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/index/index.h>
38 #include <common/kernel-ctl/kernel-ctl.h>
39 #include <common/sessiond-comm/relayd.h>
40 #include <common/sessiond-comm/sessiond-comm.h>
41 #include <common/kernel-consumer/kernel-consumer.h>
42 #include <common/relayd/relayd.h>
43 #include <common/ust-consumer/ust-consumer.h>
44 #include <common/consumer-timer.h>
47 #include "consumer-stream.h"
48 #include "consumer-testpoint.h"
50 struct lttng_consumer_global_data consumer_data
= {
53 .type
= LTTNG_CONSUMER_UNKNOWN
,
56 enum consumer_channel_action
{
59 CONSUMER_CHANNEL_QUIT
,
62 struct consumer_channel_msg
{
63 enum consumer_channel_action action
;
64 struct lttng_consumer_channel
*chan
; /* add */
65 uint64_t key
; /* del */
69 * Flag to inform the polling thread to quit when all fd hung up. Updated by
70 * the consumer_thread_receive_fds when it notices that all fds has hung up.
71 * Also updated by the signal handler (consumer_should_exit()). Read by the
74 volatile int consumer_quit
;
77 * Global hash table containing respectively metadata and data streams. The
78 * stream element in this ht should only be updated by the metadata poll thread
79 * for the metadata and the data poll thread for the data.
81 static struct lttng_ht
*metadata_ht
;
82 static struct lttng_ht
*data_ht
;
85 * Notify a thread lttng pipe to poll back again. This usually means that some
86 * global state has changed so we just send back the thread in a poll wait
89 static void notify_thread_lttng_pipe(struct lttng_pipe
*pipe
)
91 struct lttng_consumer_stream
*null_stream
= NULL
;
95 (void) lttng_pipe_write(pipe
, &null_stream
, sizeof(null_stream
));
98 static void notify_health_quit_pipe(int *pipe
)
102 ret
= lttng_write(pipe
[1], "4", 1);
104 PERROR("write consumer health quit");
108 static void notify_channel_pipe(struct lttng_consumer_local_data
*ctx
,
109 struct lttng_consumer_channel
*chan
,
111 enum consumer_channel_action action
)
113 struct consumer_channel_msg msg
;
116 memset(&msg
, 0, sizeof(msg
));
121 ret
= lttng_write(ctx
->consumer_channel_pipe
[1], &msg
, sizeof(msg
));
122 if (ret
< sizeof(msg
)) {
123 PERROR("notify_channel_pipe write error");
127 void notify_thread_del_channel(struct lttng_consumer_local_data
*ctx
,
130 notify_channel_pipe(ctx
, NULL
, key
, CONSUMER_CHANNEL_DEL
);
133 static int read_channel_pipe(struct lttng_consumer_local_data
*ctx
,
134 struct lttng_consumer_channel
**chan
,
136 enum consumer_channel_action
*action
)
138 struct consumer_channel_msg msg
;
141 ret
= lttng_read(ctx
->consumer_channel_pipe
[0], &msg
, sizeof(msg
));
142 if (ret
< sizeof(msg
)) {
146 *action
= msg
.action
;
154 * Cleanup the stream list of a channel. Those streams are not yet globally
157 static void clean_channel_stream_list(struct lttng_consumer_channel
*channel
)
159 struct lttng_consumer_stream
*stream
, *stmp
;
163 /* Delete streams that might have been left in the stream list. */
164 cds_list_for_each_entry_safe(stream
, stmp
, &channel
->streams
.head
,
166 cds_list_del(&stream
->send_node
);
168 * Once a stream is added to this list, the buffers were created so we
169 * have a guarantee that this call will succeed. Setting the monitor
170 * mode to 0 so we don't lock nor try to delete the stream from the
174 consumer_stream_destroy(stream
, NULL
);
179 * Find a stream. The consumer_data.lock must be locked during this
182 static struct lttng_consumer_stream
*find_stream(uint64_t key
,
185 struct lttng_ht_iter iter
;
186 struct lttng_ht_node_u64
*node
;
187 struct lttng_consumer_stream
*stream
= NULL
;
191 /* -1ULL keys are lookup failures */
192 if (key
== (uint64_t) -1ULL) {
198 lttng_ht_lookup(ht
, &key
, &iter
);
199 node
= lttng_ht_iter_get_node_u64(&iter
);
201 stream
= caa_container_of(node
, struct lttng_consumer_stream
, node
);
209 static void steal_stream_key(uint64_t key
, struct lttng_ht
*ht
)
211 struct lttng_consumer_stream
*stream
;
214 stream
= find_stream(key
, ht
);
216 stream
->key
= (uint64_t) -1ULL;
218 * We don't want the lookup to match, but we still need
219 * to iterate on this stream when iterating over the hash table. Just
220 * change the node key.
222 stream
->node
.key
= (uint64_t) -1ULL;
228 * Return a channel object for the given key.
230 * RCU read side lock MUST be acquired before calling this function and
231 * protects the channel ptr.
233 struct lttng_consumer_channel
*consumer_find_channel(uint64_t key
)
235 struct lttng_ht_iter iter
;
236 struct lttng_ht_node_u64
*node
;
237 struct lttng_consumer_channel
*channel
= NULL
;
239 /* -1ULL keys are lookup failures */
240 if (key
== (uint64_t) -1ULL) {
244 lttng_ht_lookup(consumer_data
.channel_ht
, &key
, &iter
);
245 node
= lttng_ht_iter_get_node_u64(&iter
);
247 channel
= caa_container_of(node
, struct lttng_consumer_channel
, node
);
254 * There is a possibility that the consumer does not have enough time between
255 * the close of the channel on the session daemon and the cleanup in here thus
256 * once we have a channel add with an existing key, we know for sure that this
257 * channel will eventually get cleaned up by all streams being closed.
259 * This function just nullifies the already existing channel key.
261 static void steal_channel_key(uint64_t key
)
263 struct lttng_consumer_channel
*channel
;
266 channel
= consumer_find_channel(key
);
268 channel
->key
= (uint64_t) -1ULL;
270 * We don't want the lookup to match, but we still need to iterate on
271 * this channel when iterating over the hash table. Just change the
274 channel
->node
.key
= (uint64_t) -1ULL;
279 static void free_channel_rcu(struct rcu_head
*head
)
281 struct lttng_ht_node_u64
*node
=
282 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
283 struct lttng_consumer_channel
*channel
=
284 caa_container_of(node
, struct lttng_consumer_channel
, node
);
290 * RCU protected relayd socket pair free.
292 static void free_relayd_rcu(struct rcu_head
*head
)
294 struct lttng_ht_node_u64
*node
=
295 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
296 struct consumer_relayd_sock_pair
*relayd
=
297 caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
300 * Close all sockets. This is done in the call RCU since we don't want the
301 * socket fds to be reassigned thus potentially creating bad state of the
304 * We do not have to lock the control socket mutex here since at this stage
305 * there is no one referencing to this relayd object.
307 (void) relayd_close(&relayd
->control_sock
);
308 (void) relayd_close(&relayd
->data_sock
);
314 * Destroy and free relayd socket pair object.
316 void consumer_destroy_relayd(struct consumer_relayd_sock_pair
*relayd
)
319 struct lttng_ht_iter iter
;
321 if (relayd
== NULL
) {
325 DBG("Consumer destroy and close relayd socket pair");
327 iter
.iter
.node
= &relayd
->node
.node
;
328 ret
= lttng_ht_del(consumer_data
.relayd_ht
, &iter
);
330 /* We assume the relayd is being or is destroyed */
334 /* RCU free() call */
335 call_rcu(&relayd
->node
.head
, free_relayd_rcu
);
339 * Remove a channel from the global list protected by a mutex. This function is
340 * also responsible for freeing its data structures.
342 void consumer_del_channel(struct lttng_consumer_channel
*channel
)
345 struct lttng_ht_iter iter
;
347 DBG("Consumer delete channel key %" PRIu64
, channel
->key
);
349 pthread_mutex_lock(&consumer_data
.lock
);
350 pthread_mutex_lock(&channel
->lock
);
352 /* Destroy streams that might have been left in the stream list. */
353 clean_channel_stream_list(channel
);
355 if (channel
->live_timer_enabled
== 1) {
356 consumer_timer_live_stop(channel
);
359 switch (consumer_data
.type
) {
360 case LTTNG_CONSUMER_KERNEL
:
362 case LTTNG_CONSUMER32_UST
:
363 case LTTNG_CONSUMER64_UST
:
364 lttng_ustconsumer_del_channel(channel
);
367 ERR("Unknown consumer_data type");
373 iter
.iter
.node
= &channel
->node
.node
;
374 ret
= lttng_ht_del(consumer_data
.channel_ht
, &iter
);
378 call_rcu(&channel
->node
.head
, free_channel_rcu
);
380 pthread_mutex_unlock(&channel
->lock
);
381 pthread_mutex_unlock(&consumer_data
.lock
);
385 * Iterate over the relayd hash table and destroy each element. Finally,
386 * destroy the whole hash table.
388 static void cleanup_relayd_ht(void)
390 struct lttng_ht_iter iter
;
391 struct consumer_relayd_sock_pair
*relayd
;
395 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
397 consumer_destroy_relayd(relayd
);
402 lttng_ht_destroy(consumer_data
.relayd_ht
);
406 * Update the end point status of all streams having the given network sequence
407 * index (relayd index).
409 * It's atomically set without having the stream mutex locked which is fine
410 * because we handle the write/read race with a pipe wakeup for each thread.
412 static void update_endpoint_status_by_netidx(uint64_t net_seq_idx
,
413 enum consumer_endpoint_status status
)
415 struct lttng_ht_iter iter
;
416 struct lttng_consumer_stream
*stream
;
418 DBG("Consumer set delete flag on stream by idx %" PRIu64
, net_seq_idx
);
422 /* Let's begin with metadata */
423 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
424 if (stream
->net_seq_idx
== net_seq_idx
) {
425 uatomic_set(&stream
->endpoint_status
, status
);
426 DBG("Delete flag set to metadata stream %d", stream
->wait_fd
);
430 /* Follow up by the data streams */
431 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
432 if (stream
->net_seq_idx
== net_seq_idx
) {
433 uatomic_set(&stream
->endpoint_status
, status
);
434 DBG("Delete flag set to data stream %d", stream
->wait_fd
);
441 * Cleanup a relayd object by flagging every associated streams for deletion,
442 * destroying the object meaning removing it from the relayd hash table,
443 * closing the sockets and freeing the memory in a RCU call.
445 * If a local data context is available, notify the threads that the streams'
446 * state have changed.
448 static void cleanup_relayd(struct consumer_relayd_sock_pair
*relayd
,
449 struct lttng_consumer_local_data
*ctx
)
455 DBG("Cleaning up relayd sockets");
457 /* Save the net sequence index before destroying the object */
458 netidx
= relayd
->net_seq_idx
;
461 * Delete the relayd from the relayd hash table, close the sockets and free
462 * the object in a RCU call.
464 consumer_destroy_relayd(relayd
);
466 /* Set inactive endpoint to all streams */
467 update_endpoint_status_by_netidx(netidx
, CONSUMER_ENDPOINT_INACTIVE
);
470 * With a local data context, notify the threads that the streams' state
471 * have changed. The write() action on the pipe acts as an "implicit"
472 * memory barrier ordering the updates of the end point status from the
473 * read of this status which happens AFTER receiving this notify.
476 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
477 notify_thread_lttng_pipe(ctx
->consumer_metadata_pipe
);
482 * Flag a relayd socket pair for destruction. Destroy it if the refcount
485 * RCU read side lock MUST be aquired before calling this function.
487 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair
*relayd
)
491 /* Set destroy flag for this object */
492 uatomic_set(&relayd
->destroy_flag
, 1);
494 /* Destroy the relayd if refcount is 0 */
495 if (uatomic_read(&relayd
->refcount
) == 0) {
496 consumer_destroy_relayd(relayd
);
501 * Completly destroy stream from every visiable data structure and the given
504 * One this call returns, the stream object is not longer usable nor visible.
506 void consumer_del_stream(struct lttng_consumer_stream
*stream
,
509 consumer_stream_destroy(stream
, ht
);
513 * XXX naming of del vs destroy is all mixed up.
515 void consumer_del_stream_for_data(struct lttng_consumer_stream
*stream
)
517 consumer_stream_destroy(stream
, data_ht
);
520 void consumer_del_stream_for_metadata(struct lttng_consumer_stream
*stream
)
522 consumer_stream_destroy(stream
, metadata_ht
);
525 struct lttng_consumer_stream
*consumer_allocate_stream(uint64_t channel_key
,
527 enum lttng_consumer_stream_state state
,
528 const char *channel_name
,
535 enum consumer_channel_type type
,
536 unsigned int monitor
)
539 struct lttng_consumer_stream
*stream
;
541 stream
= zmalloc(sizeof(*stream
));
542 if (stream
== NULL
) {
543 PERROR("malloc struct lttng_consumer_stream");
550 stream
->key
= stream_key
;
552 stream
->out_fd_offset
= 0;
553 stream
->output_written
= 0;
554 stream
->state
= state
;
557 stream
->net_seq_idx
= relayd_id
;
558 stream
->session_id
= session_id
;
559 stream
->monitor
= monitor
;
560 stream
->endpoint_status
= CONSUMER_ENDPOINT_ACTIVE
;
561 stream
->index_fd
= -1;
562 pthread_mutex_init(&stream
->lock
, NULL
);
564 /* If channel is the metadata, flag this stream as metadata. */
565 if (type
== CONSUMER_CHANNEL_TYPE_METADATA
) {
566 stream
->metadata_flag
= 1;
567 /* Metadata is flat out. */
568 strncpy(stream
->name
, DEFAULT_METADATA_NAME
, sizeof(stream
->name
));
569 /* Live rendez-vous point. */
570 pthread_cond_init(&stream
->metadata_rdv
, NULL
);
571 pthread_mutex_init(&stream
->metadata_rdv_lock
, NULL
);
573 /* Format stream name to <channel_name>_<cpu_number> */
574 ret
= snprintf(stream
->name
, sizeof(stream
->name
), "%s_%d",
577 PERROR("snprintf stream name");
582 /* Key is always the wait_fd for streams. */
583 lttng_ht_node_init_u64(&stream
->node
, stream
->key
);
585 /* Init node per channel id key */
586 lttng_ht_node_init_u64(&stream
->node_channel_id
, channel_key
);
588 /* Init session id node with the stream session id */
589 lttng_ht_node_init_u64(&stream
->node_session_id
, stream
->session_id
);
591 DBG3("Allocated stream %s (key %" PRIu64
", chan_key %" PRIu64
592 " relayd_id %" PRIu64
", session_id %" PRIu64
,
593 stream
->name
, stream
->key
, channel_key
,
594 stream
->net_seq_idx
, stream
->session_id
);
610 * Add a stream to the global list protected by a mutex.
612 int consumer_add_data_stream(struct lttng_consumer_stream
*stream
)
614 struct lttng_ht
*ht
= data_ht
;
620 DBG3("Adding consumer stream %" PRIu64
, stream
->key
);
622 pthread_mutex_lock(&consumer_data
.lock
);
623 pthread_mutex_lock(&stream
->chan
->lock
);
624 pthread_mutex_lock(&stream
->chan
->timer_lock
);
625 pthread_mutex_lock(&stream
->lock
);
628 /* Steal stream identifier to avoid having streams with the same key */
629 steal_stream_key(stream
->key
, ht
);
631 lttng_ht_add_unique_u64(ht
, &stream
->node
);
633 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
634 &stream
->node_channel_id
);
637 * Add stream to the stream_list_ht of the consumer data. No need to steal
638 * the key since the HT does not use it and we allow to add redundant keys
641 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
644 * When nb_init_stream_left reaches 0, we don't need to trigger any action
645 * in terms of destroying the associated channel, because the action that
646 * causes the count to become 0 also causes a stream to be added. The
647 * channel deletion will thus be triggered by the following removal of this
650 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
651 /* Increment refcount before decrementing nb_init_stream_left */
653 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
656 /* Update consumer data once the node is inserted. */
657 consumer_data
.stream_count
++;
658 consumer_data
.need_update
= 1;
661 pthread_mutex_unlock(&stream
->lock
);
662 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
663 pthread_mutex_unlock(&stream
->chan
->lock
);
664 pthread_mutex_unlock(&consumer_data
.lock
);
669 void consumer_del_data_stream(struct lttng_consumer_stream
*stream
)
671 consumer_del_stream(stream
, data_ht
);
675 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
676 * be acquired before calling this.
678 static int add_relayd(struct consumer_relayd_sock_pair
*relayd
)
681 struct lttng_ht_node_u64
*node
;
682 struct lttng_ht_iter iter
;
686 lttng_ht_lookup(consumer_data
.relayd_ht
,
687 &relayd
->net_seq_idx
, &iter
);
688 node
= lttng_ht_iter_get_node_u64(&iter
);
692 lttng_ht_add_unique_u64(consumer_data
.relayd_ht
, &relayd
->node
);
699 * Allocate and return a consumer relayd socket.
701 struct consumer_relayd_sock_pair
*consumer_allocate_relayd_sock_pair(
702 uint64_t net_seq_idx
)
704 struct consumer_relayd_sock_pair
*obj
= NULL
;
706 /* net sequence index of -1 is a failure */
707 if (net_seq_idx
== (uint64_t) -1ULL) {
711 obj
= zmalloc(sizeof(struct consumer_relayd_sock_pair
));
713 PERROR("zmalloc relayd sock");
717 obj
->net_seq_idx
= net_seq_idx
;
719 obj
->destroy_flag
= 0;
720 obj
->control_sock
.sock
.fd
= -1;
721 obj
->data_sock
.sock
.fd
= -1;
722 lttng_ht_node_init_u64(&obj
->node
, obj
->net_seq_idx
);
723 pthread_mutex_init(&obj
->ctrl_sock_mutex
, NULL
);
730 * Find a relayd socket pair in the global consumer data.
732 * Return the object if found else NULL.
733 * RCU read-side lock must be held across this call and while using the
736 struct consumer_relayd_sock_pair
*consumer_find_relayd(uint64_t key
)
738 struct lttng_ht_iter iter
;
739 struct lttng_ht_node_u64
*node
;
740 struct consumer_relayd_sock_pair
*relayd
= NULL
;
742 /* Negative keys are lookup failures */
743 if (key
== (uint64_t) -1ULL) {
747 lttng_ht_lookup(consumer_data
.relayd_ht
, &key
,
749 node
= lttng_ht_iter_get_node_u64(&iter
);
751 relayd
= caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
759 * Find a relayd and send the stream
761 * Returns 0 on success, < 0 on error
763 int consumer_send_relayd_stream(struct lttng_consumer_stream
*stream
,
767 struct consumer_relayd_sock_pair
*relayd
;
770 assert(stream
->net_seq_idx
!= -1ULL);
773 /* The stream is not metadata. Get relayd reference if exists. */
775 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
776 if (relayd
!= NULL
) {
777 /* Add stream on the relayd */
778 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
779 ret
= relayd_add_stream(&relayd
->control_sock
, stream
->name
,
780 path
, &stream
->relayd_stream_id
,
781 stream
->chan
->tracefile_size
, stream
->chan
->tracefile_count
);
782 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
787 uatomic_inc(&relayd
->refcount
);
788 stream
->sent_to_relayd
= 1;
790 ERR("Stream %" PRIu64
" relayd ID %" PRIu64
" unknown. Can't send it.",
791 stream
->key
, stream
->net_seq_idx
);
796 DBG("Stream %s with key %" PRIu64
" sent to relayd id %" PRIu64
,
797 stream
->name
, stream
->key
, stream
->net_seq_idx
);
805 * Find a relayd and send the streams sent message
807 * Returns 0 on success, < 0 on error
809 int consumer_send_relayd_streams_sent(uint64_t net_seq_idx
)
812 struct consumer_relayd_sock_pair
*relayd
;
814 assert(net_seq_idx
!= -1ULL);
816 /* The stream is not metadata. Get relayd reference if exists. */
818 relayd
= consumer_find_relayd(net_seq_idx
);
819 if (relayd
!= NULL
) {
820 /* Add stream on the relayd */
821 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
822 ret
= relayd_streams_sent(&relayd
->control_sock
);
823 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
828 ERR("Relayd ID %" PRIu64
" unknown. Can't send streams_sent.",
835 DBG("All streams sent relayd id %" PRIu64
, net_seq_idx
);
843 * Find a relayd and close the stream
845 void close_relayd_stream(struct lttng_consumer_stream
*stream
)
847 struct consumer_relayd_sock_pair
*relayd
;
849 /* The stream is not metadata. Get relayd reference if exists. */
851 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
853 consumer_stream_relayd_close(stream
, relayd
);
859 * Handle stream for relayd transmission if the stream applies for network
860 * streaming where the net sequence index is set.
862 * Return destination file descriptor or negative value on error.
864 static int write_relayd_stream_header(struct lttng_consumer_stream
*stream
,
865 size_t data_size
, unsigned long padding
,
866 struct consumer_relayd_sock_pair
*relayd
)
869 struct lttcomm_relayd_data_hdr data_hdr
;
875 /* Reset data header */
876 memset(&data_hdr
, 0, sizeof(data_hdr
));
878 if (stream
->metadata_flag
) {
879 /* Caller MUST acquire the relayd control socket lock */
880 ret
= relayd_send_metadata(&relayd
->control_sock
, data_size
);
885 /* Metadata are always sent on the control socket. */
886 outfd
= relayd
->control_sock
.sock
.fd
;
888 /* Set header with stream information */
889 data_hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
890 data_hdr
.data_size
= htobe32(data_size
);
891 data_hdr
.padding_size
= htobe32(padding
);
893 * Note that net_seq_num below is assigned with the *current* value of
894 * next_net_seq_num and only after that the next_net_seq_num will be
895 * increment. This is why when issuing a command on the relayd using
896 * this next value, 1 should always be substracted in order to compare
897 * the last seen sequence number on the relayd side to the last sent.
899 data_hdr
.net_seq_num
= htobe64(stream
->next_net_seq_num
);
900 /* Other fields are zeroed previously */
902 ret
= relayd_send_data_hdr(&relayd
->data_sock
, &data_hdr
,
908 ++stream
->next_net_seq_num
;
910 /* Set to go on data socket */
911 outfd
= relayd
->data_sock
.sock
.fd
;
919 * Allocate and return a new lttng_consumer_channel object using the given key
920 * to initialize the hash table node.
922 * On error, return NULL.
924 struct lttng_consumer_channel
*consumer_allocate_channel(uint64_t key
,
926 const char *pathname
,
931 enum lttng_event_output output
,
932 uint64_t tracefile_size
,
933 uint64_t tracefile_count
,
934 uint64_t session_id_per_pid
,
935 unsigned int monitor
,
936 unsigned int live_timer_interval
)
938 struct lttng_consumer_channel
*channel
;
940 channel
= zmalloc(sizeof(*channel
));
941 if (channel
== NULL
) {
942 PERROR("malloc struct lttng_consumer_channel");
947 channel
->refcount
= 0;
948 channel
->session_id
= session_id
;
949 channel
->session_id_per_pid
= session_id_per_pid
;
952 channel
->relayd_id
= relayd_id
;
953 channel
->tracefile_size
= tracefile_size
;
954 channel
->tracefile_count
= tracefile_count
;
955 channel
->monitor
= monitor
;
956 channel
->live_timer_interval
= live_timer_interval
;
957 pthread_mutex_init(&channel
->lock
, NULL
);
958 pthread_mutex_init(&channel
->timer_lock
, NULL
);
961 case LTTNG_EVENT_SPLICE
:
962 channel
->output
= CONSUMER_CHANNEL_SPLICE
;
964 case LTTNG_EVENT_MMAP
:
965 channel
->output
= CONSUMER_CHANNEL_MMAP
;
975 * In monitor mode, the streams associated with the channel will be put in
976 * a special list ONLY owned by this channel. So, the refcount is set to 1
977 * here meaning that the channel itself has streams that are referenced.
979 * On a channel deletion, once the channel is no longer visible, the
980 * refcount is decremented and checked for a zero value to delete it. With
981 * streams in no monitor mode, it will now be safe to destroy the channel.
983 if (!channel
->monitor
) {
984 channel
->refcount
= 1;
987 strncpy(channel
->pathname
, pathname
, sizeof(channel
->pathname
));
988 channel
->pathname
[sizeof(channel
->pathname
) - 1] = '\0';
990 strncpy(channel
->name
, name
, sizeof(channel
->name
));
991 channel
->name
[sizeof(channel
->name
) - 1] = '\0';
993 lttng_ht_node_init_u64(&channel
->node
, channel
->key
);
995 channel
->wait_fd
= -1;
997 CDS_INIT_LIST_HEAD(&channel
->streams
.head
);
999 DBG("Allocated channel (key %" PRIu64
")", channel
->key
)
1006 * Add a channel to the global list protected by a mutex.
1008 * Always return 0 indicating success.
1010 int consumer_add_channel(struct lttng_consumer_channel
*channel
,
1011 struct lttng_consumer_local_data
*ctx
)
1013 pthread_mutex_lock(&consumer_data
.lock
);
1014 pthread_mutex_lock(&channel
->lock
);
1015 pthread_mutex_lock(&channel
->timer_lock
);
1018 * This gives us a guarantee that the channel we are about to add to the
1019 * channel hash table will be unique. See this function comment on the why
1020 * we need to steel the channel key at this stage.
1022 steal_channel_key(channel
->key
);
1025 lttng_ht_add_unique_u64(consumer_data
.channel_ht
, &channel
->node
);
1028 pthread_mutex_unlock(&channel
->timer_lock
);
1029 pthread_mutex_unlock(&channel
->lock
);
1030 pthread_mutex_unlock(&consumer_data
.lock
);
1032 if (channel
->wait_fd
!= -1 && channel
->type
== CONSUMER_CHANNEL_TYPE_DATA
) {
1033 notify_channel_pipe(ctx
, channel
, -1, CONSUMER_CHANNEL_ADD
);
1040 * Allocate the pollfd structure and the local view of the out fds to avoid
1041 * doing a lookup in the linked list and concurrency issues when writing is
1042 * needed. Called with consumer_data.lock held.
1044 * Returns the number of fds in the structures.
1046 static int update_poll_array(struct lttng_consumer_local_data
*ctx
,
1047 struct pollfd
**pollfd
, struct lttng_consumer_stream
**local_stream
,
1048 struct lttng_ht
*ht
)
1051 struct lttng_ht_iter iter
;
1052 struct lttng_consumer_stream
*stream
;
1057 assert(local_stream
);
1059 DBG("Updating poll fd array");
1061 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1063 * Only active streams with an active end point can be added to the
1064 * poll set and local stream storage of the thread.
1066 * There is a potential race here for endpoint_status to be updated
1067 * just after the check. However, this is OK since the stream(s) will
1068 * be deleted once the thread is notified that the end point state has
1069 * changed where this function will be called back again.
1071 if (stream
->state
!= LTTNG_CONSUMER_ACTIVE_STREAM
||
1072 stream
->endpoint_status
== CONSUMER_ENDPOINT_INACTIVE
) {
1076 * This clobbers way too much the debug output. Uncomment that if you
1077 * need it for debugging purposes.
1079 * DBG("Active FD %d", stream->wait_fd);
1081 (*pollfd
)[i
].fd
= stream
->wait_fd
;
1082 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1083 local_stream
[i
] = stream
;
1089 * Insert the consumer_data_pipe at the end of the array and don't
1090 * increment i so nb_fd is the number of real FD.
1092 (*pollfd
)[i
].fd
= lttng_pipe_get_readfd(ctx
->consumer_data_pipe
);
1093 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1098 * Poll on the should_quit pipe and the command socket return -1 on error and
1099 * should exit, 0 if data is available on the command socket
1101 int lttng_consumer_poll_socket(struct pollfd
*consumer_sockpoll
)
1106 num_rdy
= poll(consumer_sockpoll
, 2, -1);
1107 if (num_rdy
== -1) {
1109 * Restart interrupted system call.
1111 if (errno
== EINTR
) {
1114 PERROR("Poll error");
1117 if (consumer_sockpoll
[0].revents
& (POLLIN
| POLLPRI
)) {
1118 DBG("consumer_should_quit wake up");
1128 * Set the error socket.
1130 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data
*ctx
,
1133 ctx
->consumer_error_socket
= sock
;
1137 * Set the command socket path.
1139 void lttng_consumer_set_command_sock_path(
1140 struct lttng_consumer_local_data
*ctx
, char *sock
)
1142 ctx
->consumer_command_sock_path
= sock
;
1146 * Send return code to the session daemon.
1147 * If the socket is not defined, we return 0, it is not a fatal error
1149 int lttng_consumer_send_error(struct lttng_consumer_local_data
*ctx
, int cmd
)
1151 if (ctx
->consumer_error_socket
> 0) {
1152 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
1153 sizeof(enum lttcomm_sessiond_command
));
1160 * Close all the tracefiles and stream fds and MUST be called when all
1161 * instances are destroyed i.e. when all threads were joined and are ended.
1163 void lttng_consumer_cleanup(void)
1165 struct lttng_ht_iter iter
;
1166 struct lttng_consumer_channel
*channel
;
1170 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
, channel
,
1172 consumer_del_channel(channel
);
1177 lttng_ht_destroy(consumer_data
.channel_ht
);
1179 cleanup_relayd_ht();
1181 lttng_ht_destroy(consumer_data
.stream_per_chan_id_ht
);
1184 * This HT contains streams that are freed by either the metadata thread or
1185 * the data thread so we do *nothing* on the hash table and simply destroy
1188 lttng_ht_destroy(consumer_data
.stream_list_ht
);
1192 * Called from signal handler.
1194 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
1199 ret
= lttng_write(ctx
->consumer_should_quit
[1], "4", 1);
1201 PERROR("write consumer quit");
1204 DBG("Consumer flag that it should quit");
1207 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1210 int outfd
= stream
->out_fd
;
1213 * This does a blocking write-and-wait on any page that belongs to the
1214 * subbuffer prior to the one we just wrote.
1215 * Don't care about error values, as these are just hints and ways to
1216 * limit the amount of page cache used.
1218 if (orig_offset
< stream
->max_sb_size
) {
1221 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1222 stream
->max_sb_size
,
1223 SYNC_FILE_RANGE_WAIT_BEFORE
1224 | SYNC_FILE_RANGE_WRITE
1225 | SYNC_FILE_RANGE_WAIT_AFTER
);
1227 * Give hints to the kernel about how we access the file:
1228 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1231 * We need to call fadvise again after the file grows because the
1232 * kernel does not seem to apply fadvise to non-existing parts of the
1235 * Call fadvise _after_ having waited for the page writeback to
1236 * complete because the dirty page writeback semantic is not well
1237 * defined. So it can be expected to lead to lower throughput in
1240 posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1241 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1245 * Initialise the necessary environnement :
1246 * - create a new context
1247 * - create the poll_pipe
1248 * - create the should_quit pipe (for signal handler)
1249 * - create the thread pipe (for splice)
1251 * Takes a function pointer as argument, this function is called when data is
1252 * available on a buffer. This function is responsible to do the
1253 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1254 * buffer configuration and then kernctl_put_next_subbuf at the end.
1256 * Returns a pointer to the new context or NULL on error.
1258 struct lttng_consumer_local_data
*lttng_consumer_create(
1259 enum lttng_consumer_type type
,
1260 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1261 struct lttng_consumer_local_data
*ctx
),
1262 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1263 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1264 int (*update_stream
)(uint64_t stream_key
, uint32_t state
))
1267 struct lttng_consumer_local_data
*ctx
;
1269 assert(consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1270 consumer_data
.type
== type
);
1271 consumer_data
.type
= type
;
1273 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
1275 PERROR("allocating context");
1279 ctx
->consumer_error_socket
= -1;
1280 ctx
->consumer_metadata_socket
= -1;
1281 pthread_mutex_init(&ctx
->metadata_socket_lock
, NULL
);
1282 /* assign the callbacks */
1283 ctx
->on_buffer_ready
= buffer_ready
;
1284 ctx
->on_recv_channel
= recv_channel
;
1285 ctx
->on_recv_stream
= recv_stream
;
1286 ctx
->on_update_stream
= update_stream
;
1288 ctx
->consumer_data_pipe
= lttng_pipe_open(0);
1289 if (!ctx
->consumer_data_pipe
) {
1290 goto error_poll_pipe
;
1293 ret
= pipe(ctx
->consumer_should_quit
);
1295 PERROR("Error creating recv pipe");
1296 goto error_quit_pipe
;
1299 ret
= pipe(ctx
->consumer_thread_pipe
);
1301 PERROR("Error creating thread pipe");
1302 goto error_thread_pipe
;
1305 ret
= pipe(ctx
->consumer_channel_pipe
);
1307 PERROR("Error creating channel pipe");
1308 goto error_channel_pipe
;
1311 ctx
->consumer_metadata_pipe
= lttng_pipe_open(0);
1312 if (!ctx
->consumer_metadata_pipe
) {
1313 goto error_metadata_pipe
;
1316 ret
= utils_create_pipe(ctx
->consumer_splice_metadata_pipe
);
1318 goto error_splice_pipe
;
1324 lttng_pipe_destroy(ctx
->consumer_metadata_pipe
);
1325 error_metadata_pipe
:
1326 utils_close_pipe(ctx
->consumer_channel_pipe
);
1328 utils_close_pipe(ctx
->consumer_thread_pipe
);
1330 utils_close_pipe(ctx
->consumer_should_quit
);
1332 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1340 * Iterate over all streams of the hashtable and free them properly.
1342 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1344 struct lttng_ht_iter iter
;
1345 struct lttng_consumer_stream
*stream
;
1352 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1354 * Ignore return value since we are currently cleaning up so any error
1357 (void) consumer_del_stream(stream
, ht
);
1361 lttng_ht_destroy(ht
);
1365 * Iterate over all streams of the metadata hashtable and free them
1368 static void destroy_metadata_stream_ht(struct lttng_ht
*ht
)
1370 struct lttng_ht_iter iter
;
1371 struct lttng_consumer_stream
*stream
;
1378 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1380 * Ignore return value since we are currently cleaning up so any error
1383 (void) consumer_del_metadata_stream(stream
, ht
);
1387 lttng_ht_destroy(ht
);
1391 * Close all fds associated with the instance and free the context.
1393 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1397 DBG("Consumer destroying it. Closing everything.");
1399 destroy_data_stream_ht(data_ht
);
1400 destroy_metadata_stream_ht(metadata_ht
);
1402 ret
= close(ctx
->consumer_error_socket
);
1406 ret
= close(ctx
->consumer_metadata_socket
);
1410 utils_close_pipe(ctx
->consumer_thread_pipe
);
1411 utils_close_pipe(ctx
->consumer_channel_pipe
);
1412 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1413 lttng_pipe_destroy(ctx
->consumer_metadata_pipe
);
1414 utils_close_pipe(ctx
->consumer_should_quit
);
1415 utils_close_pipe(ctx
->consumer_splice_metadata_pipe
);
1417 unlink(ctx
->consumer_command_sock_path
);
1422 * Write the metadata stream id on the specified file descriptor.
1424 static int write_relayd_metadata_id(int fd
,
1425 struct lttng_consumer_stream
*stream
,
1426 struct consumer_relayd_sock_pair
*relayd
, unsigned long padding
)
1429 struct lttcomm_relayd_metadata_payload hdr
;
1431 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1432 hdr
.padding_size
= htobe32(padding
);
1433 ret
= lttng_write(fd
, (void *) &hdr
, sizeof(hdr
));
1434 if (ret
< sizeof(hdr
)) {
1436 * This error means that the fd's end is closed so ignore the perror
1437 * not to clubber the error output since this can happen in a normal
1440 if (errno
!= EPIPE
) {
1441 PERROR("write metadata stream id");
1443 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1445 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1446 * handle writting the missing part so report that as an error and
1447 * don't lie to the caller.
1452 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1453 stream
->relayd_stream_id
, padding
);
1460 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1461 * core function for writing trace buffers to either the local filesystem or
1464 * It must be called with the stream lock held.
1466 * Careful review MUST be put if any changes occur!
1468 * Returns the number of bytes written
1470 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1471 struct lttng_consumer_local_data
*ctx
,
1472 struct lttng_consumer_stream
*stream
, unsigned long len
,
1473 unsigned long padding
,
1474 struct ctf_packet_index
*index
)
1476 unsigned long mmap_offset
;
1479 off_t orig_offset
= stream
->out_fd_offset
;
1480 /* Default is on the disk */
1481 int outfd
= stream
->out_fd
;
1482 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1483 unsigned int relayd_hang_up
= 0;
1485 /* RCU lock for the relayd pointer */
1488 /* Flag that the current stream if set for network streaming. */
1489 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1490 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1491 if (relayd
== NULL
) {
1497 /* get the offset inside the fd to mmap */
1498 switch (consumer_data
.type
) {
1499 case LTTNG_CONSUMER_KERNEL
:
1500 mmap_base
= stream
->mmap_base
;
1501 ret
= kernctl_get_mmap_read_offset(stream
->wait_fd
, &mmap_offset
);
1504 PERROR("tracer ctl get_mmap_read_offset");
1508 case LTTNG_CONSUMER32_UST
:
1509 case LTTNG_CONSUMER64_UST
:
1510 mmap_base
= lttng_ustctl_get_mmap_base(stream
);
1512 ERR("read mmap get mmap base for stream %s", stream
->name
);
1516 ret
= lttng_ustctl_get_mmap_read_offset(stream
, &mmap_offset
);
1518 PERROR("tracer ctl get_mmap_read_offset");
1524 ERR("Unknown consumer_data type");
1528 /* Handle stream on the relayd if the output is on the network */
1530 unsigned long netlen
= len
;
1533 * Lock the control socket for the complete duration of the function
1534 * since from this point on we will use the socket.
1536 if (stream
->metadata_flag
) {
1537 /* Metadata requires the control socket. */
1538 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1539 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1542 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1547 /* Use the returned socket. */
1550 /* Write metadata stream id before payload */
1551 if (stream
->metadata_flag
) {
1552 ret
= write_relayd_metadata_id(outfd
, stream
, relayd
, padding
);
1559 /* No streaming, we have to set the len with the full padding */
1563 * Check if we need to change the tracefile before writing the packet.
1565 if (stream
->chan
->tracefile_size
> 0 &&
1566 (stream
->tracefile_size_current
+ len
) >
1567 stream
->chan
->tracefile_size
) {
1568 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1569 stream
->name
, stream
->chan
->tracefile_size
,
1570 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1571 stream
->out_fd
, &(stream
->tracefile_count_current
),
1574 ERR("Rotating output file");
1577 outfd
= stream
->out_fd
;
1579 if (stream
->index_fd
>= 0) {
1580 ret
= index_create_file(stream
->chan
->pathname
,
1581 stream
->name
, stream
->uid
, stream
->gid
,
1582 stream
->chan
->tracefile_size
,
1583 stream
->tracefile_count_current
);
1587 stream
->index_fd
= ret
;
1590 /* Reset current size because we just perform a rotation. */
1591 stream
->tracefile_size_current
= 0;
1592 stream
->out_fd_offset
= 0;
1595 stream
->tracefile_size_current
+= len
;
1597 index
->offset
= htobe64(stream
->out_fd_offset
);
1602 * This call guarantee that len or less is returned. It's impossible to
1603 * receive a ret value that is bigger than len.
1605 ret
= lttng_write(outfd
, mmap_base
+ mmap_offset
, len
);
1606 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, len
);
1607 if (ret
< 0 || ((size_t) ret
!= len
)) {
1609 * Report error to caller if nothing was written else at least send the
1617 /* Socket operation failed. We consider the relayd dead */
1618 if (errno
== EPIPE
|| errno
== EINVAL
|| errno
== EBADF
) {
1620 * This is possible if the fd is closed on the other side
1621 * (outfd) or any write problem. It can be verbose a bit for a
1622 * normal execution if for instance the relayd is stopped
1623 * abruptly. This can happen so set this to a DBG statement.
1625 DBG("Consumer mmap write detected relayd hang up");
1627 /* Unhandled error, print it and stop function right now. */
1628 PERROR("Error in write mmap (ret %zd != len %lu)", ret
, len
);
1632 stream
->output_written
+= ret
;
1634 /* This call is useless on a socket so better save a syscall. */
1636 /* This won't block, but will start writeout asynchronously */
1637 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, len
,
1638 SYNC_FILE_RANGE_WRITE
);
1639 stream
->out_fd_offset
+= len
;
1641 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1645 * This is a special case that the relayd has closed its socket. Let's
1646 * cleanup the relayd object and all associated streams.
1648 if (relayd
&& relayd_hang_up
) {
1649 cleanup_relayd(relayd
, ctx
);
1653 /* Unlock only if ctrl socket used */
1654 if (relayd
&& stream
->metadata_flag
) {
1655 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1663 * Splice the data from the ring buffer to the tracefile.
1665 * It must be called with the stream lock held.
1667 * Returns the number of bytes spliced.
1669 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1670 struct lttng_consumer_local_data
*ctx
,
1671 struct lttng_consumer_stream
*stream
, unsigned long len
,
1672 unsigned long padding
,
1673 struct ctf_packet_index
*index
)
1675 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1677 off_t orig_offset
= stream
->out_fd_offset
;
1678 int fd
= stream
->wait_fd
;
1679 /* Default is on the disk */
1680 int outfd
= stream
->out_fd
;
1681 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1683 unsigned int relayd_hang_up
= 0;
1685 switch (consumer_data
.type
) {
1686 case LTTNG_CONSUMER_KERNEL
:
1688 case LTTNG_CONSUMER32_UST
:
1689 case LTTNG_CONSUMER64_UST
:
1690 /* Not supported for user space tracing */
1693 ERR("Unknown consumer_data type");
1697 /* RCU lock for the relayd pointer */
1700 /* Flag that the current stream if set for network streaming. */
1701 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1702 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1703 if (relayd
== NULL
) {
1710 * Choose right pipe for splice. Metadata and trace data are handled by
1711 * different threads hence the use of two pipes in order not to race or
1712 * corrupt the written data.
1714 if (stream
->metadata_flag
) {
1715 splice_pipe
= ctx
->consumer_splice_metadata_pipe
;
1717 splice_pipe
= ctx
->consumer_thread_pipe
;
1720 /* Write metadata stream id before payload */
1722 unsigned long total_len
= len
;
1724 if (stream
->metadata_flag
) {
1726 * Lock the control socket for the complete duration of the function
1727 * since from this point on we will use the socket.
1729 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1731 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
, relayd
,
1739 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1742 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1748 /* Use the returned socket. */
1751 /* No streaming, we have to set the len with the full padding */
1755 * Check if we need to change the tracefile before writing the packet.
1757 if (stream
->chan
->tracefile_size
> 0 &&
1758 (stream
->tracefile_size_current
+ len
) >
1759 stream
->chan
->tracefile_size
) {
1760 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1761 stream
->name
, stream
->chan
->tracefile_size
,
1762 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1763 stream
->out_fd
, &(stream
->tracefile_count_current
),
1767 ERR("Rotating output file");
1770 outfd
= stream
->out_fd
;
1772 if (stream
->index_fd
>= 0) {
1773 ret
= index_create_file(stream
->chan
->pathname
,
1774 stream
->name
, stream
->uid
, stream
->gid
,
1775 stream
->chan
->tracefile_size
,
1776 stream
->tracefile_count_current
);
1781 stream
->index_fd
= ret
;
1784 /* Reset current size because we just perform a rotation. */
1785 stream
->tracefile_size_current
= 0;
1786 stream
->out_fd_offset
= 0;
1789 stream
->tracefile_size_current
+= len
;
1790 index
->offset
= htobe64(stream
->out_fd_offset
);
1794 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1795 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1796 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1797 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1798 DBG("splice chan to pipe, ret %zd", ret_splice
);
1799 if (ret_splice
< 0) {
1802 PERROR("Error in relay splice");
1806 /* Handle stream on the relayd if the output is on the network */
1807 if (relayd
&& stream
->metadata_flag
) {
1808 size_t metadata_payload_size
=
1809 sizeof(struct lttcomm_relayd_metadata_payload
);
1811 /* Update counter to fit the spliced data */
1812 ret_splice
+= metadata_payload_size
;
1813 len
+= metadata_payload_size
;
1815 * We do this so the return value can match the len passed as
1816 * argument to this function.
1818 written
-= metadata_payload_size
;
1821 /* Splice data out */
1822 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1823 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1824 DBG("Consumer splice pipe to file, ret %zd", ret_splice
);
1825 if (ret_splice
< 0) {
1830 } else if (ret_splice
> len
) {
1832 * We don't expect this code path to be executed but you never know
1833 * so this is an extra protection agains a buggy splice().
1836 written
+= ret_splice
;
1837 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice
,
1841 /* All good, update current len and continue. */
1845 /* This call is useless on a socket so better save a syscall. */
1847 /* This won't block, but will start writeout asynchronously */
1848 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1849 SYNC_FILE_RANGE_WRITE
);
1850 stream
->out_fd_offset
+= ret_splice
;
1852 stream
->output_written
+= ret_splice
;
1853 written
+= ret_splice
;
1855 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1860 * This is a special case that the relayd has closed its socket. Let's
1861 * cleanup the relayd object and all associated streams.
1863 if (relayd
&& relayd_hang_up
) {
1864 cleanup_relayd(relayd
, ctx
);
1865 /* Skip splice error so the consumer does not fail */
1870 /* send the appropriate error description to sessiond */
1873 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1876 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1879 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
1884 if (relayd
&& stream
->metadata_flag
) {
1885 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1893 * Take a snapshot for a specific fd
1895 * Returns 0 on success, < 0 on error
1897 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
1899 switch (consumer_data
.type
) {
1900 case LTTNG_CONSUMER_KERNEL
:
1901 return lttng_kconsumer_take_snapshot(stream
);
1902 case LTTNG_CONSUMER32_UST
:
1903 case LTTNG_CONSUMER64_UST
:
1904 return lttng_ustconsumer_take_snapshot(stream
);
1906 ERR("Unknown consumer_data type");
1913 * Get the produced position
1915 * Returns 0 on success, < 0 on error
1917 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
1920 switch (consumer_data
.type
) {
1921 case LTTNG_CONSUMER_KERNEL
:
1922 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
1923 case LTTNG_CONSUMER32_UST
:
1924 case LTTNG_CONSUMER64_UST
:
1925 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
1927 ERR("Unknown consumer_data type");
1933 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
1934 int sock
, struct pollfd
*consumer_sockpoll
)
1936 switch (consumer_data
.type
) {
1937 case LTTNG_CONSUMER_KERNEL
:
1938 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1939 case LTTNG_CONSUMER32_UST
:
1940 case LTTNG_CONSUMER64_UST
:
1941 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1943 ERR("Unknown consumer_data type");
1949 void lttng_consumer_close_all_metadata(void)
1951 switch (consumer_data
.type
) {
1952 case LTTNG_CONSUMER_KERNEL
:
1954 * The Kernel consumer has a different metadata scheme so we don't
1955 * close anything because the stream will be closed by the session
1959 case LTTNG_CONSUMER32_UST
:
1960 case LTTNG_CONSUMER64_UST
:
1962 * Close all metadata streams. The metadata hash table is passed and
1963 * this call iterates over it by closing all wakeup fd. This is safe
1964 * because at this point we are sure that the metadata producer is
1965 * either dead or blocked.
1967 lttng_ustconsumer_close_all_metadata(metadata_ht
);
1970 ERR("Unknown consumer_data type");
1976 * Clean up a metadata stream and free its memory.
1978 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
1979 struct lttng_ht
*ht
)
1981 struct lttng_consumer_channel
*free_chan
= NULL
;
1985 * This call should NEVER receive regular stream. It must always be
1986 * metadata stream and this is crucial for data structure synchronization.
1988 assert(stream
->metadata_flag
);
1990 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
1992 pthread_mutex_lock(&consumer_data
.lock
);
1993 pthread_mutex_lock(&stream
->chan
->lock
);
1994 pthread_mutex_lock(&stream
->lock
);
1996 /* Remove any reference to that stream. */
1997 consumer_stream_delete(stream
, ht
);
1999 /* Close down everything including the relayd if one. */
2000 consumer_stream_close(stream
);
2001 /* Destroy tracer buffers of the stream. */
2002 consumer_stream_destroy_buffers(stream
);
2004 /* Atomically decrement channel refcount since other threads can use it. */
2005 if (!uatomic_sub_return(&stream
->chan
->refcount
, 1)
2006 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
2007 /* Go for channel deletion! */
2008 free_chan
= stream
->chan
;
2012 * Nullify the stream reference so it is not used after deletion. The
2013 * channel lock MUST be acquired before being able to check for a NULL
2016 stream
->chan
->metadata_stream
= NULL
;
2018 pthread_mutex_unlock(&stream
->lock
);
2019 pthread_mutex_unlock(&stream
->chan
->lock
);
2020 pthread_mutex_unlock(&consumer_data
.lock
);
2023 consumer_del_channel(free_chan
);
2026 consumer_stream_free(stream
);
2030 * Action done with the metadata stream when adding it to the consumer internal
2031 * data structures to handle it.
2033 int consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
2035 struct lttng_ht
*ht
= metadata_ht
;
2037 struct lttng_ht_iter iter
;
2038 struct lttng_ht_node_u64
*node
;
2043 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2045 pthread_mutex_lock(&consumer_data
.lock
);
2046 pthread_mutex_lock(&stream
->chan
->lock
);
2047 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2048 pthread_mutex_lock(&stream
->lock
);
2051 * From here, refcounts are updated so be _careful_ when returning an error
2058 * Lookup the stream just to make sure it does not exist in our internal
2059 * state. This should NEVER happen.
2061 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2062 node
= lttng_ht_iter_get_node_u64(&iter
);
2066 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2067 * in terms of destroying the associated channel, because the action that
2068 * causes the count to become 0 also causes a stream to be added. The
2069 * channel deletion will thus be triggered by the following removal of this
2072 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2073 /* Increment refcount before decrementing nb_init_stream_left */
2075 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2078 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2080 lttng_ht_add_unique_u64(consumer_data
.stream_per_chan_id_ht
,
2081 &stream
->node_channel_id
);
2084 * Add stream to the stream_list_ht of the consumer data. No need to steal
2085 * the key since the HT does not use it and we allow to add redundant keys
2088 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2092 pthread_mutex_unlock(&stream
->lock
);
2093 pthread_mutex_unlock(&stream
->chan
->lock
);
2094 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2095 pthread_mutex_unlock(&consumer_data
.lock
);
2100 * Delete data stream that are flagged for deletion (endpoint_status).
2102 static void validate_endpoint_status_data_stream(void)
2104 struct lttng_ht_iter iter
;
2105 struct lttng_consumer_stream
*stream
;
2107 DBG("Consumer delete flagged data stream");
2110 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2111 /* Validate delete flag of the stream */
2112 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2115 /* Delete it right now */
2116 consumer_del_stream(stream
, data_ht
);
2122 * Delete metadata stream that are flagged for deletion (endpoint_status).
2124 static void validate_endpoint_status_metadata_stream(
2125 struct lttng_poll_event
*pollset
)
2127 struct lttng_ht_iter iter
;
2128 struct lttng_consumer_stream
*stream
;
2130 DBG("Consumer delete flagged metadata stream");
2135 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2136 /* Validate delete flag of the stream */
2137 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2141 * Remove from pollset so the metadata thread can continue without
2142 * blocking on a deleted stream.
2144 lttng_poll_del(pollset
, stream
->wait_fd
);
2146 /* Delete it right now */
2147 consumer_del_metadata_stream(stream
, metadata_ht
);
2153 * Thread polls on metadata file descriptor and write them on disk or on the
2156 void *consumer_thread_metadata_poll(void *data
)
2158 int ret
, i
, pollfd
, err
= -1;
2159 uint32_t revents
, nb_fd
;
2160 struct lttng_consumer_stream
*stream
= NULL
;
2161 struct lttng_ht_iter iter
;
2162 struct lttng_ht_node_u64
*node
;
2163 struct lttng_poll_event events
;
2164 struct lttng_consumer_local_data
*ctx
= data
;
2167 rcu_register_thread();
2169 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2171 if (testpoint(consumerd_thread_metadata
)) {
2172 goto error_testpoint
;
2175 health_code_update();
2177 DBG("Thread metadata poll started");
2179 /* Size is set to 1 for the consumer_metadata pipe */
2180 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2182 ERR("Poll set creation failed");
2186 ret
= lttng_poll_add(&events
,
2187 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2193 DBG("Metadata main loop started");
2196 health_code_update();
2198 /* Only the metadata pipe is set */
2199 if (LTTNG_POLL_GETNB(&events
) == 0 && consumer_quit
== 1) {
2200 err
= 0; /* All is OK */
2205 DBG("Metadata poll wait with %d fd(s)", LTTNG_POLL_GETNB(&events
));
2206 health_poll_entry();
2207 ret
= lttng_poll_wait(&events
, -1);
2209 DBG("Metadata event catched in thread");
2211 if (errno
== EINTR
) {
2212 ERR("Poll EINTR catched");
2220 /* From here, the event is a metadata wait fd */
2221 for (i
= 0; i
< nb_fd
; i
++) {
2222 health_code_update();
2224 revents
= LTTNG_POLL_GETEV(&events
, i
);
2225 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2227 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2228 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2229 DBG("Metadata thread pipe hung up");
2231 * Remove the pipe from the poll set and continue the loop
2232 * since their might be data to consume.
2234 lttng_poll_del(&events
,
2235 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2236 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2238 } else if (revents
& LPOLLIN
) {
2241 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2242 &stream
, sizeof(stream
));
2243 if (pipe_len
< sizeof(stream
)) {
2244 PERROR("read metadata stream");
2246 * Continue here to handle the rest of the streams.
2251 /* A NULL stream means that the state has changed. */
2252 if (stream
== NULL
) {
2253 /* Check for deleted streams. */
2254 validate_endpoint_status_metadata_stream(&events
);
2258 DBG("Adding metadata stream %d to poll set",
2261 /* Add metadata stream to the global poll events list */
2262 lttng_poll_add(&events
, stream
->wait_fd
,
2263 LPOLLIN
| LPOLLPRI
| LPOLLHUP
);
2266 /* Handle other stream */
2272 uint64_t tmp_id
= (uint64_t) pollfd
;
2274 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2276 node
= lttng_ht_iter_get_node_u64(&iter
);
2279 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2282 /* Check for error event */
2283 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2284 DBG("Metadata fd %d is hup|err.", pollfd
);
2285 if (!stream
->hangup_flush_done
2286 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2287 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2288 DBG("Attempting to flush and consume the UST buffers");
2289 lttng_ustconsumer_on_stream_hangup(stream
);
2291 /* We just flushed the stream now read it. */
2293 health_code_update();
2295 len
= ctx
->on_buffer_ready(stream
, ctx
);
2297 * We don't check the return value here since if we get
2298 * a negative len, it means an error occured thus we
2299 * simply remove it from the poll set and free the
2305 lttng_poll_del(&events
, stream
->wait_fd
);
2307 * This call update the channel states, closes file descriptors
2308 * and securely free the stream.
2310 consumer_del_metadata_stream(stream
, metadata_ht
);
2311 } else if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2312 /* Get the data out of the metadata file descriptor */
2313 DBG("Metadata available on fd %d", pollfd
);
2314 assert(stream
->wait_fd
== pollfd
);
2317 health_code_update();
2319 len
= ctx
->on_buffer_ready(stream
, ctx
);
2321 * We don't check the return value here since if we get
2322 * a negative len, it means an error occured thus we
2323 * simply remove it from the poll set and free the
2328 /* It's ok to have an unavailable sub-buffer */
2329 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2330 /* Clean up stream from consumer and free it. */
2331 lttng_poll_del(&events
, stream
->wait_fd
);
2332 consumer_del_metadata_stream(stream
, metadata_ht
);
2336 /* Release RCU lock for the stream looked up */
2345 DBG("Metadata poll thread exiting");
2347 lttng_poll_clean(&events
);
2352 ERR("Health error occurred in %s", __func__
);
2354 health_unregister(health_consumerd
);
2355 rcu_unregister_thread();
2360 * This thread polls the fds in the set to consume the data and write
2361 * it to tracefile if necessary.
2363 void *consumer_thread_data_poll(void *data
)
2365 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2366 struct pollfd
*pollfd
= NULL
;
2367 /* local view of the streams */
2368 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2369 /* local view of consumer_data.fds_count */
2371 struct lttng_consumer_local_data
*ctx
= data
;
2374 rcu_register_thread();
2376 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2378 if (testpoint(consumerd_thread_data
)) {
2379 goto error_testpoint
;
2382 health_code_update();
2384 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
*));
2385 if (local_stream
== NULL
) {
2386 PERROR("local_stream malloc");
2391 health_code_update();
2397 * the fds set has been updated, we need to update our
2398 * local array as well
2400 pthread_mutex_lock(&consumer_data
.lock
);
2401 if (consumer_data
.need_update
) {
2406 local_stream
= NULL
;
2408 /* allocate for all fds + 1 for the consumer_data_pipe */
2409 pollfd
= zmalloc((consumer_data
.stream_count
+ 1) * sizeof(struct pollfd
));
2410 if (pollfd
== NULL
) {
2411 PERROR("pollfd malloc");
2412 pthread_mutex_unlock(&consumer_data
.lock
);
2416 /* allocate for all fds + 1 for the consumer_data_pipe */
2417 local_stream
= zmalloc((consumer_data
.stream_count
+ 1) *
2418 sizeof(struct lttng_consumer_stream
*));
2419 if (local_stream
== NULL
) {
2420 PERROR("local_stream malloc");
2421 pthread_mutex_unlock(&consumer_data
.lock
);
2424 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2427 ERR("Error in allocating pollfd or local_outfds");
2428 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2429 pthread_mutex_unlock(&consumer_data
.lock
);
2433 consumer_data
.need_update
= 0;
2435 pthread_mutex_unlock(&consumer_data
.lock
);
2437 /* No FDs and consumer_quit, consumer_cleanup the thread */
2438 if (nb_fd
== 0 && consumer_quit
== 1) {
2439 err
= 0; /* All is OK */
2442 /* poll on the array of fds */
2444 DBG("polling on %d fd", nb_fd
+ 1);
2445 health_poll_entry();
2446 num_rdy
= poll(pollfd
, nb_fd
+ 1, -1);
2448 DBG("poll num_rdy : %d", num_rdy
);
2449 if (num_rdy
== -1) {
2451 * Restart interrupted system call.
2453 if (errno
== EINTR
) {
2456 PERROR("Poll error");
2457 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2459 } else if (num_rdy
== 0) {
2460 DBG("Polling thread timed out");
2465 * If the consumer_data_pipe triggered poll go directly to the
2466 * beginning of the loop to update the array. We want to prioritize
2467 * array update over low-priority reads.
2469 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2470 ssize_t pipe_readlen
;
2472 DBG("consumer_data_pipe wake up");
2473 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2474 &new_stream
, sizeof(new_stream
));
2475 if (pipe_readlen
< sizeof(new_stream
)) {
2476 PERROR("Consumer data pipe");
2477 /* Continue so we can at least handle the current stream(s). */
2482 * If the stream is NULL, just ignore it. It's also possible that
2483 * the sessiond poll thread changed the consumer_quit state and is
2484 * waking us up to test it.
2486 if (new_stream
== NULL
) {
2487 validate_endpoint_status_data_stream();
2491 /* Continue to update the local streams and handle prio ones */
2495 /* Take care of high priority channels first. */
2496 for (i
= 0; i
< nb_fd
; i
++) {
2497 health_code_update();
2499 if (local_stream
[i
] == NULL
) {
2502 if (pollfd
[i
].revents
& POLLPRI
) {
2503 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2505 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2506 /* it's ok to have an unavailable sub-buffer */
2507 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2508 /* Clean the stream and free it. */
2509 consumer_del_stream(local_stream
[i
], data_ht
);
2510 local_stream
[i
] = NULL
;
2511 } else if (len
> 0) {
2512 local_stream
[i
]->data_read
= 1;
2518 * If we read high prio channel in this loop, try again
2519 * for more high prio data.
2525 /* Take care of low priority channels. */
2526 for (i
= 0; i
< nb_fd
; i
++) {
2527 health_code_update();
2529 if (local_stream
[i
] == NULL
) {
2532 if ((pollfd
[i
].revents
& POLLIN
) ||
2533 local_stream
[i
]->hangup_flush_done
) {
2534 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2535 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2536 /* it's ok to have an unavailable sub-buffer */
2537 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2538 /* Clean the stream and free it. */
2539 consumer_del_stream(local_stream
[i
], data_ht
);
2540 local_stream
[i
] = NULL
;
2541 } else if (len
> 0) {
2542 local_stream
[i
]->data_read
= 1;
2547 /* Handle hangup and errors */
2548 for (i
= 0; i
< nb_fd
; i
++) {
2549 health_code_update();
2551 if (local_stream
[i
] == NULL
) {
2554 if (!local_stream
[i
]->hangup_flush_done
2555 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2556 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2557 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2558 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2560 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2561 /* Attempt read again, for the data we just flushed. */
2562 local_stream
[i
]->data_read
= 1;
2565 * If the poll flag is HUP/ERR/NVAL and we have
2566 * read no data in this pass, we can remove the
2567 * stream from its hash table.
2569 if ((pollfd
[i
].revents
& POLLHUP
)) {
2570 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2571 if (!local_stream
[i
]->data_read
) {
2572 consumer_del_stream(local_stream
[i
], data_ht
);
2573 local_stream
[i
] = NULL
;
2576 } else if (pollfd
[i
].revents
& POLLERR
) {
2577 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2578 if (!local_stream
[i
]->data_read
) {
2579 consumer_del_stream(local_stream
[i
], data_ht
);
2580 local_stream
[i
] = NULL
;
2583 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2584 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2585 if (!local_stream
[i
]->data_read
) {
2586 consumer_del_stream(local_stream
[i
], data_ht
);
2587 local_stream
[i
] = NULL
;
2591 if (local_stream
[i
] != NULL
) {
2592 local_stream
[i
]->data_read
= 0;
2599 DBG("polling thread exiting");
2604 * Close the write side of the pipe so epoll_wait() in
2605 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2606 * read side of the pipe. If we close them both, epoll_wait strangely does
2607 * not return and could create a endless wait period if the pipe is the
2608 * only tracked fd in the poll set. The thread will take care of closing
2611 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2616 ERR("Health error occurred in %s", __func__
);
2618 health_unregister(health_consumerd
);
2620 rcu_unregister_thread();
2625 * Close wake-up end of each stream belonging to the channel. This will
2626 * allow the poll() on the stream read-side to detect when the
2627 * write-side (application) finally closes them.
2630 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2632 struct lttng_ht
*ht
;
2633 struct lttng_consumer_stream
*stream
;
2634 struct lttng_ht_iter iter
;
2636 ht
= consumer_data
.stream_per_chan_id_ht
;
2639 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2640 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2641 ht
->match_fct
, &channel
->key
,
2642 &iter
.iter
, stream
, node_channel_id
.node
) {
2644 * Protect against teardown with mutex.
2646 pthread_mutex_lock(&stream
->lock
);
2647 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2650 switch (consumer_data
.type
) {
2651 case LTTNG_CONSUMER_KERNEL
:
2653 case LTTNG_CONSUMER32_UST
:
2654 case LTTNG_CONSUMER64_UST
:
2655 if (stream
->metadata_flag
) {
2656 /* Safe and protected by the stream lock. */
2657 lttng_ustconsumer_close_metadata(stream
->chan
);
2660 * Note: a mutex is taken internally within
2661 * liblttng-ust-ctl to protect timer wakeup_fd
2662 * use from concurrent close.
2664 lttng_ustconsumer_close_stream_wakeup(stream
);
2668 ERR("Unknown consumer_data type");
2672 pthread_mutex_unlock(&stream
->lock
);
2677 static void destroy_channel_ht(struct lttng_ht
*ht
)
2679 struct lttng_ht_iter iter
;
2680 struct lttng_consumer_channel
*channel
;
2688 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2689 ret
= lttng_ht_del(ht
, &iter
);
2694 lttng_ht_destroy(ht
);
2698 * This thread polls the channel fds to detect when they are being
2699 * closed. It closes all related streams if the channel is detected as
2700 * closed. It is currently only used as a shim layer for UST because the
2701 * consumerd needs to keep the per-stream wakeup end of pipes open for
2704 void *consumer_thread_channel_poll(void *data
)
2706 int ret
, i
, pollfd
, err
= -1;
2707 uint32_t revents
, nb_fd
;
2708 struct lttng_consumer_channel
*chan
= NULL
;
2709 struct lttng_ht_iter iter
;
2710 struct lttng_ht_node_u64
*node
;
2711 struct lttng_poll_event events
;
2712 struct lttng_consumer_local_data
*ctx
= data
;
2713 struct lttng_ht
*channel_ht
;
2715 rcu_register_thread();
2717 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
2719 if (testpoint(consumerd_thread_channel
)) {
2720 goto error_testpoint
;
2723 health_code_update();
2725 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2727 /* ENOMEM at this point. Better to bail out. */
2731 DBG("Thread channel poll started");
2733 /* Size is set to 1 for the consumer_channel pipe */
2734 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2736 ERR("Poll set creation failed");
2740 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2746 DBG("Channel main loop started");
2749 health_code_update();
2751 /* Only the channel pipe is set */
2752 if (LTTNG_POLL_GETNB(&events
) == 0 && consumer_quit
== 1) {
2753 err
= 0; /* All is OK */
2758 DBG("Channel poll wait with %d fd(s)", LTTNG_POLL_GETNB(&events
));
2759 health_poll_entry();
2760 ret
= lttng_poll_wait(&events
, -1);
2762 DBG("Channel event catched in thread");
2764 if (errno
== EINTR
) {
2765 ERR("Poll EINTR catched");
2773 /* From here, the event is a channel wait fd */
2774 for (i
= 0; i
< nb_fd
; i
++) {
2775 health_code_update();
2777 revents
= LTTNG_POLL_GETEV(&events
, i
);
2778 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2780 /* Just don't waste time if no returned events for the fd */
2784 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2785 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2786 DBG("Channel thread pipe hung up");
2788 * Remove the pipe from the poll set and continue the loop
2789 * since their might be data to consume.
2791 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2793 } else if (revents
& LPOLLIN
) {
2794 enum consumer_channel_action action
;
2797 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
2799 ERR("Error reading channel pipe");
2804 case CONSUMER_CHANNEL_ADD
:
2805 DBG("Adding channel %d to poll set",
2808 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
2811 lttng_ht_add_unique_u64(channel_ht
,
2812 &chan
->wait_fd_node
);
2814 /* Add channel to the global poll events list */
2815 lttng_poll_add(&events
, chan
->wait_fd
,
2816 LPOLLIN
| LPOLLPRI
);
2818 case CONSUMER_CHANNEL_DEL
:
2821 * This command should never be called if the channel
2822 * has streams monitored by either the data or metadata
2823 * thread. The consumer only notify this thread with a
2824 * channel del. command if it receives a destroy
2825 * channel command from the session daemon that send it
2826 * if a command prior to the GET_CHANNEL failed.
2830 chan
= consumer_find_channel(key
);
2833 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
2836 lttng_poll_del(&events
, chan
->wait_fd
);
2837 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
2838 ret
= lttng_ht_del(channel_ht
, &iter
);
2841 switch (consumer_data
.type
) {
2842 case LTTNG_CONSUMER_KERNEL
:
2844 case LTTNG_CONSUMER32_UST
:
2845 case LTTNG_CONSUMER64_UST
:
2846 health_code_update();
2847 /* Destroy streams that might have been left in the stream list. */
2848 clean_channel_stream_list(chan
);
2851 ERR("Unknown consumer_data type");
2856 * Release our own refcount. Force channel deletion even if
2857 * streams were not initialized.
2859 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
2860 consumer_del_channel(chan
);
2865 case CONSUMER_CHANNEL_QUIT
:
2867 * Remove the pipe from the poll set and continue the loop
2868 * since their might be data to consume.
2870 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2873 ERR("Unknown action");
2878 /* Handle other stream */
2884 uint64_t tmp_id
= (uint64_t) pollfd
;
2886 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
2888 node
= lttng_ht_iter_get_node_u64(&iter
);
2891 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
2894 /* Check for error event */
2895 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2896 DBG("Channel fd %d is hup|err.", pollfd
);
2898 lttng_poll_del(&events
, chan
->wait_fd
);
2899 ret
= lttng_ht_del(channel_ht
, &iter
);
2903 * This will close the wait fd for each stream associated to
2904 * this channel AND monitored by the data/metadata thread thus
2905 * will be clean by the right thread.
2907 consumer_close_channel_streams(chan
);
2909 /* Release our own refcount */
2910 if (!uatomic_sub_return(&chan
->refcount
, 1)
2911 && !uatomic_read(&chan
->nb_init_stream_left
)) {
2912 consumer_del_channel(chan
);
2916 /* Release RCU lock for the channel looked up */
2924 lttng_poll_clean(&events
);
2926 destroy_channel_ht(channel_ht
);
2929 DBG("Channel poll thread exiting");
2932 ERR("Health error occurred in %s", __func__
);
2934 health_unregister(health_consumerd
);
2935 rcu_unregister_thread();
2939 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
2940 struct pollfd
*sockpoll
, int client_socket
)
2947 if (lttng_consumer_poll_socket(sockpoll
) < 0) {
2951 DBG("Metadata connection on client_socket");
2953 /* Blocking call, waiting for transmission */
2954 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
2955 if (ctx
->consumer_metadata_socket
< 0) {
2956 WARN("On accept metadata");
2967 * This thread listens on the consumerd socket and receives the file
2968 * descriptors from the session daemon.
2970 void *consumer_thread_sessiond_poll(void *data
)
2972 int sock
= -1, client_socket
, ret
, err
= -1;
2974 * structure to poll for incoming data on communication socket avoids
2975 * making blocking sockets.
2977 struct pollfd consumer_sockpoll
[2];
2978 struct lttng_consumer_local_data
*ctx
= data
;
2980 rcu_register_thread();
2982 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
2984 if (testpoint(consumerd_thread_sessiond
)) {
2985 goto error_testpoint
;
2988 health_code_update();
2990 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
2991 unlink(ctx
->consumer_command_sock_path
);
2992 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
2993 if (client_socket
< 0) {
2994 ERR("Cannot create command socket");
2998 ret
= lttcomm_listen_unix_sock(client_socket
);
3003 DBG("Sending ready command to lttng-sessiond");
3004 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3005 /* return < 0 on error, but == 0 is not fatal */
3007 ERR("Error sending ready command to lttng-sessiond");
3011 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3012 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3013 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3014 consumer_sockpoll
[1].fd
= client_socket
;
3015 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3017 if (lttng_consumer_poll_socket(consumer_sockpoll
) < 0) {
3020 DBG("Connection on client_socket");
3022 /* Blocking call, waiting for transmission */
3023 sock
= lttcomm_accept_unix_sock(client_socket
);
3030 * Setup metadata socket which is the second socket connection on the
3031 * command unix socket.
3033 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3038 /* This socket is not useful anymore. */
3039 ret
= close(client_socket
);
3041 PERROR("close client_socket");
3045 /* update the polling structure to poll on the established socket */
3046 consumer_sockpoll
[1].fd
= sock
;
3047 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3050 health_code_update();
3052 health_poll_entry();
3053 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3058 DBG("Incoming command on sock");
3059 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3060 if (ret
== -ENOENT
) {
3061 DBG("Received STOP command");
3066 * This could simply be a session daemon quitting. Don't output
3069 DBG("Communication interrupted on command socket");
3073 if (consumer_quit
) {
3074 DBG("consumer_thread_receive_fds received quit from signal");
3075 err
= 0; /* All is OK */
3078 DBG("received command on sock");
3084 DBG("Consumer thread sessiond poll exiting");
3087 * Close metadata streams since the producer is the session daemon which
3090 * NOTE: for now, this only applies to the UST tracer.
3092 lttng_consumer_close_all_metadata();
3095 * when all fds have hung up, the polling thread
3101 * Notify the data poll thread to poll back again and test the
3102 * consumer_quit state that we just set so to quit gracefully.
3104 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3106 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3108 notify_health_quit_pipe(health_quit_pipe
);
3110 /* Cleaning up possibly open sockets. */
3114 PERROR("close sock sessiond poll");
3117 if (client_socket
>= 0) {
3118 ret
= close(client_socket
);
3120 PERROR("close client_socket sessiond poll");
3127 ERR("Health error occurred in %s", __func__
);
3129 health_unregister(health_consumerd
);
3131 rcu_unregister_thread();
3135 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3136 struct lttng_consumer_local_data
*ctx
)
3140 pthread_mutex_lock(&stream
->lock
);
3141 if (stream
->metadata_flag
) {
3142 pthread_mutex_lock(&stream
->metadata_rdv_lock
);
3145 switch (consumer_data
.type
) {
3146 case LTTNG_CONSUMER_KERNEL
:
3147 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
3149 case LTTNG_CONSUMER32_UST
:
3150 case LTTNG_CONSUMER64_UST
:
3151 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
3154 ERR("Unknown consumer_data type");
3160 if (stream
->metadata_flag
) {
3161 pthread_cond_broadcast(&stream
->metadata_rdv
);
3162 pthread_mutex_unlock(&stream
->metadata_rdv_lock
);
3164 pthread_mutex_unlock(&stream
->lock
);
3168 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3170 switch (consumer_data
.type
) {
3171 case LTTNG_CONSUMER_KERNEL
:
3172 return lttng_kconsumer_on_recv_stream(stream
);
3173 case LTTNG_CONSUMER32_UST
:
3174 case LTTNG_CONSUMER64_UST
:
3175 return lttng_ustconsumer_on_recv_stream(stream
);
3177 ERR("Unknown consumer_data type");
3184 * Allocate and set consumer data hash tables.
3186 int lttng_consumer_init(void)
3188 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3189 if (!consumer_data
.channel_ht
) {
3193 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3194 if (!consumer_data
.relayd_ht
) {
3198 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3199 if (!consumer_data
.stream_list_ht
) {
3203 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3204 if (!consumer_data
.stream_per_chan_id_ht
) {
3208 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3213 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3225 * Process the ADD_RELAYD command receive by a consumer.
3227 * This will create a relayd socket pair and add it to the relayd hash table.
3228 * The caller MUST acquire a RCU read side lock before calling it.
3230 int consumer_add_relayd_socket(uint64_t net_seq_idx
, int sock_type
,
3231 struct lttng_consumer_local_data
*ctx
, int sock
,
3232 struct pollfd
*consumer_sockpoll
,
3233 struct lttcomm_relayd_sock
*relayd_sock
, uint64_t sessiond_id
,
3234 uint64_t relayd_session_id
)
3236 int fd
= -1, ret
= -1, relayd_created
= 0;
3237 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3238 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3241 assert(relayd_sock
);
3243 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", net_seq_idx
);
3245 /* Get relayd reference if exists. */
3246 relayd
= consumer_find_relayd(net_seq_idx
);
3247 if (relayd
== NULL
) {
3248 assert(sock_type
== LTTNG_STREAM_CONTROL
);
3249 /* Not found. Allocate one. */
3250 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3251 if (relayd
== NULL
) {
3253 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3256 relayd
->sessiond_session_id
= sessiond_id
;
3261 * This code path MUST continue to the consumer send status message to
3262 * we can notify the session daemon and continue our work without
3263 * killing everything.
3267 * relayd key should never be found for control socket.
3269 assert(sock_type
!= LTTNG_STREAM_CONTROL
);
3272 /* First send a status message before receiving the fds. */
3273 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3275 /* Somehow, the session daemon is not responding anymore. */
3276 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3277 goto error_nosignal
;
3280 /* Poll on consumer socket. */
3281 if (lttng_consumer_poll_socket(consumer_sockpoll
) < 0) {
3282 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3284 goto error_nosignal
;
3287 /* Get relayd socket from session daemon */
3288 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3289 if (ret
!= sizeof(fd
)) {
3291 fd
= -1; /* Just in case it gets set with an invalid value. */
3294 * Failing to receive FDs might indicate a major problem such as
3295 * reaching a fd limit during the receive where the kernel returns a
3296 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3297 * don't take any chances and stop everything.
3299 * XXX: Feature request #558 will fix that and avoid this possible
3300 * issue when reaching the fd limit.
3302 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3303 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3307 /* Copy socket information and received FD */
3308 switch (sock_type
) {
3309 case LTTNG_STREAM_CONTROL
:
3310 /* Copy received lttcomm socket */
3311 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3312 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3313 /* Handle create_sock error. */
3315 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3319 * Close the socket created internally by
3320 * lttcomm_create_sock, so we can replace it by the one
3321 * received from sessiond.
3323 if (close(relayd
->control_sock
.sock
.fd
)) {
3327 /* Assign new file descriptor */
3328 relayd
->control_sock
.sock
.fd
= fd
;
3329 fd
= -1; /* For error path */
3330 /* Assign version values. */
3331 relayd
->control_sock
.major
= relayd_sock
->major
;
3332 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3334 relayd
->relayd_session_id
= relayd_session_id
;
3337 case LTTNG_STREAM_DATA
:
3338 /* Copy received lttcomm socket */
3339 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3340 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3341 /* Handle create_sock error. */
3343 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3347 * Close the socket created internally by
3348 * lttcomm_create_sock, so we can replace it by the one
3349 * received from sessiond.
3351 if (close(relayd
->data_sock
.sock
.fd
)) {
3355 /* Assign new file descriptor */
3356 relayd
->data_sock
.sock
.fd
= fd
;
3357 fd
= -1; /* for eventual error paths */
3358 /* Assign version values. */
3359 relayd
->data_sock
.major
= relayd_sock
->major
;
3360 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3363 ERR("Unknown relayd socket type (%d)", sock_type
);
3365 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3369 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3370 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3371 relayd
->net_seq_idx
, fd
);
3373 /* We successfully added the socket. Send status back. */
3374 ret
= consumer_send_status_msg(sock
, ret_code
);
3376 /* Somehow, the session daemon is not responding anymore. */
3377 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3378 goto error_nosignal
;
3382 * Add relayd socket pair to consumer data hashtable. If object already
3383 * exists or on error, the function gracefully returns.
3391 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3392 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3396 /* Close received socket if valid. */
3399 PERROR("close received socket");
3403 if (relayd_created
) {
3411 * Try to lock the stream mutex.
3413 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
3415 static int stream_try_lock(struct lttng_consumer_stream
*stream
)
3422 * Try to lock the stream mutex. On failure, we know that the stream is
3423 * being used else where hence there is data still being extracted.
3425 ret
= pthread_mutex_trylock(&stream
->lock
);
3427 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
3439 * Search for a relayd associated to the session id and return the reference.
3441 * A rcu read side lock MUST be acquire before calling this function and locked
3442 * until the relayd object is no longer necessary.
3444 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3446 struct lttng_ht_iter iter
;
3447 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3449 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3450 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3453 * Check by sessiond id which is unique here where the relayd session
3454 * id might not be when having multiple relayd.
3456 if (relayd
->sessiond_session_id
== id
) {
3457 /* Found the relayd. There can be only one per id. */
3469 * Check if for a given session id there is still data needed to be extract
3472 * Return 1 if data is pending or else 0 meaning ready to be read.
3474 int consumer_data_pending(uint64_t id
)
3477 struct lttng_ht_iter iter
;
3478 struct lttng_ht
*ht
;
3479 struct lttng_consumer_stream
*stream
;
3480 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3481 int (*data_pending
)(struct lttng_consumer_stream
*);
3483 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3486 pthread_mutex_lock(&consumer_data
.lock
);
3488 switch (consumer_data
.type
) {
3489 case LTTNG_CONSUMER_KERNEL
:
3490 data_pending
= lttng_kconsumer_data_pending
;
3492 case LTTNG_CONSUMER32_UST
:
3493 case LTTNG_CONSUMER64_UST
:
3494 data_pending
= lttng_ustconsumer_data_pending
;
3497 ERR("Unknown consumer data type");
3501 /* Ease our life a bit */
3502 ht
= consumer_data
.stream_list_ht
;
3504 relayd
= find_relayd_by_session_id(id
);
3506 /* Send init command for data pending. */
3507 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3508 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3509 relayd
->relayd_session_id
);
3510 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3512 /* Communication error thus the relayd so no data pending. */
3513 goto data_not_pending
;
3517 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3518 ht
->hash_fct(&id
, lttng_ht_seed
),
3520 &iter
.iter
, stream
, node_session_id
.node
) {
3521 /* If this call fails, the stream is being used hence data pending. */
3522 ret
= stream_try_lock(stream
);
3528 * A removed node from the hash table indicates that the stream has
3529 * been deleted thus having a guarantee that the buffers are closed
3530 * on the consumer side. However, data can still be transmitted
3531 * over the network so don't skip the relayd check.
3533 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3536 * An empty output file is not valid. We need at least one packet
3537 * generated per stream, even if it contains no event, so it
3538 * contains at least one packet header.
3540 if (stream
->output_written
== 0) {
3541 pthread_mutex_unlock(&stream
->lock
);
3544 /* Check the stream if there is data in the buffers. */
3545 ret
= data_pending(stream
);
3547 pthread_mutex_unlock(&stream
->lock
);
3554 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3555 if (stream
->metadata_flag
) {
3556 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3557 stream
->relayd_stream_id
);
3559 ret
= relayd_data_pending(&relayd
->control_sock
,
3560 stream
->relayd_stream_id
,
3561 stream
->next_net_seq_num
- 1);
3563 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3565 pthread_mutex_unlock(&stream
->lock
);
3569 pthread_mutex_unlock(&stream
->lock
);
3573 unsigned int is_data_inflight
= 0;
3575 /* Send init command for data pending. */
3576 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3577 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3578 relayd
->relayd_session_id
, &is_data_inflight
);
3579 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3581 goto data_not_pending
;
3583 if (is_data_inflight
) {
3589 * Finding _no_ node in the hash table and no inflight data means that the
3590 * stream(s) have been removed thus data is guaranteed to be available for
3591 * analysis from the trace files.
3595 /* Data is available to be read by a viewer. */
3596 pthread_mutex_unlock(&consumer_data
.lock
);
3601 /* Data is still being extracted from buffers. */
3602 pthread_mutex_unlock(&consumer_data
.lock
);
3608 * Send a ret code status message to the sessiond daemon.
3610 * Return the sendmsg() return value.
3612 int consumer_send_status_msg(int sock
, int ret_code
)
3614 struct lttcomm_consumer_status_msg msg
;
3616 memset(&msg
, 0, sizeof(msg
));
3617 msg
.ret_code
= ret_code
;
3619 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3623 * Send a channel status message to the sessiond daemon.
3625 * Return the sendmsg() return value.
3627 int consumer_send_status_channel(int sock
,
3628 struct lttng_consumer_channel
*channel
)
3630 struct lttcomm_consumer_status_channel msg
;
3634 memset(&msg
, 0, sizeof(msg
));
3636 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3638 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3639 msg
.key
= channel
->key
;
3640 msg
.stream_count
= channel
->streams
.count
;
3643 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3647 * Using a maximum stream size with the produced and consumed position of a
3648 * stream, computes the new consumed position to be as close as possible to the
3649 * maximum possible stream size.
3651 * If maximum stream size is lower than the possible buffer size (produced -
3652 * consumed), the consumed_pos given is returned untouched else the new value
3655 unsigned long consumer_get_consumed_maxsize(unsigned long consumed_pos
,
3656 unsigned long produced_pos
, uint64_t max_stream_size
)
3658 if (max_stream_size
&& max_stream_size
< (produced_pos
- consumed_pos
)) {
3659 /* Offset from the produced position to get the latest buffers. */
3660 return produced_pos
- max_stream_size
;
3663 return consumed_pos
;