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(
536 struct lttng_consumer_channel
*channel
,
537 uint64_t channel_key
,
539 enum lttng_consumer_stream_state state
,
540 const char *channel_name
,
547 enum consumer_channel_type type
,
548 unsigned int monitor
)
551 struct lttng_consumer_stream
*stream
;
553 stream
= zmalloc(sizeof(*stream
));
554 if (stream
== NULL
) {
555 PERROR("malloc struct lttng_consumer_stream");
561 stream
->chan
= channel
;
562 stream
->key
= stream_key
;
564 stream
->out_fd_offset
= 0;
565 stream
->output_written
= 0;
566 stream
->state
= state
;
569 stream
->relayd_id
= relayd_id
;
570 stream
->session_id
= session_id
;
571 stream
->monitor
= monitor
;
572 stream
->endpoint_status
= CONSUMER_ENDPOINT_ACTIVE
;
573 stream
->index_file
= NULL
;
574 stream
->last_sequence_number
= -1ULL;
575 pthread_mutex_init(&stream
->lock
, NULL
);
576 pthread_mutex_init(&stream
->metadata_timer_lock
, NULL
);
578 /* If channel is the metadata, flag this stream as metadata. */
579 if (type
== CONSUMER_CHANNEL_TYPE_METADATA
) {
580 stream
->metadata_flag
= 1;
581 /* Metadata is flat out. */
582 strncpy(stream
->name
, DEFAULT_METADATA_NAME
, sizeof(stream
->name
));
583 /* Live rendez-vous point. */
584 pthread_cond_init(&stream
->metadata_rdv
, NULL
);
585 pthread_mutex_init(&stream
->metadata_rdv_lock
, NULL
);
587 /* Format stream name to <channel_name>_<cpu_number> */
588 ret
= snprintf(stream
->name
, sizeof(stream
->name
), "%s_%d",
591 PERROR("snprintf stream name");
596 /* Key is always the wait_fd for streams. */
597 lttng_ht_node_init_u64(&stream
->node
, stream
->key
);
599 /* Init node per channel id key */
600 lttng_ht_node_init_u64(&stream
->node_channel_id
, channel_key
);
602 /* Init session id node with the stream session id */
603 lttng_ht_node_init_u64(&stream
->node_session_id
, stream
->session_id
);
605 DBG3("Allocated stream %s (key %" PRIu64
", chan_key %" PRIu64
606 " relayd_id %" PRIu64
", session_id %" PRIu64
,
607 stream
->name
, stream
->key
, channel_key
,
608 stream
->relayd_id
, stream
->session_id
);
624 * Add a stream to the global list protected by a mutex.
626 int consumer_add_data_stream(struct lttng_consumer_stream
*stream
)
628 struct lttng_ht
*ht
= data_ht
;
634 DBG3("Adding consumer stream %" PRIu64
, stream
->key
);
636 pthread_mutex_lock(&consumer_data
.lock
);
637 pthread_mutex_lock(&stream
->chan
->lock
);
638 pthread_mutex_lock(&stream
->chan
->timer_lock
);
639 pthread_mutex_lock(&stream
->lock
);
642 /* Steal stream identifier to avoid having streams with the same key */
643 steal_stream_key(stream
->key
, ht
);
645 lttng_ht_add_unique_u64(ht
, &stream
->node
);
647 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
648 &stream
->node_channel_id
);
651 * Add stream to the stream_list_ht of the consumer data. No need to steal
652 * the key since the HT does not use it and we allow to add redundant keys
655 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
658 * When nb_init_stream_left reaches 0, we don't need to trigger any action
659 * in terms of destroying the associated channel, because the action that
660 * causes the count to become 0 also causes a stream to be added. The
661 * channel deletion will thus be triggered by the following removal of this
664 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
665 /* Increment refcount before decrementing nb_init_stream_left */
667 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
670 /* Update consumer data once the node is inserted. */
671 consumer_data
.stream_count
++;
672 consumer_data
.need_update
= 1;
675 pthread_mutex_unlock(&stream
->lock
);
676 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
677 pthread_mutex_unlock(&stream
->chan
->lock
);
678 pthread_mutex_unlock(&consumer_data
.lock
);
683 void consumer_del_data_stream(struct lttng_consumer_stream
*stream
)
685 consumer_del_stream(stream
, data_ht
);
689 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
690 * be acquired before calling this.
692 static int add_relayd(struct consumer_relayd_sock_pair
*relayd
)
695 struct lttng_ht_node_u64
*node
;
696 struct lttng_ht_iter iter
;
700 lttng_ht_lookup(consumer_data
.relayd_ht
,
702 node
= lttng_ht_iter_get_node_u64(&iter
);
706 lttng_ht_add_unique_u64(consumer_data
.relayd_ht
, &relayd
->node
);
713 * Allocate and return a consumer relayd socket.
715 static struct consumer_relayd_sock_pair
*consumer_allocate_relayd_sock_pair(
718 struct consumer_relayd_sock_pair
*obj
= NULL
;
720 /* net sequence index of -1 is a failure */
721 if (relayd_id
== (uint64_t) -1ULL) {
725 obj
= zmalloc(sizeof(struct consumer_relayd_sock_pair
));
727 PERROR("zmalloc relayd sock");
733 obj
->destroy_flag
= 0;
734 obj
->control_sock
.sock
.fd
= -1;
735 obj
->data_sock
.sock
.fd
= -1;
736 lttng_ht_node_init_u64(&obj
->node
, obj
->id
);
737 pthread_mutex_init(&obj
->ctrl_sock_mutex
, NULL
);
744 * Find a relayd socket pair in the global consumer data.
746 * Return the object if found else NULL.
747 * RCU read-side lock must be held across this call and while using the
750 struct consumer_relayd_sock_pair
*consumer_find_relayd(uint64_t key
)
752 struct lttng_ht_iter iter
;
753 struct lttng_ht_node_u64
*node
;
754 struct consumer_relayd_sock_pair
*relayd
= NULL
;
756 /* Negative keys are lookup failures */
757 if (key
== (uint64_t) -1ULL) {
761 lttng_ht_lookup(consumer_data
.relayd_ht
, &key
,
763 node
= lttng_ht_iter_get_node_u64(&iter
);
765 relayd
= caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
773 * Find a relayd and send the stream
775 * Returns 0 on success, < 0 on error
777 int consumer_send_relayd_stream(struct lttng_consumer_stream
*stream
,
781 struct consumer_relayd_sock_pair
*relayd
;
784 assert(stream
->relayd_id
!= -1ULL);
787 /* The stream is not metadata. Get relayd reference if exists. */
789 relayd
= consumer_find_relayd(stream
->relayd_id
);
790 if (relayd
!= NULL
) {
791 /* Add stream on the relayd */
792 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
793 ret
= relayd_add_stream(&relayd
->control_sock
, stream
->name
,
794 path
, &stream
->relayd_stream_id
,
795 stream
->chan
->tracefile_size
, stream
->chan
->tracefile_count
);
796 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
798 ERR("Relayd add stream failed. Cleaning up relayd %" PRIu64
".", relayd
->id
);
799 lttng_consumer_cleanup_relayd(relayd
);
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
->relayd_id
);
812 DBG("Stream %s with key %" PRIu64
" sent to relayd id %" PRIu64
,
813 stream
->name
, stream
->key
, stream
->relayd_id
);
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 relayd_id
)
828 struct consumer_relayd_sock_pair
*relayd
;
830 assert(relayd_id
!= -1ULL);
832 /* The stream is not metadata. Get relayd reference if exists. */
834 relayd
= consumer_find_relayd(relayd_id
);
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
);
841 ERR("Relayd streams sent failed. Cleaning up relayd %" PRIu64
".", relayd
->id
);
842 lttng_consumer_cleanup_relayd(relayd
);
846 ERR("Relayd ID %" PRIu64
" unknown. Can't send streams_sent.",
853 DBG("All streams sent relayd id %" PRIu64
, relayd_id
);
861 * Find a relayd and close the stream
863 void close_relayd_stream(struct lttng_consumer_stream
*stream
)
865 struct consumer_relayd_sock_pair
*relayd
;
867 /* The stream is not metadata. Get relayd reference if exists. */
869 relayd
= consumer_find_relayd(stream
->relayd_id
);
871 consumer_stream_relayd_close(stream
, relayd
);
877 * Handle stream for relayd transmission if the stream applies for network
878 * streaming where the net sequence index is set.
880 * Return destination file descriptor or negative value on error.
882 static int write_relayd_stream_header(struct lttng_consumer_stream
*stream
,
883 size_t data_size
, unsigned long padding
,
884 struct consumer_relayd_sock_pair
*relayd
)
887 struct lttcomm_relayd_data_hdr data_hdr
;
893 /* Reset data header */
894 memset(&data_hdr
, 0, sizeof(data_hdr
));
896 if (stream
->metadata_flag
) {
897 /* Caller MUST acquire the relayd control socket lock */
898 ret
= relayd_send_metadata(&relayd
->control_sock
, data_size
);
903 /* Metadata are always sent on the control socket. */
904 outfd
= relayd
->control_sock
.sock
.fd
;
906 /* Set header with stream information */
907 data_hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
908 data_hdr
.data_size
= htobe32(data_size
);
909 data_hdr
.padding_size
= htobe32(padding
);
911 * Note that net_seq_num below is assigned with the *current* value of
912 * next_net_seq_num and only after that the next_net_seq_num will be
913 * increment. This is why when issuing a command on the relayd using
914 * this next value, 1 should always be substracted in order to compare
915 * the last seen sequence number on the relayd side to the last sent.
917 data_hdr
.net_seq_num
= htobe64(stream
->next_net_seq_num
);
918 /* Other fields are zeroed previously */
920 ret
= relayd_send_data_hdr(&relayd
->data_sock
, &data_hdr
,
926 ++stream
->next_net_seq_num
;
928 /* Set to go on data socket */
929 outfd
= relayd
->data_sock
.sock
.fd
;
937 * Allocate and return a new lttng_consumer_channel object using the given key
938 * to initialize the hash table node.
940 * On error, return NULL.
942 struct lttng_consumer_channel
*consumer_allocate_channel(uint64_t key
,
944 const char *pathname
,
949 enum lttng_event_output output
,
950 uint64_t tracefile_size
,
951 uint64_t tracefile_count
,
952 uint64_t session_id_per_pid
,
953 unsigned int monitor
,
954 unsigned int live_timer_interval
,
955 const char *root_shm_path
,
956 const char *shm_path
)
958 struct lttng_consumer_channel
*channel
;
960 channel
= zmalloc(sizeof(*channel
));
961 if (channel
== NULL
) {
962 PERROR("malloc struct lttng_consumer_channel");
967 channel
->refcount
= 0;
968 channel
->session_id
= session_id
;
969 channel
->session_id_per_pid
= session_id_per_pid
;
972 channel
->relayd_id
= relayd_id
;
973 channel
->tracefile_size
= tracefile_size
;
974 channel
->tracefile_count
= tracefile_count
;
975 channel
->monitor
= monitor
;
976 channel
->live_timer_interval
= live_timer_interval
;
977 pthread_mutex_init(&channel
->lock
, NULL
);
978 pthread_mutex_init(&channel
->timer_lock
, NULL
);
981 case LTTNG_EVENT_SPLICE
:
982 channel
->output
= CONSUMER_CHANNEL_SPLICE
;
984 case LTTNG_EVENT_MMAP
:
985 channel
->output
= CONSUMER_CHANNEL_MMAP
;
995 * In monitor mode, the streams associated with the channel will be put in
996 * a special list ONLY owned by this channel. So, the refcount is set to 1
997 * here meaning that the channel itself has streams that are referenced.
999 * On a channel deletion, once the channel is no longer visible, the
1000 * refcount is decremented and checked for a zero value to delete it. With
1001 * streams in no monitor mode, it will now be safe to destroy the channel.
1003 if (!channel
->monitor
) {
1004 channel
->refcount
= 1;
1007 strncpy(channel
->pathname
, pathname
, sizeof(channel
->pathname
));
1008 channel
->pathname
[sizeof(channel
->pathname
) - 1] = '\0';
1010 strncpy(channel
->name
, name
, sizeof(channel
->name
));
1011 channel
->name
[sizeof(channel
->name
) - 1] = '\0';
1013 if (root_shm_path
) {
1014 strncpy(channel
->root_shm_path
, root_shm_path
, sizeof(channel
->root_shm_path
));
1015 channel
->root_shm_path
[sizeof(channel
->root_shm_path
) - 1] = '\0';
1018 strncpy(channel
->shm_path
, shm_path
, sizeof(channel
->shm_path
));
1019 channel
->shm_path
[sizeof(channel
->shm_path
) - 1] = '\0';
1022 lttng_ht_node_init_u64(&channel
->node
, channel
->key
);
1024 channel
->wait_fd
= -1;
1026 CDS_INIT_LIST_HEAD(&channel
->streams
.head
);
1028 DBG("Allocated channel (key %" PRIu64
")", channel
->key
);
1035 * Add a channel to the global list protected by a mutex.
1037 * Always return 0 indicating success.
1039 int consumer_add_channel(struct lttng_consumer_channel
*channel
,
1040 struct lttng_consumer_local_data
*ctx
)
1042 pthread_mutex_lock(&consumer_data
.lock
);
1043 pthread_mutex_lock(&channel
->lock
);
1044 pthread_mutex_lock(&channel
->timer_lock
);
1047 * This gives us a guarantee that the channel we are about to add to the
1048 * channel hash table will be unique. See this function comment on the why
1049 * we need to steel the channel key at this stage.
1051 steal_channel_key(channel
->key
);
1054 lttng_ht_add_unique_u64(consumer_data
.channel_ht
, &channel
->node
);
1057 pthread_mutex_unlock(&channel
->timer_lock
);
1058 pthread_mutex_unlock(&channel
->lock
);
1059 pthread_mutex_unlock(&consumer_data
.lock
);
1061 if (channel
->wait_fd
!= -1 && channel
->type
== CONSUMER_CHANNEL_TYPE_DATA
) {
1062 notify_channel_pipe(ctx
, channel
, -1, CONSUMER_CHANNEL_ADD
);
1069 * Allocate the pollfd structure and the local view of the out fds to avoid
1070 * doing a lookup in the linked list and concurrency issues when writing is
1071 * needed. Called with consumer_data.lock held.
1073 * Returns the number of fds in the structures.
1075 static int update_poll_array(struct lttng_consumer_local_data
*ctx
,
1076 struct pollfd
**pollfd
, struct lttng_consumer_stream
**local_stream
,
1077 struct lttng_ht
*ht
, int *nb_inactive_fd
)
1080 struct lttng_ht_iter iter
;
1081 struct lttng_consumer_stream
*stream
;
1086 assert(local_stream
);
1088 DBG("Updating poll fd array");
1089 *nb_inactive_fd
= 0;
1091 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1093 * Only active streams with an active end point can be added to the
1094 * poll set and local stream storage of the thread.
1096 * There is a potential race here for endpoint_status to be updated
1097 * just after the check. However, this is OK since the stream(s) will
1098 * be deleted once the thread is notified that the end point state has
1099 * changed where this function will be called back again.
1101 * We track the number of inactive FDs because they still need to be
1102 * closed by the polling thread after a wakeup on the data_pipe or
1105 if (stream
->state
!= LTTNG_CONSUMER_ACTIVE_STREAM
||
1106 stream
->endpoint_status
== CONSUMER_ENDPOINT_INACTIVE
) {
1107 (*nb_inactive_fd
)++;
1111 * This clobbers way too much the debug output. Uncomment that if you
1112 * need it for debugging purposes.
1114 * DBG("Active FD %d", stream->wait_fd);
1116 (*pollfd
)[i
].fd
= stream
->wait_fd
;
1117 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1118 local_stream
[i
] = stream
;
1124 * Insert the consumer_data_pipe at the end of the array and don't
1125 * increment i so nb_fd is the number of real FD.
1127 (*pollfd
)[i
].fd
= lttng_pipe_get_readfd(ctx
->consumer_data_pipe
);
1128 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1130 (*pollfd
)[i
+ 1].fd
= lttng_pipe_get_readfd(ctx
->consumer_wakeup_pipe
);
1131 (*pollfd
)[i
+ 1].events
= POLLIN
| POLLPRI
;
1136 * Poll on the should_quit pipe and the command socket return -1 on
1137 * error, 1 if should exit, 0 if data is available on the command socket
1139 int lttng_consumer_poll_socket(struct pollfd
*consumer_sockpoll
)
1144 num_rdy
= poll(consumer_sockpoll
, 2, -1);
1145 if (num_rdy
== -1) {
1147 * Restart interrupted system call.
1149 if (errno
== EINTR
) {
1152 PERROR("Poll error");
1155 if (consumer_sockpoll
[0].revents
& (POLLIN
| POLLPRI
)) {
1156 DBG("consumer_should_quit wake up");
1163 * Set the error socket.
1165 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data
*ctx
,
1168 ctx
->consumer_error_socket
= sock
;
1172 * Set the command socket path.
1174 void lttng_consumer_set_command_sock_path(
1175 struct lttng_consumer_local_data
*ctx
, char *sock
)
1177 ctx
->consumer_command_sock_path
= sock
;
1181 * Send return code to the session daemon.
1182 * If the socket is not defined, we return 0, it is not a fatal error
1184 int lttng_consumer_send_error(struct lttng_consumer_local_data
*ctx
, int cmd
)
1186 if (ctx
->consumer_error_socket
> 0) {
1187 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
1188 sizeof(enum lttcomm_sessiond_command
));
1195 * Close all the tracefiles and stream fds and MUST be called when all
1196 * instances are destroyed i.e. when all threads were joined and are ended.
1198 void lttng_consumer_cleanup(void)
1200 struct lttng_ht_iter iter
;
1201 struct lttng_consumer_channel
*channel
;
1205 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
, channel
,
1207 consumer_del_channel(channel
);
1212 lttng_ht_destroy(consumer_data
.channel_ht
);
1214 cleanup_relayd_ht();
1216 lttng_ht_destroy(consumer_data
.stream_per_chan_id_ht
);
1219 * This HT contains streams that are freed by either the metadata thread or
1220 * the data thread so we do *nothing* on the hash table and simply destroy
1223 lttng_ht_destroy(consumer_data
.stream_list_ht
);
1227 * Called from signal handler.
1229 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
1234 ret
= lttng_write(ctx
->consumer_should_quit
[1], "4", 1);
1236 PERROR("write consumer quit");
1239 DBG("Consumer flag that it should quit");
1244 * Flush pending writes to trace output disk file.
1247 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1251 int outfd
= stream
->out_fd
;
1254 * This does a blocking write-and-wait on any page that belongs to the
1255 * subbuffer prior to the one we just wrote.
1256 * Don't care about error values, as these are just hints and ways to
1257 * limit the amount of page cache used.
1259 if (orig_offset
< stream
->max_sb_size
) {
1262 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1263 stream
->max_sb_size
,
1264 SYNC_FILE_RANGE_WAIT_BEFORE
1265 | SYNC_FILE_RANGE_WRITE
1266 | SYNC_FILE_RANGE_WAIT_AFTER
);
1268 * Give hints to the kernel about how we access the file:
1269 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1272 * We need to call fadvise again after the file grows because the
1273 * kernel does not seem to apply fadvise to non-existing parts of the
1276 * Call fadvise _after_ having waited for the page writeback to
1277 * complete because the dirty page writeback semantic is not well
1278 * defined. So it can be expected to lead to lower throughput in
1281 ret
= posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1282 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1283 if (ret
&& ret
!= -ENOSYS
) {
1285 PERROR("posix_fadvise on fd %i", outfd
);
1290 * Initialise the necessary environnement :
1291 * - create a new context
1292 * - create the poll_pipe
1293 * - create the should_quit pipe (for signal handler)
1294 * - create the thread pipe (for splice)
1296 * Takes a function pointer as argument, this function is called when data is
1297 * available on a buffer. This function is responsible to do the
1298 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1299 * buffer configuration and then kernctl_put_next_subbuf at the end.
1301 * Returns a pointer to the new context or NULL on error.
1303 struct lttng_consumer_local_data
*lttng_consumer_create(
1304 enum lttng_consumer_type type
,
1305 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1306 struct lttng_consumer_local_data
*ctx
),
1307 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1308 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1309 int (*update_stream
)(uint64_t stream_key
, uint32_t state
))
1312 struct lttng_consumer_local_data
*ctx
;
1314 assert(consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1315 consumer_data
.type
== type
);
1316 consumer_data
.type
= type
;
1318 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
1320 PERROR("allocating context");
1324 ctx
->consumer_error_socket
= -1;
1325 ctx
->consumer_metadata_socket
= -1;
1326 pthread_mutex_init(&ctx
->metadata_socket_lock
, NULL
);
1327 /* assign the callbacks */
1328 ctx
->on_buffer_ready
= buffer_ready
;
1329 ctx
->on_recv_channel
= recv_channel
;
1330 ctx
->on_recv_stream
= recv_stream
;
1331 ctx
->on_update_stream
= update_stream
;
1333 ctx
->consumer_data_pipe
= lttng_pipe_open(0);
1334 if (!ctx
->consumer_data_pipe
) {
1335 goto error_poll_pipe
;
1338 ctx
->consumer_wakeup_pipe
= lttng_pipe_open(0);
1339 if (!ctx
->consumer_wakeup_pipe
) {
1340 goto error_wakeup_pipe
;
1343 ret
= pipe(ctx
->consumer_should_quit
);
1345 PERROR("Error creating recv pipe");
1346 goto error_quit_pipe
;
1349 ret
= pipe(ctx
->consumer_channel_pipe
);
1351 PERROR("Error creating channel pipe");
1352 goto error_channel_pipe
;
1355 ctx
->consumer_metadata_pipe
= lttng_pipe_open(0);
1356 if (!ctx
->consumer_metadata_pipe
) {
1357 goto error_metadata_pipe
;
1362 error_metadata_pipe
:
1363 utils_close_pipe(ctx
->consumer_channel_pipe
);
1365 utils_close_pipe(ctx
->consumer_should_quit
);
1367 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1369 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1377 * Iterate over all streams of the hashtable and free them properly.
1379 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1381 struct lttng_ht_iter iter
;
1382 struct lttng_consumer_stream
*stream
;
1389 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1391 * Ignore return value since we are currently cleaning up so any error
1394 (void) consumer_del_stream(stream
, ht
);
1398 lttng_ht_destroy(ht
);
1402 * Iterate over all streams of the metadata hashtable and free them
1405 static void destroy_metadata_stream_ht(struct lttng_ht
*ht
)
1407 struct lttng_ht_iter iter
;
1408 struct lttng_consumer_stream
*stream
;
1415 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1417 * Ignore return value since we are currently cleaning up so any error
1420 (void) consumer_del_metadata_stream(stream
, ht
);
1424 lttng_ht_destroy(ht
);
1428 * Close all fds associated with the instance and free the context.
1430 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1434 DBG("Consumer destroying it. Closing everything.");
1440 destroy_data_stream_ht(data_ht
);
1441 destroy_metadata_stream_ht(metadata_ht
);
1443 ret
= close(ctx
->consumer_error_socket
);
1447 ret
= close(ctx
->consumer_metadata_socket
);
1451 utils_close_pipe(ctx
->consumer_channel_pipe
);
1452 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1453 lttng_pipe_destroy(ctx
->consumer_metadata_pipe
);
1454 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1455 utils_close_pipe(ctx
->consumer_should_quit
);
1457 unlink(ctx
->consumer_command_sock_path
);
1462 * Write the metadata stream id on the specified file descriptor.
1464 static int write_relayd_metadata_id(int fd
,
1465 struct lttng_consumer_stream
*stream
,
1466 struct consumer_relayd_sock_pair
*relayd
, unsigned long padding
)
1469 struct lttcomm_relayd_metadata_payload hdr
;
1471 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1472 hdr
.padding_size
= htobe32(padding
);
1473 ret
= lttng_write(fd
, (void *) &hdr
, sizeof(hdr
));
1474 if (ret
< sizeof(hdr
)) {
1476 * This error means that the fd's end is closed so ignore the PERROR
1477 * not to clubber the error output since this can happen in a normal
1480 if (errno
!= EPIPE
) {
1481 PERROR("write metadata stream id");
1483 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1485 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1486 * handle writting the missing part so report that as an error and
1487 * don't lie to the caller.
1492 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1493 stream
->relayd_stream_id
, padding
);
1500 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1501 * core function for writing trace buffers to either the local filesystem or
1504 * It must be called with the stream lock held.
1506 * Careful review MUST be put if any changes occur!
1508 * Returns the number of bytes written
1510 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1511 struct lttng_consumer_local_data
*ctx
,
1512 struct lttng_consumer_stream
*stream
,
1513 const struct lttng_buffer_view
*buffer
,
1514 unsigned long padding
,
1515 struct ctf_packet_index
*index
)
1518 off_t orig_offset
= stream
->out_fd_offset
;
1519 /* Default is on the disk */
1520 int outfd
= stream
->out_fd
;
1521 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1522 unsigned int relayd_hang_up
= 0;
1523 const size_t subbuf_content_size
= buffer
->size
- padding
;
1526 /* RCU lock for the relayd pointer */
1529 /* Flag that the current stream if set for network streaming. */
1530 if (stream
->relayd_id
!= (uint64_t) -1ULL) {
1531 relayd
= consumer_find_relayd(stream
->relayd_id
);
1532 if (relayd
== NULL
) {
1538 /* Handle stream on the relayd if the output is on the network */
1540 unsigned long netlen
= subbuf_content_size
;
1543 * Lock the control socket for the complete duration of the function
1544 * since from this point on we will use the socket.
1546 if (stream
->metadata_flag
) {
1547 /* Metadata requires the control socket. */
1548 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1549 if (stream
->reset_metadata_flag
) {
1550 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1551 stream
->relayd_stream_id
,
1552 stream
->metadata_version
);
1557 stream
->reset_metadata_flag
= 0;
1559 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1562 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1567 /* Use the returned socket. */
1570 /* Write metadata stream id before payload */
1571 if (stream
->metadata_flag
) {
1572 ret
= write_relayd_metadata_id(outfd
, stream
, relayd
, padding
);
1579 write_len
= subbuf_content_size
;
1581 /* No streaming; we have to write the full padding. */
1582 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1583 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1585 ERR("Reset metadata file");
1588 stream
->reset_metadata_flag
= 0;
1592 * Check if we need to change the tracefile before writing the packet.
1594 if (stream
->chan
->tracefile_size
> 0 &&
1595 (stream
->tracefile_size_current
+ buffer
->size
) >
1596 stream
->chan
->tracefile_size
) {
1597 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1598 stream
->name
, stream
->chan
->tracefile_size
,
1599 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1600 stream
->out_fd
, &(stream
->tracefile_count_current
),
1603 ERR("Rotating output file");
1606 outfd
= stream
->out_fd
;
1608 if (stream
->index_file
) {
1609 lttng_index_file_put(stream
->index_file
);
1610 stream
->index_file
= lttng_index_file_create(stream
->chan
->pathname
,
1611 stream
->name
, stream
->uid
, stream
->gid
,
1612 stream
->chan
->tracefile_size
,
1613 stream
->tracefile_count_current
,
1614 CTF_INDEX_MAJOR
, CTF_INDEX_MINOR
);
1615 if (!stream
->index_file
) {
1620 /* Reset current size because we just perform a rotation. */
1621 stream
->tracefile_size_current
= 0;
1622 stream
->out_fd_offset
= 0;
1625 stream
->tracefile_size_current
+= buffer
->size
;
1627 index
->offset
= htobe64(stream
->out_fd_offset
);
1630 write_len
= buffer
->size
;
1634 * This call guarantee that len or less is returned. It's impossible to
1635 * receive a ret value that is bigger than len.
1637 ret
= lttng_write(outfd
, buffer
->data
, write_len
);
1638 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, write_len
);
1639 if (ret
< 0 || ((size_t) ret
!= write_len
)) {
1641 * Report error to caller if nothing was written else at least send the
1649 /* Socket operation failed. We consider the relayd dead */
1650 if (errno
== EPIPE
|| errno
== EINVAL
|| errno
== EBADF
) {
1652 * This is possible if the fd is closed on the other side
1653 * (outfd) or any write problem. It can be verbose a bit for a
1654 * normal execution if for instance the relayd is stopped
1655 * abruptly. This can happen so set this to a DBG statement.
1657 DBG("Consumer mmap write detected relayd hang up");
1659 /* Unhandled error, print it and stop function right now. */
1660 PERROR("Error in write mmap (ret %zd != write_len %zu)", ret
,
1665 stream
->output_written
+= ret
;
1667 /* This call is useless on a socket so better save a syscall. */
1669 /* This won't block, but will start writeout asynchronously */
1670 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, write_len
,
1671 SYNC_FILE_RANGE_WRITE
);
1672 stream
->out_fd_offset
+= write_len
;
1673 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1678 * This is a special case that the relayd has closed its socket. Let's
1679 * cleanup the relayd object and all associated streams.
1681 if (relayd
&& relayd_hang_up
) {
1682 ERR("Relayd hangup. Cleaning up relayd %" PRIu64
".", relayd
->id
);
1683 lttng_consumer_cleanup_relayd(relayd
);
1687 /* Unlock only if ctrl socket used */
1688 if (relayd
&& stream
->metadata_flag
) {
1689 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1697 * Splice the data from the ring buffer to the tracefile.
1699 * It must be called with the stream lock held.
1701 * Returns the number of bytes spliced.
1703 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1704 struct lttng_consumer_local_data
*ctx
,
1705 struct lttng_consumer_stream
*stream
, unsigned long len
,
1706 unsigned long padding
,
1707 struct ctf_packet_index
*index
)
1709 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1711 off_t orig_offset
= stream
->out_fd_offset
;
1712 int fd
= stream
->wait_fd
;
1713 /* Default is on the disk */
1714 int outfd
= stream
->out_fd
;
1715 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1717 unsigned int relayd_hang_up
= 0;
1719 switch (consumer_data
.type
) {
1720 case LTTNG_CONSUMER_KERNEL
:
1722 case LTTNG_CONSUMER32_UST
:
1723 case LTTNG_CONSUMER64_UST
:
1724 /* Not supported for user space tracing */
1727 ERR("Unknown consumer_data type");
1731 /* RCU lock for the relayd pointer */
1734 /* Flag that the current stream if set for network streaming. */
1735 if (stream
->relayd_id
!= (uint64_t) -1ULL) {
1736 relayd
= consumer_find_relayd(stream
->relayd_id
);
1737 if (relayd
== NULL
) {
1742 splice_pipe
= stream
->splice_pipe
;
1744 /* Write metadata stream id before payload */
1746 unsigned long total_len
= len
;
1748 if (stream
->metadata_flag
) {
1750 * Lock the control socket for the complete duration of the function
1751 * since from this point on we will use the socket.
1753 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1755 if (stream
->reset_metadata_flag
) {
1756 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1757 stream
->relayd_stream_id
,
1758 stream
->metadata_version
);
1763 stream
->reset_metadata_flag
= 0;
1765 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
, relayd
,
1773 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1776 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1782 /* Use the returned socket. */
1785 /* No streaming, we have to set the len with the full padding */
1788 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1789 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1791 ERR("Reset metadata file");
1794 stream
->reset_metadata_flag
= 0;
1797 * Check if we need to change the tracefile before writing the packet.
1799 if (stream
->chan
->tracefile_size
> 0 &&
1800 (stream
->tracefile_size_current
+ len
) >
1801 stream
->chan
->tracefile_size
) {
1802 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1803 stream
->name
, stream
->chan
->tracefile_size
,
1804 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1805 stream
->out_fd
, &(stream
->tracefile_count_current
),
1809 ERR("Rotating output file");
1812 outfd
= stream
->out_fd
;
1814 if (stream
->index_file
) {
1815 lttng_index_file_put(stream
->index_file
);
1816 stream
->index_file
= lttng_index_file_create(stream
->chan
->pathname
,
1817 stream
->name
, stream
->uid
, stream
->gid
,
1818 stream
->chan
->tracefile_size
,
1819 stream
->tracefile_count_current
,
1820 CTF_INDEX_MAJOR
, CTF_INDEX_MINOR
);
1821 if (!stream
->index_file
) {
1826 /* Reset current size because we just perform a rotation. */
1827 stream
->tracefile_size_current
= 0;
1828 stream
->out_fd_offset
= 0;
1831 stream
->tracefile_size_current
+= len
;
1832 index
->offset
= htobe64(stream
->out_fd_offset
);
1836 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1837 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1838 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1839 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1840 DBG("splice chan to pipe, ret %zd", ret_splice
);
1841 if (ret_splice
< 0) {
1844 PERROR("Error in relay splice");
1848 /* Handle stream on the relayd if the output is on the network */
1849 if (relayd
&& stream
->metadata_flag
) {
1850 size_t metadata_payload_size
=
1851 sizeof(struct lttcomm_relayd_metadata_payload
);
1853 /* Update counter to fit the spliced data */
1854 ret_splice
+= metadata_payload_size
;
1855 len
+= metadata_payload_size
;
1857 * We do this so the return value can match the len passed as
1858 * argument to this function.
1860 written
-= metadata_payload_size
;
1863 /* Splice data out */
1864 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1865 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1866 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
1868 if (ret_splice
< 0) {
1873 } else if (ret_splice
> len
) {
1875 * We don't expect this code path to be executed but you never know
1876 * so this is an extra protection agains a buggy splice().
1879 written
+= ret_splice
;
1880 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice
,
1884 /* All good, update current len and continue. */
1888 /* This call is useless on a socket so better save a syscall. */
1890 /* This won't block, but will start writeout asynchronously */
1891 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1892 SYNC_FILE_RANGE_WRITE
);
1893 stream
->out_fd_offset
+= ret_splice
;
1895 stream
->output_written
+= ret_splice
;
1896 written
+= ret_splice
;
1899 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1905 * This is a special case that the relayd has closed its socket. Let's
1906 * cleanup the relayd object and all associated streams.
1908 if (relayd
&& relayd_hang_up
) {
1909 ERR("Relayd hangup. Cleaning up relayd %" PRIu64
".", relayd
->id
);
1910 lttng_consumer_cleanup_relayd(relayd
);
1911 /* Skip splice error so the consumer does not fail */
1916 /* send the appropriate error description to sessiond */
1919 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1922 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1925 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
1930 if (relayd
&& stream
->metadata_flag
) {
1931 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1939 * Take a snapshot for a specific fd
1941 * Returns 0 on success, < 0 on error
1943 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
1945 switch (consumer_data
.type
) {
1946 case LTTNG_CONSUMER_KERNEL
:
1947 return lttng_kconsumer_take_snapshot(stream
);
1948 case LTTNG_CONSUMER32_UST
:
1949 case LTTNG_CONSUMER64_UST
:
1950 return lttng_ustconsumer_take_snapshot(stream
);
1952 ERR("Unknown consumer_data type");
1959 * Get the produced position
1961 * Returns 0 on success, < 0 on error
1963 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
1966 switch (consumer_data
.type
) {
1967 case LTTNG_CONSUMER_KERNEL
:
1968 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
1969 case LTTNG_CONSUMER32_UST
:
1970 case LTTNG_CONSUMER64_UST
:
1971 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
1973 ERR("Unknown consumer_data type");
1979 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
1980 int sock
, struct pollfd
*consumer_sockpoll
)
1982 switch (consumer_data
.type
) {
1983 case LTTNG_CONSUMER_KERNEL
:
1984 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1985 case LTTNG_CONSUMER32_UST
:
1986 case LTTNG_CONSUMER64_UST
:
1987 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1989 ERR("Unknown consumer_data type");
1995 void lttng_consumer_close_all_metadata(void)
1997 switch (consumer_data
.type
) {
1998 case LTTNG_CONSUMER_KERNEL
:
2000 * The Kernel consumer has a different metadata scheme so we don't
2001 * close anything because the stream will be closed by the session
2005 case LTTNG_CONSUMER32_UST
:
2006 case LTTNG_CONSUMER64_UST
:
2008 * Close all metadata streams. The metadata hash table is passed and
2009 * this call iterates over it by closing all wakeup fd. This is safe
2010 * because at this point we are sure that the metadata producer is
2011 * either dead or blocked.
2013 lttng_ustconsumer_close_all_metadata(metadata_ht
);
2016 ERR("Unknown consumer_data type");
2022 * Clean up a metadata stream and free its memory.
2024 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
2025 struct lttng_ht
*ht
)
2027 struct lttng_consumer_channel
*free_chan
= NULL
;
2031 * This call should NEVER receive regular stream. It must always be
2032 * metadata stream and this is crucial for data structure synchronization.
2034 assert(stream
->metadata_flag
);
2036 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
2038 pthread_mutex_lock(&consumer_data
.lock
);
2039 pthread_mutex_lock(&stream
->chan
->lock
);
2040 pthread_mutex_lock(&stream
->lock
);
2041 if (stream
->chan
->metadata_cache
) {
2042 /* Only applicable to userspace consumers. */
2043 pthread_mutex_lock(&stream
->chan
->metadata_cache
->lock
);
2046 /* Remove any reference to that stream. */
2047 consumer_stream_delete(stream
, ht
);
2049 /* Close down everything including the relayd if one. */
2050 consumer_stream_close(stream
);
2051 /* Destroy tracer buffers of the stream. */
2052 consumer_stream_destroy_buffers(stream
);
2054 /* Atomically decrement channel refcount since other threads can use it. */
2055 if (!uatomic_sub_return(&stream
->chan
->refcount
, 1)
2056 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
2057 /* Go for channel deletion! */
2058 free_chan
= stream
->chan
;
2062 * Nullify the stream reference so it is not used after deletion. The
2063 * channel lock MUST be acquired before being able to check for a NULL
2066 stream
->chan
->metadata_stream
= NULL
;
2068 if (stream
->chan
->metadata_cache
) {
2069 pthread_mutex_unlock(&stream
->chan
->metadata_cache
->lock
);
2071 pthread_mutex_unlock(&stream
->lock
);
2072 pthread_mutex_unlock(&stream
->chan
->lock
);
2073 pthread_mutex_unlock(&consumer_data
.lock
);
2076 consumer_del_channel(free_chan
);
2079 consumer_stream_free(stream
);
2083 * Action done with the metadata stream when adding it to the consumer internal
2084 * data structures to handle it.
2086 int consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
2088 struct lttng_ht
*ht
= metadata_ht
;
2090 struct lttng_ht_iter iter
;
2091 struct lttng_ht_node_u64
*node
;
2096 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2098 pthread_mutex_lock(&consumer_data
.lock
);
2099 pthread_mutex_lock(&stream
->chan
->lock
);
2100 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2101 pthread_mutex_lock(&stream
->lock
);
2104 * From here, refcounts are updated so be _careful_ when returning an error
2111 * Lookup the stream just to make sure it does not exist in our internal
2112 * state. This should NEVER happen.
2114 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2115 node
= lttng_ht_iter_get_node_u64(&iter
);
2119 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2120 * in terms of destroying the associated channel, because the action that
2121 * causes the count to become 0 also causes a stream to be added. The
2122 * channel deletion will thus be triggered by the following removal of this
2125 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2126 /* Increment refcount before decrementing nb_init_stream_left */
2128 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2131 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2133 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
2134 &stream
->node_channel_id
);
2137 * Add stream to the stream_list_ht of the consumer data. No need to steal
2138 * the key since the HT does not use it and we allow to add redundant keys
2141 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2145 pthread_mutex_unlock(&stream
->lock
);
2146 pthread_mutex_unlock(&stream
->chan
->lock
);
2147 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2148 pthread_mutex_unlock(&consumer_data
.lock
);
2153 * Delete data stream that are flagged for deletion (endpoint_status).
2155 static void validate_endpoint_status_data_stream(void)
2157 struct lttng_ht_iter iter
;
2158 struct lttng_consumer_stream
*stream
;
2160 DBG("Consumer delete flagged data stream");
2163 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2164 /* Validate delete flag of the stream */
2165 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2168 /* Delete it right now */
2169 consumer_del_stream(stream
, data_ht
);
2175 * Delete metadata stream that are flagged for deletion (endpoint_status).
2177 static void validate_endpoint_status_metadata_stream(
2178 struct lttng_poll_event
*pollset
)
2180 struct lttng_ht_iter iter
;
2181 struct lttng_consumer_stream
*stream
;
2183 DBG("Consumer delete flagged metadata stream");
2188 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2189 /* Validate delete flag of the stream */
2190 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2194 * Remove from pollset so the metadata thread can continue without
2195 * blocking on a deleted stream.
2197 lttng_poll_del(pollset
, stream
->wait_fd
);
2199 /* Delete it right now */
2200 consumer_del_metadata_stream(stream
, metadata_ht
);
2206 * Thread polls on metadata file descriptor and write them on disk or on the
2209 void *consumer_thread_metadata_poll(void *data
)
2211 int ret
, i
, pollfd
, err
= -1;
2212 uint32_t revents
, nb_fd
;
2213 struct lttng_consumer_stream
*stream
= NULL
;
2214 struct lttng_ht_iter iter
;
2215 struct lttng_ht_node_u64
*node
;
2216 struct lttng_poll_event events
;
2217 struct lttng_consumer_local_data
*ctx
= data
;
2220 rcu_register_thread();
2222 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2224 if (testpoint(consumerd_thread_metadata
)) {
2225 goto error_testpoint
;
2228 health_code_update();
2230 DBG("Thread metadata poll started");
2232 /* Size is set to 1 for the consumer_metadata pipe */
2233 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2235 ERR("Poll set creation failed");
2239 ret
= lttng_poll_add(&events
,
2240 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2246 DBG("Metadata main loop started");
2250 health_code_update();
2251 health_poll_entry();
2252 DBG("Metadata poll wait");
2253 ret
= lttng_poll_wait(&events
, -1);
2254 DBG("Metadata poll return from wait with %d fd(s)",
2255 LTTNG_POLL_GETNB(&events
));
2257 DBG("Metadata event caught in thread");
2259 if (errno
== EINTR
) {
2260 ERR("Poll EINTR caught");
2263 if (LTTNG_POLL_GETNB(&events
) == 0) {
2264 err
= 0; /* All is OK */
2271 /* From here, the event is a metadata wait fd */
2272 for (i
= 0; i
< nb_fd
; i
++) {
2273 health_code_update();
2275 revents
= LTTNG_POLL_GETEV(&events
, i
);
2276 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2279 /* No activity for this FD (poll implementation). */
2283 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2284 if (revents
& LPOLLIN
) {
2287 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2288 &stream
, sizeof(stream
));
2289 if (pipe_len
< sizeof(stream
)) {
2291 PERROR("read metadata stream");
2294 * Remove the pipe from the poll set and continue the loop
2295 * since their might be data to consume.
2297 lttng_poll_del(&events
,
2298 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2299 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2303 /* A NULL stream means that the state has changed. */
2304 if (stream
== NULL
) {
2305 /* Check for deleted streams. */
2306 validate_endpoint_status_metadata_stream(&events
);
2310 DBG("Adding metadata stream %d to poll set",
2313 /* Add metadata stream to the global poll events list */
2314 lttng_poll_add(&events
, stream
->wait_fd
,
2315 LPOLLIN
| LPOLLPRI
| LPOLLHUP
);
2316 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2317 DBG("Metadata thread pipe hung up");
2319 * Remove the pipe from the poll set and continue the loop
2320 * since their might be data to consume.
2322 lttng_poll_del(&events
,
2323 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2324 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2327 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2331 /* Handle other stream */
2337 uint64_t tmp_id
= (uint64_t) pollfd
;
2339 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2341 node
= lttng_ht_iter_get_node_u64(&iter
);
2344 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2347 if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2348 /* Get the data out of the metadata file descriptor */
2349 DBG("Metadata available on fd %d", pollfd
);
2350 assert(stream
->wait_fd
== pollfd
);
2353 health_code_update();
2355 len
= ctx
->on_buffer_ready(stream
, ctx
);
2357 * We don't check the return value here since if we get
2358 * a negative len, it means an error occurred thus we
2359 * simply remove it from the poll set and free the
2364 /* It's ok to have an unavailable sub-buffer */
2365 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2366 /* Clean up stream from consumer and free it. */
2367 lttng_poll_del(&events
, stream
->wait_fd
);
2368 consumer_del_metadata_stream(stream
, metadata_ht
);
2370 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2371 DBG("Metadata fd %d is hup|err.", pollfd
);
2372 if (!stream
->hangup_flush_done
2373 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2374 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2375 DBG("Attempting to flush and consume the UST buffers");
2376 lttng_ustconsumer_on_stream_hangup(stream
);
2378 /* We just flushed the stream now read it. */
2380 health_code_update();
2382 len
= ctx
->on_buffer_ready(stream
, ctx
);
2384 * We don't check the return value here since if we get
2385 * a negative len, it means an error occurred thus we
2386 * simply remove it from the poll set and free the
2392 lttng_poll_del(&events
, stream
->wait_fd
);
2394 * This call update the channel states, closes file descriptors
2395 * and securely free the stream.
2397 consumer_del_metadata_stream(stream
, metadata_ht
);
2399 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2403 /* Release RCU lock for the stream looked up */
2411 DBG("Metadata poll thread exiting");
2413 lttng_poll_clean(&events
);
2418 ERR("Health error occurred in %s", __func__
);
2420 health_unregister(health_consumerd
);
2421 rcu_unregister_thread();
2426 * This thread polls the fds in the set to consume the data and write
2427 * it to tracefile if necessary.
2429 void *consumer_thread_data_poll(void *data
)
2431 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2432 struct pollfd
*pollfd
= NULL
;
2433 /* local view of the streams */
2434 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2435 /* local view of consumer_data.fds_count */
2437 /* Number of FDs with CONSUMER_ENDPOINT_INACTIVE but still open. */
2438 int nb_inactive_fd
= 0;
2439 struct lttng_consumer_local_data
*ctx
= data
;
2442 rcu_register_thread();
2444 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2446 if (testpoint(consumerd_thread_data
)) {
2447 goto error_testpoint
;
2450 health_code_update();
2452 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
*));
2453 if (local_stream
== NULL
) {
2454 PERROR("local_stream malloc");
2459 health_code_update();
2465 * the fds set has been updated, we need to update our
2466 * local array as well
2468 pthread_mutex_lock(&consumer_data
.lock
);
2469 if (consumer_data
.need_update
) {
2474 local_stream
= NULL
;
2477 * Allocate for all fds +1 for the consumer_data_pipe and +1 for
2480 pollfd
= zmalloc((consumer_data
.stream_count
+ 2) * sizeof(struct pollfd
));
2481 if (pollfd
== NULL
) {
2482 PERROR("pollfd malloc");
2483 pthread_mutex_unlock(&consumer_data
.lock
);
2487 local_stream
= zmalloc((consumer_data
.stream_count
+ 2) *
2488 sizeof(struct lttng_consumer_stream
*));
2489 if (local_stream
== NULL
) {
2490 PERROR("local_stream malloc");
2491 pthread_mutex_unlock(&consumer_data
.lock
);
2494 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2495 data_ht
, &nb_inactive_fd
);
2497 ERR("Error in allocating pollfd or local_outfds");
2498 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2499 pthread_mutex_unlock(&consumer_data
.lock
);
2503 consumer_data
.need_update
= 0;
2505 pthread_mutex_unlock(&consumer_data
.lock
);
2507 /* No FDs and consumer_quit, consumer_cleanup the thread */
2508 if (nb_fd
== 0 && consumer_quit
== 1 && nb_inactive_fd
== 0) {
2509 err
= 0; /* All is OK */
2512 /* poll on the array of fds */
2514 DBG("polling on %d fd", nb_fd
+ 2);
2515 health_poll_entry();
2516 num_rdy
= poll(pollfd
, nb_fd
+ 2, -1);
2518 DBG("poll num_rdy : %d", num_rdy
);
2519 if (num_rdy
== -1) {
2521 * Restart interrupted system call.
2523 if (errno
== EINTR
) {
2526 PERROR("Poll error");
2527 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2529 } else if (num_rdy
== 0) {
2530 DBG("Polling thread timed out");
2535 * If the consumer_data_pipe triggered poll go directly to the
2536 * beginning of the loop to update the array. We want to prioritize
2537 * array update over low-priority reads.
2539 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2540 ssize_t pipe_readlen
;
2542 DBG("consumer_data_pipe wake up");
2543 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2544 &new_stream
, sizeof(new_stream
));
2545 if (pipe_readlen
< sizeof(new_stream
)) {
2546 PERROR("Consumer data pipe");
2547 /* Continue so we can at least handle the current stream(s). */
2552 * If the stream is NULL, just ignore it. It's also possible that
2553 * the sessiond poll thread changed the consumer_quit state and is
2554 * waking us up to test it.
2556 if (new_stream
== NULL
) {
2557 validate_endpoint_status_data_stream();
2561 /* Continue to update the local streams and handle prio ones */
2565 /* Handle wakeup pipe. */
2566 if (pollfd
[nb_fd
+ 1].revents
& (POLLIN
| POLLPRI
)) {
2568 ssize_t pipe_readlen
;
2570 pipe_readlen
= lttng_pipe_read(ctx
->consumer_wakeup_pipe
, &dummy
,
2572 if (pipe_readlen
< 0) {
2573 PERROR("Consumer data wakeup pipe");
2575 /* We've been awakened to handle stream(s). */
2576 ctx
->has_wakeup
= 0;
2579 /* Take care of high priority channels first. */
2580 for (i
= 0; i
< nb_fd
; i
++) {
2581 health_code_update();
2583 if (local_stream
[i
] == NULL
) {
2586 if (pollfd
[i
].revents
& POLLPRI
) {
2587 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2589 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2590 /* it's ok to have an unavailable sub-buffer */
2591 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2592 /* Clean the stream and free it. */
2593 consumer_del_stream(local_stream
[i
], data_ht
);
2594 local_stream
[i
] = NULL
;
2595 } else if (len
> 0) {
2596 local_stream
[i
]->data_read
= 1;
2602 * If we read high prio channel in this loop, try again
2603 * for more high prio data.
2609 /* Take care of low priority channels. */
2610 for (i
= 0; i
< nb_fd
; i
++) {
2611 health_code_update();
2613 if (local_stream
[i
] == NULL
) {
2616 if ((pollfd
[i
].revents
& POLLIN
) ||
2617 local_stream
[i
]->hangup_flush_done
||
2618 local_stream
[i
]->has_data
) {
2619 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2620 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2621 /* it's ok to have an unavailable sub-buffer */
2622 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2623 /* Clean the stream and free it. */
2624 consumer_del_stream(local_stream
[i
], data_ht
);
2625 local_stream
[i
] = NULL
;
2626 } else if (len
> 0) {
2627 local_stream
[i
]->data_read
= 1;
2632 /* Handle hangup and errors */
2633 for (i
= 0; i
< nb_fd
; i
++) {
2634 health_code_update();
2636 if (local_stream
[i
] == NULL
) {
2639 if (!local_stream
[i
]->hangup_flush_done
2640 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2641 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2642 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2643 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2645 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2646 /* Attempt read again, for the data we just flushed. */
2647 local_stream
[i
]->data_read
= 1;
2650 * If the poll flag is HUP/ERR/NVAL and we have
2651 * read no data in this pass, we can remove the
2652 * stream from its hash table.
2654 if ((pollfd
[i
].revents
& POLLHUP
)) {
2655 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2656 if (!local_stream
[i
]->data_read
) {
2657 consumer_del_stream(local_stream
[i
], data_ht
);
2658 local_stream
[i
] = NULL
;
2661 } else if (pollfd
[i
].revents
& POLLERR
) {
2662 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2663 if (!local_stream
[i
]->data_read
) {
2664 consumer_del_stream(local_stream
[i
], data_ht
);
2665 local_stream
[i
] = NULL
;
2668 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2669 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2670 if (!local_stream
[i
]->data_read
) {
2671 consumer_del_stream(local_stream
[i
], data_ht
);
2672 local_stream
[i
] = NULL
;
2676 if (local_stream
[i
] != NULL
) {
2677 local_stream
[i
]->data_read
= 0;
2684 DBG("polling thread exiting");
2689 * Close the write side of the pipe so epoll_wait() in
2690 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2691 * read side of the pipe. If we close them both, epoll_wait strangely does
2692 * not return and could create a endless wait period if the pipe is the
2693 * only tracked fd in the poll set. The thread will take care of closing
2696 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2701 ERR("Health error occurred in %s", __func__
);
2703 health_unregister(health_consumerd
);
2705 rcu_unregister_thread();
2710 * Close wake-up end of each stream belonging to the channel. This will
2711 * allow the poll() on the stream read-side to detect when the
2712 * write-side (application) finally closes them.
2715 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2717 struct lttng_ht
*ht
;
2718 struct lttng_consumer_stream
*stream
;
2719 struct lttng_ht_iter iter
;
2721 ht
= consumer_data
.stream_per_chan_id_ht
;
2724 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2725 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2726 ht
->match_fct
, &channel
->key
,
2727 &iter
.iter
, stream
, node_channel_id
.node
) {
2729 * Protect against teardown with mutex.
2731 pthread_mutex_lock(&stream
->lock
);
2732 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2735 switch (consumer_data
.type
) {
2736 case LTTNG_CONSUMER_KERNEL
:
2738 case LTTNG_CONSUMER32_UST
:
2739 case LTTNG_CONSUMER64_UST
:
2740 if (stream
->metadata_flag
) {
2741 /* Safe and protected by the stream lock. */
2742 lttng_ustconsumer_close_metadata(stream
->chan
);
2745 * Note: a mutex is taken internally within
2746 * liblttng-ust-ctl to protect timer wakeup_fd
2747 * use from concurrent close.
2749 lttng_ustconsumer_close_stream_wakeup(stream
);
2753 ERR("Unknown consumer_data type");
2757 pthread_mutex_unlock(&stream
->lock
);
2762 static void destroy_channel_ht(struct lttng_ht
*ht
)
2764 struct lttng_ht_iter iter
;
2765 struct lttng_consumer_channel
*channel
;
2773 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2774 ret
= lttng_ht_del(ht
, &iter
);
2779 lttng_ht_destroy(ht
);
2783 * This thread polls the channel fds to detect when they are being
2784 * closed. It closes all related streams if the channel is detected as
2785 * closed. It is currently only used as a shim layer for UST because the
2786 * consumerd needs to keep the per-stream wakeup end of pipes open for
2789 void *consumer_thread_channel_poll(void *data
)
2791 int ret
, i
, pollfd
, err
= -1;
2792 uint32_t revents
, nb_fd
;
2793 struct lttng_consumer_channel
*chan
= NULL
;
2794 struct lttng_ht_iter iter
;
2795 struct lttng_ht_node_u64
*node
;
2796 struct lttng_poll_event events
;
2797 struct lttng_consumer_local_data
*ctx
= data
;
2798 struct lttng_ht
*channel_ht
;
2800 rcu_register_thread();
2802 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
2804 if (testpoint(consumerd_thread_channel
)) {
2805 goto error_testpoint
;
2808 health_code_update();
2810 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2812 /* ENOMEM at this point. Better to bail out. */
2816 DBG("Thread channel poll started");
2818 /* Size is set to 1 for the consumer_channel pipe */
2819 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2821 ERR("Poll set creation failed");
2825 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2831 DBG("Channel main loop started");
2835 health_code_update();
2836 DBG("Channel poll wait");
2837 health_poll_entry();
2838 ret
= lttng_poll_wait(&events
, -1);
2839 DBG("Channel poll return from wait with %d fd(s)",
2840 LTTNG_POLL_GETNB(&events
));
2842 DBG("Channel event caught in thread");
2844 if (errno
== EINTR
) {
2845 ERR("Poll EINTR caught");
2848 if (LTTNG_POLL_GETNB(&events
) == 0) {
2849 err
= 0; /* All is OK */
2856 /* From here, the event is a channel wait fd */
2857 for (i
= 0; i
< nb_fd
; i
++) {
2858 health_code_update();
2860 revents
= LTTNG_POLL_GETEV(&events
, i
);
2861 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2864 /* No activity for this FD (poll implementation). */
2868 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2869 if (revents
& LPOLLIN
) {
2870 enum consumer_channel_action action
;
2873 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
2876 ERR("Error reading channel pipe");
2878 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2883 case CONSUMER_CHANNEL_ADD
:
2884 DBG("Adding channel %d to poll set",
2887 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
2890 lttng_ht_add_unique_u64(channel_ht
,
2891 &chan
->wait_fd_node
);
2893 /* Add channel to the global poll events list */
2894 lttng_poll_add(&events
, chan
->wait_fd
,
2895 LPOLLERR
| LPOLLHUP
);
2897 case CONSUMER_CHANNEL_DEL
:
2900 * This command should never be called if the channel
2901 * has streams monitored by either the data or metadata
2902 * thread. The consumer only notify this thread with a
2903 * channel del. command if it receives a destroy
2904 * channel command from the session daemon that send it
2905 * if a command prior to the GET_CHANNEL failed.
2909 chan
= consumer_find_channel(key
);
2912 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
2915 lttng_poll_del(&events
, chan
->wait_fd
);
2916 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
2917 ret
= lttng_ht_del(channel_ht
, &iter
);
2920 switch (consumer_data
.type
) {
2921 case LTTNG_CONSUMER_KERNEL
:
2923 case LTTNG_CONSUMER32_UST
:
2924 case LTTNG_CONSUMER64_UST
:
2925 health_code_update();
2926 /* Destroy streams that might have been left in the stream list. */
2927 clean_channel_stream_list(chan
);
2930 ERR("Unknown consumer_data type");
2935 * Release our own refcount. Force channel deletion even if
2936 * streams were not initialized.
2938 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
2939 consumer_del_channel(chan
);
2944 case CONSUMER_CHANNEL_QUIT
:
2946 * Remove the pipe from the poll set and continue the loop
2947 * since their might be data to consume.
2949 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2952 ERR("Unknown action");
2955 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2956 DBG("Channel thread pipe hung up");
2958 * Remove the pipe from the poll set and continue the loop
2959 * since their might be data to consume.
2961 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2964 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2968 /* Handle other stream */
2974 uint64_t tmp_id
= (uint64_t) pollfd
;
2976 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
2978 node
= lttng_ht_iter_get_node_u64(&iter
);
2981 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
2984 /* Check for error event */
2985 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2986 DBG("Channel fd %d is hup|err.", pollfd
);
2988 lttng_poll_del(&events
, chan
->wait_fd
);
2989 ret
= lttng_ht_del(channel_ht
, &iter
);
2993 * This will close the wait fd for each stream associated to
2994 * this channel AND monitored by the data/metadata thread thus
2995 * will be clean by the right thread.
2997 consumer_close_channel_streams(chan
);
2999 /* Release our own refcount */
3000 if (!uatomic_sub_return(&chan
->refcount
, 1)
3001 && !uatomic_read(&chan
->nb_init_stream_left
)) {
3002 consumer_del_channel(chan
);
3005 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3010 /* Release RCU lock for the channel looked up */
3018 lttng_poll_clean(&events
);
3020 destroy_channel_ht(channel_ht
);
3023 DBG("Channel poll thread exiting");
3026 ERR("Health error occurred in %s", __func__
);
3028 health_unregister(health_consumerd
);
3029 rcu_unregister_thread();
3033 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
3034 struct pollfd
*sockpoll
, int client_socket
)
3041 ret
= lttng_consumer_poll_socket(sockpoll
);
3045 DBG("Metadata connection on client_socket");
3047 /* Blocking call, waiting for transmission */
3048 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
3049 if (ctx
->consumer_metadata_socket
< 0) {
3050 WARN("On accept metadata");
3061 * This thread listens on the consumerd socket and receives the file
3062 * descriptors from the session daemon.
3064 void *consumer_thread_sessiond_poll(void *data
)
3066 int sock
= -1, client_socket
, ret
, err
= -1;
3068 * structure to poll for incoming data on communication socket avoids
3069 * making blocking sockets.
3071 struct pollfd consumer_sockpoll
[2];
3072 struct lttng_consumer_local_data
*ctx
= data
;
3074 rcu_register_thread();
3076 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
3078 if (testpoint(consumerd_thread_sessiond
)) {
3079 goto error_testpoint
;
3082 health_code_update();
3084 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
3085 unlink(ctx
->consumer_command_sock_path
);
3086 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
3087 if (client_socket
< 0) {
3088 ERR("Cannot create command socket");
3092 ret
= lttcomm_listen_unix_sock(client_socket
);
3097 DBG("Sending ready command to lttng-sessiond");
3098 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3099 /* return < 0 on error, but == 0 is not fatal */
3101 ERR("Error sending ready command to lttng-sessiond");
3105 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3106 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3107 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3108 consumer_sockpoll
[1].fd
= client_socket
;
3109 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3111 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3119 DBG("Connection on client_socket");
3121 /* Blocking call, waiting for transmission */
3122 sock
= lttcomm_accept_unix_sock(client_socket
);
3129 * Setup metadata socket which is the second socket connection on the
3130 * command unix socket.
3132 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3141 /* This socket is not useful anymore. */
3142 ret
= close(client_socket
);
3144 PERROR("close client_socket");
3148 /* update the polling structure to poll on the established socket */
3149 consumer_sockpoll
[1].fd
= sock
;
3150 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3153 health_code_update();
3155 health_poll_entry();
3156 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3165 DBG("Incoming command on sock");
3166 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3169 * This could simply be a session daemon quitting. Don't output
3172 DBG("Communication interrupted on command socket");
3176 if (consumer_quit
) {
3177 DBG("consumer_thread_receive_fds received quit from signal");
3178 err
= 0; /* All is OK */
3181 DBG("received command on sock");
3187 DBG("Consumer thread sessiond poll exiting");
3190 * Close metadata streams since the producer is the session daemon which
3193 * NOTE: for now, this only applies to the UST tracer.
3195 lttng_consumer_close_all_metadata();
3198 * when all fds have hung up, the polling thread
3204 * Notify the data poll thread to poll back again and test the
3205 * consumer_quit state that we just set so to quit gracefully.
3207 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3209 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3211 notify_health_quit_pipe(health_quit_pipe
);
3213 /* Cleaning up possibly open sockets. */
3217 PERROR("close sock sessiond poll");
3220 if (client_socket
>= 0) {
3221 ret
= close(client_socket
);
3223 PERROR("close client_socket sessiond poll");
3230 ERR("Health error occurred in %s", __func__
);
3232 health_unregister(health_consumerd
);
3234 rcu_unregister_thread();
3238 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3239 struct lttng_consumer_local_data
*ctx
)
3243 pthread_mutex_lock(&stream
->lock
);
3244 if (stream
->metadata_flag
) {
3245 pthread_mutex_lock(&stream
->metadata_rdv_lock
);
3248 switch (consumer_data
.type
) {
3249 case LTTNG_CONSUMER_KERNEL
:
3250 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
3252 case LTTNG_CONSUMER32_UST
:
3253 case LTTNG_CONSUMER64_UST
:
3254 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
3257 ERR("Unknown consumer_data type");
3263 if (stream
->metadata_flag
) {
3264 pthread_cond_broadcast(&stream
->metadata_rdv
);
3265 pthread_mutex_unlock(&stream
->metadata_rdv_lock
);
3267 pthread_mutex_unlock(&stream
->lock
);
3271 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3273 switch (consumer_data
.type
) {
3274 case LTTNG_CONSUMER_KERNEL
:
3275 return lttng_kconsumer_on_recv_stream(stream
);
3276 case LTTNG_CONSUMER32_UST
:
3277 case LTTNG_CONSUMER64_UST
:
3278 return lttng_ustconsumer_on_recv_stream(stream
);
3280 ERR("Unknown consumer_data type");
3287 * Allocate and set consumer data hash tables.
3289 int lttng_consumer_init(void)
3291 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3292 if (!consumer_data
.channel_ht
) {
3296 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3297 if (!consumer_data
.relayd_ht
) {
3301 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3302 if (!consumer_data
.stream_list_ht
) {
3306 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3307 if (!consumer_data
.stream_per_chan_id_ht
) {
3311 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3316 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3328 * Process the ADD_RELAYD command receive by a consumer.
3330 * This will create a relayd socket pair and add it to the relayd hash table.
3331 * The caller MUST acquire a RCU read side lock before calling it.
3333 void consumer_add_relayd_socket(uint64_t relayd_id
, int sock_type
,
3334 struct lttng_consumer_local_data
*ctx
, int sock
,
3335 struct pollfd
*consumer_sockpoll
,
3336 struct lttcomm_relayd_sock
*relayd_sock
, uint64_t sessiond_id
,
3337 uint64_t relayd_session_id
)
3339 int fd
= -1, ret
= -1, relayd_created
= 0;
3340 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3341 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3344 assert(relayd_sock
);
3346 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", relayd_id
);
3348 /* Get relayd reference if exists. */
3349 relayd
= consumer_find_relayd(relayd_id
);
3350 if (relayd
== NULL
) {
3351 assert(sock_type
== LTTNG_STREAM_CONTROL
);
3352 /* Not found. Allocate one. */
3353 relayd
= consumer_allocate_relayd_sock_pair(relayd_id
);
3354 if (relayd
== NULL
) {
3356 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3359 relayd
->sessiond_session_id
= sessiond_id
;
3364 * This code path MUST continue to the consumer send status message to
3365 * we can notify the session daemon and continue our work without
3366 * killing everything.
3370 * relayd key should never be found for control socket.
3372 assert(sock_type
!= LTTNG_STREAM_CONTROL
);
3375 /* First send a status message before receiving the fds. */
3376 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3378 /* Somehow, the session daemon is not responding anymore. */
3379 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3380 goto error_nosignal
;
3383 /* Poll on consumer socket. */
3384 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3386 /* Needing to exit in the middle of a command: error. */
3387 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3389 goto error_nosignal
;
3392 /* Get relayd socket from session daemon */
3393 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3394 if (ret
!= sizeof(fd
)) {
3396 fd
= -1; /* Just in case it gets set with an invalid value. */
3399 * Failing to receive FDs might indicate a major problem such as
3400 * reaching a fd limit during the receive where the kernel returns a
3401 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3402 * don't take any chances and stop everything.
3404 * XXX: Feature request #558 will fix that and avoid this possible
3405 * issue when reaching the fd limit.
3407 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3408 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3412 /* Copy socket information and received FD */
3413 switch (sock_type
) {
3414 case LTTNG_STREAM_CONTROL
:
3415 /* Copy received lttcomm socket */
3416 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3417 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3418 /* Handle create_sock error. */
3420 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3424 * Close the socket created internally by
3425 * lttcomm_create_sock, so we can replace it by the one
3426 * received from sessiond.
3428 if (close(relayd
->control_sock
.sock
.fd
)) {
3432 /* Assign new file descriptor */
3433 relayd
->control_sock
.sock
.fd
= fd
;
3434 fd
= -1; /* For error path */
3435 /* Assign version values. */
3436 relayd
->control_sock
.major
= relayd_sock
->major
;
3437 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3439 relayd
->relayd_session_id
= relayd_session_id
;
3442 case LTTNG_STREAM_DATA
:
3443 /* Copy received lttcomm socket */
3444 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3445 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3446 /* Handle create_sock error. */
3448 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3452 * Close the socket created internally by
3453 * lttcomm_create_sock, so we can replace it by the one
3454 * received from sessiond.
3456 if (close(relayd
->data_sock
.sock
.fd
)) {
3460 /* Assign new file descriptor */
3461 relayd
->data_sock
.sock
.fd
= fd
;
3462 fd
= -1; /* for eventual error paths */
3463 /* Assign version values. */
3464 relayd
->data_sock
.major
= relayd_sock
->major
;
3465 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3468 ERR("Unknown relayd socket type (%d)", sock_type
);
3470 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3474 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3475 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3478 /* We successfully added the socket. Send status back. */
3479 ret
= consumer_send_status_msg(sock
, ret_code
);
3481 /* Somehow, the session daemon is not responding anymore. */
3482 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3483 goto error_nosignal
;
3487 * Add relayd socket pair to consumer data hashtable. If object already
3488 * exists or on error, the function gracefully returns.
3497 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3498 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3502 /* Close received socket if valid. */
3505 PERROR("close received socket");
3509 if (relayd_created
) {
3515 * Search for a relayd associated to the session id and return the reference.
3517 * A rcu read side lock MUST be acquire before calling this function and locked
3518 * until the relayd object is no longer necessary.
3520 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3522 struct lttng_ht_iter iter
;
3523 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3525 /* Iterate over all relayd since they are indexed by relayd_id. */
3526 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3529 * Check by sessiond id which is unique here where the relayd session
3530 * id might not be when having multiple relayd.
3532 if (relayd
->sessiond_session_id
== id
) {
3533 /* Found the relayd. There can be only one per id. */
3545 * Check if for a given session id there is still data needed to be extract
3548 * Return 1 if data is pending or else 0 meaning ready to be read.
3550 int consumer_data_pending(uint64_t id
)
3553 struct lttng_ht_iter iter
;
3554 struct lttng_ht
*ht
;
3555 struct lttng_consumer_stream
*stream
;
3556 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3557 int (*data_pending
)(struct lttng_consumer_stream
*);
3559 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3562 pthread_mutex_lock(&consumer_data
.lock
);
3564 switch (consumer_data
.type
) {
3565 case LTTNG_CONSUMER_KERNEL
:
3566 data_pending
= lttng_kconsumer_data_pending
;
3568 case LTTNG_CONSUMER32_UST
:
3569 case LTTNG_CONSUMER64_UST
:
3570 data_pending
= lttng_ustconsumer_data_pending
;
3573 ERR("Unknown consumer data type");
3577 /* Ease our life a bit */
3578 ht
= consumer_data
.stream_list_ht
;
3580 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3581 ht
->hash_fct(&id
, lttng_ht_seed
),
3583 &iter
.iter
, stream
, node_session_id
.node
) {
3584 pthread_mutex_lock(&stream
->lock
);
3587 * A removed node from the hash table indicates that the stream has
3588 * been deleted thus having a guarantee that the buffers are closed
3589 * on the consumer side. However, data can still be transmitted
3590 * over the network so don't skip the relayd check.
3592 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3594 /* Check the stream if there is data in the buffers. */
3595 ret
= data_pending(stream
);
3597 DBG("Data is pending locally on stream %" PRIu64
, stream
->key
);
3598 pthread_mutex_unlock(&stream
->lock
);
3603 pthread_mutex_unlock(&stream
->lock
);
3606 relayd
= find_relayd_by_session_id(id
);
3608 unsigned int is_data_inflight
= 0;
3610 /* Send init command for data pending. */
3611 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3612 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3613 relayd
->relayd_session_id
);
3615 /* Communication error thus the relayd so no data pending. */
3616 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3617 ERR("Relayd begin data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->id
);
3618 lttng_consumer_cleanup_relayd(relayd
);
3619 goto data_not_pending
;
3622 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3623 ht
->hash_fct(&id
, lttng_ht_seed
),
3625 &iter
.iter
, stream
, node_session_id
.node
) {
3626 if (stream
->metadata_flag
) {
3627 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3628 stream
->relayd_stream_id
);
3630 ret
= relayd_data_pending(&relayd
->control_sock
,
3631 stream
->relayd_stream_id
,
3632 stream
->next_net_seq_num
- 1);
3635 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3639 ERR("Relayd data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->id
);
3640 lttng_consumer_cleanup_relayd(relayd
);
3641 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3642 goto data_not_pending
;
3646 /* Send end command for data pending. */
3647 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3648 relayd
->relayd_session_id
, &is_data_inflight
);
3649 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3651 ERR("Relayd end data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->id
);
3652 lttng_consumer_cleanup_relayd(relayd
);
3653 goto data_not_pending
;
3655 if (is_data_inflight
) {
3656 DBG("Data is in flight on relayd %" PRIu64
, relayd
->id
);
3662 * Finding _no_ node in the hash table and no inflight data means that the
3663 * stream(s) have been removed thus data is guaranteed to be available for
3664 * analysis from the trace files.
3668 /* Data is available to be read by a viewer. */
3669 pthread_mutex_unlock(&consumer_data
.lock
);
3674 /* Data is still being extracted from buffers. */
3675 pthread_mutex_unlock(&consumer_data
.lock
);
3681 * Send a ret code status message to the sessiond daemon.
3683 * Return the sendmsg() return value.
3685 int consumer_send_status_msg(int sock
, int ret_code
)
3687 struct lttcomm_consumer_status_msg msg
;
3689 memset(&msg
, 0, sizeof(msg
));
3690 msg
.ret_code
= ret_code
;
3692 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3696 * Send a channel status message to the sessiond daemon.
3698 * Return the sendmsg() return value.
3700 int consumer_send_status_channel(int sock
,
3701 struct lttng_consumer_channel
*channel
)
3703 struct lttcomm_consumer_status_channel msg
;
3707 memset(&msg
, 0, sizeof(msg
));
3709 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3711 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3712 msg
.key
= channel
->key
;
3713 msg
.stream_count
= channel
->streams
.count
;
3716 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3719 unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos
,
3720 unsigned long produced_pos
, uint64_t nb_packets_per_stream
,
3721 uint64_t max_sb_size
)
3723 unsigned long start_pos
;
3725 if (!nb_packets_per_stream
) {
3726 return consumed_pos
; /* Grab everything */
3728 start_pos
= produced_pos
- offset_align_floor(produced_pos
, max_sb_size
);
3729 start_pos
-= max_sb_size
* nb_packets_per_stream
;
3730 if ((long) (start_pos
- consumed_pos
) < 0) {
3731 return consumed_pos
; /* Grab everything */