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 relayd id.
422 * It's atomically set without having the stream mutex locked which is fine
423 * because we handle the write/read race with a pipe wakeup for each thread.
425 static void update_endpoint_status_by_netidx(uint64_t relayd_id
,
426 enum consumer_endpoint_status status
)
428 struct lttng_ht_iter iter
;
429 struct lttng_consumer_stream
*stream
;
431 DBG("Consumer set delete flag on stream by idx %" PRIu64
, relayd_id
);
435 /* Let's begin with metadata */
436 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
437 if (stream
->relayd_id
== relayd_id
) {
438 uatomic_set(&stream
->endpoint_status
, status
);
439 DBG("Delete flag set to metadata stream %d", stream
->wait_fd
);
443 /* Follow up by the data streams */
444 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
445 if (stream
->relayd_id
== relayd_id
) {
446 uatomic_set(&stream
->endpoint_status
, status
);
447 DBG("Delete flag set to data stream %d", stream
->wait_fd
);
454 * Cleanup a relayd object by flagging every associated streams for deletion,
455 * destroying the object meaning removing it from the relayd hash table,
456 * closing the sockets and freeing the memory in a RCU call.
458 * If a local data context is available, notify the threads that the streams'
459 * state have changed.
461 void lttng_consumer_cleanup_relayd(struct consumer_relayd_sock_pair
*relayd
)
467 DBG("Cleaning up relayd object ID %"PRIu64
, relayd
->id
);
469 /* Save the net sequence index before destroying the object */
473 * Delete the relayd from the relayd hash table, close the sockets and free
474 * the object in a RCU call.
476 consumer_destroy_relayd(relayd
);
478 /* Set inactive endpoint to all streams */
479 update_endpoint_status_by_netidx(netidx
, CONSUMER_ENDPOINT_INACTIVE
);
482 * With a local data context, notify the threads that the streams' state
483 * have changed. The write() action on the pipe acts as an "implicit"
484 * memory barrier ordering the updates of the end point status from the
485 * read of this status which happens AFTER receiving this notify.
487 notify_thread_lttng_pipe(relayd
->ctx
->consumer_data_pipe
);
488 notify_thread_lttng_pipe(relayd
->ctx
->consumer_metadata_pipe
);
492 * Flag a relayd socket pair for destruction. Destroy it if the refcount
495 * RCU read side lock MUST be aquired before calling this function.
497 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair
*relayd
)
501 /* Set destroy flag for this object */
502 uatomic_set(&relayd
->destroy_flag
, 1);
504 /* Destroy the relayd if refcount is 0 */
505 if (uatomic_read(&relayd
->refcount
) == 0) {
506 consumer_destroy_relayd(relayd
);
511 * Completly destroy stream from every visiable data structure and the given
514 * One this call returns, the stream object is not longer usable nor visible.
516 void consumer_del_stream(struct lttng_consumer_stream
*stream
,
519 consumer_stream_destroy(stream
, ht
);
523 * XXX naming of del vs destroy is all mixed up.
525 void consumer_del_stream_for_data(struct lttng_consumer_stream
*stream
)
527 consumer_stream_destroy(stream
, data_ht
);
530 void consumer_del_stream_for_metadata(struct lttng_consumer_stream
*stream
)
532 consumer_stream_destroy(stream
, metadata_ht
);
535 struct lttng_consumer_stream
*consumer_allocate_stream(uint64_t channel_key
,
537 enum lttng_consumer_stream_state state
,
538 const char *channel_name
,
545 enum consumer_channel_type type
,
546 unsigned int monitor
)
549 struct lttng_consumer_stream
*stream
;
551 stream
= zmalloc(sizeof(*stream
));
552 if (stream
== NULL
) {
553 PERROR("malloc struct lttng_consumer_stream");
560 stream
->key
= stream_key
;
562 stream
->out_fd_offset
= 0;
563 stream
->output_written
= 0;
564 stream
->state
= state
;
567 stream
->relayd_id
= relayd_id
;
568 stream
->session_id
= session_id
;
569 stream
->monitor
= monitor
;
570 stream
->endpoint_status
= CONSUMER_ENDPOINT_ACTIVE
;
571 stream
->index_file
= NULL
;
572 stream
->last_sequence_number
= -1ULL;
573 pthread_mutex_init(&stream
->lock
, NULL
);
574 pthread_mutex_init(&stream
->metadata_timer_lock
, NULL
);
576 /* If channel is the metadata, flag this stream as metadata. */
577 if (type
== CONSUMER_CHANNEL_TYPE_METADATA
) {
578 stream
->metadata_flag
= 1;
579 /* Metadata is flat out. */
580 strncpy(stream
->name
, DEFAULT_METADATA_NAME
, sizeof(stream
->name
));
581 /* Live rendez-vous point. */
582 pthread_cond_init(&stream
->metadata_rdv
, NULL
);
583 pthread_mutex_init(&stream
->metadata_rdv_lock
, NULL
);
585 /* Format stream name to <channel_name>_<cpu_number> */
586 ret
= snprintf(stream
->name
, sizeof(stream
->name
), "%s_%d",
589 PERROR("snprintf stream name");
594 /* Key is always the wait_fd for streams. */
595 lttng_ht_node_init_u64(&stream
->node
, stream
->key
);
597 /* Init node per channel id key */
598 lttng_ht_node_init_u64(&stream
->node_channel_id
, channel_key
);
600 /* Init session id node with the stream session id */
601 lttng_ht_node_init_u64(&stream
->node_session_id
, stream
->session_id
);
603 DBG3("Allocated stream %s (key %" PRIu64
", chan_key %" PRIu64
604 " relayd_id %" PRIu64
", session_id %" PRIu64
,
605 stream
->name
, stream
->key
, channel_key
,
606 stream
->relayd_id
, stream
->session_id
);
622 * Add a stream to the global list protected by a mutex.
624 int consumer_add_data_stream(struct lttng_consumer_stream
*stream
)
626 struct lttng_ht
*ht
= data_ht
;
632 DBG3("Adding consumer stream %" PRIu64
, stream
->key
);
634 pthread_mutex_lock(&consumer_data
.lock
);
635 pthread_mutex_lock(&stream
->chan
->lock
);
636 pthread_mutex_lock(&stream
->chan
->timer_lock
);
637 pthread_mutex_lock(&stream
->lock
);
640 /* Steal stream identifier to avoid having streams with the same key */
641 steal_stream_key(stream
->key
, ht
);
643 lttng_ht_add_unique_u64(ht
, &stream
->node
);
645 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
646 &stream
->node_channel_id
);
649 * Add stream to the stream_list_ht of the consumer data. No need to steal
650 * the key since the HT does not use it and we allow to add redundant keys
653 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
656 * When nb_init_stream_left reaches 0, we don't need to trigger any action
657 * in terms of destroying the associated channel, because the action that
658 * causes the count to become 0 also causes a stream to be added. The
659 * channel deletion will thus be triggered by the following removal of this
662 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
663 /* Increment refcount before decrementing nb_init_stream_left */
665 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
668 /* Update consumer data once the node is inserted. */
669 consumer_data
.stream_count
++;
670 consumer_data
.need_update
= 1;
673 pthread_mutex_unlock(&stream
->lock
);
674 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
675 pthread_mutex_unlock(&stream
->chan
->lock
);
676 pthread_mutex_unlock(&consumer_data
.lock
);
681 void consumer_del_data_stream(struct lttng_consumer_stream
*stream
)
683 consumer_del_stream(stream
, data_ht
);
687 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
688 * be acquired before calling this.
690 static int add_relayd(struct consumer_relayd_sock_pair
*relayd
)
693 struct lttng_ht_node_u64
*node
;
694 struct lttng_ht_iter iter
;
698 lttng_ht_lookup(consumer_data
.relayd_ht
,
700 node
= lttng_ht_iter_get_node_u64(&iter
);
704 lttng_ht_add_unique_u64(consumer_data
.relayd_ht
, &relayd
->node
);
711 * Allocate and return a consumer relayd socket.
713 static struct consumer_relayd_sock_pair
*consumer_allocate_relayd_sock_pair(
716 struct consumer_relayd_sock_pair
*obj
= NULL
;
718 /* net sequence index of -1 is a failure */
719 if (relayd_id
== (uint64_t) -1ULL) {
723 obj
= zmalloc(sizeof(struct consumer_relayd_sock_pair
));
725 PERROR("zmalloc relayd sock");
731 obj
->destroy_flag
= 0;
732 obj
->control_sock
.sock
.fd
= -1;
733 obj
->data_sock
.sock
.fd
= -1;
734 lttng_ht_node_init_u64(&obj
->node
, obj
->id
);
735 pthread_mutex_init(&obj
->ctrl_sock_mutex
, NULL
);
742 * Find a relayd socket pair in the global consumer data.
744 * Return the object if found else NULL.
745 * RCU read-side lock must be held across this call and while using the
748 struct consumer_relayd_sock_pair
*consumer_find_relayd(uint64_t key
)
750 struct lttng_ht_iter iter
;
751 struct lttng_ht_node_u64
*node
;
752 struct consumer_relayd_sock_pair
*relayd
= NULL
;
754 /* Negative keys are lookup failures */
755 if (key
== (uint64_t) -1ULL) {
759 lttng_ht_lookup(consumer_data
.relayd_ht
, &key
,
761 node
= lttng_ht_iter_get_node_u64(&iter
);
763 relayd
= caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
771 * Find a relayd and send the stream
773 * Returns 0 on success, < 0 on error
775 int consumer_send_relayd_stream(struct lttng_consumer_stream
*stream
,
779 struct consumer_relayd_sock_pair
*relayd
;
782 assert(stream
->relayd_id
!= -1ULL);
785 /* The stream is not metadata. Get relayd reference if exists. */
787 relayd
= consumer_find_relayd(stream
->relayd_id
);
788 if (relayd
!= NULL
) {
789 /* Add stream on the relayd */
790 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
791 ret
= relayd_add_stream(&relayd
->control_sock
, stream
->name
,
792 path
, &stream
->relayd_stream_id
,
793 stream
->chan
->tracefile_size
, stream
->chan
->tracefile_count
);
794 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
796 ERR("Relayd add stream failed. Cleaning up relayd %" PRIu64
".", relayd
->id
);
797 lttng_consumer_cleanup_relayd(relayd
);
801 uatomic_inc(&relayd
->refcount
);
802 stream
->sent_to_relayd
= 1;
804 ERR("Stream %" PRIu64
" relayd ID %" PRIu64
" unknown. Can't send it.",
805 stream
->key
, stream
->relayd_id
);
810 DBG("Stream %s with key %" PRIu64
" sent to relayd id %" PRIu64
,
811 stream
->name
, stream
->key
, stream
->relayd_id
);
819 * Find a relayd and send the streams sent message
821 * Returns 0 on success, < 0 on error
823 int consumer_send_relayd_streams_sent(uint64_t relayd_id
)
826 struct consumer_relayd_sock_pair
*relayd
;
828 assert(relayd_id
!= -1ULL);
830 /* The stream is not metadata. Get relayd reference if exists. */
832 relayd
= consumer_find_relayd(relayd_id
);
833 if (relayd
!= NULL
) {
834 /* Add stream on the relayd */
835 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
836 ret
= relayd_streams_sent(&relayd
->control_sock
);
837 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
839 ERR("Relayd streams sent failed. Cleaning up relayd %" PRIu64
".", relayd
->id
);
840 lttng_consumer_cleanup_relayd(relayd
);
844 ERR("Relayd ID %" PRIu64
" unknown. Can't send streams_sent.",
851 DBG("All streams sent relayd id %" PRIu64
, relayd_id
);
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
->relayd_id
);
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
,
1513 unsigned long padding
,
1514 struct ctf_packet_index
*index
)
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
->relayd_id
!= (uint64_t) -1ULL) {
1528 relayd
= consumer_find_relayd(stream
->relayd_id
);
1529 if (relayd
== NULL
) {
1535 /* Handle stream on the relayd if the output is on the network */
1537 unsigned long netlen
= len
;
1540 * Lock the control socket for the complete duration of the function
1541 * since from this point on we will use the socket.
1543 if (stream
->metadata_flag
) {
1544 /* Metadata requires the control socket. */
1545 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1546 if (stream
->reset_metadata_flag
) {
1547 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1548 stream
->relayd_stream_id
,
1549 stream
->metadata_version
);
1554 stream
->reset_metadata_flag
= 0;
1556 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1559 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1564 /* Use the returned socket. */
1567 /* Write metadata stream id before payload */
1568 if (stream
->metadata_flag
) {
1569 ret
= write_relayd_metadata_id(outfd
, stream
, relayd
, padding
);
1576 /* No streaming, we have to set the len with the full padding */
1579 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1580 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1582 ERR("Reset metadata file");
1585 stream
->reset_metadata_flag
= 0;
1589 * Check if we need to change the tracefile before writing the packet.
1591 if (stream
->chan
->tracefile_size
> 0 &&
1592 (stream
->tracefile_size_current
+ len
) >
1593 stream
->chan
->tracefile_size
) {
1594 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1595 stream
->name
, stream
->chan
->tracefile_size
,
1596 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1597 stream
->out_fd
, &(stream
->tracefile_count_current
),
1600 ERR("Rotating output file");
1603 outfd
= stream
->out_fd
;
1605 if (stream
->index_file
) {
1606 lttng_index_file_put(stream
->index_file
);
1607 stream
->index_file
= lttng_index_file_create(stream
->chan
->pathname
,
1608 stream
->name
, stream
->uid
, stream
->gid
,
1609 stream
->chan
->tracefile_size
,
1610 stream
->tracefile_count_current
,
1611 CTF_INDEX_MAJOR
, CTF_INDEX_MINOR
);
1612 if (!stream
->index_file
) {
1617 /* Reset current size because we just perform a rotation. */
1618 stream
->tracefile_size_current
= 0;
1619 stream
->out_fd_offset
= 0;
1622 stream
->tracefile_size_current
+= len
;
1624 index
->offset
= htobe64(stream
->out_fd_offset
);
1629 * This call guarantee that len or less is returned. It's impossible to
1630 * receive a ret value that is bigger than len.
1632 ret
= lttng_write(outfd
, buffer
, len
);
1633 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, len
);
1634 if (ret
< 0 || ((size_t) ret
!= len
)) {
1636 * Report error to caller if nothing was written else at least send the
1644 /* Socket operation failed. We consider the relayd dead */
1645 if (errno
== EPIPE
|| errno
== EINVAL
|| errno
== EBADF
) {
1647 * This is possible if the fd is closed on the other side
1648 * (outfd) or any write problem. It can be verbose a bit for a
1649 * normal execution if for instance the relayd is stopped
1650 * abruptly. This can happen so set this to a DBG statement.
1652 DBG("Consumer mmap write detected relayd hang up");
1654 /* Unhandled error, print it and stop function right now. */
1655 PERROR("Error in write mmap (ret %zd != len %lu)", ret
, len
);
1659 stream
->output_written
+= ret
;
1661 /* This call is useless on a socket so better save a syscall. */
1663 /* This won't block, but will start writeout asynchronously */
1664 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, len
,
1665 SYNC_FILE_RANGE_WRITE
);
1666 stream
->out_fd_offset
+= len
;
1667 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1672 * This is a special case that the relayd has closed its socket. Let's
1673 * cleanup the relayd object and all associated streams.
1675 if (relayd
&& relayd_hang_up
) {
1676 ERR("Relayd hangup. Cleaning up relayd %" PRIu64
".", relayd
->id
);
1677 lttng_consumer_cleanup_relayd(relayd
);
1681 /* Unlock only if ctrl socket used */
1682 if (relayd
&& stream
->metadata_flag
) {
1683 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1691 * Splice the data from the ring buffer to the tracefile.
1693 * It must be called with the stream lock held.
1695 * Returns the number of bytes spliced.
1697 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1698 struct lttng_consumer_local_data
*ctx
,
1699 struct lttng_consumer_stream
*stream
, unsigned long len
,
1700 unsigned long padding
,
1701 struct ctf_packet_index
*index
)
1703 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1705 off_t orig_offset
= stream
->out_fd_offset
;
1706 int fd
= stream
->wait_fd
;
1707 /* Default is on the disk */
1708 int outfd
= stream
->out_fd
;
1709 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1711 unsigned int relayd_hang_up
= 0;
1713 switch (consumer_data
.type
) {
1714 case LTTNG_CONSUMER_KERNEL
:
1716 case LTTNG_CONSUMER32_UST
:
1717 case LTTNG_CONSUMER64_UST
:
1718 /* Not supported for user space tracing */
1721 ERR("Unknown consumer_data type");
1725 /* RCU lock for the relayd pointer */
1728 /* Flag that the current stream if set for network streaming. */
1729 if (stream
->relayd_id
!= (uint64_t) -1ULL) {
1730 relayd
= consumer_find_relayd(stream
->relayd_id
);
1731 if (relayd
== NULL
) {
1736 splice_pipe
= stream
->splice_pipe
;
1738 /* Write metadata stream id before payload */
1740 unsigned long total_len
= len
;
1742 if (stream
->metadata_flag
) {
1744 * Lock the control socket for the complete duration of the function
1745 * since from this point on we will use the socket.
1747 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1749 if (stream
->reset_metadata_flag
) {
1750 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1751 stream
->relayd_stream_id
,
1752 stream
->metadata_version
);
1757 stream
->reset_metadata_flag
= 0;
1759 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
, relayd
,
1767 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1770 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1776 /* Use the returned socket. */
1779 /* No streaming, we have to set the len with the full padding */
1782 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1783 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1785 ERR("Reset metadata file");
1788 stream
->reset_metadata_flag
= 0;
1791 * Check if we need to change the tracefile before writing the packet.
1793 if (stream
->chan
->tracefile_size
> 0 &&
1794 (stream
->tracefile_size_current
+ len
) >
1795 stream
->chan
->tracefile_size
) {
1796 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1797 stream
->name
, stream
->chan
->tracefile_size
,
1798 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1799 stream
->out_fd
, &(stream
->tracefile_count_current
),
1803 ERR("Rotating output file");
1806 outfd
= stream
->out_fd
;
1808 if (stream
->index_file
) {
1809 lttng_index_file_put(stream
->index_file
);
1810 stream
->index_file
= lttng_index_file_create(stream
->chan
->pathname
,
1811 stream
->name
, stream
->uid
, stream
->gid
,
1812 stream
->chan
->tracefile_size
,
1813 stream
->tracefile_count_current
,
1814 CTF_INDEX_MAJOR
, CTF_INDEX_MINOR
);
1815 if (!stream
->index_file
) {
1820 /* Reset current size because we just perform a rotation. */
1821 stream
->tracefile_size_current
= 0;
1822 stream
->out_fd_offset
= 0;
1825 stream
->tracefile_size_current
+= len
;
1826 index
->offset
= htobe64(stream
->out_fd_offset
);
1830 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1831 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1832 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1833 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1834 DBG("splice chan to pipe, ret %zd", ret_splice
);
1835 if (ret_splice
< 0) {
1838 PERROR("Error in relay splice");
1842 /* Handle stream on the relayd if the output is on the network */
1843 if (relayd
&& stream
->metadata_flag
) {
1844 size_t metadata_payload_size
=
1845 sizeof(struct lttcomm_relayd_metadata_payload
);
1847 /* Update counter to fit the spliced data */
1848 ret_splice
+= metadata_payload_size
;
1849 len
+= metadata_payload_size
;
1851 * We do this so the return value can match the len passed as
1852 * argument to this function.
1854 written
-= metadata_payload_size
;
1857 /* Splice data out */
1858 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1859 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1860 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
1862 if (ret_splice
< 0) {
1867 } else if (ret_splice
> len
) {
1869 * We don't expect this code path to be executed but you never know
1870 * so this is an extra protection agains a buggy splice().
1873 written
+= ret_splice
;
1874 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice
,
1878 /* All good, update current len and continue. */
1882 /* This call is useless on a socket so better save a syscall. */
1884 /* This won't block, but will start writeout asynchronously */
1885 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1886 SYNC_FILE_RANGE_WRITE
);
1887 stream
->out_fd_offset
+= ret_splice
;
1889 stream
->output_written
+= ret_splice
;
1890 written
+= ret_splice
;
1893 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1899 * This is a special case that the relayd has closed its socket. Let's
1900 * cleanup the relayd object and all associated streams.
1902 if (relayd
&& relayd_hang_up
) {
1903 ERR("Relayd hangup. Cleaning up relayd %" PRIu64
".", relayd
->id
);
1904 lttng_consumer_cleanup_relayd(relayd
);
1905 /* Skip splice error so the consumer does not fail */
1910 /* send the appropriate error description to sessiond */
1913 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1916 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1919 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
1924 if (relayd
&& stream
->metadata_flag
) {
1925 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1933 * Take a snapshot for a specific fd
1935 * Returns 0 on success, < 0 on error
1937 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
1939 switch (consumer_data
.type
) {
1940 case LTTNG_CONSUMER_KERNEL
:
1941 return lttng_kconsumer_take_snapshot(stream
);
1942 case LTTNG_CONSUMER32_UST
:
1943 case LTTNG_CONSUMER64_UST
:
1944 return lttng_ustconsumer_take_snapshot(stream
);
1946 ERR("Unknown consumer_data type");
1953 * Get the produced position
1955 * Returns 0 on success, < 0 on error
1957 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
1960 switch (consumer_data
.type
) {
1961 case LTTNG_CONSUMER_KERNEL
:
1962 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
1963 case LTTNG_CONSUMER32_UST
:
1964 case LTTNG_CONSUMER64_UST
:
1965 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
1967 ERR("Unknown consumer_data type");
1973 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
1974 int sock
, struct pollfd
*consumer_sockpoll
)
1976 switch (consumer_data
.type
) {
1977 case LTTNG_CONSUMER_KERNEL
:
1978 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1979 case LTTNG_CONSUMER32_UST
:
1980 case LTTNG_CONSUMER64_UST
:
1981 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1983 ERR("Unknown consumer_data type");
1989 void lttng_consumer_close_all_metadata(void)
1991 switch (consumer_data
.type
) {
1992 case LTTNG_CONSUMER_KERNEL
:
1994 * The Kernel consumer has a different metadata scheme so we don't
1995 * close anything because the stream will be closed by the session
1999 case LTTNG_CONSUMER32_UST
:
2000 case LTTNG_CONSUMER64_UST
:
2002 * Close all metadata streams. The metadata hash table is passed and
2003 * this call iterates over it by closing all wakeup fd. This is safe
2004 * because at this point we are sure that the metadata producer is
2005 * either dead or blocked.
2007 lttng_ustconsumer_close_all_metadata(metadata_ht
);
2010 ERR("Unknown consumer_data type");
2016 * Clean up a metadata stream and free its memory.
2018 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
2019 struct lttng_ht
*ht
)
2021 struct lttng_consumer_channel
*free_chan
= NULL
;
2025 * This call should NEVER receive regular stream. It must always be
2026 * metadata stream and this is crucial for data structure synchronization.
2028 assert(stream
->metadata_flag
);
2030 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
2032 pthread_mutex_lock(&consumer_data
.lock
);
2033 pthread_mutex_lock(&stream
->chan
->lock
);
2034 pthread_mutex_lock(&stream
->lock
);
2035 if (stream
->chan
->metadata_cache
) {
2036 /* Only applicable to userspace consumers. */
2037 pthread_mutex_lock(&stream
->chan
->metadata_cache
->lock
);
2040 /* Remove any reference to that stream. */
2041 consumer_stream_delete(stream
, ht
);
2043 /* Close down everything including the relayd if one. */
2044 consumer_stream_close(stream
);
2045 /* Destroy tracer buffers of the stream. */
2046 consumer_stream_destroy_buffers(stream
);
2048 /* Atomically decrement channel refcount since other threads can use it. */
2049 if (!uatomic_sub_return(&stream
->chan
->refcount
, 1)
2050 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
2051 /* Go for channel deletion! */
2052 free_chan
= stream
->chan
;
2056 * Nullify the stream reference so it is not used after deletion. The
2057 * channel lock MUST be acquired before being able to check for a NULL
2060 stream
->chan
->metadata_stream
= NULL
;
2062 if (stream
->chan
->metadata_cache
) {
2063 pthread_mutex_unlock(&stream
->chan
->metadata_cache
->lock
);
2065 pthread_mutex_unlock(&stream
->lock
);
2066 pthread_mutex_unlock(&stream
->chan
->lock
);
2067 pthread_mutex_unlock(&consumer_data
.lock
);
2070 consumer_del_channel(free_chan
);
2073 consumer_stream_free(stream
);
2077 * Action done with the metadata stream when adding it to the consumer internal
2078 * data structures to handle it.
2080 int consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
2082 struct lttng_ht
*ht
= metadata_ht
;
2084 struct lttng_ht_iter iter
;
2085 struct lttng_ht_node_u64
*node
;
2090 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2092 pthread_mutex_lock(&consumer_data
.lock
);
2093 pthread_mutex_lock(&stream
->chan
->lock
);
2094 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2095 pthread_mutex_lock(&stream
->lock
);
2098 * From here, refcounts are updated so be _careful_ when returning an error
2105 * Lookup the stream just to make sure it does not exist in our internal
2106 * state. This should NEVER happen.
2108 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2109 node
= lttng_ht_iter_get_node_u64(&iter
);
2113 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2114 * in terms of destroying the associated channel, because the action that
2115 * causes the count to become 0 also causes a stream to be added. The
2116 * channel deletion will thus be triggered by the following removal of this
2119 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2120 /* Increment refcount before decrementing nb_init_stream_left */
2122 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2125 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2127 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
2128 &stream
->node_channel_id
);
2131 * Add stream to the stream_list_ht of the consumer data. No need to steal
2132 * the key since the HT does not use it and we allow to add redundant keys
2135 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2139 pthread_mutex_unlock(&stream
->lock
);
2140 pthread_mutex_unlock(&stream
->chan
->lock
);
2141 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2142 pthread_mutex_unlock(&consumer_data
.lock
);
2147 * Delete data stream that are flagged for deletion (endpoint_status).
2149 static void validate_endpoint_status_data_stream(void)
2151 struct lttng_ht_iter iter
;
2152 struct lttng_consumer_stream
*stream
;
2154 DBG("Consumer delete flagged data stream");
2157 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2158 /* Validate delete flag of the stream */
2159 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2162 /* Delete it right now */
2163 consumer_del_stream(stream
, data_ht
);
2169 * Delete metadata stream that are flagged for deletion (endpoint_status).
2171 static void validate_endpoint_status_metadata_stream(
2172 struct lttng_poll_event
*pollset
)
2174 struct lttng_ht_iter iter
;
2175 struct lttng_consumer_stream
*stream
;
2177 DBG("Consumer delete flagged metadata stream");
2182 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2183 /* Validate delete flag of the stream */
2184 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2188 * Remove from pollset so the metadata thread can continue without
2189 * blocking on a deleted stream.
2191 lttng_poll_del(pollset
, stream
->wait_fd
);
2193 /* Delete it right now */
2194 consumer_del_metadata_stream(stream
, metadata_ht
);
2200 * Thread polls on metadata file descriptor and write them on disk or on the
2203 void *consumer_thread_metadata_poll(void *data
)
2205 int ret
, i
, pollfd
, err
= -1;
2206 uint32_t revents
, nb_fd
;
2207 struct lttng_consumer_stream
*stream
= NULL
;
2208 struct lttng_ht_iter iter
;
2209 struct lttng_ht_node_u64
*node
;
2210 struct lttng_poll_event events
;
2211 struct lttng_consumer_local_data
*ctx
= data
;
2214 rcu_register_thread();
2216 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2218 if (testpoint(consumerd_thread_metadata
)) {
2219 goto error_testpoint
;
2222 health_code_update();
2224 DBG("Thread metadata poll started");
2226 /* Size is set to 1 for the consumer_metadata pipe */
2227 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2229 ERR("Poll set creation failed");
2233 ret
= lttng_poll_add(&events
,
2234 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2240 DBG("Metadata main loop started");
2244 health_code_update();
2245 health_poll_entry();
2246 DBG("Metadata poll wait");
2247 ret
= lttng_poll_wait(&events
, -1);
2248 DBG("Metadata poll return from wait with %d fd(s)",
2249 LTTNG_POLL_GETNB(&events
));
2251 DBG("Metadata event caught in thread");
2253 if (errno
== EINTR
) {
2254 ERR("Poll EINTR caught");
2257 if (LTTNG_POLL_GETNB(&events
) == 0) {
2258 err
= 0; /* All is OK */
2265 /* From here, the event is a metadata wait fd */
2266 for (i
= 0; i
< nb_fd
; i
++) {
2267 health_code_update();
2269 revents
= LTTNG_POLL_GETEV(&events
, i
);
2270 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2273 /* No activity for this FD (poll implementation). */
2277 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2278 if (revents
& LPOLLIN
) {
2281 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2282 &stream
, sizeof(stream
));
2283 if (pipe_len
< sizeof(stream
)) {
2285 PERROR("read metadata stream");
2288 * Remove the pipe from the poll set and continue the loop
2289 * since their might be data to consume.
2291 lttng_poll_del(&events
,
2292 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2293 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2297 /* A NULL stream means that the state has changed. */
2298 if (stream
== NULL
) {
2299 /* Check for deleted streams. */
2300 validate_endpoint_status_metadata_stream(&events
);
2304 DBG("Adding metadata stream %d to poll set",
2307 /* Add metadata stream to the global poll events list */
2308 lttng_poll_add(&events
, stream
->wait_fd
,
2309 LPOLLIN
| LPOLLPRI
| LPOLLHUP
);
2310 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2311 DBG("Metadata thread pipe hung up");
2313 * Remove the pipe from the poll set and continue the loop
2314 * since their might be data to consume.
2316 lttng_poll_del(&events
,
2317 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2318 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2321 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2325 /* Handle other stream */
2331 uint64_t tmp_id
= (uint64_t) pollfd
;
2333 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2335 node
= lttng_ht_iter_get_node_u64(&iter
);
2338 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2341 if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2342 /* Get the data out of the metadata file descriptor */
2343 DBG("Metadata available on fd %d", pollfd
);
2344 assert(stream
->wait_fd
== pollfd
);
2347 health_code_update();
2349 len
= ctx
->on_buffer_ready(stream
, ctx
);
2351 * We don't check the return value here since if we get
2352 * a negative len, it means an error occurred thus we
2353 * simply remove it from the poll set and free the
2358 /* It's ok to have an unavailable sub-buffer */
2359 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2360 /* Clean up stream from consumer and free it. */
2361 lttng_poll_del(&events
, stream
->wait_fd
);
2362 consumer_del_metadata_stream(stream
, metadata_ht
);
2364 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2365 DBG("Metadata fd %d is hup|err.", pollfd
);
2366 if (!stream
->hangup_flush_done
2367 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2368 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2369 DBG("Attempting to flush and consume the UST buffers");
2370 lttng_ustconsumer_on_stream_hangup(stream
);
2372 /* We just flushed the stream now read it. */
2374 health_code_update();
2376 len
= ctx
->on_buffer_ready(stream
, ctx
);
2378 * We don't check the return value here since if we get
2379 * a negative len, it means an error occurred thus we
2380 * simply remove it from the poll set and free the
2386 lttng_poll_del(&events
, stream
->wait_fd
);
2388 * This call update the channel states, closes file descriptors
2389 * and securely free the stream.
2391 consumer_del_metadata_stream(stream
, metadata_ht
);
2393 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2397 /* Release RCU lock for the stream looked up */
2405 DBG("Metadata poll thread exiting");
2407 lttng_poll_clean(&events
);
2412 ERR("Health error occurred in %s", __func__
);
2414 health_unregister(health_consumerd
);
2415 rcu_unregister_thread();
2420 * This thread polls the fds in the set to consume the data and write
2421 * it to tracefile if necessary.
2423 void *consumer_thread_data_poll(void *data
)
2425 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2426 struct pollfd
*pollfd
= NULL
;
2427 /* local view of the streams */
2428 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2429 /* local view of consumer_data.fds_count */
2431 /* Number of FDs with CONSUMER_ENDPOINT_INACTIVE but still open. */
2432 int nb_inactive_fd
= 0;
2433 struct lttng_consumer_local_data
*ctx
= data
;
2436 rcu_register_thread();
2438 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2440 if (testpoint(consumerd_thread_data
)) {
2441 goto error_testpoint
;
2444 health_code_update();
2446 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
*));
2447 if (local_stream
== NULL
) {
2448 PERROR("local_stream malloc");
2453 health_code_update();
2459 * the fds set has been updated, we need to update our
2460 * local array as well
2462 pthread_mutex_lock(&consumer_data
.lock
);
2463 if (consumer_data
.need_update
) {
2468 local_stream
= NULL
;
2471 * Allocate for all fds +1 for the consumer_data_pipe and +1 for
2474 pollfd
= zmalloc((consumer_data
.stream_count
+ 2) * sizeof(struct pollfd
));
2475 if (pollfd
== NULL
) {
2476 PERROR("pollfd malloc");
2477 pthread_mutex_unlock(&consumer_data
.lock
);
2481 local_stream
= zmalloc((consumer_data
.stream_count
+ 2) *
2482 sizeof(struct lttng_consumer_stream
*));
2483 if (local_stream
== NULL
) {
2484 PERROR("local_stream malloc");
2485 pthread_mutex_unlock(&consumer_data
.lock
);
2488 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2489 data_ht
, &nb_inactive_fd
);
2491 ERR("Error in allocating pollfd or local_outfds");
2492 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2493 pthread_mutex_unlock(&consumer_data
.lock
);
2497 consumer_data
.need_update
= 0;
2499 pthread_mutex_unlock(&consumer_data
.lock
);
2501 /* No FDs and consumer_quit, consumer_cleanup the thread */
2502 if (nb_fd
== 0 && consumer_quit
== 1 && nb_inactive_fd
== 0) {
2503 err
= 0; /* All is OK */
2506 /* poll on the array of fds */
2508 DBG("polling on %d fd", nb_fd
+ 2);
2509 health_poll_entry();
2510 num_rdy
= poll(pollfd
, nb_fd
+ 2, -1);
2512 DBG("poll num_rdy : %d", num_rdy
);
2513 if (num_rdy
== -1) {
2515 * Restart interrupted system call.
2517 if (errno
== EINTR
) {
2520 PERROR("Poll error");
2521 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2523 } else if (num_rdy
== 0) {
2524 DBG("Polling thread timed out");
2529 * If the consumer_data_pipe triggered poll go directly to the
2530 * beginning of the loop to update the array. We want to prioritize
2531 * array update over low-priority reads.
2533 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2534 ssize_t pipe_readlen
;
2536 DBG("consumer_data_pipe wake up");
2537 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2538 &new_stream
, sizeof(new_stream
));
2539 if (pipe_readlen
< sizeof(new_stream
)) {
2540 PERROR("Consumer data pipe");
2541 /* Continue so we can at least handle the current stream(s). */
2546 * If the stream is NULL, just ignore it. It's also possible that
2547 * the sessiond poll thread changed the consumer_quit state and is
2548 * waking us up to test it.
2550 if (new_stream
== NULL
) {
2551 validate_endpoint_status_data_stream();
2555 /* Continue to update the local streams and handle prio ones */
2559 /* Handle wakeup pipe. */
2560 if (pollfd
[nb_fd
+ 1].revents
& (POLLIN
| POLLPRI
)) {
2562 ssize_t pipe_readlen
;
2564 pipe_readlen
= lttng_pipe_read(ctx
->consumer_wakeup_pipe
, &dummy
,
2566 if (pipe_readlen
< 0) {
2567 PERROR("Consumer data wakeup pipe");
2569 /* We've been awakened to handle stream(s). */
2570 ctx
->has_wakeup
= 0;
2573 /* Take care of high priority channels first. */
2574 for (i
= 0; i
< nb_fd
; i
++) {
2575 health_code_update();
2577 if (local_stream
[i
] == NULL
) {
2580 if (pollfd
[i
].revents
& POLLPRI
) {
2581 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2583 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2584 /* it's ok to have an unavailable sub-buffer */
2585 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2586 /* Clean the stream and free it. */
2587 consumer_del_stream(local_stream
[i
], data_ht
);
2588 local_stream
[i
] = NULL
;
2589 } else if (len
> 0) {
2590 local_stream
[i
]->data_read
= 1;
2596 * If we read high prio channel in this loop, try again
2597 * for more high prio data.
2603 /* Take care of low priority channels. */
2604 for (i
= 0; i
< nb_fd
; i
++) {
2605 health_code_update();
2607 if (local_stream
[i
] == NULL
) {
2610 if ((pollfd
[i
].revents
& POLLIN
) ||
2611 local_stream
[i
]->hangup_flush_done
||
2612 local_stream
[i
]->has_data
) {
2613 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2614 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2615 /* it's ok to have an unavailable sub-buffer */
2616 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2617 /* Clean the stream and free it. */
2618 consumer_del_stream(local_stream
[i
], data_ht
);
2619 local_stream
[i
] = NULL
;
2620 } else if (len
> 0) {
2621 local_stream
[i
]->data_read
= 1;
2626 /* Handle hangup and errors */
2627 for (i
= 0; i
< nb_fd
; i
++) {
2628 health_code_update();
2630 if (local_stream
[i
] == NULL
) {
2633 if (!local_stream
[i
]->hangup_flush_done
2634 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2635 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2636 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2637 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2639 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2640 /* Attempt read again, for the data we just flushed. */
2641 local_stream
[i
]->data_read
= 1;
2644 * If the poll flag is HUP/ERR/NVAL and we have
2645 * read no data in this pass, we can remove the
2646 * stream from its hash table.
2648 if ((pollfd
[i
].revents
& POLLHUP
)) {
2649 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2650 if (!local_stream
[i
]->data_read
) {
2651 consumer_del_stream(local_stream
[i
], data_ht
);
2652 local_stream
[i
] = NULL
;
2655 } else if (pollfd
[i
].revents
& POLLERR
) {
2656 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2657 if (!local_stream
[i
]->data_read
) {
2658 consumer_del_stream(local_stream
[i
], data_ht
);
2659 local_stream
[i
] = NULL
;
2662 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2663 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2664 if (!local_stream
[i
]->data_read
) {
2665 consumer_del_stream(local_stream
[i
], data_ht
);
2666 local_stream
[i
] = NULL
;
2670 if (local_stream
[i
] != NULL
) {
2671 local_stream
[i
]->data_read
= 0;
2678 DBG("polling thread exiting");
2683 * Close the write side of the pipe so epoll_wait() in
2684 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2685 * read side of the pipe. If we close them both, epoll_wait strangely does
2686 * not return and could create a endless wait period if the pipe is the
2687 * only tracked fd in the poll set. The thread will take care of closing
2690 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2695 ERR("Health error occurred in %s", __func__
);
2697 health_unregister(health_consumerd
);
2699 rcu_unregister_thread();
2704 * Close wake-up end of each stream belonging to the channel. This will
2705 * allow the poll() on the stream read-side to detect when the
2706 * write-side (application) finally closes them.
2709 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2711 struct lttng_ht
*ht
;
2712 struct lttng_consumer_stream
*stream
;
2713 struct lttng_ht_iter iter
;
2715 ht
= consumer_data
.stream_per_chan_id_ht
;
2718 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2719 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2720 ht
->match_fct
, &channel
->key
,
2721 &iter
.iter
, stream
, node_channel_id
.node
) {
2723 * Protect against teardown with mutex.
2725 pthread_mutex_lock(&stream
->lock
);
2726 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2729 switch (consumer_data
.type
) {
2730 case LTTNG_CONSUMER_KERNEL
:
2732 case LTTNG_CONSUMER32_UST
:
2733 case LTTNG_CONSUMER64_UST
:
2734 if (stream
->metadata_flag
) {
2735 /* Safe and protected by the stream lock. */
2736 lttng_ustconsumer_close_metadata(stream
->chan
);
2739 * Note: a mutex is taken internally within
2740 * liblttng-ust-ctl to protect timer wakeup_fd
2741 * use from concurrent close.
2743 lttng_ustconsumer_close_stream_wakeup(stream
);
2747 ERR("Unknown consumer_data type");
2751 pthread_mutex_unlock(&stream
->lock
);
2756 static void destroy_channel_ht(struct lttng_ht
*ht
)
2758 struct lttng_ht_iter iter
;
2759 struct lttng_consumer_channel
*channel
;
2767 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2768 ret
= lttng_ht_del(ht
, &iter
);
2773 lttng_ht_destroy(ht
);
2777 * This thread polls the channel fds to detect when they are being
2778 * closed. It closes all related streams if the channel is detected as
2779 * closed. It is currently only used as a shim layer for UST because the
2780 * consumerd needs to keep the per-stream wakeup end of pipes open for
2783 void *consumer_thread_channel_poll(void *data
)
2785 int ret
, i
, pollfd
, err
= -1;
2786 uint32_t revents
, nb_fd
;
2787 struct lttng_consumer_channel
*chan
= NULL
;
2788 struct lttng_ht_iter iter
;
2789 struct lttng_ht_node_u64
*node
;
2790 struct lttng_poll_event events
;
2791 struct lttng_consumer_local_data
*ctx
= data
;
2792 struct lttng_ht
*channel_ht
;
2794 rcu_register_thread();
2796 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
2798 if (testpoint(consumerd_thread_channel
)) {
2799 goto error_testpoint
;
2802 health_code_update();
2804 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2806 /* ENOMEM at this point. Better to bail out. */
2810 DBG("Thread channel poll started");
2812 /* Size is set to 1 for the consumer_channel pipe */
2813 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2815 ERR("Poll set creation failed");
2819 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2825 DBG("Channel main loop started");
2829 health_code_update();
2830 DBG("Channel poll wait");
2831 health_poll_entry();
2832 ret
= lttng_poll_wait(&events
, -1);
2833 DBG("Channel poll return from wait with %d fd(s)",
2834 LTTNG_POLL_GETNB(&events
));
2836 DBG("Channel event caught in thread");
2838 if (errno
== EINTR
) {
2839 ERR("Poll EINTR caught");
2842 if (LTTNG_POLL_GETNB(&events
) == 0) {
2843 err
= 0; /* All is OK */
2850 /* From here, the event is a channel wait fd */
2851 for (i
= 0; i
< nb_fd
; i
++) {
2852 health_code_update();
2854 revents
= LTTNG_POLL_GETEV(&events
, i
);
2855 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2858 /* No activity for this FD (poll implementation). */
2862 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2863 if (revents
& LPOLLIN
) {
2864 enum consumer_channel_action action
;
2867 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
2870 ERR("Error reading channel pipe");
2872 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2877 case CONSUMER_CHANNEL_ADD
:
2878 DBG("Adding channel %d to poll set",
2881 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
2884 lttng_ht_add_unique_u64(channel_ht
,
2885 &chan
->wait_fd_node
);
2887 /* Add channel to the global poll events list */
2888 lttng_poll_add(&events
, chan
->wait_fd
,
2889 LPOLLERR
| LPOLLHUP
);
2891 case CONSUMER_CHANNEL_DEL
:
2894 * This command should never be called if the channel
2895 * has streams monitored by either the data or metadata
2896 * thread. The consumer only notify this thread with a
2897 * channel del. command if it receives a destroy
2898 * channel command from the session daemon that send it
2899 * if a command prior to the GET_CHANNEL failed.
2903 chan
= consumer_find_channel(key
);
2906 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
2909 lttng_poll_del(&events
, chan
->wait_fd
);
2910 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
2911 ret
= lttng_ht_del(channel_ht
, &iter
);
2914 switch (consumer_data
.type
) {
2915 case LTTNG_CONSUMER_KERNEL
:
2917 case LTTNG_CONSUMER32_UST
:
2918 case LTTNG_CONSUMER64_UST
:
2919 health_code_update();
2920 /* Destroy streams that might have been left in the stream list. */
2921 clean_channel_stream_list(chan
);
2924 ERR("Unknown consumer_data type");
2929 * Release our own refcount. Force channel deletion even if
2930 * streams were not initialized.
2932 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
2933 consumer_del_channel(chan
);
2938 case CONSUMER_CHANNEL_QUIT
:
2940 * Remove the pipe from the poll set and continue the loop
2941 * since their might be data to consume.
2943 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2946 ERR("Unknown action");
2949 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2950 DBG("Channel thread pipe hung up");
2952 * Remove the pipe from the poll set and continue the loop
2953 * since their might be data to consume.
2955 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2958 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2962 /* Handle other stream */
2968 uint64_t tmp_id
= (uint64_t) pollfd
;
2970 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
2972 node
= lttng_ht_iter_get_node_u64(&iter
);
2975 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
2978 /* Check for error event */
2979 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2980 DBG("Channel fd %d is hup|err.", pollfd
);
2982 lttng_poll_del(&events
, chan
->wait_fd
);
2983 ret
= lttng_ht_del(channel_ht
, &iter
);
2987 * This will close the wait fd for each stream associated to
2988 * this channel AND monitored by the data/metadata thread thus
2989 * will be clean by the right thread.
2991 consumer_close_channel_streams(chan
);
2993 /* Release our own refcount */
2994 if (!uatomic_sub_return(&chan
->refcount
, 1)
2995 && !uatomic_read(&chan
->nb_init_stream_left
)) {
2996 consumer_del_channel(chan
);
2999 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3004 /* Release RCU lock for the channel looked up */
3012 lttng_poll_clean(&events
);
3014 destroy_channel_ht(channel_ht
);
3017 DBG("Channel poll thread exiting");
3020 ERR("Health error occurred in %s", __func__
);
3022 health_unregister(health_consumerd
);
3023 rcu_unregister_thread();
3027 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
3028 struct pollfd
*sockpoll
, int client_socket
)
3035 ret
= lttng_consumer_poll_socket(sockpoll
);
3039 DBG("Metadata connection on client_socket");
3041 /* Blocking call, waiting for transmission */
3042 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
3043 if (ctx
->consumer_metadata_socket
< 0) {
3044 WARN("On accept metadata");
3055 * This thread listens on the consumerd socket and receives the file
3056 * descriptors from the session daemon.
3058 void *consumer_thread_sessiond_poll(void *data
)
3060 int sock
= -1, client_socket
, ret
, err
= -1;
3062 * structure to poll for incoming data on communication socket avoids
3063 * making blocking sockets.
3065 struct pollfd consumer_sockpoll
[2];
3066 struct lttng_consumer_local_data
*ctx
= data
;
3068 rcu_register_thread();
3070 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
3072 if (testpoint(consumerd_thread_sessiond
)) {
3073 goto error_testpoint
;
3076 health_code_update();
3078 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
3079 unlink(ctx
->consumer_command_sock_path
);
3080 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
3081 if (client_socket
< 0) {
3082 ERR("Cannot create command socket");
3086 ret
= lttcomm_listen_unix_sock(client_socket
);
3091 DBG("Sending ready command to lttng-sessiond");
3092 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3093 /* return < 0 on error, but == 0 is not fatal */
3095 ERR("Error sending ready command to lttng-sessiond");
3099 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3100 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3101 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3102 consumer_sockpoll
[1].fd
= client_socket
;
3103 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3105 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3113 DBG("Connection on client_socket");
3115 /* Blocking call, waiting for transmission */
3116 sock
= lttcomm_accept_unix_sock(client_socket
);
3123 * Setup metadata socket which is the second socket connection on the
3124 * command unix socket.
3126 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3135 /* This socket is not useful anymore. */
3136 ret
= close(client_socket
);
3138 PERROR("close client_socket");
3142 /* update the polling structure to poll on the established socket */
3143 consumer_sockpoll
[1].fd
= sock
;
3144 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3147 health_code_update();
3149 health_poll_entry();
3150 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3159 DBG("Incoming command on sock");
3160 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3163 * This could simply be a session daemon quitting. Don't output
3166 DBG("Communication interrupted on command socket");
3170 if (consumer_quit
) {
3171 DBG("consumer_thread_receive_fds received quit from signal");
3172 err
= 0; /* All is OK */
3175 DBG("received command on sock");
3181 DBG("Consumer thread sessiond poll exiting");
3184 * Close metadata streams since the producer is the session daemon which
3187 * NOTE: for now, this only applies to the UST tracer.
3189 lttng_consumer_close_all_metadata();
3192 * when all fds have hung up, the polling thread
3198 * Notify the data poll thread to poll back again and test the
3199 * consumer_quit state that we just set so to quit gracefully.
3201 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3203 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3205 notify_health_quit_pipe(health_quit_pipe
);
3207 /* Cleaning up possibly open sockets. */
3211 PERROR("close sock sessiond poll");
3214 if (client_socket
>= 0) {
3215 ret
= close(client_socket
);
3217 PERROR("close client_socket sessiond poll");
3224 ERR("Health error occurred in %s", __func__
);
3226 health_unregister(health_consumerd
);
3228 rcu_unregister_thread();
3232 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3233 struct lttng_consumer_local_data
*ctx
)
3237 pthread_mutex_lock(&stream
->lock
);
3238 if (stream
->metadata_flag
) {
3239 pthread_mutex_lock(&stream
->metadata_rdv_lock
);
3242 switch (consumer_data
.type
) {
3243 case LTTNG_CONSUMER_KERNEL
:
3244 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
3246 case LTTNG_CONSUMER32_UST
:
3247 case LTTNG_CONSUMER64_UST
:
3248 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
3251 ERR("Unknown consumer_data type");
3257 if (stream
->metadata_flag
) {
3258 pthread_cond_broadcast(&stream
->metadata_rdv
);
3259 pthread_mutex_unlock(&stream
->metadata_rdv_lock
);
3261 pthread_mutex_unlock(&stream
->lock
);
3265 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3267 switch (consumer_data
.type
) {
3268 case LTTNG_CONSUMER_KERNEL
:
3269 return lttng_kconsumer_on_recv_stream(stream
);
3270 case LTTNG_CONSUMER32_UST
:
3271 case LTTNG_CONSUMER64_UST
:
3272 return lttng_ustconsumer_on_recv_stream(stream
);
3274 ERR("Unknown consumer_data type");
3281 * Allocate and set consumer data hash tables.
3283 int lttng_consumer_init(void)
3285 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3286 if (!consumer_data
.channel_ht
) {
3290 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3291 if (!consumer_data
.relayd_ht
) {
3295 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3296 if (!consumer_data
.stream_list_ht
) {
3300 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3301 if (!consumer_data
.stream_per_chan_id_ht
) {
3305 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3310 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3322 * Process the ADD_RELAYD command receive by a consumer.
3324 * This will create a relayd socket pair and add it to the relayd hash table.
3325 * The caller MUST acquire a RCU read side lock before calling it.
3327 void consumer_add_relayd_socket(uint64_t relayd_id
, int sock_type
,
3328 struct lttng_consumer_local_data
*ctx
, int sock
,
3329 struct pollfd
*consumer_sockpoll
,
3330 struct lttcomm_relayd_sock
*relayd_sock
, uint64_t sessiond_id
,
3331 uint64_t relayd_session_id
)
3333 int fd
= -1, ret
= -1, relayd_created
= 0;
3334 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3335 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3338 assert(relayd_sock
);
3340 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", relayd_id
);
3342 /* Get relayd reference if exists. */
3343 relayd
= consumer_find_relayd(relayd_id
);
3344 if (relayd
== NULL
) {
3345 assert(sock_type
== LTTNG_STREAM_CONTROL
);
3346 /* Not found. Allocate one. */
3347 relayd
= consumer_allocate_relayd_sock_pair(relayd_id
);
3348 if (relayd
== NULL
) {
3350 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3353 relayd
->sessiond_session_id
= sessiond_id
;
3358 * This code path MUST continue to the consumer send status message to
3359 * we can notify the session daemon and continue our work without
3360 * killing everything.
3364 * relayd key should never be found for control socket.
3366 assert(sock_type
!= LTTNG_STREAM_CONTROL
);
3369 /* First send a status message before receiving the fds. */
3370 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3372 /* Somehow, the session daemon is not responding anymore. */
3373 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3374 goto error_nosignal
;
3377 /* Poll on consumer socket. */
3378 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3380 /* Needing to exit in the middle of a command: error. */
3381 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3383 goto error_nosignal
;
3386 /* Get relayd socket from session daemon */
3387 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3388 if (ret
!= sizeof(fd
)) {
3390 fd
= -1; /* Just in case it gets set with an invalid value. */
3393 * Failing to receive FDs might indicate a major problem such as
3394 * reaching a fd limit during the receive where the kernel returns a
3395 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3396 * don't take any chances and stop everything.
3398 * XXX: Feature request #558 will fix that and avoid this possible
3399 * issue when reaching the fd limit.
3401 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3402 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3406 /* Copy socket information and received FD */
3407 switch (sock_type
) {
3408 case LTTNG_STREAM_CONTROL
:
3409 /* Copy received lttcomm socket */
3410 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3411 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3412 /* Handle create_sock error. */
3414 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3418 * Close the socket created internally by
3419 * lttcomm_create_sock, so we can replace it by the one
3420 * received from sessiond.
3422 if (close(relayd
->control_sock
.sock
.fd
)) {
3426 /* Assign new file descriptor */
3427 relayd
->control_sock
.sock
.fd
= fd
;
3428 fd
= -1; /* For error path */
3429 /* Assign version values. */
3430 relayd
->control_sock
.major
= relayd_sock
->major
;
3431 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3433 relayd
->relayd_session_id
= relayd_session_id
;
3436 case LTTNG_STREAM_DATA
:
3437 /* Copy received lttcomm socket */
3438 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3439 ret
= lttcomm_create_sock(&relayd
->data_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
->data_sock
.sock
.fd
)) {
3454 /* Assign new file descriptor */
3455 relayd
->data_sock
.sock
.fd
= fd
;
3456 fd
= -1; /* for eventual error paths */
3457 /* Assign version values. */
3458 relayd
->data_sock
.major
= relayd_sock
->major
;
3459 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3462 ERR("Unknown relayd socket type (%d)", sock_type
);
3464 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3468 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3469 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3472 /* We successfully added the socket. Send status back. */
3473 ret
= consumer_send_status_msg(sock
, ret_code
);
3475 /* Somehow, the session daemon is not responding anymore. */
3476 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3477 goto error_nosignal
;
3481 * Add relayd socket pair to consumer data hashtable. If object already
3482 * exists or on error, the function gracefully returns.
3491 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3492 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3496 /* Close received socket if valid. */
3499 PERROR("close received socket");
3503 if (relayd_created
) {
3509 * Search for a relayd associated to the session id and return the reference.
3511 * A rcu read side lock MUST be acquire before calling this function and locked
3512 * until the relayd object is no longer necessary.
3514 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3516 struct lttng_ht_iter iter
;
3517 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3519 /* Iterate over all relayd since they are indexed by relayd_id. */
3520 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3523 * Check by sessiond id which is unique here where the relayd session
3524 * id might not be when having multiple relayd.
3526 if (relayd
->sessiond_session_id
== id
) {
3527 /* Found the relayd. There can be only one per id. */
3539 * Check if for a given session id there is still data needed to be extract
3542 * Return 1 if data is pending or else 0 meaning ready to be read.
3544 int consumer_data_pending(uint64_t id
)
3547 struct lttng_ht_iter iter
;
3548 struct lttng_ht
*ht
;
3549 struct lttng_consumer_stream
*stream
;
3550 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3551 int (*data_pending
)(struct lttng_consumer_stream
*);
3553 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3556 pthread_mutex_lock(&consumer_data
.lock
);
3558 switch (consumer_data
.type
) {
3559 case LTTNG_CONSUMER_KERNEL
:
3560 data_pending
= lttng_kconsumer_data_pending
;
3562 case LTTNG_CONSUMER32_UST
:
3563 case LTTNG_CONSUMER64_UST
:
3564 data_pending
= lttng_ustconsumer_data_pending
;
3567 ERR("Unknown consumer data type");
3571 /* Ease our life a bit */
3572 ht
= consumer_data
.stream_list_ht
;
3574 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3575 ht
->hash_fct(&id
, lttng_ht_seed
),
3577 &iter
.iter
, stream
, node_session_id
.node
) {
3578 pthread_mutex_lock(&stream
->lock
);
3581 * A removed node from the hash table indicates that the stream has
3582 * been deleted thus having a guarantee that the buffers are closed
3583 * on the consumer side. However, data can still be transmitted
3584 * over the network so don't skip the relayd check.
3586 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3588 /* Check the stream if there is data in the buffers. */
3589 ret
= data_pending(stream
);
3591 DBG("Data is pending locally on stream %" PRIu64
, stream
->key
);
3592 pthread_mutex_unlock(&stream
->lock
);
3597 pthread_mutex_unlock(&stream
->lock
);
3600 relayd
= find_relayd_by_session_id(id
);
3602 unsigned int is_data_inflight
= 0;
3604 /* Send init command for data pending. */
3605 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3606 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3607 relayd
->relayd_session_id
);
3609 /* Communication error thus the relayd so no data pending. */
3610 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3611 ERR("Relayd begin data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->id
);
3612 lttng_consumer_cleanup_relayd(relayd
);
3613 goto data_not_pending
;
3616 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3617 ht
->hash_fct(&id
, lttng_ht_seed
),
3619 &iter
.iter
, stream
, node_session_id
.node
) {
3620 if (stream
->metadata_flag
) {
3621 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3622 stream
->relayd_stream_id
);
3624 ret
= relayd_data_pending(&relayd
->control_sock
,
3625 stream
->relayd_stream_id
,
3626 stream
->next_net_seq_num
- 1);
3629 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3633 ERR("Relayd data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->id
);
3634 lttng_consumer_cleanup_relayd(relayd
);
3635 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3636 goto data_not_pending
;
3640 /* Send end command for data pending. */
3641 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3642 relayd
->relayd_session_id
, &is_data_inflight
);
3643 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3645 ERR("Relayd end data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->id
);
3646 lttng_consumer_cleanup_relayd(relayd
);
3647 goto data_not_pending
;
3649 if (is_data_inflight
) {
3650 DBG("Data is in flight on relayd %" PRIu64
, relayd
->id
);
3656 * Finding _no_ node in the hash table and no inflight data means that the
3657 * stream(s) have been removed thus data is guaranteed to be available for
3658 * analysis from the trace files.
3662 /* Data is available to be read by a viewer. */
3663 pthread_mutex_unlock(&consumer_data
.lock
);
3668 /* Data is still being extracted from buffers. */
3669 pthread_mutex_unlock(&consumer_data
.lock
);
3675 * Send a ret code status message to the sessiond daemon.
3677 * Return the sendmsg() return value.
3679 int consumer_send_status_msg(int sock
, int ret_code
)
3681 struct lttcomm_consumer_status_msg msg
;
3683 memset(&msg
, 0, sizeof(msg
));
3684 msg
.ret_code
= ret_code
;
3686 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3690 * Send a channel status message to the sessiond daemon.
3692 * Return the sendmsg() return value.
3694 int consumer_send_status_channel(int sock
,
3695 struct lttng_consumer_channel
*channel
)
3697 struct lttcomm_consumer_status_channel msg
;
3701 memset(&msg
, 0, sizeof(msg
));
3703 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3705 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3706 msg
.key
= channel
->key
;
3707 msg
.stream_count
= channel
->streams
.count
;
3710 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3713 unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos
,
3714 unsigned long produced_pos
, uint64_t nb_packets_per_stream
,
3715 uint64_t max_sb_size
)
3717 unsigned long start_pos
;
3719 if (!nb_packets_per_stream
) {
3720 return consumed_pos
; /* Grab everything */
3722 start_pos
= produced_pos
- offset_align_floor(produced_pos
, max_sb_size
);
3723 start_pos
-= max_sb_size
* nb_packets_per_stream
;
3724 if ((long) (start_pos
- consumed_pos
) < 0) {
3725 return consumed_pos
; /* Grab everything */