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
1099 * error, 1 if 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");
1125 * Set the error socket.
1127 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data
*ctx
,
1130 ctx
->consumer_error_socket
= sock
;
1134 * Set the command socket path.
1136 void lttng_consumer_set_command_sock_path(
1137 struct lttng_consumer_local_data
*ctx
, char *sock
)
1139 ctx
->consumer_command_sock_path
= sock
;
1143 * Send return code to the session daemon.
1144 * If the socket is not defined, we return 0, it is not a fatal error
1146 int lttng_consumer_send_error(struct lttng_consumer_local_data
*ctx
, int cmd
)
1148 if (ctx
->consumer_error_socket
> 0) {
1149 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
1150 sizeof(enum lttcomm_sessiond_command
));
1157 * Close all the tracefiles and stream fds and MUST be called when all
1158 * instances are destroyed i.e. when all threads were joined and are ended.
1160 void lttng_consumer_cleanup(void)
1162 struct lttng_ht_iter iter
;
1163 struct lttng_consumer_channel
*channel
;
1167 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
, channel
,
1169 consumer_del_channel(channel
);
1174 lttng_ht_destroy(consumer_data
.channel_ht
);
1176 cleanup_relayd_ht();
1178 lttng_ht_destroy(consumer_data
.stream_per_chan_id_ht
);
1181 * This HT contains streams that are freed by either the metadata thread or
1182 * the data thread so we do *nothing* on the hash table and simply destroy
1185 lttng_ht_destroy(consumer_data
.stream_list_ht
);
1189 * Called from signal handler.
1191 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
1196 ret
= lttng_write(ctx
->consumer_should_quit
[1], "4", 1);
1198 PERROR("write consumer quit");
1201 DBG("Consumer flag that it should quit");
1204 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1207 int outfd
= stream
->out_fd
;
1210 * This does a blocking write-and-wait on any page that belongs to the
1211 * subbuffer prior to the one we just wrote.
1212 * Don't care about error values, as these are just hints and ways to
1213 * limit the amount of page cache used.
1215 if (orig_offset
< stream
->max_sb_size
) {
1218 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1219 stream
->max_sb_size
,
1220 SYNC_FILE_RANGE_WAIT_BEFORE
1221 | SYNC_FILE_RANGE_WRITE
1222 | SYNC_FILE_RANGE_WAIT_AFTER
);
1224 * Give hints to the kernel about how we access the file:
1225 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1228 * We need to call fadvise again after the file grows because the
1229 * kernel does not seem to apply fadvise to non-existing parts of the
1232 * Call fadvise _after_ having waited for the page writeback to
1233 * complete because the dirty page writeback semantic is not well
1234 * defined. So it can be expected to lead to lower throughput in
1237 posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1238 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1242 * Initialise the necessary environnement :
1243 * - create a new context
1244 * - create the poll_pipe
1245 * - create the should_quit pipe (for signal handler)
1246 * - create the thread pipe (for splice)
1248 * Takes a function pointer as argument, this function is called when data is
1249 * available on a buffer. This function is responsible to do the
1250 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1251 * buffer configuration and then kernctl_put_next_subbuf at the end.
1253 * Returns a pointer to the new context or NULL on error.
1255 struct lttng_consumer_local_data
*lttng_consumer_create(
1256 enum lttng_consumer_type type
,
1257 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1258 struct lttng_consumer_local_data
*ctx
),
1259 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1260 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1261 int (*update_stream
)(uint64_t stream_key
, uint32_t state
))
1264 struct lttng_consumer_local_data
*ctx
;
1266 assert(consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1267 consumer_data
.type
== type
);
1268 consumer_data
.type
= type
;
1270 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
1272 PERROR("allocating context");
1276 ctx
->consumer_error_socket
= -1;
1277 ctx
->consumer_metadata_socket
= -1;
1278 pthread_mutex_init(&ctx
->metadata_socket_lock
, NULL
);
1279 /* assign the callbacks */
1280 ctx
->on_buffer_ready
= buffer_ready
;
1281 ctx
->on_recv_channel
= recv_channel
;
1282 ctx
->on_recv_stream
= recv_stream
;
1283 ctx
->on_update_stream
= update_stream
;
1285 ctx
->consumer_data_pipe
= lttng_pipe_open(0);
1286 if (!ctx
->consumer_data_pipe
) {
1287 goto error_poll_pipe
;
1290 ret
= pipe(ctx
->consumer_should_quit
);
1292 PERROR("Error creating recv pipe");
1293 goto error_quit_pipe
;
1296 ret
= pipe(ctx
->consumer_thread_pipe
);
1298 PERROR("Error creating thread pipe");
1299 goto error_thread_pipe
;
1302 ret
= pipe(ctx
->consumer_channel_pipe
);
1304 PERROR("Error creating channel pipe");
1305 goto error_channel_pipe
;
1308 ctx
->consumer_metadata_pipe
= lttng_pipe_open(0);
1309 if (!ctx
->consumer_metadata_pipe
) {
1310 goto error_metadata_pipe
;
1313 ret
= utils_create_pipe(ctx
->consumer_splice_metadata_pipe
);
1315 goto error_splice_pipe
;
1321 lttng_pipe_destroy(ctx
->consumer_metadata_pipe
);
1322 error_metadata_pipe
:
1323 utils_close_pipe(ctx
->consumer_channel_pipe
);
1325 utils_close_pipe(ctx
->consumer_thread_pipe
);
1327 utils_close_pipe(ctx
->consumer_should_quit
);
1329 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1337 * Iterate over all streams of the hashtable and free them properly.
1339 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1341 struct lttng_ht_iter iter
;
1342 struct lttng_consumer_stream
*stream
;
1349 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1351 * Ignore return value since we are currently cleaning up so any error
1354 (void) consumer_del_stream(stream
, ht
);
1358 lttng_ht_destroy(ht
);
1362 * Iterate over all streams of the metadata hashtable and free them
1365 static void destroy_metadata_stream_ht(struct lttng_ht
*ht
)
1367 struct lttng_ht_iter iter
;
1368 struct lttng_consumer_stream
*stream
;
1375 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1377 * Ignore return value since we are currently cleaning up so any error
1380 (void) consumer_del_metadata_stream(stream
, ht
);
1384 lttng_ht_destroy(ht
);
1388 * Close all fds associated with the instance and free the context.
1390 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1394 DBG("Consumer destroying it. Closing everything.");
1396 destroy_data_stream_ht(data_ht
);
1397 destroy_metadata_stream_ht(metadata_ht
);
1399 ret
= close(ctx
->consumer_error_socket
);
1403 ret
= close(ctx
->consumer_metadata_socket
);
1407 utils_close_pipe(ctx
->consumer_thread_pipe
);
1408 utils_close_pipe(ctx
->consumer_channel_pipe
);
1409 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1410 lttng_pipe_destroy(ctx
->consumer_metadata_pipe
);
1411 utils_close_pipe(ctx
->consumer_should_quit
);
1412 utils_close_pipe(ctx
->consumer_splice_metadata_pipe
);
1414 unlink(ctx
->consumer_command_sock_path
);
1419 * Write the metadata stream id on the specified file descriptor.
1421 static int write_relayd_metadata_id(int fd
,
1422 struct lttng_consumer_stream
*stream
,
1423 struct consumer_relayd_sock_pair
*relayd
, unsigned long padding
)
1426 struct lttcomm_relayd_metadata_payload hdr
;
1428 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1429 hdr
.padding_size
= htobe32(padding
);
1430 ret
= lttng_write(fd
, (void *) &hdr
, sizeof(hdr
));
1431 if (ret
< sizeof(hdr
)) {
1433 * This error means that the fd's end is closed so ignore the perror
1434 * not to clubber the error output since this can happen in a normal
1437 if (errno
!= EPIPE
) {
1438 PERROR("write metadata stream id");
1440 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1442 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1443 * handle writting the missing part so report that as an error and
1444 * don't lie to the caller.
1449 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1450 stream
->relayd_stream_id
, padding
);
1457 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1458 * core function for writing trace buffers to either the local filesystem or
1461 * It must be called with the stream lock held.
1463 * Careful review MUST be put if any changes occur!
1465 * Returns the number of bytes written
1467 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1468 struct lttng_consumer_local_data
*ctx
,
1469 struct lttng_consumer_stream
*stream
, unsigned long len
,
1470 unsigned long padding
,
1471 struct ctf_packet_index
*index
)
1473 unsigned long mmap_offset
;
1476 off_t orig_offset
= stream
->out_fd_offset
;
1477 /* Default is on the disk */
1478 int outfd
= stream
->out_fd
;
1479 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1480 unsigned int relayd_hang_up
= 0;
1482 /* RCU lock for the relayd pointer */
1485 /* Flag that the current stream if set for network streaming. */
1486 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1487 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1488 if (relayd
== NULL
) {
1494 /* get the offset inside the fd to mmap */
1495 switch (consumer_data
.type
) {
1496 case LTTNG_CONSUMER_KERNEL
:
1497 mmap_base
= stream
->mmap_base
;
1498 ret
= kernctl_get_mmap_read_offset(stream
->wait_fd
, &mmap_offset
);
1501 PERROR("tracer ctl get_mmap_read_offset");
1505 case LTTNG_CONSUMER32_UST
:
1506 case LTTNG_CONSUMER64_UST
:
1507 mmap_base
= lttng_ustctl_get_mmap_base(stream
);
1509 ERR("read mmap get mmap base for stream %s", stream
->name
);
1513 ret
= lttng_ustctl_get_mmap_read_offset(stream
, &mmap_offset
);
1515 PERROR("tracer ctl get_mmap_read_offset");
1521 ERR("Unknown consumer_data type");
1525 /* Handle stream on the relayd if the output is on the network */
1527 unsigned long netlen
= len
;
1530 * Lock the control socket for the complete duration of the function
1531 * since from this point on we will use the socket.
1533 if (stream
->metadata_flag
) {
1534 /* Metadata requires the control socket. */
1535 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1536 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1539 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1544 /* Use the returned socket. */
1547 /* Write metadata stream id before payload */
1548 if (stream
->metadata_flag
) {
1549 ret
= write_relayd_metadata_id(outfd
, stream
, relayd
, padding
);
1556 /* No streaming, we have to set the len with the full padding */
1560 * Check if we need to change the tracefile before writing the packet.
1562 if (stream
->chan
->tracefile_size
> 0 &&
1563 (stream
->tracefile_size_current
+ len
) >
1564 stream
->chan
->tracefile_size
) {
1565 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1566 stream
->name
, stream
->chan
->tracefile_size
,
1567 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1568 stream
->out_fd
, &(stream
->tracefile_count_current
),
1571 ERR("Rotating output file");
1574 outfd
= stream
->out_fd
;
1576 if (stream
->index_fd
>= 0) {
1577 ret
= index_create_file(stream
->chan
->pathname
,
1578 stream
->name
, stream
->uid
, stream
->gid
,
1579 stream
->chan
->tracefile_size
,
1580 stream
->tracefile_count_current
);
1584 stream
->index_fd
= ret
;
1587 /* Reset current size because we just perform a rotation. */
1588 stream
->tracefile_size_current
= 0;
1589 stream
->out_fd_offset
= 0;
1592 stream
->tracefile_size_current
+= len
;
1594 index
->offset
= htobe64(stream
->out_fd_offset
);
1599 * This call guarantee that len or less is returned. It's impossible to
1600 * receive a ret value that is bigger than len.
1602 ret
= lttng_write(outfd
, mmap_base
+ mmap_offset
, len
);
1603 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, len
);
1604 if (ret
< 0 || ((size_t) ret
!= len
)) {
1606 * Report error to caller if nothing was written else at least send the
1614 /* Socket operation failed. We consider the relayd dead */
1615 if (errno
== EPIPE
|| errno
== EINVAL
|| errno
== EBADF
) {
1617 * This is possible if the fd is closed on the other side
1618 * (outfd) or any write problem. It can be verbose a bit for a
1619 * normal execution if for instance the relayd is stopped
1620 * abruptly. This can happen so set this to a DBG statement.
1622 DBG("Consumer mmap write detected relayd hang up");
1624 /* Unhandled error, print it and stop function right now. */
1625 PERROR("Error in write mmap (ret %zd != len %lu)", ret
, len
);
1629 stream
->output_written
+= ret
;
1631 /* This call is useless on a socket so better save a syscall. */
1633 /* This won't block, but will start writeout asynchronously */
1634 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, len
,
1635 SYNC_FILE_RANGE_WRITE
);
1636 stream
->out_fd_offset
+= len
;
1638 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1642 * This is a special case that the relayd has closed its socket. Let's
1643 * cleanup the relayd object and all associated streams.
1645 if (relayd
&& relayd_hang_up
) {
1646 cleanup_relayd(relayd
, ctx
);
1650 /* Unlock only if ctrl socket used */
1651 if (relayd
&& stream
->metadata_flag
) {
1652 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1660 * Splice the data from the ring buffer to the tracefile.
1662 * It must be called with the stream lock held.
1664 * Returns the number of bytes spliced.
1666 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1667 struct lttng_consumer_local_data
*ctx
,
1668 struct lttng_consumer_stream
*stream
, unsigned long len
,
1669 unsigned long padding
,
1670 struct ctf_packet_index
*index
)
1672 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1674 off_t orig_offset
= stream
->out_fd_offset
;
1675 int fd
= stream
->wait_fd
;
1676 /* Default is on the disk */
1677 int outfd
= stream
->out_fd
;
1678 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1680 unsigned int relayd_hang_up
= 0;
1682 switch (consumer_data
.type
) {
1683 case LTTNG_CONSUMER_KERNEL
:
1685 case LTTNG_CONSUMER32_UST
:
1686 case LTTNG_CONSUMER64_UST
:
1687 /* Not supported for user space tracing */
1690 ERR("Unknown consumer_data type");
1694 /* RCU lock for the relayd pointer */
1697 /* Flag that the current stream if set for network streaming. */
1698 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1699 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1700 if (relayd
== NULL
) {
1707 * Choose right pipe for splice. Metadata and trace data are handled by
1708 * different threads hence the use of two pipes in order not to race or
1709 * corrupt the written data.
1711 if (stream
->metadata_flag
) {
1712 splice_pipe
= ctx
->consumer_splice_metadata_pipe
;
1714 splice_pipe
= ctx
->consumer_thread_pipe
;
1717 /* Write metadata stream id before payload */
1719 unsigned long total_len
= len
;
1721 if (stream
->metadata_flag
) {
1723 * Lock the control socket for the complete duration of the function
1724 * since from this point on we will use the socket.
1726 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1728 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
, relayd
,
1736 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1739 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1745 /* Use the returned socket. */
1748 /* No streaming, we have to set the len with the full padding */
1752 * Check if we need to change the tracefile before writing the packet.
1754 if (stream
->chan
->tracefile_size
> 0 &&
1755 (stream
->tracefile_size_current
+ len
) >
1756 stream
->chan
->tracefile_size
) {
1757 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1758 stream
->name
, stream
->chan
->tracefile_size
,
1759 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1760 stream
->out_fd
, &(stream
->tracefile_count_current
),
1764 ERR("Rotating output file");
1767 outfd
= stream
->out_fd
;
1769 if (stream
->index_fd
>= 0) {
1770 ret
= index_create_file(stream
->chan
->pathname
,
1771 stream
->name
, stream
->uid
, stream
->gid
,
1772 stream
->chan
->tracefile_size
,
1773 stream
->tracefile_count_current
);
1778 stream
->index_fd
= ret
;
1781 /* Reset current size because we just perform a rotation. */
1782 stream
->tracefile_size_current
= 0;
1783 stream
->out_fd_offset
= 0;
1786 stream
->tracefile_size_current
+= len
;
1787 index
->offset
= htobe64(stream
->out_fd_offset
);
1791 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1792 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1793 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1794 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1795 DBG("splice chan to pipe, ret %zd", ret_splice
);
1796 if (ret_splice
< 0) {
1799 PERROR("Error in relay splice");
1803 /* Handle stream on the relayd if the output is on the network */
1804 if (relayd
&& stream
->metadata_flag
) {
1805 size_t metadata_payload_size
=
1806 sizeof(struct lttcomm_relayd_metadata_payload
);
1808 /* Update counter to fit the spliced data */
1809 ret_splice
+= metadata_payload_size
;
1810 len
+= metadata_payload_size
;
1812 * We do this so the return value can match the len passed as
1813 * argument to this function.
1815 written
-= metadata_payload_size
;
1818 /* Splice data out */
1819 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1820 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1821 DBG("Consumer splice pipe to file, ret %zd", ret_splice
);
1822 if (ret_splice
< 0) {
1827 } else if (ret_splice
> len
) {
1829 * We don't expect this code path to be executed but you never know
1830 * so this is an extra protection agains a buggy splice().
1833 written
+= ret_splice
;
1834 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice
,
1838 /* All good, update current len and continue. */
1842 /* This call is useless on a socket so better save a syscall. */
1844 /* This won't block, but will start writeout asynchronously */
1845 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1846 SYNC_FILE_RANGE_WRITE
);
1847 stream
->out_fd_offset
+= ret_splice
;
1849 stream
->output_written
+= ret_splice
;
1850 written
+= ret_splice
;
1852 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1857 * This is a special case that the relayd has closed its socket. Let's
1858 * cleanup the relayd object and all associated streams.
1860 if (relayd
&& relayd_hang_up
) {
1861 cleanup_relayd(relayd
, ctx
);
1862 /* Skip splice error so the consumer does not fail */
1867 /* send the appropriate error description to sessiond */
1870 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1873 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1876 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
1881 if (relayd
&& stream
->metadata_flag
) {
1882 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1890 * Take a snapshot for a specific fd
1892 * Returns 0 on success, < 0 on error
1894 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
1896 switch (consumer_data
.type
) {
1897 case LTTNG_CONSUMER_KERNEL
:
1898 return lttng_kconsumer_take_snapshot(stream
);
1899 case LTTNG_CONSUMER32_UST
:
1900 case LTTNG_CONSUMER64_UST
:
1901 return lttng_ustconsumer_take_snapshot(stream
);
1903 ERR("Unknown consumer_data type");
1910 * Get the produced position
1912 * Returns 0 on success, < 0 on error
1914 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
1917 switch (consumer_data
.type
) {
1918 case LTTNG_CONSUMER_KERNEL
:
1919 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
1920 case LTTNG_CONSUMER32_UST
:
1921 case LTTNG_CONSUMER64_UST
:
1922 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
1924 ERR("Unknown consumer_data type");
1930 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
1931 int sock
, struct pollfd
*consumer_sockpoll
)
1933 switch (consumer_data
.type
) {
1934 case LTTNG_CONSUMER_KERNEL
:
1935 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1936 case LTTNG_CONSUMER32_UST
:
1937 case LTTNG_CONSUMER64_UST
:
1938 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1940 ERR("Unknown consumer_data type");
1946 void lttng_consumer_close_all_metadata(void)
1948 switch (consumer_data
.type
) {
1949 case LTTNG_CONSUMER_KERNEL
:
1951 * The Kernel consumer has a different metadata scheme so we don't
1952 * close anything because the stream will be closed by the session
1956 case LTTNG_CONSUMER32_UST
:
1957 case LTTNG_CONSUMER64_UST
:
1959 * Close all metadata streams. The metadata hash table is passed and
1960 * this call iterates over it by closing all wakeup fd. This is safe
1961 * because at this point we are sure that the metadata producer is
1962 * either dead or blocked.
1964 lttng_ustconsumer_close_all_metadata(metadata_ht
);
1967 ERR("Unknown consumer_data type");
1973 * Clean up a metadata stream and free its memory.
1975 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
1976 struct lttng_ht
*ht
)
1978 struct lttng_consumer_channel
*free_chan
= NULL
;
1982 * This call should NEVER receive regular stream. It must always be
1983 * metadata stream and this is crucial for data structure synchronization.
1985 assert(stream
->metadata_flag
);
1987 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
1989 pthread_mutex_lock(&consumer_data
.lock
);
1990 pthread_mutex_lock(&stream
->chan
->lock
);
1991 pthread_mutex_lock(&stream
->lock
);
1993 /* Remove any reference to that stream. */
1994 consumer_stream_delete(stream
, ht
);
1996 /* Close down everything including the relayd if one. */
1997 consumer_stream_close(stream
);
1998 /* Destroy tracer buffers of the stream. */
1999 consumer_stream_destroy_buffers(stream
);
2001 /* Atomically decrement channel refcount since other threads can use it. */
2002 if (!uatomic_sub_return(&stream
->chan
->refcount
, 1)
2003 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
2004 /* Go for channel deletion! */
2005 free_chan
= stream
->chan
;
2009 * Nullify the stream reference so it is not used after deletion. The
2010 * channel lock MUST be acquired before being able to check for a NULL
2013 stream
->chan
->metadata_stream
= NULL
;
2015 pthread_mutex_unlock(&stream
->lock
);
2016 pthread_mutex_unlock(&stream
->chan
->lock
);
2017 pthread_mutex_unlock(&consumer_data
.lock
);
2020 consumer_del_channel(free_chan
);
2023 consumer_stream_free(stream
);
2027 * Action done with the metadata stream when adding it to the consumer internal
2028 * data structures to handle it.
2030 int consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
2032 struct lttng_ht
*ht
= metadata_ht
;
2034 struct lttng_ht_iter iter
;
2035 struct lttng_ht_node_u64
*node
;
2040 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2042 pthread_mutex_lock(&consumer_data
.lock
);
2043 pthread_mutex_lock(&stream
->chan
->lock
);
2044 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2045 pthread_mutex_lock(&stream
->lock
);
2048 * From here, refcounts are updated so be _careful_ when returning an error
2055 * Lookup the stream just to make sure it does not exist in our internal
2056 * state. This should NEVER happen.
2058 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2059 node
= lttng_ht_iter_get_node_u64(&iter
);
2063 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2064 * in terms of destroying the associated channel, because the action that
2065 * causes the count to become 0 also causes a stream to be added. The
2066 * channel deletion will thus be triggered by the following removal of this
2069 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2070 /* Increment refcount before decrementing nb_init_stream_left */
2072 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2075 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2077 lttng_ht_add_unique_u64(consumer_data
.stream_per_chan_id_ht
,
2078 &stream
->node_channel_id
);
2081 * Add stream to the stream_list_ht of the consumer data. No need to steal
2082 * the key since the HT does not use it and we allow to add redundant keys
2085 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2089 pthread_mutex_unlock(&stream
->lock
);
2090 pthread_mutex_unlock(&stream
->chan
->lock
);
2091 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2092 pthread_mutex_unlock(&consumer_data
.lock
);
2097 * Delete data stream that are flagged for deletion (endpoint_status).
2099 static void validate_endpoint_status_data_stream(void)
2101 struct lttng_ht_iter iter
;
2102 struct lttng_consumer_stream
*stream
;
2104 DBG("Consumer delete flagged data stream");
2107 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2108 /* Validate delete flag of the stream */
2109 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2112 /* Delete it right now */
2113 consumer_del_stream(stream
, data_ht
);
2119 * Delete metadata stream that are flagged for deletion (endpoint_status).
2121 static void validate_endpoint_status_metadata_stream(
2122 struct lttng_poll_event
*pollset
)
2124 struct lttng_ht_iter iter
;
2125 struct lttng_consumer_stream
*stream
;
2127 DBG("Consumer delete flagged metadata stream");
2132 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2133 /* Validate delete flag of the stream */
2134 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2138 * Remove from pollset so the metadata thread can continue without
2139 * blocking on a deleted stream.
2141 lttng_poll_del(pollset
, stream
->wait_fd
);
2143 /* Delete it right now */
2144 consumer_del_metadata_stream(stream
, metadata_ht
);
2150 * Thread polls on metadata file descriptor and write them on disk or on the
2153 void *consumer_thread_metadata_poll(void *data
)
2155 int ret
, i
, pollfd
, err
= -1;
2156 uint32_t revents
, nb_fd
;
2157 struct lttng_consumer_stream
*stream
= NULL
;
2158 struct lttng_ht_iter iter
;
2159 struct lttng_ht_node_u64
*node
;
2160 struct lttng_poll_event events
;
2161 struct lttng_consumer_local_data
*ctx
= data
;
2164 rcu_register_thread();
2166 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2168 if (testpoint(consumerd_thread_metadata
)) {
2169 goto error_testpoint
;
2172 health_code_update();
2174 DBG("Thread metadata poll started");
2176 /* Size is set to 1 for the consumer_metadata pipe */
2177 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2179 ERR("Poll set creation failed");
2183 ret
= lttng_poll_add(&events
,
2184 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2190 DBG("Metadata main loop started");
2193 health_code_update();
2195 /* Only the metadata pipe is set */
2196 if (LTTNG_POLL_GETNB(&events
) == 0 && consumer_quit
== 1) {
2197 err
= 0; /* All is OK */
2202 DBG("Metadata poll wait with %d fd(s)", LTTNG_POLL_GETNB(&events
));
2203 health_poll_entry();
2204 ret
= lttng_poll_wait(&events
, -1);
2206 DBG("Metadata event catched in thread");
2208 if (errno
== EINTR
) {
2209 ERR("Poll EINTR catched");
2217 /* From here, the event is a metadata wait fd */
2218 for (i
= 0; i
< nb_fd
; i
++) {
2219 health_code_update();
2221 revents
= LTTNG_POLL_GETEV(&events
, i
);
2222 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2224 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2225 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2226 DBG("Metadata thread pipe hung up");
2228 * Remove the pipe from the poll set and continue the loop
2229 * since their might be data to consume.
2231 lttng_poll_del(&events
,
2232 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2233 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2235 } else if (revents
& LPOLLIN
) {
2238 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2239 &stream
, sizeof(stream
));
2240 if (pipe_len
< sizeof(stream
)) {
2241 PERROR("read metadata stream");
2243 * Continue here to handle the rest of the streams.
2248 /* A NULL stream means that the state has changed. */
2249 if (stream
== NULL
) {
2250 /* Check for deleted streams. */
2251 validate_endpoint_status_metadata_stream(&events
);
2255 DBG("Adding metadata stream %d to poll set",
2258 /* Add metadata stream to the global poll events list */
2259 lttng_poll_add(&events
, stream
->wait_fd
,
2260 LPOLLIN
| LPOLLPRI
| LPOLLHUP
);
2263 /* Handle other stream */
2269 uint64_t tmp_id
= (uint64_t) pollfd
;
2271 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2273 node
= lttng_ht_iter_get_node_u64(&iter
);
2276 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2279 /* Check for error event */
2280 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2281 DBG("Metadata fd %d is hup|err.", pollfd
);
2282 if (!stream
->hangup_flush_done
2283 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2284 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2285 DBG("Attempting to flush and consume the UST buffers");
2286 lttng_ustconsumer_on_stream_hangup(stream
);
2288 /* We just flushed the stream now read it. */
2290 health_code_update();
2292 len
= ctx
->on_buffer_ready(stream
, ctx
);
2294 * We don't check the return value here since if we get
2295 * a negative len, it means an error occured thus we
2296 * simply remove it from the poll set and free the
2302 lttng_poll_del(&events
, stream
->wait_fd
);
2304 * This call update the channel states, closes file descriptors
2305 * and securely free the stream.
2307 consumer_del_metadata_stream(stream
, metadata_ht
);
2308 } else if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2309 /* Get the data out of the metadata file descriptor */
2310 DBG("Metadata available on fd %d", pollfd
);
2311 assert(stream
->wait_fd
== pollfd
);
2314 health_code_update();
2316 len
= ctx
->on_buffer_ready(stream
, ctx
);
2318 * We don't check the return value here since if we get
2319 * a negative len, it means an error occured thus we
2320 * simply remove it from the poll set and free the
2325 /* It's ok to have an unavailable sub-buffer */
2326 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2327 /* Clean up stream from consumer and free it. */
2328 lttng_poll_del(&events
, stream
->wait_fd
);
2329 consumer_del_metadata_stream(stream
, metadata_ht
);
2333 /* Release RCU lock for the stream looked up */
2342 DBG("Metadata poll thread exiting");
2344 lttng_poll_clean(&events
);
2349 ERR("Health error occurred in %s", __func__
);
2351 health_unregister(health_consumerd
);
2352 rcu_unregister_thread();
2357 * This thread polls the fds in the set to consume the data and write
2358 * it to tracefile if necessary.
2360 void *consumer_thread_data_poll(void *data
)
2362 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2363 struct pollfd
*pollfd
= NULL
;
2364 /* local view of the streams */
2365 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2366 /* local view of consumer_data.fds_count */
2368 struct lttng_consumer_local_data
*ctx
= data
;
2371 rcu_register_thread();
2373 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2375 if (testpoint(consumerd_thread_data
)) {
2376 goto error_testpoint
;
2379 health_code_update();
2381 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
*));
2382 if (local_stream
== NULL
) {
2383 PERROR("local_stream malloc");
2388 health_code_update();
2394 * the fds set has been updated, we need to update our
2395 * local array as well
2397 pthread_mutex_lock(&consumer_data
.lock
);
2398 if (consumer_data
.need_update
) {
2403 local_stream
= NULL
;
2405 /* allocate for all fds + 1 for the consumer_data_pipe */
2406 pollfd
= zmalloc((consumer_data
.stream_count
+ 1) * sizeof(struct pollfd
));
2407 if (pollfd
== NULL
) {
2408 PERROR("pollfd malloc");
2409 pthread_mutex_unlock(&consumer_data
.lock
);
2413 /* allocate for all fds + 1 for the consumer_data_pipe */
2414 local_stream
= zmalloc((consumer_data
.stream_count
+ 1) *
2415 sizeof(struct lttng_consumer_stream
*));
2416 if (local_stream
== NULL
) {
2417 PERROR("local_stream malloc");
2418 pthread_mutex_unlock(&consumer_data
.lock
);
2421 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2424 ERR("Error in allocating pollfd or local_outfds");
2425 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2426 pthread_mutex_unlock(&consumer_data
.lock
);
2430 consumer_data
.need_update
= 0;
2432 pthread_mutex_unlock(&consumer_data
.lock
);
2434 /* No FDs and consumer_quit, consumer_cleanup the thread */
2435 if (nb_fd
== 0 && consumer_quit
== 1) {
2436 err
= 0; /* All is OK */
2439 /* poll on the array of fds */
2441 DBG("polling on %d fd", nb_fd
+ 1);
2442 health_poll_entry();
2443 num_rdy
= poll(pollfd
, nb_fd
+ 1, -1);
2445 DBG("poll num_rdy : %d", num_rdy
);
2446 if (num_rdy
== -1) {
2448 * Restart interrupted system call.
2450 if (errno
== EINTR
) {
2453 PERROR("Poll error");
2454 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2456 } else if (num_rdy
== 0) {
2457 DBG("Polling thread timed out");
2462 * If the consumer_data_pipe triggered poll go directly to the
2463 * beginning of the loop to update the array. We want to prioritize
2464 * array update over low-priority reads.
2466 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2467 ssize_t pipe_readlen
;
2469 DBG("consumer_data_pipe wake up");
2470 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2471 &new_stream
, sizeof(new_stream
));
2472 if (pipe_readlen
< sizeof(new_stream
)) {
2473 PERROR("Consumer data pipe");
2474 /* Continue so we can at least handle the current stream(s). */
2479 * If the stream is NULL, just ignore it. It's also possible that
2480 * the sessiond poll thread changed the consumer_quit state and is
2481 * waking us up to test it.
2483 if (new_stream
== NULL
) {
2484 validate_endpoint_status_data_stream();
2488 /* Continue to update the local streams and handle prio ones */
2492 /* Take care of high priority channels first. */
2493 for (i
= 0; i
< nb_fd
; i
++) {
2494 health_code_update();
2496 if (local_stream
[i
] == NULL
) {
2499 if (pollfd
[i
].revents
& POLLPRI
) {
2500 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2502 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2503 /* it's ok to have an unavailable sub-buffer */
2504 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2505 /* Clean the stream and free it. */
2506 consumer_del_stream(local_stream
[i
], data_ht
);
2507 local_stream
[i
] = NULL
;
2508 } else if (len
> 0) {
2509 local_stream
[i
]->data_read
= 1;
2515 * If we read high prio channel in this loop, try again
2516 * for more high prio data.
2522 /* Take care of low priority channels. */
2523 for (i
= 0; i
< nb_fd
; i
++) {
2524 health_code_update();
2526 if (local_stream
[i
] == NULL
) {
2529 if ((pollfd
[i
].revents
& POLLIN
) ||
2530 local_stream
[i
]->hangup_flush_done
) {
2531 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2532 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2533 /* it's ok to have an unavailable sub-buffer */
2534 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2535 /* Clean the stream and free it. */
2536 consumer_del_stream(local_stream
[i
], data_ht
);
2537 local_stream
[i
] = NULL
;
2538 } else if (len
> 0) {
2539 local_stream
[i
]->data_read
= 1;
2544 /* Handle hangup and errors */
2545 for (i
= 0; i
< nb_fd
; i
++) {
2546 health_code_update();
2548 if (local_stream
[i
] == NULL
) {
2551 if (!local_stream
[i
]->hangup_flush_done
2552 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2553 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2554 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2555 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2557 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2558 /* Attempt read again, for the data we just flushed. */
2559 local_stream
[i
]->data_read
= 1;
2562 * If the poll flag is HUP/ERR/NVAL and we have
2563 * read no data in this pass, we can remove the
2564 * stream from its hash table.
2566 if ((pollfd
[i
].revents
& POLLHUP
)) {
2567 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2568 if (!local_stream
[i
]->data_read
) {
2569 consumer_del_stream(local_stream
[i
], data_ht
);
2570 local_stream
[i
] = NULL
;
2573 } else if (pollfd
[i
].revents
& POLLERR
) {
2574 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2575 if (!local_stream
[i
]->data_read
) {
2576 consumer_del_stream(local_stream
[i
], data_ht
);
2577 local_stream
[i
] = NULL
;
2580 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2581 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2582 if (!local_stream
[i
]->data_read
) {
2583 consumer_del_stream(local_stream
[i
], data_ht
);
2584 local_stream
[i
] = NULL
;
2588 if (local_stream
[i
] != NULL
) {
2589 local_stream
[i
]->data_read
= 0;
2596 DBG("polling thread exiting");
2601 * Close the write side of the pipe so epoll_wait() in
2602 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2603 * read side of the pipe. If we close them both, epoll_wait strangely does
2604 * not return and could create a endless wait period if the pipe is the
2605 * only tracked fd in the poll set. The thread will take care of closing
2608 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2613 ERR("Health error occurred in %s", __func__
);
2615 health_unregister(health_consumerd
);
2617 rcu_unregister_thread();
2622 * Close wake-up end of each stream belonging to the channel. This will
2623 * allow the poll() on the stream read-side to detect when the
2624 * write-side (application) finally closes them.
2627 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2629 struct lttng_ht
*ht
;
2630 struct lttng_consumer_stream
*stream
;
2631 struct lttng_ht_iter iter
;
2633 ht
= consumer_data
.stream_per_chan_id_ht
;
2636 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2637 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2638 ht
->match_fct
, &channel
->key
,
2639 &iter
.iter
, stream
, node_channel_id
.node
) {
2641 * Protect against teardown with mutex.
2643 pthread_mutex_lock(&stream
->lock
);
2644 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2647 switch (consumer_data
.type
) {
2648 case LTTNG_CONSUMER_KERNEL
:
2650 case LTTNG_CONSUMER32_UST
:
2651 case LTTNG_CONSUMER64_UST
:
2652 if (stream
->metadata_flag
) {
2653 /* Safe and protected by the stream lock. */
2654 lttng_ustconsumer_close_metadata(stream
->chan
);
2657 * Note: a mutex is taken internally within
2658 * liblttng-ust-ctl to protect timer wakeup_fd
2659 * use from concurrent close.
2661 lttng_ustconsumer_close_stream_wakeup(stream
);
2665 ERR("Unknown consumer_data type");
2669 pthread_mutex_unlock(&stream
->lock
);
2674 static void destroy_channel_ht(struct lttng_ht
*ht
)
2676 struct lttng_ht_iter iter
;
2677 struct lttng_consumer_channel
*channel
;
2685 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2686 ret
= lttng_ht_del(ht
, &iter
);
2691 lttng_ht_destroy(ht
);
2695 * This thread polls the channel fds to detect when they are being
2696 * closed. It closes all related streams if the channel is detected as
2697 * closed. It is currently only used as a shim layer for UST because the
2698 * consumerd needs to keep the per-stream wakeup end of pipes open for
2701 void *consumer_thread_channel_poll(void *data
)
2703 int ret
, i
, pollfd
, err
= -1;
2704 uint32_t revents
, nb_fd
;
2705 struct lttng_consumer_channel
*chan
= NULL
;
2706 struct lttng_ht_iter iter
;
2707 struct lttng_ht_node_u64
*node
;
2708 struct lttng_poll_event events
;
2709 struct lttng_consumer_local_data
*ctx
= data
;
2710 struct lttng_ht
*channel_ht
;
2712 rcu_register_thread();
2714 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
2716 if (testpoint(consumerd_thread_channel
)) {
2717 goto error_testpoint
;
2720 health_code_update();
2722 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2724 /* ENOMEM at this point. Better to bail out. */
2728 DBG("Thread channel poll started");
2730 /* Size is set to 1 for the consumer_channel pipe */
2731 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2733 ERR("Poll set creation failed");
2737 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2743 DBG("Channel main loop started");
2746 health_code_update();
2748 /* Only the channel pipe is set */
2749 if (LTTNG_POLL_GETNB(&events
) == 0 && consumer_quit
== 1) {
2750 err
= 0; /* All is OK */
2755 DBG("Channel poll wait with %d fd(s)", LTTNG_POLL_GETNB(&events
));
2756 health_poll_entry();
2757 ret
= lttng_poll_wait(&events
, -1);
2759 DBG("Channel event catched in thread");
2761 if (errno
== EINTR
) {
2762 ERR("Poll EINTR catched");
2770 /* From here, the event is a channel wait fd */
2771 for (i
= 0; i
< nb_fd
; i
++) {
2772 health_code_update();
2774 revents
= LTTNG_POLL_GETEV(&events
, i
);
2775 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2777 /* Just don't waste time if no returned events for the fd */
2781 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2782 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2783 DBG("Channel thread pipe hung up");
2785 * Remove the pipe from the poll set and continue the loop
2786 * since their might be data to consume.
2788 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2790 } else if (revents
& LPOLLIN
) {
2791 enum consumer_channel_action action
;
2794 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
2796 ERR("Error reading channel pipe");
2801 case CONSUMER_CHANNEL_ADD
:
2802 DBG("Adding channel %d to poll set",
2805 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
2808 lttng_ht_add_unique_u64(channel_ht
,
2809 &chan
->wait_fd_node
);
2811 /* Add channel to the global poll events list */
2812 lttng_poll_add(&events
, chan
->wait_fd
,
2813 LPOLLIN
| LPOLLPRI
);
2815 case CONSUMER_CHANNEL_DEL
:
2818 * This command should never be called if the channel
2819 * has streams monitored by either the data or metadata
2820 * thread. The consumer only notify this thread with a
2821 * channel del. command if it receives a destroy
2822 * channel command from the session daemon that send it
2823 * if a command prior to the GET_CHANNEL failed.
2827 chan
= consumer_find_channel(key
);
2830 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
2833 lttng_poll_del(&events
, chan
->wait_fd
);
2834 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
2835 ret
= lttng_ht_del(channel_ht
, &iter
);
2838 switch (consumer_data
.type
) {
2839 case LTTNG_CONSUMER_KERNEL
:
2841 case LTTNG_CONSUMER32_UST
:
2842 case LTTNG_CONSUMER64_UST
:
2843 health_code_update();
2844 /* Destroy streams that might have been left in the stream list. */
2845 clean_channel_stream_list(chan
);
2848 ERR("Unknown consumer_data type");
2853 * Release our own refcount. Force channel deletion even if
2854 * streams were not initialized.
2856 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
2857 consumer_del_channel(chan
);
2862 case CONSUMER_CHANNEL_QUIT
:
2864 * Remove the pipe from the poll set and continue the loop
2865 * since their might be data to consume.
2867 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2870 ERR("Unknown action");
2875 /* Handle other stream */
2881 uint64_t tmp_id
= (uint64_t) pollfd
;
2883 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
2885 node
= lttng_ht_iter_get_node_u64(&iter
);
2888 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
2891 /* Check for error event */
2892 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2893 DBG("Channel fd %d is hup|err.", pollfd
);
2895 lttng_poll_del(&events
, chan
->wait_fd
);
2896 ret
= lttng_ht_del(channel_ht
, &iter
);
2900 * This will close the wait fd for each stream associated to
2901 * this channel AND monitored by the data/metadata thread thus
2902 * will be clean by the right thread.
2904 consumer_close_channel_streams(chan
);
2906 /* Release our own refcount */
2907 if (!uatomic_sub_return(&chan
->refcount
, 1)
2908 && !uatomic_read(&chan
->nb_init_stream_left
)) {
2909 consumer_del_channel(chan
);
2913 /* Release RCU lock for the channel looked up */
2921 lttng_poll_clean(&events
);
2923 destroy_channel_ht(channel_ht
);
2926 DBG("Channel poll thread exiting");
2929 ERR("Health error occurred in %s", __func__
);
2931 health_unregister(health_consumerd
);
2932 rcu_unregister_thread();
2936 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
2937 struct pollfd
*sockpoll
, int client_socket
)
2944 ret
= lttng_consumer_poll_socket(sockpoll
);
2948 DBG("Metadata connection on client_socket");
2950 /* Blocking call, waiting for transmission */
2951 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
2952 if (ctx
->consumer_metadata_socket
< 0) {
2953 WARN("On accept metadata");
2964 * This thread listens on the consumerd socket and receives the file
2965 * descriptors from the session daemon.
2967 void *consumer_thread_sessiond_poll(void *data
)
2969 int sock
= -1, client_socket
, ret
, err
= -1;
2971 * structure to poll for incoming data on communication socket avoids
2972 * making blocking sockets.
2974 struct pollfd consumer_sockpoll
[2];
2975 struct lttng_consumer_local_data
*ctx
= data
;
2977 rcu_register_thread();
2979 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
2981 if (testpoint(consumerd_thread_sessiond
)) {
2982 goto error_testpoint
;
2985 health_code_update();
2987 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
2988 unlink(ctx
->consumer_command_sock_path
);
2989 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
2990 if (client_socket
< 0) {
2991 ERR("Cannot create command socket");
2995 ret
= lttcomm_listen_unix_sock(client_socket
);
3000 DBG("Sending ready command to lttng-sessiond");
3001 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3002 /* return < 0 on error, but == 0 is not fatal */
3004 ERR("Error sending ready command to lttng-sessiond");
3008 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3009 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3010 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3011 consumer_sockpoll
[1].fd
= client_socket
;
3012 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3014 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3022 DBG("Connection on client_socket");
3024 /* Blocking call, waiting for transmission */
3025 sock
= lttcomm_accept_unix_sock(client_socket
);
3032 * Setup metadata socket which is the second socket connection on the
3033 * command unix socket.
3035 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3044 /* This socket is not useful anymore. */
3045 ret
= close(client_socket
);
3047 PERROR("close client_socket");
3051 /* update the polling structure to poll on the established socket */
3052 consumer_sockpoll
[1].fd
= sock
;
3053 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3056 health_code_update();
3058 health_poll_entry();
3059 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3068 DBG("Incoming command on sock");
3069 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3072 * This could simply be a session daemon quitting. Don't output
3075 DBG("Communication interrupted on command socket");
3079 if (consumer_quit
) {
3080 DBG("consumer_thread_receive_fds received quit from signal");
3081 err
= 0; /* All is OK */
3084 DBG("received command on sock");
3090 DBG("Consumer thread sessiond poll exiting");
3093 * Close metadata streams since the producer is the session daemon which
3096 * NOTE: for now, this only applies to the UST tracer.
3098 lttng_consumer_close_all_metadata();
3101 * when all fds have hung up, the polling thread
3107 * Notify the data poll thread to poll back again and test the
3108 * consumer_quit state that we just set so to quit gracefully.
3110 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3112 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3114 notify_health_quit_pipe(health_quit_pipe
);
3116 /* Cleaning up possibly open sockets. */
3120 PERROR("close sock sessiond poll");
3123 if (client_socket
>= 0) {
3124 ret
= close(client_socket
);
3126 PERROR("close client_socket sessiond poll");
3133 ERR("Health error occurred in %s", __func__
);
3135 health_unregister(health_consumerd
);
3137 rcu_unregister_thread();
3141 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3142 struct lttng_consumer_local_data
*ctx
)
3146 pthread_mutex_lock(&stream
->lock
);
3147 if (stream
->metadata_flag
) {
3148 pthread_mutex_lock(&stream
->metadata_rdv_lock
);
3151 switch (consumer_data
.type
) {
3152 case LTTNG_CONSUMER_KERNEL
:
3153 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
3155 case LTTNG_CONSUMER32_UST
:
3156 case LTTNG_CONSUMER64_UST
:
3157 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
3160 ERR("Unknown consumer_data type");
3166 if (stream
->metadata_flag
) {
3167 pthread_cond_broadcast(&stream
->metadata_rdv
);
3168 pthread_mutex_unlock(&stream
->metadata_rdv_lock
);
3170 pthread_mutex_unlock(&stream
->lock
);
3174 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3176 switch (consumer_data
.type
) {
3177 case LTTNG_CONSUMER_KERNEL
:
3178 return lttng_kconsumer_on_recv_stream(stream
);
3179 case LTTNG_CONSUMER32_UST
:
3180 case LTTNG_CONSUMER64_UST
:
3181 return lttng_ustconsumer_on_recv_stream(stream
);
3183 ERR("Unknown consumer_data type");
3190 * Allocate and set consumer data hash tables.
3192 int lttng_consumer_init(void)
3194 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3195 if (!consumer_data
.channel_ht
) {
3199 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3200 if (!consumer_data
.relayd_ht
) {
3204 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3205 if (!consumer_data
.stream_list_ht
) {
3209 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3210 if (!consumer_data
.stream_per_chan_id_ht
) {
3214 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3219 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3231 * Process the ADD_RELAYD command receive by a consumer.
3233 * This will create a relayd socket pair and add it to the relayd hash table.
3234 * The caller MUST acquire a RCU read side lock before calling it.
3236 int consumer_add_relayd_socket(uint64_t net_seq_idx
, int sock_type
,
3237 struct lttng_consumer_local_data
*ctx
, int sock
,
3238 struct pollfd
*consumer_sockpoll
,
3239 struct lttcomm_relayd_sock
*relayd_sock
, uint64_t sessiond_id
,
3240 uint64_t relayd_session_id
)
3242 int fd
= -1, ret
= -1, relayd_created
= 0;
3243 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3244 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3247 assert(relayd_sock
);
3249 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", net_seq_idx
);
3251 /* Get relayd reference if exists. */
3252 relayd
= consumer_find_relayd(net_seq_idx
);
3253 if (relayd
== NULL
) {
3254 assert(sock_type
== LTTNG_STREAM_CONTROL
);
3255 /* Not found. Allocate one. */
3256 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3257 if (relayd
== NULL
) {
3259 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3262 relayd
->sessiond_session_id
= sessiond_id
;
3267 * This code path MUST continue to the consumer send status message to
3268 * we can notify the session daemon and continue our work without
3269 * killing everything.
3273 * relayd key should never be found for control socket.
3275 assert(sock_type
!= LTTNG_STREAM_CONTROL
);
3278 /* First send a status message before receiving the fds. */
3279 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3281 /* Somehow, the session daemon is not responding anymore. */
3282 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3283 goto error_nosignal
;
3286 /* Poll on consumer socket. */
3287 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3289 /* Needing to exit in the middle of a command: error. */
3290 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3292 goto error_nosignal
;
3295 /* Get relayd socket from session daemon */
3296 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3297 if (ret
!= sizeof(fd
)) {
3299 fd
= -1; /* Just in case it gets set with an invalid value. */
3302 * Failing to receive FDs might indicate a major problem such as
3303 * reaching a fd limit during the receive where the kernel returns a
3304 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3305 * don't take any chances and stop everything.
3307 * XXX: Feature request #558 will fix that and avoid this possible
3308 * issue when reaching the fd limit.
3310 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3311 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3315 /* Copy socket information and received FD */
3316 switch (sock_type
) {
3317 case LTTNG_STREAM_CONTROL
:
3318 /* Copy received lttcomm socket */
3319 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3320 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3321 /* Handle create_sock error. */
3323 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3327 * Close the socket created internally by
3328 * lttcomm_create_sock, so we can replace it by the one
3329 * received from sessiond.
3331 if (close(relayd
->control_sock
.sock
.fd
)) {
3335 /* Assign new file descriptor */
3336 relayd
->control_sock
.sock
.fd
= fd
;
3337 fd
= -1; /* For error path */
3338 /* Assign version values. */
3339 relayd
->control_sock
.major
= relayd_sock
->major
;
3340 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3342 relayd
->relayd_session_id
= relayd_session_id
;
3345 case LTTNG_STREAM_DATA
:
3346 /* Copy received lttcomm socket */
3347 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3348 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3349 /* Handle create_sock error. */
3351 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3355 * Close the socket created internally by
3356 * lttcomm_create_sock, so we can replace it by the one
3357 * received from sessiond.
3359 if (close(relayd
->data_sock
.sock
.fd
)) {
3363 /* Assign new file descriptor */
3364 relayd
->data_sock
.sock
.fd
= fd
;
3365 fd
= -1; /* for eventual error paths */
3366 /* Assign version values. */
3367 relayd
->data_sock
.major
= relayd_sock
->major
;
3368 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3371 ERR("Unknown relayd socket type (%d)", sock_type
);
3373 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3377 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3378 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3379 relayd
->net_seq_idx
, fd
);
3381 /* We successfully added the socket. Send status back. */
3382 ret
= consumer_send_status_msg(sock
, ret_code
);
3384 /* Somehow, the session daemon is not responding anymore. */
3385 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3386 goto error_nosignal
;
3390 * Add relayd socket pair to consumer data hashtable. If object already
3391 * exists or on error, the function gracefully returns.
3399 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3400 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3404 /* Close received socket if valid. */
3407 PERROR("close received socket");
3411 if (relayd_created
) {
3419 * Try to lock the stream mutex.
3421 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
3423 static int stream_try_lock(struct lttng_consumer_stream
*stream
)
3430 * Try to lock the stream mutex. On failure, we know that the stream is
3431 * being used else where hence there is data still being extracted.
3433 ret
= pthread_mutex_trylock(&stream
->lock
);
3435 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
3447 * Search for a relayd associated to the session id and return the reference.
3449 * A rcu read side lock MUST be acquire before calling this function and locked
3450 * until the relayd object is no longer necessary.
3452 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3454 struct lttng_ht_iter iter
;
3455 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3457 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3458 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3461 * Check by sessiond id which is unique here where the relayd session
3462 * id might not be when having multiple relayd.
3464 if (relayd
->sessiond_session_id
== id
) {
3465 /* Found the relayd. There can be only one per id. */
3477 * Check if for a given session id there is still data needed to be extract
3480 * Return 1 if data is pending or else 0 meaning ready to be read.
3482 int consumer_data_pending(uint64_t id
)
3485 struct lttng_ht_iter iter
;
3486 struct lttng_ht
*ht
;
3487 struct lttng_consumer_stream
*stream
;
3488 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3489 int (*data_pending
)(struct lttng_consumer_stream
*);
3491 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3494 pthread_mutex_lock(&consumer_data
.lock
);
3496 switch (consumer_data
.type
) {
3497 case LTTNG_CONSUMER_KERNEL
:
3498 data_pending
= lttng_kconsumer_data_pending
;
3500 case LTTNG_CONSUMER32_UST
:
3501 case LTTNG_CONSUMER64_UST
:
3502 data_pending
= lttng_ustconsumer_data_pending
;
3505 ERR("Unknown consumer data type");
3509 /* Ease our life a bit */
3510 ht
= consumer_data
.stream_list_ht
;
3512 relayd
= find_relayd_by_session_id(id
);
3514 /* Send init command for data pending. */
3515 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3516 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3517 relayd
->relayd_session_id
);
3518 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3520 /* Communication error thus the relayd so no data pending. */
3521 goto data_not_pending
;
3525 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3526 ht
->hash_fct(&id
, lttng_ht_seed
),
3528 &iter
.iter
, stream
, node_session_id
.node
) {
3529 /* If this call fails, the stream is being used hence data pending. */
3530 ret
= stream_try_lock(stream
);
3536 * A removed node from the hash table indicates that the stream has
3537 * been deleted thus having a guarantee that the buffers are closed
3538 * on the consumer side. However, data can still be transmitted
3539 * over the network so don't skip the relayd check.
3541 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3544 * An empty output file is not valid. We need at least one packet
3545 * generated per stream, even if it contains no event, so it
3546 * contains at least one packet header.
3548 if (stream
->output_written
== 0) {
3549 pthread_mutex_unlock(&stream
->lock
);
3552 /* Check the stream if there is data in the buffers. */
3553 ret
= data_pending(stream
);
3555 pthread_mutex_unlock(&stream
->lock
);
3562 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3563 if (stream
->metadata_flag
) {
3564 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3565 stream
->relayd_stream_id
);
3567 ret
= relayd_data_pending(&relayd
->control_sock
,
3568 stream
->relayd_stream_id
,
3569 stream
->next_net_seq_num
- 1);
3571 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3573 pthread_mutex_unlock(&stream
->lock
);
3577 pthread_mutex_unlock(&stream
->lock
);
3581 unsigned int is_data_inflight
= 0;
3583 /* Send init command for data pending. */
3584 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3585 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3586 relayd
->relayd_session_id
, &is_data_inflight
);
3587 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3589 goto data_not_pending
;
3591 if (is_data_inflight
) {
3597 * Finding _no_ node in the hash table and no inflight data means that the
3598 * stream(s) have been removed thus data is guaranteed to be available for
3599 * analysis from the trace files.
3603 /* Data is available to be read by a viewer. */
3604 pthread_mutex_unlock(&consumer_data
.lock
);
3609 /* Data is still being extracted from buffers. */
3610 pthread_mutex_unlock(&consumer_data
.lock
);
3616 * Send a ret code status message to the sessiond daemon.
3618 * Return the sendmsg() return value.
3620 int consumer_send_status_msg(int sock
, int ret_code
)
3622 struct lttcomm_consumer_status_msg msg
;
3624 memset(&msg
, 0, sizeof(msg
));
3625 msg
.ret_code
= ret_code
;
3627 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3631 * Send a channel status message to the sessiond daemon.
3633 * Return the sendmsg() return value.
3635 int consumer_send_status_channel(int sock
,
3636 struct lttng_consumer_channel
*channel
)
3638 struct lttcomm_consumer_status_channel msg
;
3642 memset(&msg
, 0, sizeof(msg
));
3644 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3646 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3647 msg
.key
= channel
->key
;
3648 msg
.stream_count
= channel
->streams
.count
;
3651 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3655 * Using a maximum stream size with the produced and consumed position of a
3656 * stream, computes the new consumed position to be as close as possible to the
3657 * maximum possible stream size.
3659 * If maximum stream size is lower than the possible buffer size (produced -
3660 * consumed), the consumed_pos given is returned untouched else the new value
3663 unsigned long consumer_get_consumed_maxsize(unsigned long consumed_pos
,
3664 unsigned long produced_pos
, uint64_t max_stream_size
)
3666 if (max_stream_size
&& max_stream_size
< (produced_pos
- consumed_pos
)) {
3667 /* Offset from the produced position to get the latest buffers. */
3668 return produced_pos
- max_stream_size
;
3671 return consumed_pos
;