
, metadata_ht
);
527 struct lttng_consumer_stream
*consumer_allocate_stream(uint64_t channel_key
,
529 enum lttng_consumer_stream_state state
,
530 const char *channel_name
,
537 enum consumer_channel_type type
,
538 unsigned int monitor
)
541 struct lttng_consumer_stream
*stream
;
543 stream
= zmalloc(sizeof(*stream
));
544 if (stream
== NULL
) {
545 PERROR("malloc struct lttng_consumer_stream");
552 stream
->key
= stream_key
;
554 stream
->out_fd_offset
= 0;
555 stream
->output_written
= 0;
556 stream
->state
= state
;
559 stream
->net_seq_idx
= relayd_id
;
560 stream
->session_id
= session_id
;
561 stream
->monitor
= monitor
;
562 stream
->endpoint_status
= CONSUMER_ENDPOINT_ACTIVE
;
563 stream
->index_fd
= -1;
564 pthread_mutex_init(&stream
->lock
, NULL
);
565 pthread_mutex_init(&stream
->metadata_timer_lock
, NULL
);
567 /* If channel is the metadata, flag this stream as metadata. */
568 if (type
== CONSUMER_CHANNEL_TYPE_METADATA
) {
569 stream
->metadata_flag
= 1;
570 /* Metadata is flat out. */
571 strncpy(stream
->name
, DEFAULT_METADATA_NAME
, sizeof(stream
->name
));
572 /* Live rendez-vous point. */
573 pthread_cond_init(&stream
->metadata_rdv
, NULL
);
574 pthread_mutex_init(&stream
->metadata_rdv_lock
, NULL
);
576 /* Format stream name to <channel_name>_<cpu_number> */
577 ret
= snprintf(stream
->name
, sizeof(stream
->name
), "%s_%d",
580 PERROR("snprintf stream name");
585 /* Key is always the wait_fd for streams. */
586 lttng_ht_node_init_u64(&stream
->node
, stream
->key
);
588 /* Init node per channel id key */
589 lttng_ht_node_init_u64(&stream
->node_channel_id
, channel_key
);
591 /* Init session id node with the stream session id */
592 lttng_ht_node_init_u64(&stream
->node_session_id
, stream
->session_id
);
594 DBG3("Allocated stream %s (key %" PRIu64
", chan_key %" PRIu64
595 " relayd_id %" PRIu64
", session_id %" PRIu64
,
596 stream
->name
, stream
->key
, channel_key
,
597 stream
->net_seq_idx
, stream
->session_id
);
613 * Add a stream to the global list protected by a mutex.
615 int consumer_add_data_stream(struct lttng_consumer_stream
*stream
)
617 struct lttng_ht
*ht
= data_ht
;
623 DBG3("Adding consumer stream %" PRIu64
, stream
->key
);
625 pthread_mutex_lock(&consumer_data
.lock
);
626 pthread_mutex_lock(&stream
->chan
->lock
);
627 pthread_mutex_lock(&stream
->chan
->timer_lock
);
628 pthread_mutex_lock(&stream
->lock
);
631 /* Steal stream identifier to avoid having streams with the same key */
632 steal_stream_key(stream
->key
, ht
);
634 lttng_ht_add_unique_u64(ht
, &stream
->node
);
636 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
637 &stream
->node_channel_id
);
640 * Add stream to the stream_list_ht of the consumer data. No need to steal
641 * the key since the HT does not use it and we allow to add redundant keys
644 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
647 * When nb_init_stream_left reaches 0, we don't need to trigger any action
648 * in terms of destroying the associated channel, because the action that
649 * causes the count to become 0 also causes a stream to be added. The
650 * channel deletion will thus be triggered by the following removal of this
653 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
654 /* Increment refcount before decrementing nb_init_stream_left */
656 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
659 /* Update consumer data once the node is inserted. */
660 consumer_data
.stream_count
++;
661 consumer_data
.need_update
= 1;
664 pthread_mutex_unlock(&stream
->lock
);
665 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
666 pthread_mutex_unlock(&stream
->chan
->lock
);
667 pthread_mutex_unlock(&consumer_data
.lock
);
672 void consumer_del_data_stream(struct lttng_consumer_stream
*stream
)
674 consumer_del_stream(stream
, data_ht
);
678 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
679 * be acquired before calling this.
681 static int add_relayd(struct consumer_relayd_sock_pair
*relayd
)
684 struct lttng_ht_node_u64
*node
;
685 struct lttng_ht_iter iter
;
689 lttng_ht_lookup(consumer_data
.relayd_ht
,
690 &relayd
->net_seq_idx
, &iter
);
691 node
= lttng_ht_iter_get_node_u64(&iter
);
695 lttng_ht_add_unique_u64(consumer_data
.relayd_ht
, &relayd
->node
);
702 * Allocate and return a consumer relayd socket.
704 struct consumer_relayd_sock_pair
*consumer_allocate_relayd_sock_pair(
705 uint64_t net_seq_idx
)
707 struct consumer_relayd_sock_pair
*obj
= NULL
;
709 /* net sequence index of -1 is a failure */
710 if (net_seq_idx
== (uint64_t) -1ULL) {
714 obj
= zmalloc(sizeof(struct consumer_relayd_sock_pair
));
716 PERROR("zmalloc relayd sock");
720 obj
->net_seq_idx
= net_seq_idx
;
722 obj
->destroy_flag
= 0;
723 obj
->control_sock
.sock
.fd
= -1;
724 obj
->data_sock
.sock
.fd
= -1;
725 lttng_ht_node_init_u64(&obj
->node
, obj
->net_seq_idx
);
726 pthread_mutex_init(&obj
->ctrl_sock_mutex
, NULL
);
733 * Find a relayd socket pair in the global consumer data.
735 * Return the object if found else NULL.
736 * RCU read-side lock must be held across this call and while using the
739 struct consumer_relayd_sock_pair
*consumer_find_relayd(uint64_t key
)
741 struct lttng_ht_iter iter
;
742 struct lttng_ht_node_u64
*node
;
743 struct consumer_relayd_sock_pair
*relayd
= NULL
;
745 /* Negative keys are lookup failures */
746 if (key
== (uint64_t) -1ULL) {
750 lttng_ht_lookup(consumer_data
.relayd_ht
, &key
,
752 node
= lttng_ht_iter_get_node_u64(&iter
);
754 relayd
= caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
762 * Find a relayd and send the stream
764 * Returns 0 on success, < 0 on error
766 int consumer_send_relayd_stream(struct lttng_consumer_stream
*stream
,
770 struct consumer_relayd_sock_pair
*relayd
;
773 assert(stream
->net_seq_idx
!= -1ULL);
776 /* The stream is not metadata. Get relayd reference if exists. */
778 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
779 if (relayd
!= NULL
) {
780 /* Add stream on the relayd */
781 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
782 ret
= relayd_add_stream(&relayd
->control_sock
, stream
->name
,
783 path
, &stream
->relayd_stream_id
,
784 stream
->chan
->tracefile_size
, stream
->chan
->tracefile_count
);
785 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
790 uatomic_inc(&relayd
->refcount
);
791 stream
->sent_to_relayd
= 1;
793 ERR("Stream %" PRIu64
" relayd ID %" PRIu64
" unknown. Can't send it.",
794 stream
->key
, stream
->net_seq_idx
);
799 DBG("Stream %s with key %" PRIu64
" sent to relayd id %" PRIu64
,
800 stream
->name
, stream
->key
, stream
->net_seq_idx
);
808 * Find a relayd and send the streams sent message
810 * Returns 0 on success, < 0 on error
812 int consumer_send_relayd_streams_sent(uint64_t net_seq_idx
)
815 struct consumer_relayd_sock_pair
*relayd
;
817 assert(net_seq_idx
!= -1ULL);
819 /* The stream is not metadata. Get relayd reference if exists. */
821 relayd
= consumer_find_relayd(net_seq_idx
);
822 if (relayd
!= NULL
) {
823 /* Add stream on the relayd */
824 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
825 ret
= relayd_streams_sent(&relayd
->control_sock
);
826 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
831 ERR("Relayd ID %" PRIu64
" unknown. Can't send streams_sent.",
838 DBG("All streams sent relayd id %" PRIu64
, net_seq_idx
);
846 * Find a relayd and close the stream
848 void close_relayd_stream(struct lttng_consumer_stream
*stream
)
850 struct consumer_relayd_sock_pair
*relayd
;
852 /* The stream is not metadata. Get relayd reference if exists. */
854 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
856 consumer_stream_relayd_close(stream
, relayd
);
862 * Handle stream for relayd transmission if the stream applies for network
863 * streaming where the net sequence index is set.
865 * Return destination file descriptor or negative value on error.
867 static int write_relayd_stream_header(struct lttng_consumer_stream
*stream
,
868 size_t data_size
, unsigned long padding
,
869 struct consumer_relayd_sock_pair
*relayd
)
872 struct lttcomm_relayd_data_hdr data_hdr
;
878 /* Reset data header */
879 memset(&data_hdr
, 0, sizeof(data_hdr
));
881 if (stream
->metadata_flag
) {
882 /* Caller MUST acquire the relayd control socket lock */
883 ret
= relayd_send_metadata(&relayd
->control_sock
, data_size
);
888 /* Metadata are always sent on the control socket. */
889 outfd
= relayd
->control_sock
.sock
.fd
;
891 /* Set header with stream information */
892 data_hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
893 data_hdr
.data_size
= htobe32(data_size
);
894 data_hdr
.padding_size
= htobe32(padding
);
896 * Note that net_seq_num below is assigned with the *current* value of
897 * next_net_seq_num and only after that the next_net_seq_num will be
898 * increment. This is why when issuing a command on the relayd using
899 * this next value, 1 should always be substracted in order to compare
900 * the last seen sequence number on the relayd side to the last sent.
902 data_hdr
.net_seq_num
= htobe64(stream
->next_net_seq_num
);
903 /* Other fields are zeroed previously */
905 ret
= relayd_send_data_hdr(&relayd
->data_sock
, &data_hdr
,
911 ++stream
->next_net_seq_num
;
913 /* Set to go on data socket */
914 outfd
= relayd
->data_sock
.sock
.fd
;
922 * Allocate and return a new lttng_consumer_channel object using the given key
923 * to initialize the hash table node.
925 * On error, return NULL.
927 struct lttng_consumer_channel
*consumer_allocate_channel(uint64_t key
,
929 const char *pathname
,
934 enum lttng_event_output output
,
935 uint64_t tracefile_size
,
936 uint64_t tracefile_count
,
937 uint64_t session_id_per_pid
,
938 unsigned int monitor
,
939 unsigned int live_timer_interval
)
941 struct lttng_consumer_channel
*channel
;
943 channel
= zmalloc(sizeof(*channel
));
944 if (channel
== NULL
) {
945 PERROR("malloc struct lttng_consumer_channel");
950 channel
->refcount
= 0;
951 channel
->session_id
= session_id
;
952 channel
->session_id_per_pid
= session_id_per_pid
;
955 channel
->relayd_id
= relayd_id
;
956 channel
->tracefile_size
= tracefile_size
;
957 channel
->tracefile_count
= tracefile_count
;
958 channel
->monitor
= monitor
;
959 channel
->live_timer_interval
= live_timer_interval
;
960 pthread_mutex_init(&channel
->lock
, NULL
);
961 pthread_mutex_init(&channel
->timer_lock
, NULL
);
964 case LTTNG_EVENT_SPLICE
:
965 channel
->output
= CONSUMER_CHANNEL_SPLICE
;
967 case LTTNG_EVENT_MMAP
:
968 channel
->output
= CONSUMER_CHANNEL_MMAP
;
978 * In monitor mode, the streams associated with the channel will be put in
979 * a special list ONLY owned by this channel. So, the refcount is set to 1
980 * here meaning that the channel itself has streams that are referenced.
982 * On a channel deletion, once the channel is no longer visible, the
983 * refcount is decremented and checked for a zero value to delete it. With
984 * streams in no monitor mode, it will now be safe to destroy the channel.
986 if (!channel
->monitor
) {
987 channel
->refcount
= 1;
990 strncpy(channel
->pathname
, pathname
, sizeof(channel
->pathname
));
991 channel
->pathname
[sizeof(channel
->pathname
) - 1] = '\0';
993 strncpy(channel
->name
, name
, sizeof(channel
->name
));
994 channel
->name
[sizeof(channel
->name
) - 1] = '\0';
996 lttng_ht_node_init_u64(&channel
->node
, channel
->key
);
998 channel
->wait_fd
= -1;
1000 CDS_INIT_LIST_HEAD(&channel
->streams
.head
);
1002 DBG("Allocated channel (key %" PRIu64
")", channel
->key
)
1009 * Add a channel to the global list protected by a mutex.
1011 * Always return 0 indicating success.
1013 int consumer_add_channel(struct lttng_consumer_channel
*channel
,
1014 struct lttng_consumer_local_data
*ctx
)
1016 pthread_mutex_lock(&consumer_data
.lock
);
1017 pthread_mutex_lock(&channel
->lock
);
1018 pthread_mutex_lock(&channel
->timer_lock
);
1021 * This gives us a guarantee that the channel we are about to add to the
1022 * channel hash table will be unique. See this function comment on the why
1023 * we need to steel the channel key at this stage.
1025 steal_channel_key(channel
->key
);
1028 lttng_ht_add_unique_u64(consumer_data
.channel_ht
, &channel
->node
);
1031 pthread_mutex_unlock(&channel
->timer_lock
);
1032 pthread_mutex_unlock(&channel
->lock
);
1033 pthread_mutex_unlock(&consumer_data
.lock
);
1035 if (channel
->wait_fd
!= -1 && channel
->type
== CONSUMER_CHANNEL_TYPE_DATA
) {
1036 notify_channel_pipe(ctx
, channel
, -1, CONSUMER_CHANNEL_ADD
);
1043 * Allocate the pollfd structure and the local view of the out fds to avoid
1044 * doing a lookup in the linked list and concurrency issues when writing is
1045 * needed. Called with consumer_data.lock held.
1047 * Returns the number of fds in the structures.
1049 static int update_poll_array(struct lttng_consumer_local_data
*ctx
,
1050 struct pollfd
**pollfd
, struct lttng_consumer_stream
**local_stream
,
1051 struct lttng_ht
*ht
)
1054 struct lttng_ht_iter iter
;
1055 struct lttng_consumer_stream
*stream
;
1060 assert(local_stream
);
1062 DBG("Updating poll fd array");
1064 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1066 * Only active streams with an active end point can be added to the
1067 * poll set and local stream storage of the thread.
1069 * There is a potential race here for endpoint_status to be updated
1070 * just after the check. However, this is OK since the stream(s) will
1071 * be deleted once the thread is notified that the end point state has
1072 * changed where this function will be called back again.
1074 if (stream
->state
!= LTTNG_CONSUMER_ACTIVE_STREAM
||
1075 stream
->endpoint_status
== CONSUMER_ENDPOINT_INACTIVE
) {
1079 * This clobbers way too much the debug output. Uncomment that if you
1080 * need it for debugging purposes.
1082 * DBG("Active FD %d", stream->wait_fd);
1084 (*pollfd
)[i
].fd
= stream
->wait_fd
;
1085 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1086 local_stream
[i
] = stream
;
1092 * Insert the consumer_data_pipe at the end of the array and don't
1093 * increment i so nb_fd is the number of real FD.
1095 (*pollfd
)[i
].fd
= lttng_pipe_get_readfd(ctx
->consumer_data_pipe
);
1096 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1098 (*pollfd
)[i
+ 1].fd
= lttng_pipe_get_readfd(ctx
->consumer_wakeup_pipe
);
1099 (*pollfd
)[i
+ 1].events
= POLLIN
| POLLPRI
;
1104 * Poll on the should_quit pipe and the command socket return -1 on
1105 * error, 1 if should exit, 0 if data is available on the command socket
1107 int lttng_consumer_poll_socket(struct pollfd
*consumer_sockpoll
)
1112 num_rdy
= poll(consumer_sockpoll
, 2, -1);
1113 if (num_rdy
== -1) {
1115 * Restart interrupted system call.
1117 if (errno
== EINTR
) {
1120 PERROR("Poll error");
1123 if (consumer_sockpoll
[0].revents
& (POLLIN
| POLLPRI
)) {
1124 DBG("consumer_should_quit wake up");
1131 * Set the error socket.
1133 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data
*ctx
,
1136 ctx
->consumer_error_socket
= sock
;
1140 * Set the command socket path.
1142 void lttng_consumer_set_command_sock_path(
1143 struct lttng_consumer_local_data
*ctx
, char *sock
)
1145 ctx
->consumer_command_sock_path
= sock
;
1149 * Send return code to the session daemon.
1150 * If the socket is not defined, we return 0, it is not a fatal error
1152 int lttng_consumer_send_error(struct lttng_consumer_local_data
*ctx
, int cmd
)
1154 if (ctx
->consumer_error_socket
> 0) {
1155 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
1156 sizeof(enum lttcomm_sessiond_command
));
1163 * Close all the tracefiles and stream fds and MUST be called when all
1164 * instances are destroyed i.e. when all threads were joined and are ended.
1166 void lttng_consumer_cleanup(void)
1168 struct lttng_ht_iter iter
;
1169 struct lttng_consumer_channel
*channel
;
1173 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
, channel
,
1175 consumer_del_channel(channel
);
1180 lttng_ht_destroy(consumer_data
.channel_ht
);
1182 cleanup_relayd_ht();
1184 lttng_ht_destroy(consumer_data
.stream_per_chan_id_ht
);
1187 * This HT contains streams that are freed by either the metadata thread or
1188 * the data thread so we do *nothing* on the hash table and simply destroy
1191 lttng_ht_destroy(consumer_data
.stream_list_ht
);
1195 * Called from signal handler.
1197 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
1202 ret
= lttng_write(ctx
->consumer_should_quit
[1], "4", 1);
1204 PERROR("write consumer quit");
1207 DBG("Consumer flag that it should quit");
1210 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1213 int outfd
= stream
->out_fd
;
1216 * This does a blocking write-and-wait on any page that belongs to the
1217 * subbuffer prior to the one we just wrote.
1218 * Don't care about error values, as these are just hints and ways to
1219 * limit the amount of page cache used.
1221 if (orig_offset
< stream
->max_sb_size
) {
1224 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1225 stream
->max_sb_size
,
1226 SYNC_FILE_RANGE_WAIT_BEFORE
1227 | SYNC_FILE_RANGE_WRITE
1228 | SYNC_FILE_RANGE_WAIT_AFTER
);
1230 * Give hints to the kernel about how we access the file:
1231 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1234 * We need to call fadvise again after the file grows because the
1235 * kernel does not seem to apply fadvise to non-existing parts of the
1238 * Call fadvise _after_ having waited for the page writeback to
1239 * complete because the dirty page writeback semantic is not well
1240 * defined. So it can be expected to lead to lower throughput in
1243 posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1244 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1248 * Initialise the necessary environnement :
1249 * - create a new context
1250 * - create the poll_pipe
1251 * - create the should_quit pipe (for signal handler)
1252 * - create the thread pipe (for splice)
1254 * Takes a function pointer as argument, this function is called when data is
1255 * available on a buffer. This function is responsible to do the
1256 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1257 * buffer configuration and then kernctl_put_next_subbuf at the end.
1259 * Returns a pointer to the new context or NULL on error.
1261 struct lttng_consumer_local_data
*lttng_consumer_create(
1262 enum lttng_consumer_type type
,
1263 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1264 struct lttng_consumer_local_data
*ctx
),
1265 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1266 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1267 int (*update_stream
)(uint64_t stream_key
, uint32_t state
))
1270 struct lttng_consumer_local_data
*ctx
;
1272 assert(consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1273 consumer_data
.type
== type
);
1274 consumer_data
.type
= type
;
1276 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
1278 PERROR("allocating context");
1282 ctx
->consumer_error_socket
= -1;
1283 ctx
->consumer_metadata_socket
= -1;
1284 pthread_mutex_init(&ctx
->metadata_socket_lock
, NULL
);
1285 /* assign the callbacks */
1286 ctx
->on_buffer_ready
= buffer_ready
;
1287 ctx
->on_recv_channel
= recv_channel
;
1288 ctx
->on_recv_stream
= recv_stream
;
1289 ctx
->on_update_stream
= update_stream
;
1291 ctx
->consumer_data_pipe
= lttng_pipe_open(0);
1292 if (!ctx
->consumer_data_pipe
) {
1293 goto error_poll_pipe
;
1296 ctx
->consumer_wakeup_pipe
= lttng_pipe_open(0);
1297 if (!ctx
->consumer_wakeup_pipe
) {
1298 goto error_wakeup_pipe
;
1301 ret
= pipe(ctx
->consumer_should_quit
);
1303 PERROR("Error creating recv pipe");
1304 goto error_quit_pipe
;
1307 ret
= pipe(ctx
->consumer_channel_pipe
);
1309 PERROR("Error creating channel pipe");
1310 goto error_channel_pipe
;
1313 ctx
->consumer_metadata_pipe
= lttng_pipe_open(0);
1314 if (!ctx
->consumer_metadata_pipe
) {
1315 goto error_metadata_pipe
;
1320 error_metadata_pipe
:
1321 utils_close_pipe(ctx
->consumer_channel_pipe
);
1323 utils_close_pipe(ctx
->consumer_should_quit
);
1325 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1327 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1335 * Iterate over all streams of the hashtable and free them properly.
1337 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1339 struct lttng_ht_iter iter
;
1340 struct lttng_consumer_stream
*stream
;
1347 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1349 * Ignore return value since we are currently cleaning up so any error
1352 (void) consumer_del_stream(stream
, ht
);
1356 lttng_ht_destroy(ht
);
1360 * Iterate over all streams of the metadata hashtable and free them
1363 static void destroy_metadata_stream_ht(struct lttng_ht
*ht
)
1365 struct lttng_ht_iter iter
;
1366 struct lttng_consumer_stream
*stream
;
1373 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1375 * Ignore return value since we are currently cleaning up so any error
1378 (void) consumer_del_metadata_stream(stream
, ht
);
1382 lttng_ht_destroy(ht
);
1386 * Close all fds associated with the instance and free the context.
1388 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1392 DBG("Consumer destroying it. Closing everything.");
1398 destroy_data_stream_ht(data_ht
);
1399 destroy_metadata_stream_ht(metadata_ht
);
1401 ret
= close(ctx
->consumer_error_socket
);
1405 ret
= close(ctx
->consumer_metadata_socket
);
1409 utils_close_pipe(ctx
->consumer_channel_pipe
);
1410 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1411 lttng_pipe_destroy(ctx
->consumer_metadata_pipe
);
1412 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1413 utils_close_pipe(ctx
->consumer_should_quit
);
1415 unlink(ctx
->consumer_command_sock_path
);
1420 * Write the metadata stream id on the specified file descriptor.
1422 static int write_relayd_metadata_id(int fd
,
1423 struct lttng_consumer_stream
*stream
,
1424 struct consumer_relayd_sock_pair
*relayd
, unsigned long padding
)
1427 struct lttcomm_relayd_metadata_payload hdr
;
1429 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1430 hdr
.padding_size
= htobe32(padding
);
1431 ret
= lttng_write(fd
, (void *) &hdr
, sizeof(hdr
));
1432 if (ret
< sizeof(hdr
)) {
1434 * This error means that the fd's end is closed so ignore the perror
1435 * not to clubber the error output since this can happen in a normal
1438 if (errno
!= EPIPE
) {
1439 PERROR("write metadata stream id");
1441 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1443 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1444 * handle writting the missing part so report that as an error and
1445 * don't lie to the caller.
1450 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1451 stream
->relayd_stream_id
, padding
);
1458 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1459 * core function for writing trace buffers to either the local filesystem or
1462 * It must be called with the stream lock held.
1464 * Careful review MUST be put if any changes occur!
1466 * Returns the number of bytes written
1468 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1469 struct lttng_consumer_local_data
*ctx
,
1470 struct lttng_consumer_stream
*stream
, unsigned long len
,
1471 unsigned long padding
,
1472 struct ctf_packet_index
*index
)
1474 unsigned long mmap_offset
;
1477 off_t orig_offset
= stream
->out_fd_offset
;
1478 /* Default is on the disk */
1479 int outfd
= stream
->out_fd
;
1480 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1481 unsigned int relayd_hang_up
= 0;
1483 /* RCU lock for the relayd pointer */
1486 /* Flag that the current stream if set for network streaming. */
1487 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1488 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1489 if (relayd
== NULL
) {
1495 /* get the offset inside the fd to mmap */
1496 switch (consumer_data
.type
) {
1497 case LTTNG_CONSUMER_KERNEL
:
1498 mmap_base
= stream
->mmap_base
;
1499 ret
= kernctl_get_mmap_read_offset(stream
->wait_fd
, &mmap_offset
);
1502 PERROR("tracer ctl get_mmap_read_offset");
1506 case LTTNG_CONSUMER32_UST
:
1507 case LTTNG_CONSUMER64_UST
:
1508 mmap_base
= lttng_ustctl_get_mmap_base(stream
);
1510 ERR("read mmap get mmap base for stream %s", stream
->name
);
1514 ret
= lttng_ustctl_get_mmap_read_offset(stream
, &mmap_offset
);
1516 PERROR("tracer ctl get_mmap_read_offset");
1522 ERR("Unknown consumer_data type");
1526 /* Handle stream on the relayd if the output is on the network */
1528 unsigned long netlen
= len
;
1531 * Lock the control socket for the complete duration of the function
1532 * since from this point on we will use the socket.
1534 if (stream
->metadata_flag
) {
1535 /* Metadata requires the control socket. */
1536 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1537 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1540 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1545 /* Use the returned socket. */
1548 /* Write metadata stream id before payload */
1549 if (stream
->metadata_flag
) {
1550 ret
= write_relayd_metadata_id(outfd
, stream
, relayd
, padding
);
1557 /* No streaming, we have to set the len with the full padding */
1561 * Check if we need to change the tracefile before writing the packet.
1563 if (stream
->chan
->tracefile_size
> 0 &&
1564 (stream
->tracefile_size_current
+ len
) >
1565 stream
->chan
->tracefile_size
) {
1566 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1567 stream
->name
, stream
->chan
->tracefile_size
,
1568 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1569 stream
->out_fd
, &(stream
->tracefile_count_current
),
1572 ERR("Rotating output file");
1575 outfd
= stream
->out_fd
;
1577 if (stream
->index_fd
>= 0) {
1578 ret
= close(stream
->index_fd
);
1580 PERROR("Closing index");
1583 stream
->index_fd
= -1;
1584 ret
= index_create_file(stream
->chan
->pathname
,
1585 stream
->name
, stream
->uid
, stream
->gid
,
1586 stream
->chan
->tracefile_size
,
1587 stream
->tracefile_count_current
);
1591 stream
->index_fd
= ret
;
1594 /* Reset current size because we just perform a rotation. */
1595 stream
->tracefile_size_current
= 0;
1596 stream
->out_fd_offset
= 0;
1599 stream
->tracefile_size_current
+= len
;
1601 index
->offset
= htobe64(stream
->out_fd_offset
);
1606 * This call guarantee that len or less is returned. It's impossible to
1607 * receive a ret value that is bigger than len.
1609 ret
= lttng_write(outfd
, mmap_base
+ mmap_offset
, len
);
1610 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, len
);
1611 if (ret
< 0 || ((size_t) ret
!= len
)) {
1613 * Report error to caller if nothing was written else at least send the
1621 /* Socket operation failed. We consider the relayd dead */
1622 if (errno
== EPIPE
|| errno
== EINVAL
|| errno
== EBADF
) {
1624 * This is possible if the fd is closed on the other side
1625 * (outfd) or any write problem. It can be verbose a bit for a
1626 * normal execution if for instance the relayd is stopped
1627 * abruptly. This can happen so set this to a DBG statement.
1629 DBG("Consumer mmap write detected relayd hang up");
1631 /* Unhandled error, print it and stop function right now. */
1632 PERROR("Error in write mmap (ret %zd != len %lu)", ret
, len
);
1636 stream
->output_written
+= ret
;
1638 /* This call is useless on a socket so better save a syscall. */
1640 /* This won't block, but will start writeout asynchronously */
1641 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, len
,
1642 SYNC_FILE_RANGE_WRITE
);
1643 stream
->out_fd_offset
+= len
;
1645 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1649 * This is a special case that the relayd has closed its socket. Let's
1650 * cleanup the relayd object and all associated streams.
1652 if (relayd
&& relayd_hang_up
) {
1653 cleanup_relayd(relayd
, ctx
);
1657 /* Unlock only if ctrl socket used */
1658 if (relayd
&& stream
->metadata_flag
) {
1659 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1667 * Splice the data from the ring buffer to the tracefile.
1669 * It must be called with the stream lock held.
1671 * Returns the number of bytes spliced.
1673 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1674 struct lttng_consumer_local_data
*ctx
,
1675 struct lttng_consumer_stream
*stream
, unsigned long len
,
1676 unsigned long padding
,
1677 struct ctf_packet_index
*index
)
1679 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1681 off_t orig_offset
= stream
->out_fd_offset
;
1682 int fd
= stream
->wait_fd
;
1683 /* Default is on the disk */
1684 int outfd
= stream
->out_fd
;
1685 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1687 unsigned int relayd_hang_up
= 0;
1689 switch (consumer_data
.type
) {
1690 case LTTNG_CONSUMER_KERNEL
:
1692 case LTTNG_CONSUMER32_UST
:
1693 case LTTNG_CONSUMER64_UST
:
1694 /* Not supported for user space tracing */
1697 ERR("Unknown consumer_data type");
1701 /* RCU lock for the relayd pointer */
1704 /* Flag that the current stream if set for network streaming. */
1705 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1706 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1707 if (relayd
== NULL
) {
1712 splice_pipe
= stream
->splice_pipe
;
1714 /* Write metadata stream id before payload */
1716 unsigned long total_len
= len
;
1718 if (stream
->metadata_flag
) {
1720 * Lock the control socket for the complete duration of the function
1721 * since from this point on we will use the socket.
1723 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1725 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
, relayd
,
1733 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1736 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1742 /* Use the returned socket. */
1745 /* No streaming, we have to set the len with the full padding */
1749 * Check if we need to change the tracefile before writing the packet.
1751 if (stream
->chan
->tracefile_size
> 0 &&
1752 (stream
->tracefile_size_current
+ len
) >
1753 stream
->chan
->tracefile_size
) {
1754 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1755 stream
->name
, stream
->chan
->tracefile_size
,
1756 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1757 stream
->out_fd
, &(stream
->tracefile_count_current
),
1761 ERR("Rotating output file");
1764 outfd
= stream
->out_fd
;
1766 if (stream
->index_fd
>= 0) {
1767 ret
= close(stream
->index_fd
);
1769 PERROR("Closing index");
1772 stream
->index_fd
= -1;
1773 ret
= index_create_file(stream
->chan
->pathname
,
1774 stream
->name
, stream
->uid
, stream
->gid
,
1775 stream
->chan
->tracefile_size
,
1776 stream
->tracefile_count_current
);
1781 stream
->index_fd
= ret
;
1784 /* Reset current size because we just perform a rotation. */
1785 stream
->tracefile_size_current
= 0;
1786 stream
->out_fd_offset
= 0;
1789 stream
->tracefile_size_current
+= len
;
1790 index
->offset
= htobe64(stream
->out_fd_offset
);
1794 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1795 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1796 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1797 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1798 DBG("splice chan to pipe, ret %zd", ret_splice
);
1799 if (ret_splice
< 0) {
1802 PERROR("Error in relay splice");
1806 /* Handle stream on the relayd if the output is on the network */
1807 if (relayd
&& stream
->metadata_flag
) {
1808 size_t metadata_payload_size
=
1809 sizeof(struct lttcomm_relayd_metadata_payload
);
1811 /* Update counter to fit the spliced data */
1812 ret_splice
+= metadata_payload_size
;
1813 len
+= metadata_payload_size
;
1815 * We do this so the return value can match the len passed as
1816 * argument to this function.
1818 written
-= metadata_payload_size
;
1821 /* Splice data out */
1822 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1823 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1824 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
1826 if (ret_splice
< 0) {
1831 } else if (ret_splice
> len
) {
1833 * We don't expect this code path to be executed but you never know
1834 * so this is an extra protection agains a buggy splice().
1837 written
+= ret_splice
;
1838 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice
,
1842 /* All good, update current len and continue. */
1846 /* This call is useless on a socket so better save a syscall. */
1848 /* This won't block, but will start writeout asynchronously */
1849 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1850 SYNC_FILE_RANGE_WRITE
);
1851 stream
->out_fd_offset
+= ret_splice
;
1853 stream
->output_written
+= ret_splice
;
1854 written
+= ret_splice
;
1856 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1861 * This is a special case that the relayd has closed its socket. Let's
1862 * cleanup the relayd object and all associated streams.
1864 if (relayd
&& relayd_hang_up
) {
1865 cleanup_relayd(relayd
, ctx
);
1866 /* Skip splice error so the consumer does not fail */
1871 /* send the appropriate error description to sessiond */
1874 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1877 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1880 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
1885 if (relayd
&& stream
->metadata_flag
) {
1886 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1894 * Take a snapshot for a specific fd
1896 * Returns 0 on success, < 0 on error
1898 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
1900 switch (consumer_data
.type
) {
1901 case LTTNG_CONSUMER_KERNEL
:
1902 return lttng_kconsumer_take_snapshot(stream
);
1903 case LTTNG_CONSUMER32_UST
:
1904 case LTTNG_CONSUMER64_UST
:
1905 return lttng_ustconsumer_take_snapshot(stream
);
1907 ERR("Unknown consumer_data type");
1914 * Get the produced position
1916 * Returns 0 on success, < 0 on error
1918 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
1921 switch (consumer_data
.type
) {
1922 case LTTNG_CONSUMER_KERNEL
:
1923 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
1924 case LTTNG_CONSUMER32_UST
:
1925 case LTTNG_CONSUMER64_UST
:
1926 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
1928 ERR("Unknown consumer_data type");
1934 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
1935 int sock
, struct pollfd
*consumer_sockpoll
)
1937 switch (consumer_data
.type
) {
1938 case LTTNG_CONSUMER_KERNEL
:
1939 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1940 case LTTNG_CONSUMER32_UST
:
1941 case LTTNG_CONSUMER64_UST
:
1942 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1944 ERR("Unknown consumer_data type");
1950 void lttng_consumer_close_all_metadata(void)
1952 switch (consumer_data
.type
) {
1953 case LTTNG_CONSUMER_KERNEL
:
1955 * The Kernel consumer has a different metadata scheme so we don't
1956 * close anything because the stream will be closed by the session
1960 case LTTNG_CONSUMER32_UST
:
1961 case LTTNG_CONSUMER64_UST
:
1963 * Close all metadata streams. The metadata hash table is passed and
1964 * this call iterates over it by closing all wakeup fd. This is safe
1965 * because at this point we are sure that the metadata producer is
1966 * either dead or blocked.
1968 lttng_ustconsumer_close_all_metadata(metadata_ht
);
1971 ERR("Unknown consumer_data type");
1977 * Clean up a metadata stream and free its memory.
1979 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
1980 struct lttng_ht
*ht
)
1982 struct lttng_consumer_channel
*free_chan
= NULL
;
1986 * This call should NEVER receive regular stream. It must always be
1987 * metadata stream and this is crucial for data structure synchronization.
1989 assert(stream
->metadata_flag
);
1991 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
1993 pthread_mutex_lock(&consumer_data
.lock
);
1994 pthread_mutex_lock(&stream
->chan
->lock
);
1995 pthread_mutex_lock(&stream
->lock
);
1997 /* Remove any reference to that stream. */
1998 consumer_stream_delete(stream
, ht
);
2000 /* Close down everything including the relayd if one. */
2001 consumer_stream_close(stream
);
2002 /* Destroy tracer buffers of the stream. */
2003 consumer_stream_destroy_buffers(stream
);
2005 /* Atomically decrement channel refcount since other threads can use it. */
2006 if (!uatomic_sub_return(&stream
->chan
->refcount
, 1)
2007 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
2008 /* Go for channel deletion! */
2009 free_chan
= stream
->chan
;
2013 * Nullify the stream reference so it is not used after deletion. The
2014 * channel lock MUST be acquired before being able to check for a NULL
2017 stream
->chan
->metadata_stream
= NULL
;
2019 pthread_mutex_unlock(&stream
->lock
);
2020 pthread_mutex_unlock(&stream
->chan
->lock
);
2021 pthread_mutex_unlock(&consumer_data
.lock
);
2024 consumer_del_channel(free_chan
);
2027 consumer_stream_free(stream
);
2031 * Action done with the metadata stream when adding it to the consumer internal
2032 * data structures to handle it.
2034 int consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
2036 struct lttng_ht
*ht
= metadata_ht
;
2038 struct lttng_ht_iter iter
;
2039 struct lttng_ht_node_u64
*node
;
2044 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2046 pthread_mutex_lock(&consumer_data
.lock
);
2047 pthread_mutex_lock(&stream
->chan
->lock
);
2048 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2049 pthread_mutex_lock(&stream
->lock
);
2052 * From here, refcounts are updated so be _careful_ when returning an error
2059 * Lookup the stream just to make sure it does not exist in our internal
2060 * state. This should NEVER happen.
2062 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2063 node
= lttng_ht_iter_get_node_u64(&iter
);
2067 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2068 * in terms of destroying the associated channel, because the action that
2069 * causes the count to become 0 also causes a stream to be added. The
2070 * channel deletion will thus be triggered by the following removal of this
2073 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2074 /* Increment refcount before decrementing nb_init_stream_left */
2076 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2079 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2081 lttng_ht_add_unique_u64(consumer_data
.stream_per_chan_id_ht
,
2082 &stream
->node_channel_id
);
2085 * Add stream to the stream_list_ht of the consumer data. No need to steal
2086 * the key since the HT does not use it and we allow to add redundant keys
2089 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2093 pthread_mutex_unlock(&stream
->lock
);
2094 pthread_mutex_unlock(&stream
->chan
->lock
);
2095 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2096 pthread_mutex_unlock(&consumer_data
.lock
);
2101 * Delete data stream that are flagged for deletion (endpoint_status).
2103 static void validate_endpoint_status_data_stream(void)
2105 struct lttng_ht_iter iter
;
2106 struct lttng_consumer_stream
*stream
;
2108 DBG("Consumer delete flagged data stream");
2111 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2112 /* Validate delete flag of the stream */
2113 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2116 /* Delete it right now */
2117 consumer_del_stream(stream
, data_ht
);
2123 * Delete metadata stream that are flagged for deletion (endpoint_status).
2125 static void validate_endpoint_status_metadata_stream(
2126 struct lttng_poll_event
*pollset
)
2128 struct lttng_ht_iter iter
;
2129 struct lttng_consumer_stream
*stream
;
2131 DBG("Consumer delete flagged metadata stream");
2136 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2137 /* Validate delete flag of the stream */
2138 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2142 * Remove from pollset so the metadata thread can continue without
2143 * blocking on a deleted stream.
2145 lttng_poll_del(pollset
, stream
->wait_fd
);
2147 /* Delete it right now */
2148 consumer_del_metadata_stream(stream
, metadata_ht
);
2154 * Thread polls on metadata file descriptor and write them on disk or on the
2157 void *consumer_thread_metadata_poll(void *data
)
2159 int ret
, i
, pollfd
, err
= -1;
2160 uint32_t revents
, nb_fd
;
2161 struct lttng_consumer_stream
*stream
= NULL
;
2162 struct lttng_ht_iter iter
;
2163 struct lttng_ht_node_u64
*node
;
2164 struct lttng_poll_event events
;
2165 struct lttng_consumer_local_data
*ctx
= data
;
2168 rcu_register_thread();
2170 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2172 if (testpoint(consumerd_thread_metadata
)) {
2173 goto error_testpoint
;
2176 health_code_update();
2178 DBG("Thread metadata poll started");
2180 /* Size is set to 1 for the consumer_metadata pipe */
2181 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2183 ERR("Poll set creation failed");
2187 ret
= lttng_poll_add(&events
,
2188 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2194 DBG("Metadata main loop started");
2198 health_code_update();
2199 health_poll_entry();
2200 DBG("Metadata poll wait");
2201 ret
= lttng_poll_wait(&events
, -1);
2202 DBG("Metadata poll return from wait with %d fd(s)",
2203 LTTNG_POLL_GETNB(&events
));
2205 DBG("Metadata event catched in thread");
2207 if (errno
== EINTR
) {
2208 ERR("Poll EINTR catched");
2211 if (LTTNG_POLL_GETNB(&events
) == 0) {
2212 err
= 0; /* All is OK */
2219 /* From here, the event is a metadata wait fd */
2220 for (i
= 0; i
< nb_fd
; i
++) {
2221 health_code_update();
2223 revents
= LTTNG_POLL_GETEV(&events
, i
);
2224 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2227 /* No activity for this FD (poll implementation). */
2231 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2232 if (revents
& LPOLLIN
) {
2235 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2236 &stream
, sizeof(stream
));
2237 if (pipe_len
< sizeof(stream
)) {
2239 PERROR("read metadata stream");
2242 * Remove the pipe from the poll set and continue the loop
2243 * since their might be data to consume.
2245 lttng_poll_del(&events
,
2246 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2247 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2251 /* A NULL stream means that the state has changed. */
2252 if (stream
== NULL
) {
2253 /* Check for deleted streams. */
2254 validate_endpoint_status_metadata_stream(&events
);
2258 DBG("Adding metadata stream %d to poll set",
2261 /* Add metadata stream to the global poll events list */
2262 lttng_poll_add(&events
, stream
->wait_fd
,
2263 LPOLLIN
| LPOLLPRI
| LPOLLHUP
);
2264 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2265 DBG("Metadata thread pipe hung up");
2267 * Remove the pipe from the poll set and continue the loop
2268 * since their might be data to consume.
2270 lttng_poll_del(&events
,
2271 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2272 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2275 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2279 /* Handle other stream */
2285 uint64_t tmp_id
= (uint64_t) pollfd
;
2287 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2289 node
= lttng_ht_iter_get_node_u64(&iter
);
2292 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2295 if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2296 /* Get the data out of the metadata file descriptor */
2297 DBG("Metadata available on fd %d", pollfd
);
2298 assert(stream
->wait_fd
== pollfd
);
2301 health_code_update();
2303 len
= ctx
->on_buffer_ready(stream
, ctx
);
2305 * We don't check the return value here since if we get
2306 * a negative len, it means an error occured thus we
2307 * simply remove it from the poll set and free the
2312 /* It's ok to have an unavailable sub-buffer */
2313 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2314 /* Clean up stream from consumer and free it. */
2315 lttng_poll_del(&events
, stream
->wait_fd
);
2316 consumer_del_metadata_stream(stream
, metadata_ht
);
2318 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2319 DBG("Metadata fd %d is hup|err.", pollfd
);
2320 if (!stream
->hangup_flush_done
2321 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2322 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2323 DBG("Attempting to flush and consume the UST buffers");
2324 lttng_ustconsumer_on_stream_hangup(stream
);
2326 /* We just flushed the stream now read it. */
2328 health_code_update();
2330 len
= ctx
->on_buffer_ready(stream
, ctx
);
2332 * We don't check the return value here since if we get
2333 * a negative len, it means an error occured thus we
2334 * simply remove it from the poll set and free the
2340 lttng_poll_del(&events
, stream
->wait_fd
);
2342 * This call update the channel states, closes file descriptors
2343 * and securely free the stream.
2345 consumer_del_metadata_stream(stream
, metadata_ht
);
2347 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2351 /* Release RCU lock for the stream looked up */
2359 DBG("Metadata poll thread exiting");
2361 lttng_poll_clean(&events
);
2366 ERR("Health error occurred in %s", __func__
);
2368 health_unregister(health_consumerd
);
2369 rcu_unregister_thread();
2374 * This thread polls the fds in the set to consume the data and write
2375 * it to tracefile if necessary.
2377 void *consumer_thread_data_poll(void *data
)
2379 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2380 struct pollfd
*pollfd
= NULL
;
2381 /* local view of the streams */
2382 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2383 /* local view of consumer_data.fds_count */
2385 struct lttng_consumer_local_data
*ctx
= data
;
2388 rcu_register_thread();
2390 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2392 if (testpoint(consumerd_thread_data
)) {
2393 goto error_testpoint
;
2396 health_code_update();
2398 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
*));
2399 if (local_stream
== NULL
) {
2400 PERROR("local_stream malloc");
2405 health_code_update();
2411 * the fds set has been updated, we need to update our
2412 * local array as well
2414 pthread_mutex_lock(&consumer_data
.lock
);
2415 if (consumer_data
.need_update
) {
2420 local_stream
= NULL
;
2423 * Allocate for all fds +1 for the consumer_data_pipe and +1 for
2426 pollfd
= zmalloc((consumer_data
.stream_count
+ 2) * sizeof(struct pollfd
));
2427 if (pollfd
== NULL
) {
2428 PERROR("pollfd malloc");
2429 pthread_mutex_unlock(&consumer_data
.lock
);
2433 local_stream
= zmalloc((consumer_data
.stream_count
+ 2) *
2434 sizeof(struct lttng_consumer_stream
*));
2435 if (local_stream
== NULL
) {
2436 PERROR("local_stream malloc");
2437 pthread_mutex_unlock(&consumer_data
.lock
);
2440 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2443 ERR("Error in allocating pollfd or local_outfds");
2444 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2445 pthread_mutex_unlock(&consumer_data
.lock
);
2449 consumer_data
.need_update
= 0;
2451 pthread_mutex_unlock(&consumer_data
.lock
);
2453 /* No FDs and consumer_quit, consumer_cleanup the thread */
2454 if (nb_fd
== 0 && consumer_quit
== 1) {
2455 err
= 0; /* All is OK */
2458 /* poll on the array of fds */
2460 DBG("polling on %d fd", nb_fd
+ 2);
2461 health_poll_entry();
2462 num_rdy
= poll(pollfd
, nb_fd
+ 2, -1);
2464 DBG("poll num_rdy : %d", num_rdy
);
2465 if (num_rdy
== -1) {
2467 * Restart interrupted system call.
2469 if (errno
== EINTR
) {
2472 PERROR("Poll error");
2473 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2475 } else if (num_rdy
== 0) {
2476 DBG("Polling thread timed out");
2481 * If the consumer_data_pipe triggered poll go directly to the
2482 * beginning of the loop to update the array. We want to prioritize
2483 * array update over low-priority reads.
2485 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2486 ssize_t pipe_readlen
;
2488 DBG("consumer_data_pipe wake up");
2489 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2490 &new_stream
, sizeof(new_stream
));
2491 if (pipe_readlen
< sizeof(new_stream
)) {
2492 PERROR("Consumer data pipe");
2493 /* Continue so we can at least handle the current stream(s). */
2498 * If the stream is NULL, just ignore it. It's also possible that
2499 * the sessiond poll thread changed the consumer_quit state and is
2500 * waking us up to test it.
2502 if (new_stream
== NULL
) {
2503 validate_endpoint_status_data_stream();
2507 /* Continue to update the local streams and handle prio ones */
2511 /* Handle wakeup pipe. */
2512 if (pollfd
[nb_fd
+ 1].revents
& (POLLIN
| POLLPRI
)) {
2514 ssize_t pipe_readlen
;
2516 pipe_readlen
= lttng_pipe_read(ctx
->consumer_wakeup_pipe
, &dummy
,
2518 if (pipe_readlen
< 0) {
2519 PERROR("Consumer data wakeup pipe");
2521 /* We've been awakened to handle stream(s). */
2522 ctx
->has_wakeup
= 0;
2525 /* Take care of high priority channels first. */
2526 for (i
= 0; i
< nb_fd
; i
++) {
2527 health_code_update();
2529 if (local_stream
[i
] == NULL
) {
2532 if (pollfd
[i
].revents
& POLLPRI
) {
2533 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2535 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2536 /* it's ok to have an unavailable sub-buffer */
2537 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2538 /* Clean the stream and free it. */
2539 consumer_del_stream(local_stream
[i
], data_ht
);
2540 local_stream
[i
] = NULL
;
2541 } else if (len
> 0) {
2542 local_stream
[i
]->data_read
= 1;
2548 * If we read high prio channel in this loop, try again
2549 * for more high prio data.
2555 /* Take care of low priority channels. */
2556 for (i
= 0; i
< nb_fd
; i
++) {
2557 health_code_update();
2559 if (local_stream
[i
] == NULL
) {
2562 if ((pollfd
[i
].revents
& POLLIN
) ||
2563 local_stream
[i
]->hangup_flush_done
||
2564 local_stream
[i
]->has_data
) {
2565 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2566 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2567 /* it's ok to have an unavailable sub-buffer */
2568 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2569 /* Clean the stream and free it. */
2570 consumer_del_stream(local_stream
[i
], data_ht
);
2571 local_stream
[i
] = NULL
;
2572 } else if (len
> 0) {
2573 local_stream
[i
]->data_read
= 1;
2578 /* Handle hangup and errors */
2579 for (i
= 0; i
< nb_fd
; i
++) {
2580 health_code_update();
2582 if (local_stream
[i
] == NULL
) {
2585 if (!local_stream
[i
]->hangup_flush_done
2586 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2587 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2588 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2589 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2591 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2592 /* Attempt read again, for the data we just flushed. */
2593 local_stream
[i
]->data_read
= 1;
2596 * If the poll flag is HUP/ERR/NVAL and we have
2597 * read no data in this pass, we can remove the
2598 * stream from its hash table.
2600 if ((pollfd
[i
].revents
& POLLHUP
)) {
2601 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2602 if (!local_stream
[i
]->data_read
) {
2603 consumer_del_stream(local_stream
[i
], data_ht
);
2604 local_stream
[i
] = NULL
;
2607 } else if (pollfd
[i
].revents
& POLLERR
) {
2608 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2609 if (!local_stream
[i
]->data_read
) {
2610 consumer_del_stream(local_stream
[i
], data_ht
);
2611 local_stream
[i
] = NULL
;
2614 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2615 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2616 if (!local_stream
[i
]->data_read
) {
2617 consumer_del_stream(local_stream
[i
], data_ht
);
2618 local_stream
[i
] = NULL
;
2622 if (local_stream
[i
] != NULL
) {
2623 local_stream
[i
]->data_read
= 0;
2630 DBG("polling thread exiting");
2635 * Close the write side of the pipe so epoll_wait() in
2636 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2637 * read side of the pipe. If we close them both, epoll_wait strangely does
2638 * not return and could create a endless wait period if the pipe is the
2639 * only tracked fd in the poll set. The thread will take care of closing
2642 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2647 ERR("Health error occurred in %s", __func__
);
2649 health_unregister(health_consumerd
);
2651 rcu_unregister_thread();
2656 * Close wake-up end of each stream belonging to the channel. This will
2657 * allow the poll() on the stream read-side to detect when the
2658 * write-side (application) finally closes them.
2661 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2663 struct lttng_ht
*ht
;
2664 struct lttng_consumer_stream
*stream
;
2665 struct lttng_ht_iter iter
;
2667 ht
= consumer_data
.stream_per_chan_id_ht
;
2670 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2671 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2672 ht
->match_fct
, &channel
->key
,
2673 &iter
.iter
, stream
, node_channel_id
.node
) {
2675 * Protect against teardown with mutex.
2677 pthread_mutex_lock(&stream
->lock
);
2678 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2681 switch (consumer_data
.type
) {
2682 case LTTNG_CONSUMER_KERNEL
:
2684 case LTTNG_CONSUMER32_UST
:
2685 case LTTNG_CONSUMER64_UST
:
2686 if (stream
->metadata_flag
) {
2687 /* Safe and protected by the stream lock. */
2688 lttng_ustconsumer_close_metadata(stream
->chan
);
2691 * Note: a mutex is taken internally within
2692 * liblttng-ust-ctl to protect timer wakeup_fd
2693 * use from concurrent close.
2695 lttng_ustconsumer_close_stream_wakeup(stream
);
2699 ERR("Unknown consumer_data type");
2703 pthread_mutex_unlock(&stream
->lock
);
2708 static void destroy_channel_ht(struct lttng_ht
*ht
)
2710 struct lttng_ht_iter iter
;
2711 struct lttng_consumer_channel
*channel
;
2719 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2720 ret
= lttng_ht_del(ht
, &iter
);
2725 lttng_ht_destroy(ht
);
2729 * This thread polls the channel fds to detect when they are being
2730 * closed. It closes all related streams if the channel is detected as
2731 * closed. It is currently only used as a shim layer for UST because the
2732 * consumerd needs to keep the per-stream wakeup end of pipes open for
2735 void *consumer_thread_channel_poll(void *data
)
2737 int ret
, i
, pollfd
, err
= -1;
2738 uint32_t revents
, nb_fd
;
2739 struct lttng_consumer_channel
*chan
= NULL
;
2740 struct lttng_ht_iter iter
;
2741 struct lttng_ht_node_u64
*node
;
2742 struct lttng_poll_event events
;
2743 struct lttng_consumer_local_data
*ctx
= data
;
2744 struct lttng_ht
*channel_ht
;
2746 rcu_register_thread();
2748 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
2750 if (testpoint(consumerd_thread_channel
)) {
2751 goto error_testpoint
;
2754 health_code_update();
2756 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2758 /* ENOMEM at this point. Better to bail out. */
2762 DBG("Thread channel poll started");
2764 /* Size is set to 1 for the consumer_channel pipe */
2765 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2767 ERR("Poll set creation failed");
2771 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2777 DBG("Channel main loop started");
2781 health_code_update();
2782 DBG("Channel poll wait");
2783 health_poll_entry();
2784 ret
= lttng_poll_wait(&events
, -1);
2785 DBG("Channel poll return from wait with %d fd(s)",
2786 LTTNG_POLL_GETNB(&events
));
2788 DBG("Channel event catched in thread");
2790 if (errno
== EINTR
) {
2791 ERR("Poll EINTR catched");
2794 if (LTTNG_POLL_GETNB(&events
) == 0) {
2795 err
= 0; /* All is OK */
2802 /* From here, the event is a channel wait fd */
2803 for (i
= 0; i
< nb_fd
; i
++) {
2804 health_code_update();
2806 revents
= LTTNG_POLL_GETEV(&events
, i
);
2807 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2810 /* No activity for this FD (poll implementation). */
2814 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2815 if (revents
& LPOLLIN
) {
2816 enum consumer_channel_action action
;
2819 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
2822 ERR("Error reading channel pipe");
2824 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2829 case CONSUMER_CHANNEL_ADD
:
2830 DBG("Adding channel %d to poll set",
2833 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
2836 lttng_ht_add_unique_u64(channel_ht
,
2837 &chan
->wait_fd_node
);
2839 /* Add channel to the global poll events list */
2840 lttng_poll_add(&events
, chan
->wait_fd
,
2841 LPOLLERR
| LPOLLHUP
);
2843 case CONSUMER_CHANNEL_DEL
:
2846 * This command should never be called if the channel
2847 * has streams monitored by either the data or metadata
2848 * thread. The consumer only notify this thread with a
2849 * channel del. command if it receives a destroy
2850 * channel command from the session daemon that send it
2851 * if a command prior to the GET_CHANNEL failed.
2855 chan
= consumer_find_channel(key
);
2858 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
2861 lttng_poll_del(&events
, chan
->wait_fd
);
2862 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
2863 ret
= lttng_ht_del(channel_ht
, &iter
);
2866 switch (consumer_data
.type
) {
2867 case LTTNG_CONSUMER_KERNEL
:
2869 case LTTNG_CONSUMER32_UST
:
2870 case LTTNG_CONSUMER64_UST
:
2871 health_code_update();
2872 /* Destroy streams that might have been left in the stream list. */
2873 clean_channel_stream_list(chan
);
2876 ERR("Unknown consumer_data type");
2881 * Release our own refcount. Force channel deletion even if
2882 * streams were not initialized.
2884 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
2885 consumer_del_channel(chan
);
2890 case CONSUMER_CHANNEL_QUIT
:
2892 * Remove the pipe from the poll set and continue the loop
2893 * since their might be data to consume.
2895 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2898 ERR("Unknown action");
2901 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2902 DBG("Channel thread pipe hung up");
2904 * Remove the pipe from the poll set and continue the loop
2905 * since their might be data to consume.
2907 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2910 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2914 /* Handle other stream */
2920 uint64_t tmp_id
= (uint64_t) pollfd
;
2922 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
2924 node
= lttng_ht_iter_get_node_u64(&iter
);
2927 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
2930 /* Check for error event */
2931 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2932 DBG("Channel fd %d is hup|err.", pollfd
);
2934 lttng_poll_del(&events
, chan
->wait_fd
);
2935 ret
= lttng_ht_del(channel_ht
, &iter
);
2939 * This will close the wait fd for each stream associated to
2940 * this channel AND monitored by the data/metadata thread thus
2941 * will be clean by the right thread.
2943 consumer_close_channel_streams(chan
);
2945 /* Release our own refcount */
2946 if (!uatomic_sub_return(&chan
->refcount
, 1)
2947 && !uatomic_read(&chan
->nb_init_stream_left
)) {
2948 consumer_del_channel(chan
);
2951 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2956 /* Release RCU lock for the channel looked up */
2964 lttng_poll_clean(&events
);
2966 destroy_channel_ht(channel_ht
);
2969 DBG("Channel poll thread exiting");
2972 ERR("Health error occurred in %s", __func__
);
2974 health_unregister(health_consumerd
);
2975 rcu_unregister_thread();
2979 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
2980 struct pollfd
*sockpoll
, int client_socket
)
2987 ret
= lttng_consumer_poll_socket(sockpoll
);
2991 DBG("Metadata connection on client_socket");
2993 /* Blocking call, waiting for transmission */
2994 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
2995 if (ctx
->consumer_metadata_socket
< 0) {
2996 WARN("On accept metadata");
3007 * This thread listens on the consumerd socket and receives the file
3008 * descriptors from the session daemon.
3010 void *consumer_thread_sessiond_poll(void *data
)
3012 int sock
= -1, client_socket
, ret
, err
= -1;
3014 * structure to poll for incoming data on communication socket avoids
3015 * making blocking sockets.
3017 struct pollfd consumer_sockpoll
[2];
3018 struct lttng_consumer_local_data
*ctx
= data
;
3020 rcu_register_thread();
3022 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
3024 if (testpoint(consumerd_thread_sessiond
)) {
3025 goto error_testpoint
;
3028 health_code_update();
3030 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
3031 unlink(ctx
->consumer_command_sock_path
);
3032 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
3033 if (client_socket
< 0) {
3034 ERR("Cannot create command socket");
3038 ret
= lttcomm_listen_unix_sock(client_socket
);
3043 DBG("Sending ready command to lttng-sessiond");
3044 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3045 /* return < 0 on error, but == 0 is not fatal */
3047 ERR("Error sending ready command to lttng-sessiond");
3051 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3052 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3053 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3054 consumer_sockpoll
[1].fd
= client_socket
;
3055 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3057 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3065 DBG("Connection on client_socket");
3067 /* Blocking call, waiting for transmission */
3068 sock
= lttcomm_accept_unix_sock(client_socket
);
3075 * Setup metadata socket which is the second socket connection on the
3076 * command unix socket.
3078 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3087 /* This socket is not useful anymore. */
3088 ret
= close(client_socket
);
3090 PERROR("close client_socket");
3094 /* update the polling structure to poll on the established socket */
3095 consumer_sockpoll
[1].fd
= sock
;
3096 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3099 health_code_update();
3101 health_poll_entry();
3102 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3111 DBG("Incoming command on sock");
3112 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3115 * This could simply be a session daemon quitting. Don't output
3118 DBG("Communication interrupted on command socket");
3122 if (consumer_quit
) {
3123 DBG("consumer_thread_receive_fds received quit from signal");
3124 err
= 0; /* All is OK */
3127 DBG("received command on sock");
3133 DBG("Consumer thread sessiond poll exiting");
3136 * Close metadata streams since the producer is the session daemon which
3139 * NOTE: for now, this only applies to the UST tracer.
3141 lttng_consumer_close_all_metadata();
3144 * when all fds have hung up, the polling thread
3150 * Notify the data poll thread to poll back again and test the
3151 * consumer_quit state that we just set so to quit gracefully.
3153 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3155 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3157 notify_health_quit_pipe(health_quit_pipe
);
3159 /* Cleaning up possibly open sockets. */
3163 PERROR("close sock sessiond poll");
3166 if (client_socket
>= 0) {
3167 ret
= close(client_socket
);
3169 PERROR("close client_socket sessiond poll");
3176 ERR("Health error occurred in %s", __func__
);
3178 health_unregister(health_consumerd
);
3180 rcu_unregister_thread();
3184 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3185 struct lttng_consumer_local_data
*ctx
)
3189 pthread_mutex_lock(&stream
->lock
);
3190 if (stream
->metadata_flag
) {
3191 pthread_mutex_lock(&stream
->metadata_rdv_lock
);
3194 switch (consumer_data
.type
) {
3195 case LTTNG_CONSUMER_KERNEL
:
3196 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
3198 case LTTNG_CONSUMER32_UST
:
3199 case LTTNG_CONSUMER64_UST
:
3200 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
3203 ERR("Unknown consumer_data type");
3209 if (stream
->metadata_flag
) {
3210 pthread_cond_broadcast(&stream
->metadata_rdv
);
3211 pthread_mutex_unlock(&stream
->metadata_rdv_lock
);
3213 pthread_mutex_unlock(&stream
->lock
);
3217 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3219 switch (consumer_data
.type
) {
3220 case LTTNG_CONSUMER_KERNEL
:
3221 return lttng_kconsumer_on_recv_stream(stream
);
3222 case LTTNG_CONSUMER32_UST
:
3223 case LTTNG_CONSUMER64_UST
:
3224 return lttng_ustconsumer_on_recv_stream(stream
);
3226 ERR("Unknown consumer_data type");
3233 * Allocate and set consumer data hash tables.
3235 int lttng_consumer_init(void)
3237 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3238 if (!consumer_data
.channel_ht
) {
3242 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3243 if (!consumer_data
.relayd_ht
) {
3247 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3248 if (!consumer_data
.stream_list_ht
) {
3252 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3253 if (!consumer_data
.stream_per_chan_id_ht
) {
3257 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3262 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3274 * Process the ADD_RELAYD command receive by a consumer.
3276 * This will create a relayd socket pair and add it to the relayd hash table.
3277 * The caller MUST acquire a RCU read side lock before calling it.
3279 int consumer_add_relayd_socket(uint64_t net_seq_idx
, int sock_type
,
3280 struct lttng_consumer_local_data
*ctx
, int sock
,
3281 struct pollfd
*consumer_sockpoll
,
3282 struct lttcomm_relayd_sock
*relayd_sock
, uint64_t sessiond_id
,
3283 uint64_t relayd_session_id
)
3285 int fd
= -1, ret
= -1, relayd_created
= 0;
3286 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3287 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3290 assert(relayd_sock
);
3292 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", net_seq_idx
);
3294 /* Get relayd reference if exists. */
3295 relayd
= consumer_find_relayd(net_seq_idx
);
3296 if (relayd
== NULL
) {
3297 assert(sock_type
== LTTNG_STREAM_CONTROL
);
3298 /* Not found. Allocate one. */
3299 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3300 if (relayd
== NULL
) {
3302 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3305 relayd
->sessiond_session_id
= sessiond_id
;
3310 * This code path MUST continue to the consumer send status message to
3311 * we can notify the session daemon and continue our work without
3312 * killing everything.
3316 * relayd key should never be found for control socket.
3318 assert(sock_type
!= LTTNG_STREAM_CONTROL
);
3321 /* First send a status message before receiving the fds. */
3322 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3324 /* Somehow, the session daemon is not responding anymore. */
3325 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3326 goto error_nosignal
;
3329 /* Poll on consumer socket. */
3330 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3332 /* Needing to exit in the middle of a command: error. */
3333 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3335 goto error_nosignal
;
3338 /* Get relayd socket from session daemon */
3339 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3340 if (ret
!= sizeof(fd
)) {
3342 fd
= -1; /* Just in case it gets set with an invalid value. */
3345 * Failing to receive FDs might indicate a major problem such as
3346 * reaching a fd limit during the receive where the kernel returns a
3347 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3348 * don't take any chances and stop everything.
3350 * XXX: Feature request #558 will fix that and avoid this possible
3351 * issue when reaching the fd limit.
3353 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3354 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3358 /* Copy socket information and received FD */
3359 switch (sock_type
) {
3360 case LTTNG_STREAM_CONTROL
:
3361 /* Copy received lttcomm socket */
3362 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3363 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3364 /* Handle create_sock error. */
3366 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3370 * Close the socket created internally by
3371 * lttcomm_create_sock, so we can replace it by the one
3372 * received from sessiond.
3374 if (close(relayd
->control_sock
.sock
.fd
)) {
3378 /* Assign new file descriptor */
3379 relayd
->control_sock
.sock
.fd
= fd
;
3380 fd
= -1; /* For error path */
3381 /* Assign version values. */
3382 relayd
->control_sock
.major
= relayd_sock
->major
;
3383 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3385 relayd
->relayd_session_id
= relayd_session_id
;
3388 case LTTNG_STREAM_DATA
:
3389 /* Copy received lttcomm socket */
3390 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3391 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3392 /* Handle create_sock error. */
3394 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3398 * Close the socket created internally by
3399 * lttcomm_create_sock, so we can replace it by the one
3400 * received from sessiond.
3402 if (close(relayd
->data_sock
.sock
.fd
)) {
3406 /* Assign new file descriptor */
3407 relayd
->data_sock
.sock
.fd
= fd
;
3408 fd
= -1; /* for eventual error paths */
3409 /* Assign version values. */
3410 relayd
->data_sock
.major
= relayd_sock
->major
;
3411 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3414 ERR("Unknown relayd socket type (%d)", sock_type
);
3416 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3420 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3421 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3422 relayd
->net_seq_idx
, fd
);
3424 /* We successfully added the socket. Send status back. */
3425 ret
= consumer_send_status_msg(sock
, ret_code
);
3427 /* Somehow, the session daemon is not responding anymore. */
3428 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3429 goto error_nosignal
;
3433 * Add relayd socket pair to consumer data hashtable. If object already
3434 * exists or on error, the function gracefully returns.
3442 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3443 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3447 /* Close received socket if valid. */
3450 PERROR("close received socket");
3454 if (relayd_created
) {
3462 * Try to lock the stream mutex.
3464 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
3466 static int stream_try_lock(struct lttng_consumer_stream
*stream
)
3473 * Try to lock the stream mutex. On failure, we know that the stream is
3474 * being used else where hence there is data still being extracted.
3476 ret
= pthread_mutex_trylock(&stream
->lock
);
3478 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
3490 * Search for a relayd associated to the session id and return the reference.
3492 * A rcu read side lock MUST be acquire before calling this function and locked
3493 * until the relayd object is no longer necessary.
3495 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3497 struct lttng_ht_iter iter
;
3498 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3500 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3501 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3504 * Check by sessiond id which is unique here where the relayd session
3505 * id might not be when having multiple relayd.
3507 if (relayd
->sessiond_session_id
== id
) {
3508 /* Found the relayd. There can be only one per id. */
3520 * Check if for a given session id there is still data needed to be extract
3523 * Return 1 if data is pending or else 0 meaning ready to be read.
3525 int consumer_data_pending(uint64_t id
)
3528 struct lttng_ht_iter iter
;
3529 struct lttng_ht
*ht
;
3530 struct lttng_consumer_stream
*stream
;
3531 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3532 int (*data_pending
)(struct lttng_consumer_stream
*);
3534 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3537 pthread_mutex_lock(&consumer_data
.lock
);
3539 switch (consumer_data
.type
) {
3540 case LTTNG_CONSUMER_KERNEL
:
3541 data_pending
= lttng_kconsumer_data_pending
;
3543 case LTTNG_CONSUMER32_UST
:
3544 case LTTNG_CONSUMER64_UST
:
3545 data_pending
= lttng_ustconsumer_data_pending
;
3548 ERR("Unknown consumer data type");
3552 /* Ease our life a bit */
3553 ht
= consumer_data
.stream_list_ht
;
3555 relayd
= find_relayd_by_session_id(id
);
3557 /* Send init command for data pending. */
3558 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3559 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3560 relayd
->relayd_session_id
);
3561 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3563 /* Communication error thus the relayd so no data pending. */
3564 goto data_not_pending
;
3568 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3569 ht
->hash_fct(&id
, lttng_ht_seed
),
3571 &iter
.iter
, stream
, node_session_id
.node
) {
3572 /* If this call fails, the stream is being used hence data pending. */
3573 ret
= stream_try_lock(stream
);
3579 * A removed node from the hash table indicates that the stream has
3580 * been deleted thus having a guarantee that the buffers are closed
3581 * on the consumer side. However, data can still be transmitted
3582 * over the network so don't skip the relayd check.
3584 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3586 /* Check the stream if there is data in the buffers. */
3587 ret
= data_pending(stream
);
3589 pthread_mutex_unlock(&stream
->lock
);
3596 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3597 if (stream
->metadata_flag
) {
3598 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3599 stream
->relayd_stream_id
);
3601 ret
= relayd_data_pending(&relayd
->control_sock
,
3602 stream
->relayd_stream_id
,
3603 stream
->next_net_seq_num
- 1);
3605 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3607 pthread_mutex_unlock(&stream
->lock
);
3611 pthread_mutex_unlock(&stream
->lock
);
3615 unsigned int is_data_inflight
= 0;
3617 /* Send init command for data pending. */
3618 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3619 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3620 relayd
->relayd_session_id
, &is_data_inflight
);
3621 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3623 goto data_not_pending
;
3625 if (is_data_inflight
) {
3631 * Finding _no_ node in the hash table and no inflight data means that the
3632 * stream(s) have been removed thus data is guaranteed to be available for
3633 * analysis from the trace files.
3637 /* Data is available to be read by a viewer. */
3638 pthread_mutex_unlock(&consumer_data
.lock
);
3643 /* Data is still being extracted from buffers. */
3644 pthread_mutex_unlock(&consumer_data
.lock
);
3650 * Send a ret code status message to the sessiond daemon.
3652 * Return the sendmsg() return value.
3654 int consumer_send_status_msg(int sock
, int ret_code
)
3656 struct lttcomm_consumer_status_msg msg
;
3658 memset(&msg
, 0, sizeof(msg
));
3659 msg
.ret_code
= ret_code
;
3661 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3665 * Send a channel status message to the sessiond daemon.
3667 * Return the sendmsg() return value.
3669 int consumer_send_status_channel(int sock
,
3670 struct lttng_consumer_channel
*channel
)
3672 struct lttcomm_consumer_status_channel msg
;
3676 memset(&msg
, 0, sizeof(msg
));
3678 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3680 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3681 msg
.key
= channel
->key
;
3682 msg
.stream_count
= channel
->streams
.count
;
3685 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3688 unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos
,
3689 unsigned long produced_pos
, uint64_t nb_packets_per_stream
,
3690 uint64_t max_sb_size
)
3692 unsigned long start_pos
;
3694 if (!nb_packets_per_stream
) {
3695 return consumed_pos
; /* Grab everything */
3697 start_pos
= produced_pos
- offset_align_floor(produced_pos
, max_sb_size
);
3698 start_pos
-= max_sb_size
* nb_packets_per_stream
;
3699 if ((long) (start_pos
- consumed_pos
) < 0) {
3700 return consumed_pos
; /* Grab everything */