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 */
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
);
288 switch (consumer_data
.type
) {
289 case LTTNG_CONSUMER_KERNEL
:
291 case LTTNG_CONSUMER32_UST
:
292 case LTTNG_CONSUMER64_UST
:
293 lttng_ustconsumer_free_channel(channel
);
296 ERR("Unknown consumer_data type");
303 * RCU protected relayd socket pair free.
305 static void free_relayd_rcu(struct rcu_head
*head
)
307 struct lttng_ht_node_u64
*node
=
308 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
309 struct consumer_relayd_sock_pair
*relayd
=
310 caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
313 * Close all sockets. This is done in the call RCU since we don't want the
314 * socket fds to be reassigned thus potentially creating bad state of the
317 * We do not have to lock the control socket mutex here since at this stage
318 * there is no one referencing to this relayd object.
320 (void) relayd_close(&relayd
->control_sock
);
321 (void) relayd_close(&relayd
->data_sock
);
323 pthread_mutex_destroy(&relayd
->ctrl_sock_mutex
);
328 * Destroy and free relayd socket pair object.
330 void consumer_destroy_relayd(struct consumer_relayd_sock_pair
*relayd
)
333 struct lttng_ht_iter iter
;
335 if (relayd
== NULL
) {
339 DBG("Consumer destroy and close relayd socket pair");
341 iter
.iter
.node
= &relayd
->node
.node
;
342 ret
= lttng_ht_del(consumer_data
.relayd_ht
, &iter
);
344 /* We assume the relayd is being or is destroyed */
348 /* RCU free() call */
349 call_rcu(&relayd
->node
.head
, free_relayd_rcu
);
353 * Remove a channel from the global list protected by a mutex. This function is
354 * also responsible for freeing its data structures.
356 void consumer_del_channel(struct lttng_consumer_channel
*channel
)
359 struct lttng_ht_iter iter
;
361 DBG("Consumer delete channel key %" PRIu64
, channel
->key
);
363 pthread_mutex_lock(&consumer_data
.lock
);
364 pthread_mutex_lock(&channel
->lock
);
366 /* Destroy streams that might have been left in the stream list. */
367 clean_channel_stream_list(channel
);
369 if (channel
->live_timer_enabled
== 1) {
370 consumer_timer_live_stop(channel
);
373 switch (consumer_data
.type
) {
374 case LTTNG_CONSUMER_KERNEL
:
376 case LTTNG_CONSUMER32_UST
:
377 case LTTNG_CONSUMER64_UST
:
378 lttng_ustconsumer_del_channel(channel
);
381 ERR("Unknown consumer_data type");
387 iter
.iter
.node
= &channel
->node
.node
;
388 ret
= lttng_ht_del(consumer_data
.channel_ht
, &iter
);
392 call_rcu(&channel
->node
.head
, free_channel_rcu
);
394 pthread_mutex_unlock(&channel
->lock
);
395 pthread_mutex_unlock(&consumer_data
.lock
);
399 * Iterate over the relayd hash table and destroy each element. Finally,
400 * destroy the whole hash table.
402 static void cleanup_relayd_ht(void)
404 struct lttng_ht_iter iter
;
405 struct consumer_relayd_sock_pair
*relayd
;
409 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
411 consumer_destroy_relayd(relayd
);
416 lttng_ht_destroy(consumer_data
.relayd_ht
);
420 * Update the end point status of all streams having the given network sequence
421 * index (relayd index).
423 * It's atomically set without having the stream mutex locked which is fine
424 * because we handle the write/read race with a pipe wakeup for each thread.
426 static void update_endpoint_status_by_netidx(uint64_t net_seq_idx
,
427 enum consumer_endpoint_status status
)
429 struct lttng_ht_iter iter
;
430 struct lttng_consumer_stream
*stream
;
432 DBG("Consumer set delete flag on stream by idx %" PRIu64
, net_seq_idx
);
436 /* Let's begin with metadata */
437 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
438 if (stream
->net_seq_idx
== net_seq_idx
) {
439 uatomic_set(&stream
->endpoint_status
, status
);
440 DBG("Delete flag set to metadata stream %d", stream
->wait_fd
);
444 /* Follow up by the data streams */
445 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
446 if (stream
->net_seq_idx
== net_seq_idx
) {
447 uatomic_set(&stream
->endpoint_status
, status
);
448 DBG("Delete flag set to data stream %d", stream
->wait_fd
);
455 * Cleanup a relayd object by flagging every associated streams for deletion,
456 * destroying the object meaning removing it from the relayd hash table,
457 * closing the sockets and freeing the memory in a RCU call.
459 * If a local data context is available, notify the threads that the streams'
460 * state have changed.
462 static void cleanup_relayd(struct consumer_relayd_sock_pair
*relayd
,
463 struct lttng_consumer_local_data
*ctx
)
469 DBG("Cleaning up relayd sockets");
471 /* Save the net sequence index before destroying the object */
472 netidx
= relayd
->net_seq_idx
;
475 * Delete the relayd from the relayd hash table, close the sockets and free
476 * the object in a RCU call.
478 consumer_destroy_relayd(relayd
);
480 /* Set inactive endpoint to all streams */
481 update_endpoint_status_by_netidx(netidx
, CONSUMER_ENDPOINT_INACTIVE
);
484 * With a local data context, notify the threads that the streams' state
485 * have changed. The write() action on the pipe acts as an "implicit"
486 * memory barrier ordering the updates of the end point status from the
487 * read of this status which happens AFTER receiving this notify.
490 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
491 notify_thread_lttng_pipe(ctx
->consumer_metadata_pipe
);
496 * Flag a relayd socket pair for destruction. Destroy it if the refcount
499 * RCU read side lock MUST be aquired before calling this function.
501 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair
*relayd
)
505 /* Set destroy flag for this object */
506 uatomic_set(&relayd
->destroy_flag
, 1);
508 /* Destroy the relayd if refcount is 0 */
509 if (uatomic_read(&relayd
->refcount
) == 0) {
510 consumer_destroy_relayd(relayd
);
515 * Completly destroy stream from every visiable data structure and the given
518 * One this call returns, the stream object is not longer usable nor visible.
520 void consumer_del_stream(struct lttng_consumer_stream
*stream
,
523 consumer_stream_destroy(stream
, ht
);
527 * XXX naming of del vs destroy is all mixed up.
529 void consumer_del_stream_for_data(struct lttng_consumer_stream
*stream
)
531 consumer_stream_destroy(stream
, data_ht
);
534 void consumer_del_stream_for_metadata(struct lttng_consumer_stream
*stream
)
536 consumer_stream_destroy(stream
, metadata_ht
);
539 struct lttng_consumer_stream
*consumer_allocate_stream(uint64_t channel_key
,
541 enum lttng_consumer_stream_state state
,
542 const char *channel_name
,
549 enum consumer_channel_type type
,
550 unsigned int monitor
)
553 struct lttng_consumer_stream
*stream
;
555 stream
= zmalloc(sizeof(*stream
));
556 if (stream
== NULL
) {
557 PERROR("malloc struct lttng_consumer_stream");
564 stream
->key
= stream_key
;
566 stream
->out_fd_offset
= 0;
567 stream
->output_written
= 0;
568 stream
->state
= state
;
571 stream
->net_seq_idx
= relayd_id
;
572 stream
->session_id
= session_id
;
573 stream
->monitor
= monitor
;
574 stream
->endpoint_status
= CONSUMER_ENDPOINT_ACTIVE
;
575 stream
->index_file
= NULL
;
576 stream
->last_sequence_number
= -1ULL;
577 pthread_mutex_init(&stream
->lock
, NULL
);
578 pthread_mutex_init(&stream
->metadata_timer_lock
, NULL
);
580 /* If channel is the metadata, flag this stream as metadata. */
581 if (type
== CONSUMER_CHANNEL_TYPE_METADATA
) {
582 stream
->metadata_flag
= 1;
583 /* Metadata is flat out. */
584 strncpy(stream
->name
, DEFAULT_METADATA_NAME
, sizeof(stream
->name
));
585 /* Live rendez-vous point. */
586 pthread_cond_init(&stream
->metadata_rdv
, NULL
);
587 pthread_mutex_init(&stream
->metadata_rdv_lock
, NULL
);
589 /* Format stream name to <channel_name>_<cpu_number> */
590 ret
= snprintf(stream
->name
, sizeof(stream
->name
), "%s_%d",
593 PERROR("snprintf stream name");
598 /* Key is always the wait_fd for streams. */
599 lttng_ht_node_init_u64(&stream
->node
, stream
->key
);
601 /* Init node per channel id key */
602 lttng_ht_node_init_u64(&stream
->node_channel_id
, channel_key
);
604 /* Init session id node with the stream session id */
605 lttng_ht_node_init_u64(&stream
->node_session_id
, stream
->session_id
);
607 DBG3("Allocated stream %s (key %" PRIu64
", chan_key %" PRIu64
608 " relayd_id %" PRIu64
", session_id %" PRIu64
,
609 stream
->name
, stream
->key
, channel_key
,
610 stream
->net_seq_idx
, stream
->session_id
);
626 * Add a stream to the global list protected by a mutex.
628 int consumer_add_data_stream(struct lttng_consumer_stream
*stream
)
630 struct lttng_ht
*ht
= data_ht
;
636 DBG3("Adding consumer stream %" PRIu64
, stream
->key
);
638 pthread_mutex_lock(&consumer_data
.lock
);
639 pthread_mutex_lock(&stream
->chan
->lock
);
640 pthread_mutex_lock(&stream
->chan
->timer_lock
);
641 pthread_mutex_lock(&stream
->lock
);
644 /* Steal stream identifier to avoid having streams with the same key */
645 steal_stream_key(stream
->key
, ht
);
647 lttng_ht_add_unique_u64(ht
, &stream
->node
);
649 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
650 &stream
->node_channel_id
);
653 * Add stream to the stream_list_ht of the consumer data. No need to steal
654 * the key since the HT does not use it and we allow to add redundant keys
657 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
660 * When nb_init_stream_left reaches 0, we don't need to trigger any action
661 * in terms of destroying the associated channel, because the action that
662 * causes the count to become 0 also causes a stream to be added. The
663 * channel deletion will thus be triggered by the following removal of this
666 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
667 /* Increment refcount before decrementing nb_init_stream_left */
669 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
672 /* Update consumer data once the node is inserted. */
673 consumer_data
.stream_count
++;
674 consumer_data
.need_update
= 1;
677 pthread_mutex_unlock(&stream
->lock
);
678 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
679 pthread_mutex_unlock(&stream
->chan
->lock
);
680 pthread_mutex_unlock(&consumer_data
.lock
);
685 void consumer_del_data_stream(struct lttng_consumer_stream
*stream
)
687 consumer_del_stream(stream
, data_ht
);
691 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
692 * be acquired before calling this.
694 static int add_relayd(struct consumer_relayd_sock_pair
*relayd
)
697 struct lttng_ht_node_u64
*node
;
698 struct lttng_ht_iter iter
;
702 lttng_ht_lookup(consumer_data
.relayd_ht
,
703 &relayd
->net_seq_idx
, &iter
);
704 node
= lttng_ht_iter_get_node_u64(&iter
);
708 lttng_ht_add_unique_u64(consumer_data
.relayd_ht
, &relayd
->node
);
715 * Allocate and return a consumer relayd socket.
717 static struct consumer_relayd_sock_pair
*consumer_allocate_relayd_sock_pair(
718 uint64_t net_seq_idx
)
720 struct consumer_relayd_sock_pair
*obj
= NULL
;
722 /* net sequence index of -1 is a failure */
723 if (net_seq_idx
== (uint64_t) -1ULL) {
727 obj
= zmalloc(sizeof(struct consumer_relayd_sock_pair
));
729 PERROR("zmalloc relayd sock");
733 obj
->net_seq_idx
= net_seq_idx
;
735 obj
->destroy_flag
= 0;
736 obj
->control_sock
.sock
.fd
= -1;
737 obj
->data_sock
.sock
.fd
= -1;
738 lttng_ht_node_init_u64(&obj
->node
, obj
->net_seq_idx
);
739 pthread_mutex_init(&obj
->ctrl_sock_mutex
, NULL
);
746 * Find a relayd socket pair in the global consumer data.
748 * Return the object if found else NULL.
749 * RCU read-side lock must be held across this call and while using the
752 struct consumer_relayd_sock_pair
*consumer_find_relayd(uint64_t key
)
754 struct lttng_ht_iter iter
;
755 struct lttng_ht_node_u64
*node
;
756 struct consumer_relayd_sock_pair
*relayd
= NULL
;
758 /* Negative keys are lookup failures */
759 if (key
== (uint64_t) -1ULL) {
763 lttng_ht_lookup(consumer_data
.relayd_ht
, &key
,
765 node
= lttng_ht_iter_get_node_u64(&iter
);
767 relayd
= caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
775 * Find a relayd and send the stream
777 * Returns 0 on success, < 0 on error
779 int consumer_send_relayd_stream(struct lttng_consumer_stream
*stream
,
783 struct consumer_relayd_sock_pair
*relayd
;
786 assert(stream
->net_seq_idx
!= -1ULL);
789 /* The stream is not metadata. Get relayd reference if exists. */
791 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
792 if (relayd
!= NULL
) {
793 /* Add stream on the relayd */
794 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
795 ret
= relayd_add_stream(&relayd
->control_sock
, stream
->name
,
796 path
, &stream
->relayd_stream_id
,
797 stream
->chan
->tracefile_size
, stream
->chan
->tracefile_count
);
798 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
803 uatomic_inc(&relayd
->refcount
);
804 stream
->sent_to_relayd
= 1;
806 ERR("Stream %" PRIu64
" relayd ID %" PRIu64
" unknown. Can't send it.",
807 stream
->key
, stream
->net_seq_idx
);
812 DBG("Stream %s with key %" PRIu64
" sent to relayd id %" PRIu64
,
813 stream
->name
, stream
->key
, stream
->net_seq_idx
);
821 * Find a relayd and send the streams sent message
823 * Returns 0 on success, < 0 on error
825 int consumer_send_relayd_streams_sent(uint64_t net_seq_idx
)
828 struct consumer_relayd_sock_pair
*relayd
;
830 assert(net_seq_idx
!= -1ULL);
832 /* The stream is not metadata. Get relayd reference if exists. */
834 relayd
= consumer_find_relayd(net_seq_idx
);
835 if (relayd
!= NULL
) {
836 /* Add stream on the relayd */
837 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
838 ret
= relayd_streams_sent(&relayd
->control_sock
);
839 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
844 ERR("Relayd ID %" PRIu64
" unknown. Can't send streams_sent.",
851 DBG("All streams sent relayd id %" PRIu64
, net_seq_idx
);
859 * Find a relayd and close the stream
861 void close_relayd_stream(struct lttng_consumer_stream
*stream
)
863 struct consumer_relayd_sock_pair
*relayd
;
865 /* The stream is not metadata. Get relayd reference if exists. */
867 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
869 consumer_stream_relayd_close(stream
, relayd
);
875 * Handle stream for relayd transmission if the stream applies for network
876 * streaming where the net sequence index is set.
878 * Return destination file descriptor or negative value on error.
880 static int write_relayd_stream_header(struct lttng_consumer_stream
*stream
,
881 size_t data_size
, unsigned long padding
,
882 struct consumer_relayd_sock_pair
*relayd
)
885 struct lttcomm_relayd_data_hdr data_hdr
;
891 /* Reset data header */
892 memset(&data_hdr
, 0, sizeof(data_hdr
));
894 if (stream
->metadata_flag
) {
895 /* Caller MUST acquire the relayd control socket lock */
896 ret
= relayd_send_metadata(&relayd
->control_sock
, data_size
);
901 /* Metadata are always sent on the control socket. */
902 outfd
= relayd
->control_sock
.sock
.fd
;
904 /* Set header with stream information */
905 data_hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
906 data_hdr
.data_size
= htobe32(data_size
);
907 data_hdr
.padding_size
= htobe32(padding
);
909 * Note that net_seq_num below is assigned with the *current* value of
910 * next_net_seq_num and only after that the next_net_seq_num will be
911 * increment. This is why when issuing a command on the relayd using
912 * this next value, 1 should always be substracted in order to compare
913 * the last seen sequence number on the relayd side to the last sent.
915 data_hdr
.net_seq_num
= htobe64(stream
->next_net_seq_num
);
916 /* Other fields are zeroed previously */
918 ret
= relayd_send_data_hdr(&relayd
->data_sock
, &data_hdr
,
924 ++stream
->next_net_seq_num
;
926 /* Set to go on data socket */
927 outfd
= relayd
->data_sock
.sock
.fd
;
935 * Allocate and return a new lttng_consumer_channel object using the given key
936 * to initialize the hash table node.
938 * On error, return NULL.
940 struct lttng_consumer_channel
*consumer_allocate_channel(uint64_t key
,
942 const char *pathname
,
947 enum lttng_event_output output
,
948 uint64_t tracefile_size
,
949 uint64_t tracefile_count
,
950 uint64_t session_id_per_pid
,
951 unsigned int monitor
,
952 unsigned int live_timer_interval
,
953 const char *root_shm_path
,
954 const char *shm_path
)
956 struct lttng_consumer_channel
*channel
;
958 channel
= zmalloc(sizeof(*channel
));
959 if (channel
== NULL
) {
960 PERROR("malloc struct lttng_consumer_channel");
965 channel
->refcount
= 0;
966 channel
->session_id
= session_id
;
967 channel
->session_id_per_pid
= session_id_per_pid
;
970 channel
->relayd_id
= relayd_id
;
971 channel
->tracefile_size
= tracefile_size
;
972 channel
->tracefile_count
= tracefile_count
;
973 channel
->monitor
= monitor
;
974 channel
->live_timer_interval
= live_timer_interval
;
975 pthread_mutex_init(&channel
->lock
, NULL
);
976 pthread_mutex_init(&channel
->timer_lock
, NULL
);
979 case LTTNG_EVENT_SPLICE
:
980 channel
->output
= CONSUMER_CHANNEL_SPLICE
;
982 case LTTNG_EVENT_MMAP
:
983 channel
->output
= CONSUMER_CHANNEL_MMAP
;
993 * In monitor mode, the streams associated with the channel will be put in
994 * a special list ONLY owned by this channel. So, the refcount is set to 1
995 * here meaning that the channel itself has streams that are referenced.
997 * On a channel deletion, once the channel is no longer visible, the
998 * refcount is decremented and checked for a zero value to delete it. With
999 * streams in no monitor mode, it will now be safe to destroy the channel.
1001 if (!channel
->monitor
) {
1002 channel
->refcount
= 1;
1005 strncpy(channel
->pathname
, pathname
, sizeof(channel
->pathname
));
1006 channel
->pathname
[sizeof(channel
->pathname
) - 1] = '\0';
1008 strncpy(channel
->name
, name
, sizeof(channel
->name
));
1009 channel
->name
[sizeof(channel
->name
) - 1] = '\0';
1011 if (root_shm_path
) {
1012 strncpy(channel
->root_shm_path
, root_shm_path
, sizeof(channel
->root_shm_path
));
1013 channel
->root_shm_path
[sizeof(channel
->root_shm_path
) - 1] = '\0';
1016 strncpy(channel
->shm_path
, shm_path
, sizeof(channel
->shm_path
));
1017 channel
->shm_path
[sizeof(channel
->shm_path
) - 1] = '\0';
1020 lttng_ht_node_init_u64(&channel
->node
, channel
->key
);
1022 channel
->wait_fd
= -1;
1024 CDS_INIT_LIST_HEAD(&channel
->streams
.head
);
1026 DBG("Allocated channel (key %" PRIu64
")", channel
->key
);
1033 * Add a channel to the global list protected by a mutex.
1035 * Always return 0 indicating success.
1037 int consumer_add_channel(struct lttng_consumer_channel
*channel
,
1038 struct lttng_consumer_local_data
*ctx
)
1040 pthread_mutex_lock(&consumer_data
.lock
);
1041 pthread_mutex_lock(&channel
->lock
);
1042 pthread_mutex_lock(&channel
->timer_lock
);
1045 * This gives us a guarantee that the channel we are about to add to the
1046 * channel hash table will be unique. See this function comment on the why
1047 * we need to steel the channel key at this stage.
1049 steal_channel_key(channel
->key
);
1052 lttng_ht_add_unique_u64(consumer_data
.channel_ht
, &channel
->node
);
1055 pthread_mutex_unlock(&channel
->timer_lock
);
1056 pthread_mutex_unlock(&channel
->lock
);
1057 pthread_mutex_unlock(&consumer_data
.lock
);
1059 if (channel
->wait_fd
!= -1 && channel
->type
== CONSUMER_CHANNEL_TYPE_DATA
) {
1060 notify_channel_pipe(ctx
, channel
, -1, CONSUMER_CHANNEL_ADD
);
1067 * Allocate the pollfd structure and the local view of the out fds to avoid
1068 * doing a lookup in the linked list and concurrency issues when writing is
1069 * needed. Called with consumer_data.lock held.
1071 * Returns the number of fds in the structures.
1073 static int update_poll_array(struct lttng_consumer_local_data
*ctx
,
1074 struct pollfd
**pollfd
, struct lttng_consumer_stream
**local_stream
,
1075 struct lttng_ht
*ht
, int *nb_inactive_fd
)
1078 struct lttng_ht_iter iter
;
1079 struct lttng_consumer_stream
*stream
;
1084 assert(local_stream
);
1086 DBG("Updating poll fd array");
1087 *nb_inactive_fd
= 0;
1089 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1091 * Only active streams with an active end point can be added to the
1092 * poll set and local stream storage of the thread.
1094 * There is a potential race here for endpoint_status to be updated
1095 * just after the check. However, this is OK since the stream(s) will
1096 * be deleted once the thread is notified that the end point state has
1097 * changed where this function will be called back again.
1099 * We track the number of inactive FDs because they still need to be
1100 * closed by the polling thread after a wakeup on the data_pipe or
1103 if (stream
->state
!= LTTNG_CONSUMER_ACTIVE_STREAM
||
1104 stream
->endpoint_status
== CONSUMER_ENDPOINT_INACTIVE
) {
1105 (*nb_inactive_fd
)++;
1109 * This clobbers way too much the debug output. Uncomment that if you
1110 * need it for debugging purposes.
1112 * DBG("Active FD %d", stream->wait_fd);
1114 (*pollfd
)[i
].fd
= stream
->wait_fd
;
1115 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1116 local_stream
[i
] = stream
;
1122 * Insert the consumer_data_pipe at the end of the array and don't
1123 * increment i so nb_fd is the number of real FD.
1125 (*pollfd
)[i
].fd
= lttng_pipe_get_readfd(ctx
->consumer_data_pipe
);
1126 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1128 (*pollfd
)[i
+ 1].fd
= lttng_pipe_get_readfd(ctx
->consumer_wakeup_pipe
);
1129 (*pollfd
)[i
+ 1].events
= POLLIN
| POLLPRI
;
1134 * Poll on the should_quit pipe and the command socket return -1 on
1135 * error, 1 if should exit, 0 if data is available on the command socket
1137 int lttng_consumer_poll_socket(struct pollfd
*consumer_sockpoll
)
1142 num_rdy
= poll(consumer_sockpoll
, 2, -1);
1143 if (num_rdy
== -1) {
1145 * Restart interrupted system call.
1147 if (errno
== EINTR
) {
1150 PERROR("Poll error");
1153 if (consumer_sockpoll
[0].revents
& (POLLIN
| POLLPRI
)) {
1154 DBG("consumer_should_quit wake up");
1161 * Set the error socket.
1163 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data
*ctx
,
1166 ctx
->consumer_error_socket
= sock
;
1170 * Set the command socket path.
1172 void lttng_consumer_set_command_sock_path(
1173 struct lttng_consumer_local_data
*ctx
, char *sock
)
1175 ctx
->consumer_command_sock_path
= sock
;
1179 * Send return code to the session daemon.
1180 * If the socket is not defined, we return 0, it is not a fatal error
1182 int lttng_consumer_send_error(struct lttng_consumer_local_data
*ctx
, int cmd
)
1184 if (ctx
->consumer_error_socket
> 0) {
1185 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
1186 sizeof(enum lttcomm_sessiond_command
));
1193 * Close all the tracefiles and stream fds and MUST be called when all
1194 * instances are destroyed i.e. when all threads were joined and are ended.
1196 void lttng_consumer_cleanup(void)
1198 struct lttng_ht_iter iter
;
1199 struct lttng_consumer_channel
*channel
;
1203 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
, channel
,
1205 consumer_del_channel(channel
);
1210 lttng_ht_destroy(consumer_data
.channel_ht
);
1212 cleanup_relayd_ht();
1214 lttng_ht_destroy(consumer_data
.stream_per_chan_id_ht
);
1217 * This HT contains streams that are freed by either the metadata thread or
1218 * the data thread so we do *nothing* on the hash table and simply destroy
1221 lttng_ht_destroy(consumer_data
.stream_list_ht
);
1225 * Called from signal handler.
1227 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
1232 ret
= lttng_write(ctx
->consumer_should_quit
[1], "4", 1);
1234 PERROR("write consumer quit");
1237 DBG("Consumer flag that it should quit");
1242 * Flush pending writes to trace output disk file.
1245 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1249 int outfd
= stream
->out_fd
;
1252 * This does a blocking write-and-wait on any page that belongs to the
1253 * subbuffer prior to the one we just wrote.
1254 * Don't care about error values, as these are just hints and ways to
1255 * limit the amount of page cache used.
1257 if (orig_offset
< stream
->max_sb_size
) {
1260 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1261 stream
->max_sb_size
,
1262 SYNC_FILE_RANGE_WAIT_BEFORE
1263 | SYNC_FILE_RANGE_WRITE
1264 | SYNC_FILE_RANGE_WAIT_AFTER
);
1266 * Give hints to the kernel about how we access the file:
1267 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1270 * We need to call fadvise again after the file grows because the
1271 * kernel does not seem to apply fadvise to non-existing parts of the
1274 * Call fadvise _after_ having waited for the page writeback to
1275 * complete because the dirty page writeback semantic is not well
1276 * defined. So it can be expected to lead to lower throughput in
1279 ret
= posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1280 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1281 if (ret
&& ret
!= -ENOSYS
) {
1283 PERROR("posix_fadvise on fd %i", outfd
);
1288 * Initialise the necessary environnement :
1289 * - create a new context
1290 * - create the poll_pipe
1291 * - create the should_quit pipe (for signal handler)
1292 * - create the thread pipe (for splice)
1294 * Takes a function pointer as argument, this function is called when data is
1295 * available on a buffer. This function is responsible to do the
1296 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1297 * buffer configuration and then kernctl_put_next_subbuf at the end.
1299 * Returns a pointer to the new context or NULL on error.
1301 struct lttng_consumer_local_data
*lttng_consumer_create(
1302 enum lttng_consumer_type type
,
1303 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1304 struct lttng_consumer_local_data
*ctx
),
1305 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1306 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1307 int (*update_stream
)(uint64_t stream_key
, uint32_t state
))
1310 struct lttng_consumer_local_data
*ctx
;
1312 assert(consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1313 consumer_data
.type
== type
);
1314 consumer_data
.type
= type
;
1316 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
1318 PERROR("allocating context");
1322 ctx
->consumer_error_socket
= -1;
1323 ctx
->consumer_metadata_socket
= -1;
1324 pthread_mutex_init(&ctx
->metadata_socket_lock
, NULL
);
1325 /* assign the callbacks */
1326 ctx
->on_buffer_ready
= buffer_ready
;
1327 ctx
->on_recv_channel
= recv_channel
;
1328 ctx
->on_recv_stream
= recv_stream
;
1329 ctx
->on_update_stream
= update_stream
;
1331 ctx
->consumer_data_pipe
= lttng_pipe_open(0);
1332 if (!ctx
->consumer_data_pipe
) {
1333 goto error_poll_pipe
;
1336 ctx
->consumer_wakeup_pipe
= lttng_pipe_open(0);
1337 if (!ctx
->consumer_wakeup_pipe
) {
1338 goto error_wakeup_pipe
;
1341 ret
= pipe(ctx
->consumer_should_quit
);
1343 PERROR("Error creating recv pipe");
1344 goto error_quit_pipe
;
1347 ret
= pipe(ctx
->consumer_channel_pipe
);
1349 PERROR("Error creating channel pipe");
1350 goto error_channel_pipe
;
1353 ctx
->consumer_metadata_pipe
= lttng_pipe_open(0);
1354 if (!ctx
->consumer_metadata_pipe
) {
1355 goto error_metadata_pipe
;
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_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 /* Number of FDs with CONSUMER_ENDPOINT_INACTIVE but still open. */
2460 int nb_inactive_fd
= 0;
2461 struct lttng_consumer_local_data
*ctx
= data
;
2464 rcu_register_thread();
2466 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2468 if (testpoint(consumerd_thread_data
)) {
2469 goto error_testpoint
;
2472 health_code_update();
2474 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
*));
2475 if (local_stream
== NULL
) {
2476 PERROR("local_stream malloc");
2481 health_code_update();
2487 * the fds set has been updated, we need to update our
2488 * local array as well
2490 pthread_mutex_lock(&consumer_data
.lock
);
2491 if (consumer_data
.need_update
) {
2496 local_stream
= NULL
;
2499 * Allocate for all fds +1 for the consumer_data_pipe and +1 for
2502 pollfd
= zmalloc((consumer_data
.stream_count
+ 2) * sizeof(struct pollfd
));
2503 if (pollfd
== NULL
) {
2504 PERROR("pollfd malloc");
2505 pthread_mutex_unlock(&consumer_data
.lock
);
2509 local_stream
= zmalloc((consumer_data
.stream_count
+ 2) *
2510 sizeof(struct lttng_consumer_stream
*));
2511 if (local_stream
== NULL
) {
2512 PERROR("local_stream malloc");
2513 pthread_mutex_unlock(&consumer_data
.lock
);
2516 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2517 data_ht
, &nb_inactive_fd
);
2519 ERR("Error in allocating pollfd or local_outfds");
2520 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2521 pthread_mutex_unlock(&consumer_data
.lock
);
2525 consumer_data
.need_update
= 0;
2527 pthread_mutex_unlock(&consumer_data
.lock
);
2529 /* No FDs and consumer_quit, consumer_cleanup the thread */
2530 if (nb_fd
== 0 && consumer_quit
== 1 && nb_inactive_fd
== 0) {
2531 err
= 0; /* All is OK */
2534 /* poll on the array of fds */
2536 DBG("polling on %d fd", nb_fd
+ 2);
2537 health_poll_entry();
2538 num_rdy
= poll(pollfd
, nb_fd
+ 2, -1);
2540 DBG("poll num_rdy : %d", num_rdy
);
2541 if (num_rdy
== -1) {
2543 * Restart interrupted system call.
2545 if (errno
== EINTR
) {
2548 PERROR("Poll error");
2549 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2551 } else if (num_rdy
== 0) {
2552 DBG("Polling thread timed out");
2557 * If the consumer_data_pipe triggered poll go directly to the
2558 * beginning of the loop to update the array. We want to prioritize
2559 * array update over low-priority reads.
2561 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2562 ssize_t pipe_readlen
;
2564 DBG("consumer_data_pipe wake up");
2565 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2566 &new_stream
, sizeof(new_stream
));
2567 if (pipe_readlen
< sizeof(new_stream
)) {
2568 PERROR("Consumer data pipe");
2569 /* Continue so we can at least handle the current stream(s). */
2574 * If the stream is NULL, just ignore it. It's also possible that
2575 * the sessiond poll thread changed the consumer_quit state and is
2576 * waking us up to test it.
2578 if (new_stream
== NULL
) {
2579 validate_endpoint_status_data_stream();
2583 /* Continue to update the local streams and handle prio ones */
2587 /* Handle wakeup pipe. */
2588 if (pollfd
[nb_fd
+ 1].revents
& (POLLIN
| POLLPRI
)) {
2590 ssize_t pipe_readlen
;
2592 pipe_readlen
= lttng_pipe_read(ctx
->consumer_wakeup_pipe
, &dummy
,
2594 if (pipe_readlen
< 0) {
2595 PERROR("Consumer data wakeup pipe");
2597 /* We've been awakened to handle stream(s). */
2598 ctx
->has_wakeup
= 0;
2601 /* Take care of high priority channels first. */
2602 for (i
= 0; i
< nb_fd
; i
++) {
2603 health_code_update();
2605 if (local_stream
[i
] == NULL
) {
2608 if (pollfd
[i
].revents
& POLLPRI
) {
2609 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2611 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2612 /* it's ok to have an unavailable sub-buffer */
2613 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2614 /* Clean the stream and free it. */
2615 consumer_del_stream(local_stream
[i
], data_ht
);
2616 local_stream
[i
] = NULL
;
2617 } else if (len
> 0) {
2618 local_stream
[i
]->data_read
= 1;
2624 * If we read high prio channel in this loop, try again
2625 * for more high prio data.
2631 /* Take care of low priority channels. */
2632 for (i
= 0; i
< nb_fd
; i
++) {
2633 health_code_update();
2635 if (local_stream
[i
] == NULL
) {
2638 if ((pollfd
[i
].revents
& POLLIN
) ||
2639 local_stream
[i
]->hangup_flush_done
||
2640 local_stream
[i
]->has_data
) {
2641 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2642 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2643 /* it's ok to have an unavailable sub-buffer */
2644 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2645 /* Clean the stream and free it. */
2646 consumer_del_stream(local_stream
[i
], data_ht
);
2647 local_stream
[i
] = NULL
;
2648 } else if (len
> 0) {
2649 local_stream
[i
]->data_read
= 1;
2654 /* Handle hangup and errors */
2655 for (i
= 0; i
< nb_fd
; i
++) {
2656 health_code_update();
2658 if (local_stream
[i
] == NULL
) {
2661 if (!local_stream
[i
]->hangup_flush_done
2662 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2663 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2664 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2665 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2667 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2668 /* Attempt read again, for the data we just flushed. */
2669 local_stream
[i
]->data_read
= 1;
2672 * If the poll flag is HUP/ERR/NVAL and we have
2673 * read no data in this pass, we can remove the
2674 * stream from its hash table.
2676 if ((pollfd
[i
].revents
& POLLHUP
)) {
2677 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2678 if (!local_stream
[i
]->data_read
) {
2679 consumer_del_stream(local_stream
[i
], data_ht
);
2680 local_stream
[i
] = NULL
;
2683 } else if (pollfd
[i
].revents
& POLLERR
) {
2684 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2685 if (!local_stream
[i
]->data_read
) {
2686 consumer_del_stream(local_stream
[i
], data_ht
);
2687 local_stream
[i
] = NULL
;
2690 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2691 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2692 if (!local_stream
[i
]->data_read
) {
2693 consumer_del_stream(local_stream
[i
], data_ht
);
2694 local_stream
[i
] = NULL
;
2698 if (local_stream
[i
] != NULL
) {
2699 local_stream
[i
]->data_read
= 0;
2706 DBG("polling thread exiting");
2711 * Close the write side of the pipe so epoll_wait() in
2712 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2713 * read side of the pipe. If we close them both, epoll_wait strangely does
2714 * not return and could create a endless wait period if the pipe is the
2715 * only tracked fd in the poll set. The thread will take care of closing
2718 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2723 ERR("Health error occurred in %s", __func__
);
2725 health_unregister(health_consumerd
);
2727 rcu_unregister_thread();
2732 * Close wake-up end of each stream belonging to the channel. This will
2733 * allow the poll() on the stream read-side to detect when the
2734 * write-side (application) finally closes them.
2737 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2739 struct lttng_ht
*ht
;
2740 struct lttng_consumer_stream
*stream
;
2741 struct lttng_ht_iter iter
;
2743 ht
= consumer_data
.stream_per_chan_id_ht
;
2746 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2747 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2748 ht
->match_fct
, &channel
->key
,
2749 &iter
.iter
, stream
, node_channel_id
.node
) {
2751 * Protect against teardown with mutex.
2753 pthread_mutex_lock(&stream
->lock
);
2754 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2757 switch (consumer_data
.type
) {
2758 case LTTNG_CONSUMER_KERNEL
:
2760 case LTTNG_CONSUMER32_UST
:
2761 case LTTNG_CONSUMER64_UST
:
2762 if (stream
->metadata_flag
) {
2763 /* Safe and protected by the stream lock. */
2764 lttng_ustconsumer_close_metadata(stream
->chan
);
2767 * Note: a mutex is taken internally within
2768 * liblttng-ust-ctl to protect timer wakeup_fd
2769 * use from concurrent close.
2771 lttng_ustconsumer_close_stream_wakeup(stream
);
2775 ERR("Unknown consumer_data type");
2779 pthread_mutex_unlock(&stream
->lock
);
2784 static void destroy_channel_ht(struct lttng_ht
*ht
)
2786 struct lttng_ht_iter iter
;
2787 struct lttng_consumer_channel
*channel
;
2795 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2796 ret
= lttng_ht_del(ht
, &iter
);
2801 lttng_ht_destroy(ht
);
2805 * This thread polls the channel fds to detect when they are being
2806 * closed. It closes all related streams if the channel is detected as
2807 * closed. It is currently only used as a shim layer for UST because the
2808 * consumerd needs to keep the per-stream wakeup end of pipes open for
2811 void *consumer_thread_channel_poll(void *data
)
2813 int ret
, i
, pollfd
, err
= -1;
2814 uint32_t revents
, nb_fd
;
2815 struct lttng_consumer_channel
*chan
= NULL
;
2816 struct lttng_ht_iter iter
;
2817 struct lttng_ht_node_u64
*node
;
2818 struct lttng_poll_event events
;
2819 struct lttng_consumer_local_data
*ctx
= data
;
2820 struct lttng_ht
*channel_ht
;
2822 rcu_register_thread();
2824 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
2826 if (testpoint(consumerd_thread_channel
)) {
2827 goto error_testpoint
;
2830 health_code_update();
2832 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2834 /* ENOMEM at this point. Better to bail out. */
2838 DBG("Thread channel poll started");
2840 /* Size is set to 1 for the consumer_channel pipe */
2841 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2843 ERR("Poll set creation failed");
2847 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2853 DBG("Channel main loop started");
2857 health_code_update();
2858 DBG("Channel poll wait");
2859 health_poll_entry();
2860 ret
= lttng_poll_wait(&events
, -1);
2861 DBG("Channel poll return from wait with %d fd(s)",
2862 LTTNG_POLL_GETNB(&events
));
2864 DBG("Channel event caught in thread");
2866 if (errno
== EINTR
) {
2867 ERR("Poll EINTR caught");
2870 if (LTTNG_POLL_GETNB(&events
) == 0) {
2871 err
= 0; /* All is OK */
2878 /* From here, the event is a channel wait fd */
2879 for (i
= 0; i
< nb_fd
; i
++) {
2880 health_code_update();
2882 revents
= LTTNG_POLL_GETEV(&events
, i
);
2883 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2886 /* No activity for this FD (poll implementation). */
2890 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2891 if (revents
& LPOLLIN
) {
2892 enum consumer_channel_action action
;
2895 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
2898 ERR("Error reading channel pipe");
2900 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2905 case CONSUMER_CHANNEL_ADD
:
2906 DBG("Adding channel %d to poll set",
2909 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
2912 lttng_ht_add_unique_u64(channel_ht
,
2913 &chan
->wait_fd_node
);
2915 /* Add channel to the global poll events list */
2916 lttng_poll_add(&events
, chan
->wait_fd
,
2917 LPOLLERR
| LPOLLHUP
);
2919 case CONSUMER_CHANNEL_DEL
:
2922 * This command should never be called if the channel
2923 * has streams monitored by either the data or metadata
2924 * thread. The consumer only notify this thread with a
2925 * channel del. command if it receives a destroy
2926 * channel command from the session daemon that send it
2927 * if a command prior to the GET_CHANNEL failed.
2931 chan
= consumer_find_channel(key
);
2934 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
2937 lttng_poll_del(&events
, chan
->wait_fd
);
2938 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
2939 ret
= lttng_ht_del(channel_ht
, &iter
);
2942 switch (consumer_data
.type
) {
2943 case LTTNG_CONSUMER_KERNEL
:
2945 case LTTNG_CONSUMER32_UST
:
2946 case LTTNG_CONSUMER64_UST
:
2947 health_code_update();
2948 /* Destroy streams that might have been left in the stream list. */
2949 clean_channel_stream_list(chan
);
2952 ERR("Unknown consumer_data type");
2957 * Release our own refcount. Force channel deletion even if
2958 * streams were not initialized.
2960 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
2961 consumer_del_channel(chan
);
2966 case CONSUMER_CHANNEL_QUIT
:
2968 * Remove the pipe from the poll set and continue the loop
2969 * since their might be data to consume.
2971 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2974 ERR("Unknown action");
2977 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2978 DBG("Channel thread pipe hung up");
2980 * Remove the pipe from the poll set and continue the loop
2981 * since their might be data to consume.
2983 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2986 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2990 /* Handle other stream */
2996 uint64_t tmp_id
= (uint64_t) pollfd
;
2998 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
3000 node
= lttng_ht_iter_get_node_u64(&iter
);
3003 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
3006 /* Check for error event */
3007 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
3008 DBG("Channel fd %d is hup|err.", pollfd
);
3010 lttng_poll_del(&events
, chan
->wait_fd
);
3011 ret
= lttng_ht_del(channel_ht
, &iter
);
3015 * This will close the wait fd for each stream associated to
3016 * this channel AND monitored by the data/metadata thread thus
3017 * will be clean by the right thread.
3019 consumer_close_channel_streams(chan
);
3021 /* Release our own refcount */
3022 if (!uatomic_sub_return(&chan
->refcount
, 1)
3023 && !uatomic_read(&chan
->nb_init_stream_left
)) {
3024 consumer_del_channel(chan
);
3027 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3032 /* Release RCU lock for the channel looked up */
3040 lttng_poll_clean(&events
);
3042 destroy_channel_ht(channel_ht
);
3045 DBG("Channel poll thread exiting");
3048 ERR("Health error occurred in %s", __func__
);
3050 health_unregister(health_consumerd
);
3051 rcu_unregister_thread();
3055 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
3056 struct pollfd
*sockpoll
, int client_socket
)
3063 ret
= lttng_consumer_poll_socket(sockpoll
);
3067 DBG("Metadata connection on client_socket");
3069 /* Blocking call, waiting for transmission */
3070 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
3071 if (ctx
->consumer_metadata_socket
< 0) {
3072 WARN("On accept metadata");
3083 * This thread listens on the consumerd socket and receives the file
3084 * descriptors from the session daemon.
3086 void *consumer_thread_sessiond_poll(void *data
)
3088 int sock
= -1, client_socket
, ret
, err
= -1;
3090 * structure to poll for incoming data on communication socket avoids
3091 * making blocking sockets.
3093 struct pollfd consumer_sockpoll
[2];
3094 struct lttng_consumer_local_data
*ctx
= data
;
3096 rcu_register_thread();
3098 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
3100 if (testpoint(consumerd_thread_sessiond
)) {
3101 goto error_testpoint
;
3104 health_code_update();
3106 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
3107 unlink(ctx
->consumer_command_sock_path
);
3108 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
3109 if (client_socket
< 0) {
3110 ERR("Cannot create command socket");
3114 ret
= lttcomm_listen_unix_sock(client_socket
);
3119 DBG("Sending ready command to lttng-sessiond");
3120 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3121 /* return < 0 on error, but == 0 is not fatal */
3123 ERR("Error sending ready command to lttng-sessiond");
3127 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3128 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3129 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3130 consumer_sockpoll
[1].fd
= client_socket
;
3131 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3133 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3141 DBG("Connection on client_socket");
3143 /* Blocking call, waiting for transmission */
3144 sock
= lttcomm_accept_unix_sock(client_socket
);
3151 * Setup metadata socket which is the second socket connection on the
3152 * command unix socket.
3154 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3163 /* This socket is not useful anymore. */
3164 ret
= close(client_socket
);
3166 PERROR("close client_socket");
3170 /* update the polling structure to poll on the established socket */
3171 consumer_sockpoll
[1].fd
= sock
;
3172 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3175 health_code_update();
3177 health_poll_entry();
3178 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3187 DBG("Incoming command on sock");
3188 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3191 * This could simply be a session daemon quitting. Don't output
3194 DBG("Communication interrupted on command socket");
3198 if (consumer_quit
) {
3199 DBG("consumer_thread_receive_fds received quit from signal");
3200 err
= 0; /* All is OK */
3203 DBG("received command on sock");
3209 DBG("Consumer thread sessiond poll exiting");
3212 * Close metadata streams since the producer is the session daemon which
3215 * NOTE: for now, this only applies to the UST tracer.
3217 lttng_consumer_close_all_metadata();
3220 * when all fds have hung up, the polling thread
3226 * Notify the data poll thread to poll back again and test the
3227 * consumer_quit state that we just set so to quit gracefully.
3229 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3231 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3233 notify_health_quit_pipe(health_quit_pipe
);
3235 /* Cleaning up possibly open sockets. */
3239 PERROR("close sock sessiond poll");
3242 if (client_socket
>= 0) {
3243 ret
= close(client_socket
);
3245 PERROR("close client_socket sessiond poll");
3252 ERR("Health error occurred in %s", __func__
);
3254 health_unregister(health_consumerd
);
3256 rcu_unregister_thread();
3260 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3261 struct lttng_consumer_local_data
*ctx
)
3265 pthread_mutex_lock(&stream
->lock
);
3266 if (stream
->metadata_flag
) {
3267 pthread_mutex_lock(&stream
->metadata_rdv_lock
);
3270 switch (consumer_data
.type
) {
3271 case LTTNG_CONSUMER_KERNEL
:
3272 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
3274 case LTTNG_CONSUMER32_UST
:
3275 case LTTNG_CONSUMER64_UST
:
3276 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
3279 ERR("Unknown consumer_data type");
3285 if (stream
->metadata_flag
) {
3286 pthread_cond_broadcast(&stream
->metadata_rdv
);
3287 pthread_mutex_unlock(&stream
->metadata_rdv_lock
);
3289 pthread_mutex_unlock(&stream
->lock
);
3293 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3295 switch (consumer_data
.type
) {
3296 case LTTNG_CONSUMER_KERNEL
:
3297 return lttng_kconsumer_on_recv_stream(stream
);
3298 case LTTNG_CONSUMER32_UST
:
3299 case LTTNG_CONSUMER64_UST
:
3300 return lttng_ustconsumer_on_recv_stream(stream
);
3302 ERR("Unknown consumer_data type");
3309 * Allocate and set consumer data hash tables.
3311 int lttng_consumer_init(void)
3313 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3314 if (!consumer_data
.channel_ht
) {
3318 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3319 if (!consumer_data
.relayd_ht
) {
3323 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3324 if (!consumer_data
.stream_list_ht
) {
3328 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3329 if (!consumer_data
.stream_per_chan_id_ht
) {
3333 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3338 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3350 * Process the ADD_RELAYD command receive by a consumer.
3352 * This will create a relayd socket pair and add it to the relayd hash table.
3353 * The caller MUST acquire a RCU read side lock before calling it.
3355 int consumer_add_relayd_socket(uint64_t net_seq_idx
, int sock_type
,
3356 struct lttng_consumer_local_data
*ctx
, int sock
,
3357 struct pollfd
*consumer_sockpoll
,
3358 struct lttcomm_relayd_sock
*relayd_sock
, uint64_t sessiond_id
,
3359 uint64_t relayd_session_id
)
3361 int fd
= -1, ret
= -1, relayd_created
= 0;
3362 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3363 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3366 assert(relayd_sock
);
3368 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", net_seq_idx
);
3370 /* Get relayd reference if exists. */
3371 relayd
= consumer_find_relayd(net_seq_idx
);
3372 if (relayd
== NULL
) {
3373 assert(sock_type
== LTTNG_STREAM_CONTROL
);
3374 /* Not found. Allocate one. */
3375 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3376 if (relayd
== NULL
) {
3378 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3381 relayd
->sessiond_session_id
= sessiond_id
;
3386 * This code path MUST continue to the consumer send status message to
3387 * we can notify the session daemon and continue our work without
3388 * killing everything.
3392 * relayd key should never be found for control socket.
3394 assert(sock_type
!= LTTNG_STREAM_CONTROL
);
3397 /* First send a status message before receiving the fds. */
3398 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3400 /* Somehow, the session daemon is not responding anymore. */
3401 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3402 goto error_nosignal
;
3405 /* Poll on consumer socket. */
3406 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3408 /* Needing to exit in the middle of a command: error. */
3409 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3411 goto error_nosignal
;
3414 /* Get relayd socket from session daemon */
3415 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3416 if (ret
!= sizeof(fd
)) {
3418 fd
= -1; /* Just in case it gets set with an invalid value. */
3421 * Failing to receive FDs might indicate a major problem such as
3422 * reaching a fd limit during the receive where the kernel returns a
3423 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3424 * don't take any chances and stop everything.
3426 * XXX: Feature request #558 will fix that and avoid this possible
3427 * issue when reaching the fd limit.
3429 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3430 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3434 /* Copy socket information and received FD */
3435 switch (sock_type
) {
3436 case LTTNG_STREAM_CONTROL
:
3437 /* Copy received lttcomm socket */
3438 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3439 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3440 /* Handle create_sock error. */
3442 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3446 * Close the socket created internally by
3447 * lttcomm_create_sock, so we can replace it by the one
3448 * received from sessiond.
3450 if (close(relayd
->control_sock
.sock
.fd
)) {
3454 /* Assign new file descriptor */
3455 relayd
->control_sock
.sock
.fd
= fd
;
3456 fd
= -1; /* For error path */
3457 /* Assign version values. */
3458 relayd
->control_sock
.major
= relayd_sock
->major
;
3459 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3461 relayd
->relayd_session_id
= relayd_session_id
;
3464 case LTTNG_STREAM_DATA
:
3465 /* Copy received lttcomm socket */
3466 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3467 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3468 /* Handle create_sock error. */
3470 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3474 * Close the socket created internally by
3475 * lttcomm_create_sock, so we can replace it by the one
3476 * received from sessiond.
3478 if (close(relayd
->data_sock
.sock
.fd
)) {
3482 /* Assign new file descriptor */
3483 relayd
->data_sock
.sock
.fd
= fd
;
3484 fd
= -1; /* for eventual error paths */
3485 /* Assign version values. */
3486 relayd
->data_sock
.major
= relayd_sock
->major
;
3487 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3490 ERR("Unknown relayd socket type (%d)", sock_type
);
3492 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3496 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3497 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3498 relayd
->net_seq_idx
, fd
);
3500 /* We successfully added the socket. Send status back. */
3501 ret
= consumer_send_status_msg(sock
, ret_code
);
3503 /* Somehow, the session daemon is not responding anymore. */
3504 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3505 goto error_nosignal
;
3509 * Add relayd socket pair to consumer data hashtable. If object already
3510 * exists or on error, the function gracefully returns.
3518 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3519 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3523 /* Close received socket if valid. */
3526 PERROR("close received socket");
3530 if (relayd_created
) {
3538 * Try to lock the stream mutex.
3540 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
3542 static int stream_try_lock(struct lttng_consumer_stream
*stream
)
3549 * Try to lock the stream mutex. On failure, we know that the stream is
3550 * being used else where hence there is data still being extracted.
3552 ret
= pthread_mutex_trylock(&stream
->lock
);
3554 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
3566 * Search for a relayd associated to the session id and return the reference.
3568 * A rcu read side lock MUST be acquire before calling this function and locked
3569 * until the relayd object is no longer necessary.
3571 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3573 struct lttng_ht_iter iter
;
3574 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3576 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3577 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3580 * Check by sessiond id which is unique here where the relayd session
3581 * id might not be when having multiple relayd.
3583 if (relayd
->sessiond_session_id
== id
) {
3584 /* Found the relayd. There can be only one per id. */
3596 * Check if for a given session id there is still data needed to be extract
3599 * Return 1 if data is pending or else 0 meaning ready to be read.
3601 int consumer_data_pending(uint64_t id
)
3604 struct lttng_ht_iter iter
;
3605 struct lttng_ht
*ht
;
3606 struct lttng_consumer_stream
*stream
;
3607 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3608 int (*data_pending
)(struct lttng_consumer_stream
*);
3610 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3613 pthread_mutex_lock(&consumer_data
.lock
);
3615 switch (consumer_data
.type
) {
3616 case LTTNG_CONSUMER_KERNEL
:
3617 data_pending
= lttng_kconsumer_data_pending
;
3619 case LTTNG_CONSUMER32_UST
:
3620 case LTTNG_CONSUMER64_UST
:
3621 data_pending
= lttng_ustconsumer_data_pending
;
3624 ERR("Unknown consumer data type");
3628 /* Ease our life a bit */
3629 ht
= consumer_data
.stream_list_ht
;
3631 relayd
= find_relayd_by_session_id(id
);
3633 /* Send init command for data pending. */
3634 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3635 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3636 relayd
->relayd_session_id
);
3637 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3639 /* Communication error thus the relayd so no data pending. */
3640 goto data_not_pending
;
3644 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3645 ht
->hash_fct(&id
, lttng_ht_seed
),
3647 &iter
.iter
, stream
, node_session_id
.node
) {
3648 /* If this call fails, the stream is being used hence data pending. */
3649 ret
= stream_try_lock(stream
);
3655 * A removed node from the hash table indicates that the stream has
3656 * been deleted thus having a guarantee that the buffers are closed
3657 * on the consumer side. However, data can still be transmitted
3658 * over the network so don't skip the relayd check.
3660 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3662 /* Check the stream if there is data in the buffers. */
3663 ret
= data_pending(stream
);
3665 pthread_mutex_unlock(&stream
->lock
);
3672 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3673 if (stream
->metadata_flag
) {
3674 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3675 stream
->relayd_stream_id
);
3677 ret
= relayd_data_pending(&relayd
->control_sock
,
3678 stream
->relayd_stream_id
,
3679 stream
->next_net_seq_num
- 1);
3681 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3683 pthread_mutex_unlock(&stream
->lock
);
3687 pthread_mutex_unlock(&stream
->lock
);
3691 unsigned int is_data_inflight
= 0;
3693 /* Send init command for data pending. */
3694 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3695 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3696 relayd
->relayd_session_id
, &is_data_inflight
);
3697 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3699 goto data_not_pending
;
3701 if (is_data_inflight
) {
3707 * Finding _no_ node in the hash table and no inflight data means that the
3708 * stream(s) have been removed thus data is guaranteed to be available for
3709 * analysis from the trace files.
3713 /* Data is available to be read by a viewer. */
3714 pthread_mutex_unlock(&consumer_data
.lock
);
3719 /* Data is still being extracted from buffers. */
3720 pthread_mutex_unlock(&consumer_data
.lock
);
3726 * Send a ret code status message to the sessiond daemon.
3728 * Return the sendmsg() return value.
3730 int consumer_send_status_msg(int sock
, int ret_code
)
3732 struct lttcomm_consumer_status_msg msg
;
3734 memset(&msg
, 0, sizeof(msg
));
3735 msg
.ret_code
= ret_code
;
3737 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3741 * Send a channel status message to the sessiond daemon.
3743 * Return the sendmsg() return value.
3745 int consumer_send_status_channel(int sock
,
3746 struct lttng_consumer_channel
*channel
)
3748 struct lttcomm_consumer_status_channel msg
;
3752 memset(&msg
, 0, sizeof(msg
));
3754 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3756 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3757 msg
.key
= channel
->key
;
3758 msg
.stream_count
= channel
->streams
.count
;
3761 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3764 unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos
,
3765 unsigned long produced_pos
, uint64_t nb_packets_per_stream
,
3766 uint64_t max_sb_size
)
3768 unsigned long start_pos
;
3770 if (!nb_packets_per_stream
) {
3771 return consumed_pos
; /* Grab everything */
3773 start_pos
= produced_pos
- offset_align_floor(produced_pos
, max_sb_size
);
3774 start_pos
-= max_sb_size
* nb_packets_per_stream
;
3775 if ((long) (start_pos
- consumed_pos
) < 0) {
3776 return consumed_pos
; /* Grab everything */