2 * Copyright (C) 2011 - Julien Desfossez <julien.desfossez@polymtl.ca>
3 * Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
4 * 2012 - David Goulet <dgoulet@efficios.com>
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License, version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
15 * You should have received a copy of the GNU General Public License along
16 * with this program; if not, write to the Free Software Foundation, Inc.,
17 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
27 #include <sys/socket.h>
28 #include <sys/types.h>
33 #include <bin/lttng-consumerd/health-consumerd.h>
34 #include <common/common.h>
35 #include <common/utils.h>
36 #include <common/compat/poll.h>
37 #include <common/compat/endian.h>
38 #include <common/index/index.h>
39 #include <common/kernel-ctl/kernel-ctl.h>
40 #include <common/sessiond-comm/relayd.h>
41 #include <common/sessiond-comm/sessiond-comm.h>
42 #include <common/kernel-consumer/kernel-consumer.h>
43 #include <common/relayd/relayd.h>
44 #include <common/ust-consumer/ust-consumer.h>
45 #include <common/consumer/consumer-timer.h>
46 #include <common/consumer/consumer.h>
47 #include <common/consumer/consumer-stream.h>
48 #include <common/consumer/consumer-testpoint.h>
49 #include <common/align.h>
51 struct lttng_consumer_global_data consumer_data
= {
54 .type
= LTTNG_CONSUMER_UNKNOWN
,
57 enum consumer_channel_action
{
60 CONSUMER_CHANNEL_QUIT
,
63 struct consumer_channel_msg
{
64 enum consumer_channel_action action
;
65 struct lttng_consumer_channel
*chan
; /* add */
66 uint64_t key
; /* del */
70 * Flag to inform the polling thread to quit when all fd hung up. Updated by
71 * the consumer_thread_receive_fds when it notices that all fds has hung up.
72 * Also updated by the signal handler (consumer_should_exit()). Read by the
75 volatile int consumer_quit
;
78 * Global hash table containing respectively metadata and data streams. The
79 * stream element in this ht should only be updated by the metadata poll thread
80 * for the metadata and the data poll thread for the data.
82 static struct lttng_ht
*metadata_ht
;
83 static struct lttng_ht
*data_ht
;
86 * Notify a thread lttng pipe to poll back again. This usually means that some
87 * global state has changed so we just send back the thread in a poll wait
90 static void notify_thread_lttng_pipe(struct lttng_pipe
*pipe
)
92 struct lttng_consumer_stream
*null_stream
= NULL
;
96 (void) lttng_pipe_write(pipe
, &null_stream
, sizeof(null_stream
));
99 static void notify_health_quit_pipe(int *pipe
)
103 ret
= lttng_write(pipe
[1], "4", 1);
105 PERROR("write consumer health quit");
109 static void notify_channel_pipe(struct lttng_consumer_local_data
*ctx
,
110 struct lttng_consumer_channel
*chan
,
112 enum consumer_channel_action action
)
114 struct consumer_channel_msg msg
;
117 memset(&msg
, 0, sizeof(msg
));
122 ret
= lttng_write(ctx
->consumer_channel_pipe
[1], &msg
, sizeof(msg
));
123 if (ret
< sizeof(msg
)) {
124 PERROR("notify_channel_pipe write error");
128 void notify_thread_del_channel(struct lttng_consumer_local_data
*ctx
,
131 notify_channel_pipe(ctx
, NULL
, key
, CONSUMER_CHANNEL_DEL
);
134 static int read_channel_pipe(struct lttng_consumer_local_data
*ctx
,
135 struct lttng_consumer_channel
**chan
,
137 enum consumer_channel_action
*action
)
139 struct consumer_channel_msg msg
;
142 ret
= lttng_read(ctx
->consumer_channel_pipe
[0], &msg
, sizeof(msg
));
143 if (ret
< sizeof(msg
)) {
147 *action
= msg
.action
;
155 * Cleanup the stream list of a channel. Those streams are not yet globally
158 static void clean_channel_stream_list(struct lttng_consumer_channel
*channel
)
160 struct lttng_consumer_stream
*stream
, *stmp
;
164 /* Delete streams that might have been left in the stream list. */
165 cds_list_for_each_entry_safe(stream
, stmp
, &channel
->streams
.head
,
167 cds_list_del(&stream
->send_node
);
169 * Once a stream is added to this list, the buffers were created so we
170 * have a guarantee that this call will succeed. Setting the monitor
171 * mode to 0 so we don't lock nor try to delete the stream from the
175 consumer_stream_destroy(stream
, NULL
);
180 * Find a stream. The consumer_data.lock must be locked during this
183 static struct lttng_consumer_stream
*find_stream(uint64_t key
,
186 struct lttng_ht_iter iter
;
187 struct lttng_ht_node_u64
*node
;
188 struct lttng_consumer_stream
*stream
= NULL
;
192 /* -1ULL keys are lookup failures */
193 if (key
== (uint64_t) -1ULL) {
199 lttng_ht_lookup(ht
, &key
, &iter
);
200 node
= lttng_ht_iter_get_node_u64(&iter
);
202 stream
= caa_container_of(node
, struct lttng_consumer_stream
, node
);
210 static void steal_stream_key(uint64_t key
, struct lttng_ht
*ht
)
212 struct lttng_consumer_stream
*stream
;
215 stream
= find_stream(key
, ht
);
217 stream
->key
= (uint64_t) -1ULL;
219 * We don't want the lookup to match, but we still need
220 * to iterate on this stream when iterating over the hash table. Just
221 * change the node key.
223 stream
->node
.key
= (uint64_t) -1ULL;
229 * Return a channel object for the given key.
231 * RCU read side lock MUST be acquired before calling this function and
232 * protects the channel ptr.
234 struct lttng_consumer_channel
*consumer_find_channel(uint64_t key
)
236 struct lttng_ht_iter iter
;
237 struct lttng_ht_node_u64
*node
;
238 struct lttng_consumer_channel
*channel
= NULL
;
240 /* -1ULL keys are lookup failures */
241 if (key
== (uint64_t) -1ULL) {
245 lttng_ht_lookup(consumer_data
.channel_ht
, &key
, &iter
);
246 node
= lttng_ht_iter_get_node_u64(&iter
);
248 channel
= caa_container_of(node
, struct lttng_consumer_channel
, node
);
255 * There is a possibility that the consumer does not have enough time between
256 * the close of the channel on the session daemon and the cleanup in here thus
257 * once we have a channel add with an existing key, we know for sure that this
258 * channel will eventually get cleaned up by all streams being closed.
260 * This function just nullifies the already existing channel key.
262 static void steal_channel_key(uint64_t key
)
264 struct lttng_consumer_channel
*channel
;
267 channel
= consumer_find_channel(key
);
269 channel
->key
= (uint64_t) -1ULL;
271 * We don't want the lookup to match, but we still need to iterate on
272 * this channel when iterating over the hash table. Just change the
275 channel
->node
.key
= (uint64_t) -1ULL;
280 static void free_channel_rcu(struct rcu_head
*head
)
282 struct lttng_ht_node_u64
*node
=
283 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
284 struct lttng_consumer_channel
*channel
=
285 caa_container_of(node
, struct lttng_consumer_channel
, node
);
287 switch (consumer_data
.type
) {
288 case LTTNG_CONSUMER_KERNEL
:
290 case LTTNG_CONSUMER32_UST
:
291 case LTTNG_CONSUMER64_UST
:
292 lttng_ustconsumer_free_channel(channel
);
295 ERR("Unknown consumer_data type");
302 * RCU protected relayd socket pair free.
304 static void free_relayd_rcu(struct rcu_head
*head
)
306 struct lttng_ht_node_u64
*node
=
307 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
308 struct consumer_relayd_sock_pair
*relayd
=
309 caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
312 * Close all sockets. This is done in the call RCU since we don't want the
313 * socket fds to be reassigned thus potentially creating bad state of the
316 * We do not have to lock the control socket mutex here since at this stage
317 * there is no one referencing to this relayd object.
319 (void) relayd_close(&relayd
->control_sock
);
320 (void) relayd_close(&relayd
->data_sock
);
326 * Destroy and free relayd socket pair object.
328 void consumer_destroy_relayd(struct consumer_relayd_sock_pair
*relayd
)
331 struct lttng_ht_iter iter
;
333 if (relayd
== NULL
) {
337 DBG("Consumer destroy and close relayd socket pair");
339 iter
.iter
.node
= &relayd
->node
.node
;
340 ret
= lttng_ht_del(consumer_data
.relayd_ht
, &iter
);
342 /* We assume the relayd is being or is destroyed */
346 /* RCU free() call */
347 call_rcu(&relayd
->node
.head
, free_relayd_rcu
);
351 * Remove a channel from the global list protected by a mutex. This function is
352 * also responsible for freeing its data structures.
354 void consumer_del_channel(struct lttng_consumer_channel
*channel
)
357 struct lttng_ht_iter iter
;
359 DBG("Consumer delete channel key %" PRIu64
, channel
->key
);
361 pthread_mutex_lock(&consumer_data
.lock
);
362 pthread_mutex_lock(&channel
->lock
);
364 /* Destroy streams that might have been left in the stream list. */
365 clean_channel_stream_list(channel
);
367 if (channel
->live_timer_enabled
== 1) {
368 consumer_timer_live_stop(channel
);
371 switch (consumer_data
.type
) {
372 case LTTNG_CONSUMER_KERNEL
:
374 case LTTNG_CONSUMER32_UST
:
375 case LTTNG_CONSUMER64_UST
:
376 lttng_ustconsumer_del_channel(channel
);
379 ERR("Unknown consumer_data type");
385 iter
.iter
.node
= &channel
->node
.node
;
386 ret
= lttng_ht_del(consumer_data
.channel_ht
, &iter
);
390 call_rcu(&channel
->node
.head
, free_channel_rcu
);
392 pthread_mutex_unlock(&channel
->lock
);
393 pthread_mutex_unlock(&consumer_data
.lock
);
397 * Iterate over the relayd hash table and destroy each element. Finally,
398 * destroy the whole hash table.
400 static void cleanup_relayd_ht(void)
402 struct lttng_ht_iter iter
;
403 struct consumer_relayd_sock_pair
*relayd
;
407 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
409 consumer_destroy_relayd(relayd
);
414 lttng_ht_destroy(consumer_data
.relayd_ht
);
418 * Update the end point status of all streams having the given network sequence
419 * index (relayd index).
421 * It's atomically set without having the stream mutex locked which is fine
422 * because we handle the write/read race with a pipe wakeup for each thread.
424 static void update_endpoint_status_by_netidx(uint64_t net_seq_idx
,
425 enum consumer_endpoint_status status
)
427 struct lttng_ht_iter iter
;
428 struct lttng_consumer_stream
*stream
;
430 DBG("Consumer set delete flag on stream by idx %" PRIu64
, net_seq_idx
);
434 /* Let's begin with metadata */
435 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
436 if (stream
->net_seq_idx
== net_seq_idx
) {
437 uatomic_set(&stream
->endpoint_status
, status
);
438 DBG("Delete flag set to metadata stream %d", stream
->wait_fd
);
442 /* Follow up by the data streams */
443 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
444 if (stream
->net_seq_idx
== net_seq_idx
) {
445 uatomic_set(&stream
->endpoint_status
, status
);
446 DBG("Delete flag set to data stream %d", stream
->wait_fd
);
453 * Cleanup a relayd object by flagging every associated streams for deletion,
454 * destroying the object meaning removing it from the relayd hash table,
455 * closing the sockets and freeing the memory in a RCU call.
457 * If a local data context is available, notify the threads that the streams'
458 * state have changed.
460 static void cleanup_relayd(struct consumer_relayd_sock_pair
*relayd
,
461 struct lttng_consumer_local_data
*ctx
)
467 DBG("Cleaning up relayd sockets");
469 /* Save the net sequence index before destroying the object */
470 netidx
= relayd
->net_seq_idx
;
473 * Delete the relayd from the relayd hash table, close the sockets and free
474 * the object in a RCU call.
476 consumer_destroy_relayd(relayd
);
478 /* Set inactive endpoint to all streams */
479 update_endpoint_status_by_netidx(netidx
, CONSUMER_ENDPOINT_INACTIVE
);
482 * With a local data context, notify the threads that the streams' state
483 * have changed. The write() action on the pipe acts as an "implicit"
484 * memory barrier ordering the updates of the end point status from the
485 * read of this status which happens AFTER receiving this notify.
488 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
489 notify_thread_lttng_pipe(ctx
->consumer_metadata_pipe
);
494 * Flag a relayd socket pair for destruction. Destroy it if the refcount
497 * RCU read side lock MUST be aquired before calling this function.
499 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair
*relayd
)
503 /* Set destroy flag for this object */
504 uatomic_set(&relayd
->destroy_flag
, 1);
506 /* Destroy the relayd if refcount is 0 */
507 if (uatomic_read(&relayd
->refcount
) == 0) {
508 consumer_destroy_relayd(relayd
);
513 * Completly destroy stream from every visiable data structure and the given
516 * One this call returns, the stream object is not longer usable nor visible.
518 void consumer_del_stream(struct lttng_consumer_stream
*stream
,
521 consumer_stream_destroy(stream
, ht
);
525 * XXX naming of del vs destroy is all mixed up.
527 void consumer_del_stream_for_data(struct lttng_consumer_stream
*stream
)
529 consumer_stream_destroy(stream
, data_ht
);
532 void consumer_del_stream_for_metadata(struct lttng_consumer_stream
*stream
)
534 consumer_stream_destroy(stream
, metadata_ht
);
537 struct lttng_consumer_stream
*consumer_allocate_stream(uint64_t channel_key
,
539 enum lttng_consumer_stream_state state
,
540 const char *channel_name
,
547 enum consumer_channel_type type
,
548 unsigned int monitor
)
551 struct lttng_consumer_stream
*stream
;
553 stream
= zmalloc(sizeof(*stream
));
554 if (stream
== NULL
) {
555 PERROR("malloc struct lttng_consumer_stream");
562 stream
->key
= stream_key
;
564 stream
->out_fd_offset
= 0;
565 stream
->output_written
= 0;
566 stream
->state
= state
;
569 stream
->net_seq_idx
= relayd_id
;
570 stream
->session_id
= session_id
;
571 stream
->monitor
= monitor
;
572 stream
->endpoint_status
= CONSUMER_ENDPOINT_ACTIVE
;
573 stream
->index_fd
= -1;
574 stream
->last_sequence_number
= -1ULL;
575 pthread_mutex_init(&stream
->lock
, NULL
);
576 pthread_mutex_init(&stream
->metadata_timer_lock
, NULL
);
578 /* If channel is the metadata, flag this stream as metadata. */
579 if (type
== CONSUMER_CHANNEL_TYPE_METADATA
) {
580 stream
->metadata_flag
= 1;
581 /* Metadata is flat out. */
582 strncpy(stream
->name
, DEFAULT_METADATA_NAME
, sizeof(stream
->name
));
583 /* Live rendez-vous point. */
584 pthread_cond_init(&stream
->metadata_rdv
, NULL
);
585 pthread_mutex_init(&stream
->metadata_rdv_lock
, NULL
);
587 /* Format stream name to <channel_name>_<cpu_number> */
588 ret
= snprintf(stream
->name
, sizeof(stream
->name
), "%s_%d",
591 PERROR("snprintf stream name");
596 /* Key is always the wait_fd for streams. */
597 lttng_ht_node_init_u64(&stream
->node
, stream
->key
);
599 /* Init node per channel id key */
600 lttng_ht_node_init_u64(&stream
->node_channel_id
, channel_key
);
602 /* Init session id node with the stream session id */
603 lttng_ht_node_init_u64(&stream
->node_session_id
, stream
->session_id
);
605 DBG3("Allocated stream %s (key %" PRIu64
", chan_key %" PRIu64
606 " relayd_id %" PRIu64
", session_id %" PRIu64
,
607 stream
->name
, stream
->key
, channel_key
,
608 stream
->net_seq_idx
, stream
->session_id
);
624 * Add a stream to the global list protected by a mutex.
626 int consumer_add_data_stream(struct lttng_consumer_stream
*stream
)
628 struct lttng_ht
*ht
= data_ht
;
634 DBG3("Adding consumer stream %" PRIu64
, stream
->key
);
636 pthread_mutex_lock(&consumer_data
.lock
);
637 pthread_mutex_lock(&stream
->chan
->lock
);
638 pthread_mutex_lock(&stream
->chan
->timer_lock
);
639 pthread_mutex_lock(&stream
->lock
);
642 /* Steal stream identifier to avoid having streams with the same key */
643 steal_stream_key(stream
->key
, ht
);
645 lttng_ht_add_unique_u64(ht
, &stream
->node
);
647 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
648 &stream
->node_channel_id
);
651 * Add stream to the stream_list_ht of the consumer data. No need to steal
652 * the key since the HT does not use it and we allow to add redundant keys
655 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
658 * When nb_init_stream_left reaches 0, we don't need to trigger any action
659 * in terms of destroying the associated channel, because the action that
660 * causes the count to become 0 also causes a stream to be added. The
661 * channel deletion will thus be triggered by the following removal of this
664 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
665 /* Increment refcount before decrementing nb_init_stream_left */
667 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
670 /* Update consumer data once the node is inserted. */
671 consumer_data
.stream_count
++;
672 consumer_data
.need_update
= 1;
675 pthread_mutex_unlock(&stream
->lock
);
676 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
677 pthread_mutex_unlock(&stream
->chan
->lock
);
678 pthread_mutex_unlock(&consumer_data
.lock
);
683 void consumer_del_data_stream(struct lttng_consumer_stream
*stream
)
685 consumer_del_stream(stream
, data_ht
);
689 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
690 * be acquired before calling this.
692 static int add_relayd(struct consumer_relayd_sock_pair
*relayd
)
695 struct lttng_ht_node_u64
*node
;
696 struct lttng_ht_iter iter
;
700 lttng_ht_lookup(consumer_data
.relayd_ht
,
701 &relayd
->net_seq_idx
, &iter
);
702 node
= lttng_ht_iter_get_node_u64(&iter
);
706 lttng_ht_add_unique_u64(consumer_data
.relayd_ht
, &relayd
->node
);
713 * Allocate and return a consumer relayd socket.
715 static struct consumer_relayd_sock_pair
*consumer_allocate_relayd_sock_pair(
716 uint64_t net_seq_idx
)
718 struct consumer_relayd_sock_pair
*obj
= NULL
;
720 /* net sequence index of -1 is a failure */
721 if (net_seq_idx
== (uint64_t) -1ULL) {
725 obj
= zmalloc(sizeof(struct consumer_relayd_sock_pair
));
727 PERROR("zmalloc relayd sock");
731 obj
->net_seq_idx
= net_seq_idx
;
733 obj
->destroy_flag
= 0;
734 obj
->control_sock
.sock
.fd
= -1;
735 obj
->data_sock
.sock
.fd
= -1;
736 lttng_ht_node_init_u64(&obj
->node
, obj
->net_seq_idx
);
737 pthread_mutex_init(&obj
->ctrl_sock_mutex
, NULL
);
744 * Find a relayd socket pair in the global consumer data.
746 * Return the object if found else NULL.
747 * RCU read-side lock must be held across this call and while using the
750 struct consumer_relayd_sock_pair
*consumer_find_relayd(uint64_t key
)
752 struct lttng_ht_iter iter
;
753 struct lttng_ht_node_u64
*node
;
754 struct consumer_relayd_sock_pair
*relayd
= NULL
;
756 /* Negative keys are lookup failures */
757 if (key
== (uint64_t) -1ULL) {
761 lttng_ht_lookup(consumer_data
.relayd_ht
, &key
,
763 node
= lttng_ht_iter_get_node_u64(&iter
);
765 relayd
= caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
773 * Find a relayd and send the stream
775 * Returns 0 on success, < 0 on error
777 int consumer_send_relayd_stream(struct lttng_consumer_stream
*stream
,
781 struct consumer_relayd_sock_pair
*relayd
;
784 assert(stream
->net_seq_idx
!= -1ULL);
787 /* The stream is not metadata. Get relayd reference if exists. */
789 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
790 if (relayd
!= NULL
) {
791 /* Add stream on the relayd */
792 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
793 ret
= relayd_add_stream(&relayd
->control_sock
, stream
->name
,
794 path
, &stream
->relayd_stream_id
,
795 stream
->chan
->tracefile_size
, stream
->chan
->tracefile_count
);
796 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
801 uatomic_inc(&relayd
->refcount
);
802 stream
->sent_to_relayd
= 1;
804 ERR("Stream %" PRIu64
" relayd ID %" PRIu64
" unknown. Can't send it.",
805 stream
->key
, stream
->net_seq_idx
);
810 DBG("Stream %s with key %" PRIu64
" sent to relayd id %" PRIu64
,
811 stream
->name
, stream
->key
, stream
->net_seq_idx
);
819 * Find a relayd and send the streams sent message
821 * Returns 0 on success, < 0 on error
823 int consumer_send_relayd_streams_sent(uint64_t net_seq_idx
)
826 struct consumer_relayd_sock_pair
*relayd
;
828 assert(net_seq_idx
!= -1ULL);
830 /* The stream is not metadata. Get relayd reference if exists. */
832 relayd
= consumer_find_relayd(net_seq_idx
);
833 if (relayd
!= NULL
) {
834 /* Add stream on the relayd */
835 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
836 ret
= relayd_streams_sent(&relayd
->control_sock
);
837 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
842 ERR("Relayd ID %" PRIu64
" unknown. Can't send streams_sent.",
849 DBG("All streams sent relayd id %" PRIu64
, net_seq_idx
);
857 * Find a relayd and close the stream
859 void close_relayd_stream(struct lttng_consumer_stream
*stream
)
861 struct consumer_relayd_sock_pair
*relayd
;
863 /* The stream is not metadata. Get relayd reference if exists. */
865 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
867 consumer_stream_relayd_close(stream
, relayd
);
873 * Handle stream for relayd transmission if the stream applies for network
874 * streaming where the net sequence index is set.
876 * Return destination file descriptor or negative value on error.
878 static int write_relayd_stream_header(struct lttng_consumer_stream
*stream
,
879 size_t data_size
, unsigned long padding
,
880 struct consumer_relayd_sock_pair
*relayd
)
883 struct lttcomm_relayd_data_hdr data_hdr
;
889 /* Reset data header */
890 memset(&data_hdr
, 0, sizeof(data_hdr
));
892 if (stream
->metadata_flag
) {
893 /* Caller MUST acquire the relayd control socket lock */
894 ret
= relayd_send_metadata(&relayd
->control_sock
, data_size
);
899 /* Metadata are always sent on the control socket. */
900 outfd
= relayd
->control_sock
.sock
.fd
;
902 /* Set header with stream information */
903 data_hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
904 data_hdr
.data_size
= htobe32(data_size
);
905 data_hdr
.padding_size
= htobe32(padding
);
907 * Note that net_seq_num below is assigned with the *current* value of
908 * next_net_seq_num and only after that the next_net_seq_num will be
909 * increment. This is why when issuing a command on the relayd using
910 * this next value, 1 should always be substracted in order to compare
911 * the last seen sequence number on the relayd side to the last sent.
913 data_hdr
.net_seq_num
= htobe64(stream
->next_net_seq_num
);
914 /* Other fields are zeroed previously */
916 ret
= relayd_send_data_hdr(&relayd
->data_sock
, &data_hdr
,
922 ++stream
->next_net_seq_num
;
924 /* Set to go on data socket */
925 outfd
= relayd
->data_sock
.sock
.fd
;
933 * Allocate and return a new lttng_consumer_channel object using the given key
934 * to initialize the hash table node.
936 * On error, return NULL.
938 struct lttng_consumer_channel
*consumer_allocate_channel(uint64_t key
,
940 const char *pathname
,
945 enum lttng_event_output output
,
946 uint64_t tracefile_size
,
947 uint64_t tracefile_count
,
948 uint64_t session_id_per_pid
,
949 unsigned int monitor
,
950 unsigned int live_timer_interval
,
951 const char *root_shm_path
,
952 const char *shm_path
)
954 struct lttng_consumer_channel
*channel
;
956 channel
= zmalloc(sizeof(*channel
));
957 if (channel
== NULL
) {
958 PERROR("malloc struct lttng_consumer_channel");
963 channel
->refcount
= 0;
964 channel
->session_id
= session_id
;
965 channel
->session_id_per_pid
= session_id_per_pid
;
968 channel
->relayd_id
= relayd_id
;
969 channel
->tracefile_size
= tracefile_size
;
970 channel
->tracefile_count
= tracefile_count
;
971 channel
->monitor
= monitor
;
972 channel
->live_timer_interval
= live_timer_interval
;
973 pthread_mutex_init(&channel
->lock
, NULL
);
974 pthread_mutex_init(&channel
->timer_lock
, NULL
);
977 case LTTNG_EVENT_SPLICE
:
978 channel
->output
= CONSUMER_CHANNEL_SPLICE
;
980 case LTTNG_EVENT_MMAP
:
981 channel
->output
= CONSUMER_CHANNEL_MMAP
;
991 * In monitor mode, the streams associated with the channel will be put in
992 * a special list ONLY owned by this channel. So, the refcount is set to 1
993 * here meaning that the channel itself has streams that are referenced.
995 * On a channel deletion, once the channel is no longer visible, the
996 * refcount is decremented and checked for a zero value to delete it. With
997 * streams in no monitor mode, it will now be safe to destroy the channel.
999 if (!channel
->monitor
) {
1000 channel
->refcount
= 1;
1003 strncpy(channel
->pathname
, pathname
, sizeof(channel
->pathname
));
1004 channel
->pathname
[sizeof(channel
->pathname
) - 1] = '\0';
1006 strncpy(channel
->name
, name
, sizeof(channel
->name
));
1007 channel
->name
[sizeof(channel
->name
) - 1] = '\0';
1009 if (root_shm_path
) {
1010 strncpy(channel
->root_shm_path
, root_shm_path
, sizeof(channel
->root_shm_path
));
1011 channel
->root_shm_path
[sizeof(channel
->root_shm_path
) - 1] = '\0';
1014 strncpy(channel
->shm_path
, shm_path
, sizeof(channel
->shm_path
));
1015 channel
->shm_path
[sizeof(channel
->shm_path
) - 1] = '\0';
1018 lttng_ht_node_init_u64(&channel
->node
, channel
->key
);
1020 channel
->wait_fd
= -1;
1022 CDS_INIT_LIST_HEAD(&channel
->streams
.head
);
1024 DBG("Allocated channel (key %" PRIu64
")", channel
->key
);
1031 * Add a channel to the global list protected by a mutex.
1033 * Always return 0 indicating success.
1035 int consumer_add_channel(struct lttng_consumer_channel
*channel
,
1036 struct lttng_consumer_local_data
*ctx
)
1038 pthread_mutex_lock(&consumer_data
.lock
);
1039 pthread_mutex_lock(&channel
->lock
);
1040 pthread_mutex_lock(&channel
->timer_lock
);
1043 * This gives us a guarantee that the channel we are about to add to the
1044 * channel hash table will be unique. See this function comment on the why
1045 * we need to steel the channel key at this stage.
1047 steal_channel_key(channel
->key
);
1050 lttng_ht_add_unique_u64(consumer_data
.channel_ht
, &channel
->node
);
1053 pthread_mutex_unlock(&channel
->timer_lock
);
1054 pthread_mutex_unlock(&channel
->lock
);
1055 pthread_mutex_unlock(&consumer_data
.lock
);
1057 if (channel
->wait_fd
!= -1 && channel
->type
== CONSUMER_CHANNEL_TYPE_DATA
) {
1058 notify_channel_pipe(ctx
, channel
, -1, CONSUMER_CHANNEL_ADD
);
1065 * Allocate the pollfd structure and the local view of the out fds to avoid
1066 * doing a lookup in the linked list and concurrency issues when writing is
1067 * needed. Called with consumer_data.lock held.
1069 * Returns the number of fds in the structures.
1071 static int update_poll_array(struct lttng_consumer_local_data
*ctx
,
1072 struct pollfd
**pollfd
, struct lttng_consumer_stream
**local_stream
,
1073 struct lttng_ht
*ht
)
1076 struct lttng_ht_iter iter
;
1077 struct lttng_consumer_stream
*stream
;
1082 assert(local_stream
);
1084 DBG("Updating poll fd array");
1086 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1088 * Only active streams with an active end point can be added to the
1089 * poll set and local stream storage of the thread.
1091 * There is a potential race here for endpoint_status to be updated
1092 * just after the check. However, this is OK since the stream(s) will
1093 * be deleted once the thread is notified that the end point state has
1094 * changed where this function will be called back again.
1096 if (stream
->state
!= LTTNG_CONSUMER_ACTIVE_STREAM
||
1097 stream
->endpoint_status
== CONSUMER_ENDPOINT_INACTIVE
) {
1101 * This clobbers way too much the debug output. Uncomment that if you
1102 * need it for debugging purposes.
1104 * DBG("Active FD %d", stream->wait_fd);
1106 (*pollfd
)[i
].fd
= stream
->wait_fd
;
1107 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1108 local_stream
[i
] = stream
;
1114 * Insert the consumer_data_pipe at the end of the array and don't
1115 * increment i so nb_fd is the number of real FD.
1117 (*pollfd
)[i
].fd
= lttng_pipe_get_readfd(ctx
->consumer_data_pipe
);
1118 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1120 (*pollfd
)[i
+ 1].fd
= lttng_pipe_get_readfd(ctx
->consumer_wakeup_pipe
);
1121 (*pollfd
)[i
+ 1].events
= POLLIN
| POLLPRI
;
1126 * Poll on the should_quit pipe and the command socket return -1 on
1127 * error, 1 if should exit, 0 if data is available on the command socket
1129 int lttng_consumer_poll_socket(struct pollfd
*consumer_sockpoll
)
1134 num_rdy
= poll(consumer_sockpoll
, 2, -1);
1135 if (num_rdy
== -1) {
1137 * Restart interrupted system call.
1139 if (errno
== EINTR
) {
1142 PERROR("Poll error");
1145 if (consumer_sockpoll
[0].revents
& (POLLIN
| POLLPRI
)) {
1146 DBG("consumer_should_quit wake up");
1153 * Set the error socket.
1155 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data
*ctx
,
1158 ctx
->consumer_error_socket
= sock
;
1162 * Set the command socket path.
1164 void lttng_consumer_set_command_sock_path(
1165 struct lttng_consumer_local_data
*ctx
, char *sock
)
1167 ctx
->consumer_command_sock_path
= sock
;
1171 * Send return code to the session daemon.
1172 * If the socket is not defined, we return 0, it is not a fatal error
1174 int lttng_consumer_send_error(struct lttng_consumer_local_data
*ctx
, int cmd
)
1176 if (ctx
->consumer_error_socket
> 0) {
1177 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
1178 sizeof(enum lttcomm_sessiond_command
));
1185 * Close all the tracefiles and stream fds and MUST be called when all
1186 * instances are destroyed i.e. when all threads were joined and are ended.
1188 void lttng_consumer_cleanup(void)
1190 struct lttng_ht_iter iter
;
1191 struct lttng_consumer_channel
*channel
;
1195 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
, channel
,
1197 consumer_del_channel(channel
);
1202 lttng_ht_destroy(consumer_data
.channel_ht
);
1204 cleanup_relayd_ht();
1206 lttng_ht_destroy(consumer_data
.stream_per_chan_id_ht
);
1209 * This HT contains streams that are freed by either the metadata thread or
1210 * the data thread so we do *nothing* on the hash table and simply destroy
1213 lttng_ht_destroy(consumer_data
.stream_list_ht
);
1217 * Called from signal handler.
1219 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
1224 ret
= lttng_write(ctx
->consumer_should_quit
[1], "4", 1);
1226 PERROR("write consumer quit");
1229 DBG("Consumer flag that it should quit");
1234 * Flush pending writes to trace output disk file.
1237 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1241 int outfd
= stream
->out_fd
;
1244 * This does a blocking write-and-wait on any page that belongs to the
1245 * subbuffer prior to the one we just wrote.
1246 * Don't care about error values, as these are just hints and ways to
1247 * limit the amount of page cache used.
1249 if (orig_offset
< stream
->max_sb_size
) {
1252 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1253 stream
->max_sb_size
,
1254 SYNC_FILE_RANGE_WAIT_BEFORE
1255 | SYNC_FILE_RANGE_WRITE
1256 | SYNC_FILE_RANGE_WAIT_AFTER
);
1258 * Give hints to the kernel about how we access the file:
1259 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1262 * We need to call fadvise again after the file grows because the
1263 * kernel does not seem to apply fadvise to non-existing parts of the
1266 * Call fadvise _after_ having waited for the page writeback to
1267 * complete because the dirty page writeback semantic is not well
1268 * defined. So it can be expected to lead to lower throughput in
1271 ret
= posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1272 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1273 if (ret
&& ret
!= -ENOSYS
) {
1275 PERROR("posix_fadvise on fd %i", outfd
);
1280 * Initialise the necessary environnement :
1281 * - create a new context
1282 * - create the poll_pipe
1283 * - create the should_quit pipe (for signal handler)
1284 * - create the thread pipe (for splice)
1286 * Takes a function pointer as argument, this function is called when data is
1287 * available on a buffer. This function is responsible to do the
1288 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1289 * buffer configuration and then kernctl_put_next_subbuf at the end.
1291 * Returns a pointer to the new context or NULL on error.
1293 struct lttng_consumer_local_data
*lttng_consumer_create(
1294 enum lttng_consumer_type type
,
1295 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1296 struct lttng_consumer_local_data
*ctx
),
1297 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1298 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1299 int (*update_stream
)(uint64_t stream_key
, uint32_t state
))
1302 struct lttng_consumer_local_data
*ctx
;
1304 assert(consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1305 consumer_data
.type
== type
);
1306 consumer_data
.type
= type
;
1308 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
1310 PERROR("allocating context");
1314 ctx
->consumer_error_socket
= -1;
1315 ctx
->consumer_metadata_socket
= -1;
1316 pthread_mutex_init(&ctx
->metadata_socket_lock
, NULL
);
1317 /* assign the callbacks */
1318 ctx
->on_buffer_ready
= buffer_ready
;
1319 ctx
->on_recv_channel
= recv_channel
;
1320 ctx
->on_recv_stream
= recv_stream
;
1321 ctx
->on_update_stream
= update_stream
;
1323 ctx
->consumer_data_pipe
= lttng_pipe_open(0);
1324 if (!ctx
->consumer_data_pipe
) {
1325 goto error_poll_pipe
;
1328 ctx
->consumer_wakeup_pipe
= lttng_pipe_open(0);
1329 if (!ctx
->consumer_wakeup_pipe
) {
1330 goto error_wakeup_pipe
;
1333 ret
= pipe(ctx
->consumer_should_quit
);
1335 PERROR("Error creating recv pipe");
1336 goto error_quit_pipe
;
1339 ret
= pipe(ctx
->consumer_channel_pipe
);
1341 PERROR("Error creating channel pipe");
1342 goto error_channel_pipe
;
1345 ctx
->consumer_metadata_pipe
= lttng_pipe_open(0);
1346 if (!ctx
->consumer_metadata_pipe
) {
1347 goto error_metadata_pipe
;
1352 error_metadata_pipe
:
1353 utils_close_pipe(ctx
->consumer_channel_pipe
);
1355 utils_close_pipe(ctx
->consumer_should_quit
);
1357 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1359 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1367 * Iterate over all streams of the hashtable and free them properly.
1369 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1371 struct lttng_ht_iter iter
;
1372 struct lttng_consumer_stream
*stream
;
1379 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1381 * Ignore return value since we are currently cleaning up so any error
1384 (void) consumer_del_stream(stream
, ht
);
1388 lttng_ht_destroy(ht
);
1392 * Iterate over all streams of the metadata hashtable and free them
1395 static void destroy_metadata_stream_ht(struct lttng_ht
*ht
)
1397 struct lttng_ht_iter iter
;
1398 struct lttng_consumer_stream
*stream
;
1405 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1407 * Ignore return value since we are currently cleaning up so any error
1410 (void) consumer_del_metadata_stream(stream
, ht
);
1414 lttng_ht_destroy(ht
);
1418 * Close all fds associated with the instance and free the context.
1420 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1424 DBG("Consumer destroying it. Closing everything.");
1430 destroy_data_stream_ht(data_ht
);
1431 destroy_metadata_stream_ht(metadata_ht
);
1433 ret
= close(ctx
->consumer_error_socket
);
1437 ret
= close(ctx
->consumer_metadata_socket
);
1441 utils_close_pipe(ctx
->consumer_channel_pipe
);
1442 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1443 lttng_pipe_destroy(ctx
->consumer_metadata_pipe
);
1444 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1445 utils_close_pipe(ctx
->consumer_should_quit
);
1447 unlink(ctx
->consumer_command_sock_path
);
1452 * Write the metadata stream id on the specified file descriptor.
1454 static int write_relayd_metadata_id(int fd
,
1455 struct lttng_consumer_stream
*stream
,
1456 struct consumer_relayd_sock_pair
*relayd
, unsigned long padding
)
1459 struct lttcomm_relayd_metadata_payload hdr
;
1461 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1462 hdr
.padding_size
= htobe32(padding
);
1463 ret
= lttng_write(fd
, (void *) &hdr
, sizeof(hdr
));
1464 if (ret
< sizeof(hdr
)) {
1466 * This error means that the fd's end is closed so ignore the PERROR
1467 * not to clubber the error output since this can happen in a normal
1470 if (errno
!= EPIPE
) {
1471 PERROR("write metadata stream id");
1473 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1475 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1476 * handle writting the missing part so report that as an error and
1477 * don't lie to the caller.
1482 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1483 stream
->relayd_stream_id
, padding
);
1490 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1491 * core function for writing trace buffers to either the local filesystem or
1494 * It must be called with the stream lock held.
1496 * Careful review MUST be put if any changes occur!
1498 * Returns the number of bytes written
1500 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1501 struct lttng_consumer_local_data
*ctx
,
1502 struct lttng_consumer_stream
*stream
, unsigned long len
,
1503 unsigned long padding
,
1504 struct ctf_packet_index
*index
)
1506 unsigned long mmap_offset
;
1509 off_t orig_offset
= stream
->out_fd_offset
;
1510 /* Default is on the disk */
1511 int outfd
= stream
->out_fd
;
1512 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1513 unsigned int relayd_hang_up
= 0;
1515 /* RCU lock for the relayd pointer */
1518 /* Flag that the current stream if set for network streaming. */
1519 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1520 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1521 if (relayd
== NULL
) {
1527 /* get the offset inside the fd to mmap */
1528 switch (consumer_data
.type
) {
1529 case LTTNG_CONSUMER_KERNEL
:
1530 mmap_base
= stream
->mmap_base
;
1531 ret
= kernctl_get_mmap_read_offset(stream
->wait_fd
, &mmap_offset
);
1534 PERROR("tracer ctl get_mmap_read_offset");
1538 case LTTNG_CONSUMER32_UST
:
1539 case LTTNG_CONSUMER64_UST
:
1540 mmap_base
= lttng_ustctl_get_mmap_base(stream
);
1542 ERR("read mmap get mmap base for stream %s", stream
->name
);
1546 ret
= lttng_ustctl_get_mmap_read_offset(stream
, &mmap_offset
);
1548 PERROR("tracer ctl get_mmap_read_offset");
1554 ERR("Unknown consumer_data type");
1558 /* Handle stream on the relayd if the output is on the network */
1560 unsigned long netlen
= len
;
1563 * Lock the control socket for the complete duration of the function
1564 * since from this point on we will use the socket.
1566 if (stream
->metadata_flag
) {
1567 /* Metadata requires the control socket. */
1568 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1569 if (stream
->reset_metadata_flag
) {
1570 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1571 stream
->relayd_stream_id
,
1572 stream
->metadata_version
);
1577 stream
->reset_metadata_flag
= 0;
1579 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1582 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1587 /* Use the returned socket. */
1590 /* Write metadata stream id before payload */
1591 if (stream
->metadata_flag
) {
1592 ret
= write_relayd_metadata_id(outfd
, stream
, relayd
, padding
);
1599 /* No streaming, we have to set the len with the full padding */
1602 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1603 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1605 ERR("Reset metadata file");
1608 stream
->reset_metadata_flag
= 0;
1612 * Check if we need to change the tracefile before writing the packet.
1614 if (stream
->chan
->tracefile_size
> 0 &&
1615 (stream
->tracefile_size_current
+ len
) >
1616 stream
->chan
->tracefile_size
) {
1617 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1618 stream
->name
, stream
->chan
->tracefile_size
,
1619 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1620 stream
->out_fd
, &(stream
->tracefile_count_current
),
1623 ERR("Rotating output file");
1626 outfd
= stream
->out_fd
;
1628 if (stream
->index_fd
>= 0) {
1629 ret
= close(stream
->index_fd
);
1631 PERROR("Closing index");
1634 stream
->index_fd
= -1;
1635 ret
= index_create_file(stream
->chan
->pathname
,
1636 stream
->name
, stream
->uid
, stream
->gid
,
1637 stream
->chan
->tracefile_size
,
1638 stream
->tracefile_count_current
);
1642 stream
->index_fd
= ret
;
1645 /* Reset current size because we just perform a rotation. */
1646 stream
->tracefile_size_current
= 0;
1647 stream
->out_fd_offset
= 0;
1650 stream
->tracefile_size_current
+= len
;
1652 index
->offset
= htobe64(stream
->out_fd_offset
);
1657 * This call guarantee that len or less is returned. It's impossible to
1658 * receive a ret value that is bigger than len.
1660 ret
= lttng_write(outfd
, mmap_base
+ mmap_offset
, len
);
1661 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, len
);
1662 if (ret
< 0 || ((size_t) ret
!= len
)) {
1664 * Report error to caller if nothing was written else at least send the
1672 /* Socket operation failed. We consider the relayd dead */
1673 if (errno
== EPIPE
|| errno
== EINVAL
|| errno
== EBADF
) {
1675 * This is possible if the fd is closed on the other side
1676 * (outfd) or any write problem. It can be verbose a bit for a
1677 * normal execution if for instance the relayd is stopped
1678 * abruptly. This can happen so set this to a DBG statement.
1680 DBG("Consumer mmap write detected relayd hang up");
1682 /* Unhandled error, print it and stop function right now. */
1683 PERROR("Error in write mmap (ret %zd != len %lu)", ret
, len
);
1687 stream
->output_written
+= ret
;
1689 /* This call is useless on a socket so better save a syscall. */
1691 /* This won't block, but will start writeout asynchronously */
1692 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, len
,
1693 SYNC_FILE_RANGE_WRITE
);
1694 stream
->out_fd_offset
+= len
;
1695 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1700 * This is a special case that the relayd has closed its socket. Let's
1701 * cleanup the relayd object and all associated streams.
1703 if (relayd
&& relayd_hang_up
) {
1704 cleanup_relayd(relayd
, ctx
);
1708 /* Unlock only if ctrl socket used */
1709 if (relayd
&& stream
->metadata_flag
) {
1710 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1718 * Splice the data from the ring buffer to the tracefile.
1720 * It must be called with the stream lock held.
1722 * Returns the number of bytes spliced.
1724 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1725 struct lttng_consumer_local_data
*ctx
,
1726 struct lttng_consumer_stream
*stream
, unsigned long len
,
1727 unsigned long padding
,
1728 struct ctf_packet_index
*index
)
1730 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1732 off_t orig_offset
= stream
->out_fd_offset
;
1733 int fd
= stream
->wait_fd
;
1734 /* Default is on the disk */
1735 int outfd
= stream
->out_fd
;
1736 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1738 unsigned int relayd_hang_up
= 0;
1740 switch (consumer_data
.type
) {
1741 case LTTNG_CONSUMER_KERNEL
:
1743 case LTTNG_CONSUMER32_UST
:
1744 case LTTNG_CONSUMER64_UST
:
1745 /* Not supported for user space tracing */
1748 ERR("Unknown consumer_data type");
1752 /* RCU lock for the relayd pointer */
1755 /* Flag that the current stream if set for network streaming. */
1756 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1757 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1758 if (relayd
== NULL
) {
1763 splice_pipe
= stream
->splice_pipe
;
1765 /* Write metadata stream id before payload */
1767 unsigned long total_len
= len
;
1769 if (stream
->metadata_flag
) {
1771 * Lock the control socket for the complete duration of the function
1772 * since from this point on we will use the socket.
1774 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1776 if (stream
->reset_metadata_flag
) {
1777 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1778 stream
->relayd_stream_id
,
1779 stream
->metadata_version
);
1784 stream
->reset_metadata_flag
= 0;
1786 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
, relayd
,
1794 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1797 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1803 /* Use the returned socket. */
1806 /* No streaming, we have to set the len with the full padding */
1809 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1810 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1812 ERR("Reset metadata file");
1815 stream
->reset_metadata_flag
= 0;
1818 * Check if we need to change the tracefile before writing the packet.
1820 if (stream
->chan
->tracefile_size
> 0 &&
1821 (stream
->tracefile_size_current
+ len
) >
1822 stream
->chan
->tracefile_size
) {
1823 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1824 stream
->name
, stream
->chan
->tracefile_size
,
1825 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1826 stream
->out_fd
, &(stream
->tracefile_count_current
),
1830 ERR("Rotating output file");
1833 outfd
= stream
->out_fd
;
1835 if (stream
->index_fd
>= 0) {
1836 ret
= close(stream
->index_fd
);
1838 PERROR("Closing index");
1841 stream
->index_fd
= -1;
1842 ret
= index_create_file(stream
->chan
->pathname
,
1843 stream
->name
, stream
->uid
, stream
->gid
,
1844 stream
->chan
->tracefile_size
,
1845 stream
->tracefile_count_current
);
1850 stream
->index_fd
= ret
;
1853 /* Reset current size because we just perform a rotation. */
1854 stream
->tracefile_size_current
= 0;
1855 stream
->out_fd_offset
= 0;
1858 stream
->tracefile_size_current
+= len
;
1859 index
->offset
= htobe64(stream
->out_fd_offset
);
1863 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1864 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1865 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1866 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1867 DBG("splice chan to pipe, ret %zd", ret_splice
);
1868 if (ret_splice
< 0) {
1871 PERROR("Error in relay splice");
1875 /* Handle stream on the relayd if the output is on the network */
1876 if (relayd
&& stream
->metadata_flag
) {
1877 size_t metadata_payload_size
=
1878 sizeof(struct lttcomm_relayd_metadata_payload
);
1880 /* Update counter to fit the spliced data */
1881 ret_splice
+= metadata_payload_size
;
1882 len
+= metadata_payload_size
;
1884 * We do this so the return value can match the len passed as
1885 * argument to this function.
1887 written
-= metadata_payload_size
;
1890 /* Splice data out */
1891 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1892 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1893 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
1895 if (ret_splice
< 0) {
1900 } else if (ret_splice
> len
) {
1902 * We don't expect this code path to be executed but you never know
1903 * so this is an extra protection agains a buggy splice().
1906 written
+= ret_splice
;
1907 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice
,
1911 /* All good, update current len and continue. */
1915 /* This call is useless on a socket so better save a syscall. */
1917 /* This won't block, but will start writeout asynchronously */
1918 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1919 SYNC_FILE_RANGE_WRITE
);
1920 stream
->out_fd_offset
+= ret_splice
;
1922 stream
->output_written
+= ret_splice
;
1923 written
+= ret_splice
;
1926 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1932 * This is a special case that the relayd has closed its socket. Let's
1933 * cleanup the relayd object and all associated streams.
1935 if (relayd
&& relayd_hang_up
) {
1936 cleanup_relayd(relayd
, ctx
);
1937 /* Skip splice error so the consumer does not fail */
1942 /* send the appropriate error description to sessiond */
1945 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1948 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1951 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
1956 if (relayd
&& stream
->metadata_flag
) {
1957 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1965 * Take a snapshot for a specific fd
1967 * Returns 0 on success, < 0 on error
1969 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
1971 switch (consumer_data
.type
) {
1972 case LTTNG_CONSUMER_KERNEL
:
1973 return lttng_kconsumer_take_snapshot(stream
);
1974 case LTTNG_CONSUMER32_UST
:
1975 case LTTNG_CONSUMER64_UST
:
1976 return lttng_ustconsumer_take_snapshot(stream
);
1978 ERR("Unknown consumer_data type");
1985 * Get the produced position
1987 * Returns 0 on success, < 0 on error
1989 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
1992 switch (consumer_data
.type
) {
1993 case LTTNG_CONSUMER_KERNEL
:
1994 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
1995 case LTTNG_CONSUMER32_UST
:
1996 case LTTNG_CONSUMER64_UST
:
1997 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
1999 ERR("Unknown consumer_data type");
2005 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
2006 int sock
, struct pollfd
*consumer_sockpoll
)
2008 switch (consumer_data
.type
) {
2009 case LTTNG_CONSUMER_KERNEL
:
2010 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2011 case LTTNG_CONSUMER32_UST
:
2012 case LTTNG_CONSUMER64_UST
:
2013 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2015 ERR("Unknown consumer_data type");
2021 void lttng_consumer_close_all_metadata(void)
2023 switch (consumer_data
.type
) {
2024 case LTTNG_CONSUMER_KERNEL
:
2026 * The Kernel consumer has a different metadata scheme so we don't
2027 * close anything because the stream will be closed by the session
2031 case LTTNG_CONSUMER32_UST
:
2032 case LTTNG_CONSUMER64_UST
:
2034 * Close all metadata streams. The metadata hash table is passed and
2035 * this call iterates over it by closing all wakeup fd. This is safe
2036 * because at this point we are sure that the metadata producer is
2037 * either dead or blocked.
2039 lttng_ustconsumer_close_all_metadata(metadata_ht
);
2042 ERR("Unknown consumer_data type");
2048 * Clean up a metadata stream and free its memory.
2050 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
2051 struct lttng_ht
*ht
)
2053 struct lttng_consumer_channel
*free_chan
= NULL
;
2057 * This call should NEVER receive regular stream. It must always be
2058 * metadata stream and this is crucial for data structure synchronization.
2060 assert(stream
->metadata_flag
);
2062 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
2064 pthread_mutex_lock(&consumer_data
.lock
);
2065 pthread_mutex_lock(&stream
->chan
->lock
);
2066 pthread_mutex_lock(&stream
->lock
);
2068 /* Remove any reference to that stream. */
2069 consumer_stream_delete(stream
, ht
);
2071 /* Close down everything including the relayd if one. */
2072 consumer_stream_close(stream
);
2073 /* Destroy tracer buffers of the stream. */
2074 consumer_stream_destroy_buffers(stream
);
2076 /* Atomically decrement channel refcount since other threads can use it. */
2077 if (!uatomic_sub_return(&stream
->chan
->refcount
, 1)
2078 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
2079 /* Go for channel deletion! */
2080 free_chan
= stream
->chan
;
2084 * Nullify the stream reference so it is not used after deletion. The
2085 * channel lock MUST be acquired before being able to check for a NULL
2088 stream
->chan
->metadata_stream
= NULL
;
2090 pthread_mutex_unlock(&stream
->lock
);
2091 pthread_mutex_unlock(&stream
->chan
->lock
);
2092 pthread_mutex_unlock(&consumer_data
.lock
);
2095 consumer_del_channel(free_chan
);
2098 consumer_stream_free(stream
);
2102 * Action done with the metadata stream when adding it to the consumer internal
2103 * data structures to handle it.
2105 int consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
2107 struct lttng_ht
*ht
= metadata_ht
;
2109 struct lttng_ht_iter iter
;
2110 struct lttng_ht_node_u64
*node
;
2115 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2117 pthread_mutex_lock(&consumer_data
.lock
);
2118 pthread_mutex_lock(&stream
->chan
->lock
);
2119 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2120 pthread_mutex_lock(&stream
->lock
);
2123 * From here, refcounts are updated so be _careful_ when returning an error
2130 * Lookup the stream just to make sure it does not exist in our internal
2131 * state. This should NEVER happen.
2133 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2134 node
= lttng_ht_iter_get_node_u64(&iter
);
2138 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2139 * in terms of destroying the associated channel, because the action that
2140 * causes the count to become 0 also causes a stream to be added. The
2141 * channel deletion will thus be triggered by the following removal of this
2144 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2145 /* Increment refcount before decrementing nb_init_stream_left */
2147 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2150 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2152 lttng_ht_add_unique_u64(consumer_data
.stream_per_chan_id_ht
,
2153 &stream
->node_channel_id
);
2156 * Add stream to the stream_list_ht of the consumer data. No need to steal
2157 * the key since the HT does not use it and we allow to add redundant keys
2160 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2164 pthread_mutex_unlock(&stream
->lock
);
2165 pthread_mutex_unlock(&stream
->chan
->lock
);
2166 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2167 pthread_mutex_unlock(&consumer_data
.lock
);
2172 * Delete data stream that are flagged for deletion (endpoint_status).
2174 static void validate_endpoint_status_data_stream(void)
2176 struct lttng_ht_iter iter
;
2177 struct lttng_consumer_stream
*stream
;
2179 DBG("Consumer delete flagged data stream");
2182 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2183 /* Validate delete flag of the stream */
2184 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2187 /* Delete it right now */
2188 consumer_del_stream(stream
, data_ht
);
2194 * Delete metadata stream that are flagged for deletion (endpoint_status).
2196 static void validate_endpoint_status_metadata_stream(
2197 struct lttng_poll_event
*pollset
)
2199 struct lttng_ht_iter iter
;
2200 struct lttng_consumer_stream
*stream
;
2202 DBG("Consumer delete flagged metadata stream");
2207 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2208 /* Validate delete flag of the stream */
2209 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2213 * Remove from pollset so the metadata thread can continue without
2214 * blocking on a deleted stream.
2216 lttng_poll_del(pollset
, stream
->wait_fd
);
2218 /* Delete it right now */
2219 consumer_del_metadata_stream(stream
, metadata_ht
);
2225 * Thread polls on metadata file descriptor and write them on disk or on the
2228 void *consumer_thread_metadata_poll(void *data
)
2230 int ret
, i
, pollfd
, err
= -1;
2231 uint32_t revents
, nb_fd
;
2232 struct lttng_consumer_stream
*stream
= NULL
;
2233 struct lttng_ht_iter iter
;
2234 struct lttng_ht_node_u64
*node
;
2235 struct lttng_poll_event events
;
2236 struct lttng_consumer_local_data
*ctx
= data
;
2239 rcu_register_thread();
2241 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2243 if (testpoint(consumerd_thread_metadata
)) {
2244 goto error_testpoint
;
2247 health_code_update();
2249 DBG("Thread metadata poll started");
2251 /* Size is set to 1 for the consumer_metadata pipe */
2252 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2254 ERR("Poll set creation failed");
2258 ret
= lttng_poll_add(&events
,
2259 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2265 DBG("Metadata main loop started");
2269 health_code_update();
2270 health_poll_entry();
2271 DBG("Metadata poll wait");
2272 ret
= lttng_poll_wait(&events
, -1);
2273 DBG("Metadata poll return from wait with %d fd(s)",
2274 LTTNG_POLL_GETNB(&events
));
2276 DBG("Metadata event caught in thread");
2278 if (errno
== EINTR
) {
2279 ERR("Poll EINTR caught");
2282 if (LTTNG_POLL_GETNB(&events
) == 0) {
2283 err
= 0; /* All is OK */
2290 /* From here, the event is a metadata wait fd */
2291 for (i
= 0; i
< nb_fd
; i
++) {
2292 health_code_update();
2294 revents
= LTTNG_POLL_GETEV(&events
, i
);
2295 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2298 /* No activity for this FD (poll implementation). */
2302 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2303 if (revents
& LPOLLIN
) {
2306 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2307 &stream
, sizeof(stream
));
2308 if (pipe_len
< sizeof(stream
)) {
2310 PERROR("read metadata stream");
2313 * Remove the pipe from the poll set and continue the loop
2314 * since their might be data to consume.
2316 lttng_poll_del(&events
,
2317 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2318 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2322 /* A NULL stream means that the state has changed. */
2323 if (stream
== NULL
) {
2324 /* Check for deleted streams. */
2325 validate_endpoint_status_metadata_stream(&events
);
2329 DBG("Adding metadata stream %d to poll set",
2332 /* Add metadata stream to the global poll events list */
2333 lttng_poll_add(&events
, stream
->wait_fd
,
2334 LPOLLIN
| LPOLLPRI
| LPOLLHUP
);
2335 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2336 DBG("Metadata thread pipe hung up");
2338 * Remove the pipe from the poll set and continue the loop
2339 * since their might be data to consume.
2341 lttng_poll_del(&events
,
2342 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2343 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2346 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2350 /* Handle other stream */
2356 uint64_t tmp_id
= (uint64_t) pollfd
;
2358 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2360 node
= lttng_ht_iter_get_node_u64(&iter
);
2363 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2366 if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2367 /* Get the data out of the metadata file descriptor */
2368 DBG("Metadata available on fd %d", pollfd
);
2369 assert(stream
->wait_fd
== pollfd
);
2372 health_code_update();
2374 len
= ctx
->on_buffer_ready(stream
, ctx
);
2376 * We don't check the return value here since if we get
2377 * a negative len, it means an error occurred thus we
2378 * simply remove it from the poll set and free the
2383 /* It's ok to have an unavailable sub-buffer */
2384 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2385 /* Clean up stream from consumer and free it. */
2386 lttng_poll_del(&events
, stream
->wait_fd
);
2387 consumer_del_metadata_stream(stream
, metadata_ht
);
2389 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2390 DBG("Metadata fd %d is hup|err.", pollfd
);
2391 if (!stream
->hangup_flush_done
2392 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2393 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2394 DBG("Attempting to flush and consume the UST buffers");
2395 lttng_ustconsumer_on_stream_hangup(stream
);
2397 /* We just flushed the stream now read it. */
2399 health_code_update();
2401 len
= ctx
->on_buffer_ready(stream
, ctx
);
2403 * We don't check the return value here since if we get
2404 * a negative len, it means an error occurred thus we
2405 * simply remove it from the poll set and free the
2411 lttng_poll_del(&events
, stream
->wait_fd
);
2413 * This call update the channel states, closes file descriptors
2414 * and securely free the stream.
2416 consumer_del_metadata_stream(stream
, metadata_ht
);
2418 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2422 /* Release RCU lock for the stream looked up */
2430 DBG("Metadata poll thread exiting");
2432 lttng_poll_clean(&events
);
2437 ERR("Health error occurred in %s", __func__
);
2439 health_unregister(health_consumerd
);
2440 rcu_unregister_thread();
2445 * This thread polls the fds in the set to consume the data and write
2446 * it to tracefile if necessary.
2448 void *consumer_thread_data_poll(void *data
)
2450 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2451 struct pollfd
*pollfd
= NULL
;
2452 /* local view of the streams */
2453 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2454 /* local view of consumer_data.fds_count */
2456 struct lttng_consumer_local_data
*ctx
= data
;
2459 rcu_register_thread();
2461 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2463 if (testpoint(consumerd_thread_data
)) {
2464 goto error_testpoint
;
2467 health_code_update();
2469 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
*));
2470 if (local_stream
== NULL
) {
2471 PERROR("local_stream malloc");
2476 health_code_update();
2482 * the fds set has been updated, we need to update our
2483 * local array as well
2485 pthread_mutex_lock(&consumer_data
.lock
);
2486 if (consumer_data
.need_update
) {
2491 local_stream
= NULL
;
2494 * Allocate for all fds +1 for the consumer_data_pipe and +1 for
2497 pollfd
= zmalloc((consumer_data
.stream_count
+ 2) * sizeof(struct pollfd
));
2498 if (pollfd
== NULL
) {
2499 PERROR("pollfd malloc");
2500 pthread_mutex_unlock(&consumer_data
.lock
);
2504 local_stream
= zmalloc((consumer_data
.stream_count
+ 2) *
2505 sizeof(struct lttng_consumer_stream
*));
2506 if (local_stream
== NULL
) {
2507 PERROR("local_stream malloc");
2508 pthread_mutex_unlock(&consumer_data
.lock
);
2511 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2514 ERR("Error in allocating pollfd or local_outfds");
2515 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2516 pthread_mutex_unlock(&consumer_data
.lock
);
2520 consumer_data
.need_update
= 0;
2522 pthread_mutex_unlock(&consumer_data
.lock
);
2524 /* No FDs and consumer_quit, consumer_cleanup the thread */
2525 if (nb_fd
== 0 && consumer_quit
== 1) {
2526 err
= 0; /* All is OK */
2529 /* poll on the array of fds */
2531 DBG("polling on %d fd", nb_fd
+ 2);
2532 health_poll_entry();
2533 num_rdy
= poll(pollfd
, nb_fd
+ 2, -1);
2535 DBG("poll num_rdy : %d", num_rdy
);
2536 if (num_rdy
== -1) {
2538 * Restart interrupted system call.
2540 if (errno
== EINTR
) {
2543 PERROR("Poll error");
2544 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2546 } else if (num_rdy
== 0) {
2547 DBG("Polling thread timed out");
2552 * If the consumer_data_pipe triggered poll go directly to the
2553 * beginning of the loop to update the array. We want to prioritize
2554 * array update over low-priority reads.
2556 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2557 ssize_t pipe_readlen
;
2559 DBG("consumer_data_pipe wake up");
2560 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2561 &new_stream
, sizeof(new_stream
));
2562 if (pipe_readlen
< sizeof(new_stream
)) {
2563 PERROR("Consumer data pipe");
2564 /* Continue so we can at least handle the current stream(s). */
2569 * If the stream is NULL, just ignore it. It's also possible that
2570 * the sessiond poll thread changed the consumer_quit state and is
2571 * waking us up to test it.
2573 if (new_stream
== NULL
) {
2574 validate_endpoint_status_data_stream();
2578 /* Continue to update the local streams and handle prio ones */
2582 /* Handle wakeup pipe. */
2583 if (pollfd
[nb_fd
+ 1].revents
& (POLLIN
| POLLPRI
)) {
2585 ssize_t pipe_readlen
;
2587 pipe_readlen
= lttng_pipe_read(ctx
->consumer_wakeup_pipe
, &dummy
,
2589 if (pipe_readlen
< 0) {
2590 PERROR("Consumer data wakeup pipe");
2592 /* We've been awakened to handle stream(s). */
2593 ctx
->has_wakeup
= 0;
2596 /* Take care of high priority channels first. */
2597 for (i
= 0; i
< nb_fd
; i
++) {
2598 health_code_update();
2600 if (local_stream
[i
] == NULL
) {
2603 if (pollfd
[i
].revents
& POLLPRI
) {
2604 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2606 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2607 /* it's ok to have an unavailable sub-buffer */
2608 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2609 /* Clean the stream and free it. */
2610 consumer_del_stream(local_stream
[i
], data_ht
);
2611 local_stream
[i
] = NULL
;
2612 } else if (len
> 0) {
2613 local_stream
[i
]->data_read
= 1;
2619 * If we read high prio channel in this loop, try again
2620 * for more high prio data.
2626 /* Take care of low priority channels. */
2627 for (i
= 0; i
< nb_fd
; i
++) {
2628 health_code_update();
2630 if (local_stream
[i
] == NULL
) {
2633 if ((pollfd
[i
].revents
& POLLIN
) ||
2634 local_stream
[i
]->hangup_flush_done
||
2635 local_stream
[i
]->has_data
) {
2636 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2637 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2638 /* it's ok to have an unavailable sub-buffer */
2639 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2640 /* Clean the stream and free it. */
2641 consumer_del_stream(local_stream
[i
], data_ht
);
2642 local_stream
[i
] = NULL
;
2643 } else if (len
> 0) {
2644 local_stream
[i
]->data_read
= 1;
2649 /* Handle hangup and errors */
2650 for (i
= 0; i
< nb_fd
; i
++) {
2651 health_code_update();
2653 if (local_stream
[i
] == NULL
) {
2656 if (!local_stream
[i
]->hangup_flush_done
2657 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2658 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2659 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2660 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2662 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2663 /* Attempt read again, for the data we just flushed. */
2664 local_stream
[i
]->data_read
= 1;
2667 * If the poll flag is HUP/ERR/NVAL and we have
2668 * read no data in this pass, we can remove the
2669 * stream from its hash table.
2671 if ((pollfd
[i
].revents
& POLLHUP
)) {
2672 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2673 if (!local_stream
[i
]->data_read
) {
2674 consumer_del_stream(local_stream
[i
], data_ht
);
2675 local_stream
[i
] = NULL
;
2678 } else if (pollfd
[i
].revents
& POLLERR
) {
2679 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2680 if (!local_stream
[i
]->data_read
) {
2681 consumer_del_stream(local_stream
[i
], data_ht
);
2682 local_stream
[i
] = NULL
;
2685 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2686 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2687 if (!local_stream
[i
]->data_read
) {
2688 consumer_del_stream(local_stream
[i
], data_ht
);
2689 local_stream
[i
] = NULL
;
2693 if (local_stream
[i
] != NULL
) {
2694 local_stream
[i
]->data_read
= 0;
2701 DBG("polling thread exiting");
2706 * Close the write side of the pipe so epoll_wait() in
2707 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2708 * read side of the pipe. If we close them both, epoll_wait strangely does
2709 * not return and could create a endless wait period if the pipe is the
2710 * only tracked fd in the poll set. The thread will take care of closing
2713 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2718 ERR("Health error occurred in %s", __func__
);
2720 health_unregister(health_consumerd
);
2722 rcu_unregister_thread();
2727 * Close wake-up end of each stream belonging to the channel. This will
2728 * allow the poll() on the stream read-side to detect when the
2729 * write-side (application) finally closes them.
2732 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2734 struct lttng_ht
*ht
;
2735 struct lttng_consumer_stream
*stream
;
2736 struct lttng_ht_iter iter
;
2738 ht
= consumer_data
.stream_per_chan_id_ht
;
2741 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2742 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2743 ht
->match_fct
, &channel
->key
,
2744 &iter
.iter
, stream
, node_channel_id
.node
) {
2746 * Protect against teardown with mutex.
2748 pthread_mutex_lock(&stream
->lock
);
2749 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2752 switch (consumer_data
.type
) {
2753 case LTTNG_CONSUMER_KERNEL
:
2755 case LTTNG_CONSUMER32_UST
:
2756 case LTTNG_CONSUMER64_UST
:
2757 if (stream
->metadata_flag
) {
2758 /* Safe and protected by the stream lock. */
2759 lttng_ustconsumer_close_metadata(stream
->chan
);
2762 * Note: a mutex is taken internally within
2763 * liblttng-ust-ctl to protect timer wakeup_fd
2764 * use from concurrent close.
2766 lttng_ustconsumer_close_stream_wakeup(stream
);
2770 ERR("Unknown consumer_data type");
2774 pthread_mutex_unlock(&stream
->lock
);
2779 static void destroy_channel_ht(struct lttng_ht
*ht
)
2781 struct lttng_ht_iter iter
;
2782 struct lttng_consumer_channel
*channel
;
2790 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2791 ret
= lttng_ht_del(ht
, &iter
);
2796 lttng_ht_destroy(ht
);
2800 * This thread polls the channel fds to detect when they are being
2801 * closed. It closes all related streams if the channel is detected as
2802 * closed. It is currently only used as a shim layer for UST because the
2803 * consumerd needs to keep the per-stream wakeup end of pipes open for
2806 void *consumer_thread_channel_poll(void *data
)
2808 int ret
, i
, pollfd
, err
= -1;
2809 uint32_t revents
, nb_fd
;
2810 struct lttng_consumer_channel
*chan
= NULL
;
2811 struct lttng_ht_iter iter
;
2812 struct lttng_ht_node_u64
*node
;
2813 struct lttng_poll_event events
;
2814 struct lttng_consumer_local_data
*ctx
= data
;
2815 struct lttng_ht
*channel_ht
;
2817 rcu_register_thread();
2819 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
2821 if (testpoint(consumerd_thread_channel
)) {
2822 goto error_testpoint
;
2825 health_code_update();
2827 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2829 /* ENOMEM at this point. Better to bail out. */
2833 DBG("Thread channel poll started");
2835 /* Size is set to 1 for the consumer_channel pipe */
2836 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2838 ERR("Poll set creation failed");
2842 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2848 DBG("Channel main loop started");
2852 health_code_update();
2853 DBG("Channel poll wait");
2854 health_poll_entry();
2855 ret
= lttng_poll_wait(&events
, -1);
2856 DBG("Channel poll return from wait with %d fd(s)",
2857 LTTNG_POLL_GETNB(&events
));
2859 DBG("Channel event caught in thread");
2861 if (errno
== EINTR
) {
2862 ERR("Poll EINTR caught");
2865 if (LTTNG_POLL_GETNB(&events
) == 0) {
2866 err
= 0; /* All is OK */
2873 /* From here, the event is a channel wait fd */
2874 for (i
= 0; i
< nb_fd
; i
++) {
2875 health_code_update();
2877 revents
= LTTNG_POLL_GETEV(&events
, i
);
2878 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2881 /* No activity for this FD (poll implementation). */
2885 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2886 if (revents
& LPOLLIN
) {
2887 enum consumer_channel_action action
;
2890 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
2893 ERR("Error reading channel pipe");
2895 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2900 case CONSUMER_CHANNEL_ADD
:
2901 DBG("Adding channel %d to poll set",
2904 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
2907 lttng_ht_add_unique_u64(channel_ht
,
2908 &chan
->wait_fd_node
);
2910 /* Add channel to the global poll events list */
2911 lttng_poll_add(&events
, chan
->wait_fd
,
2912 LPOLLERR
| LPOLLHUP
);
2914 case CONSUMER_CHANNEL_DEL
:
2917 * This command should never be called if the channel
2918 * has streams monitored by either the data or metadata
2919 * thread. The consumer only notify this thread with a
2920 * channel del. command if it receives a destroy
2921 * channel command from the session daemon that send it
2922 * if a command prior to the GET_CHANNEL failed.
2926 chan
= consumer_find_channel(key
);
2929 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
2932 lttng_poll_del(&events
, chan
->wait_fd
);
2933 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
2934 ret
= lttng_ht_del(channel_ht
, &iter
);
2937 switch (consumer_data
.type
) {
2938 case LTTNG_CONSUMER_KERNEL
:
2940 case LTTNG_CONSUMER32_UST
:
2941 case LTTNG_CONSUMER64_UST
:
2942 health_code_update();
2943 /* Destroy streams that might have been left in the stream list. */
2944 clean_channel_stream_list(chan
);
2947 ERR("Unknown consumer_data type");
2952 * Release our own refcount. Force channel deletion even if
2953 * streams were not initialized.
2955 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
2956 consumer_del_channel(chan
);
2961 case CONSUMER_CHANNEL_QUIT
:
2963 * Remove the pipe from the poll set and continue the loop
2964 * since their might be data to consume.
2966 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2969 ERR("Unknown action");
2972 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2973 DBG("Channel thread pipe hung up");
2975 * Remove the pipe from the poll set and continue the loop
2976 * since their might be data to consume.
2978 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2981 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2985 /* Handle other stream */
2991 uint64_t tmp_id
= (uint64_t) pollfd
;
2993 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
2995 node
= lttng_ht_iter_get_node_u64(&iter
);
2998 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
3001 /* Check for error event */
3002 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
3003 DBG("Channel fd %d is hup|err.", pollfd
);
3005 lttng_poll_del(&events
, chan
->wait_fd
);
3006 ret
= lttng_ht_del(channel_ht
, &iter
);
3010 * This will close the wait fd for each stream associated to
3011 * this channel AND monitored by the data/metadata thread thus
3012 * will be clean by the right thread.
3014 consumer_close_channel_streams(chan
);
3016 /* Release our own refcount */
3017 if (!uatomic_sub_return(&chan
->refcount
, 1)
3018 && !uatomic_read(&chan
->nb_init_stream_left
)) {
3019 consumer_del_channel(chan
);
3022 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3027 /* Release RCU lock for the channel looked up */
3035 lttng_poll_clean(&events
);
3037 destroy_channel_ht(channel_ht
);
3040 DBG("Channel poll thread exiting");
3043 ERR("Health error occurred in %s", __func__
);
3045 health_unregister(health_consumerd
);
3046 rcu_unregister_thread();
3050 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
3051 struct pollfd
*sockpoll
, int client_socket
)
3058 ret
= lttng_consumer_poll_socket(sockpoll
);
3062 DBG("Metadata connection on client_socket");
3064 /* Blocking call, waiting for transmission */
3065 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
3066 if (ctx
->consumer_metadata_socket
< 0) {
3067 WARN("On accept metadata");
3078 * This thread listens on the consumerd socket and receives the file
3079 * descriptors from the session daemon.
3081 void *consumer_thread_sessiond_poll(void *data
)
3083 int sock
= -1, client_socket
, ret
, err
= -1;
3085 * structure to poll for incoming data on communication socket avoids
3086 * making blocking sockets.
3088 struct pollfd consumer_sockpoll
[2];
3089 struct lttng_consumer_local_data
*ctx
= data
;
3091 rcu_register_thread();
3093 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
3095 if (testpoint(consumerd_thread_sessiond
)) {
3096 goto error_testpoint
;
3099 health_code_update();
3101 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
3102 unlink(ctx
->consumer_command_sock_path
);
3103 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
3104 if (client_socket
< 0) {
3105 ERR("Cannot create command socket");
3109 ret
= lttcomm_listen_unix_sock(client_socket
);
3114 DBG("Sending ready command to lttng-sessiond");
3115 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3116 /* return < 0 on error, but == 0 is not fatal */
3118 ERR("Error sending ready command to lttng-sessiond");
3122 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3123 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3124 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3125 consumer_sockpoll
[1].fd
= client_socket
;
3126 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3128 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3136 DBG("Connection on client_socket");
3138 /* Blocking call, waiting for transmission */
3139 sock
= lttcomm_accept_unix_sock(client_socket
);
3146 * Setup metadata socket which is the second socket connection on the
3147 * command unix socket.
3149 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3158 /* This socket is not useful anymore. */
3159 ret
= close(client_socket
);
3161 PERROR("close client_socket");
3165 /* update the polling structure to poll on the established socket */
3166 consumer_sockpoll
[1].fd
= sock
;
3167 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3170 health_code_update();
3172 health_poll_entry();
3173 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3182 DBG("Incoming command on sock");
3183 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3186 * This could simply be a session daemon quitting. Don't output
3189 DBG("Communication interrupted on command socket");
3193 if (consumer_quit
) {
3194 DBG("consumer_thread_receive_fds received quit from signal");
3195 err
= 0; /* All is OK */
3198 DBG("received command on sock");
3204 DBG("Consumer thread sessiond poll exiting");
3207 * Close metadata streams since the producer is the session daemon which
3210 * NOTE: for now, this only applies to the UST tracer.
3212 lttng_consumer_close_all_metadata();
3215 * when all fds have hung up, the polling thread
3221 * Notify the data poll thread to poll back again and test the
3222 * consumer_quit state that we just set so to quit gracefully.
3224 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3226 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3228 notify_health_quit_pipe(health_quit_pipe
);
3230 /* Cleaning up possibly open sockets. */
3234 PERROR("close sock sessiond poll");
3237 if (client_socket
>= 0) {
3238 ret
= close(client_socket
);
3240 PERROR("close client_socket sessiond poll");
3247 ERR("Health error occurred in %s", __func__
);
3249 health_unregister(health_consumerd
);
3251 rcu_unregister_thread();
3255 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3256 struct lttng_consumer_local_data
*ctx
)
3260 pthread_mutex_lock(&stream
->lock
);
3261 if (stream
->metadata_flag
) {
3262 pthread_mutex_lock(&stream
->metadata_rdv_lock
);
3265 switch (consumer_data
.type
) {
3266 case LTTNG_CONSUMER_KERNEL
:
3267 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
3269 case LTTNG_CONSUMER32_UST
:
3270 case LTTNG_CONSUMER64_UST
:
3271 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
3274 ERR("Unknown consumer_data type");
3280 if (stream
->metadata_flag
) {
3281 pthread_cond_broadcast(&stream
->metadata_rdv
);
3282 pthread_mutex_unlock(&stream
->metadata_rdv_lock
);
3284 pthread_mutex_unlock(&stream
->lock
);
3288 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3290 switch (consumer_data
.type
) {
3291 case LTTNG_CONSUMER_KERNEL
:
3292 return lttng_kconsumer_on_recv_stream(stream
);
3293 case LTTNG_CONSUMER32_UST
:
3294 case LTTNG_CONSUMER64_UST
:
3295 return lttng_ustconsumer_on_recv_stream(stream
);
3297 ERR("Unknown consumer_data type");
3304 * Allocate and set consumer data hash tables.
3306 int lttng_consumer_init(void)
3308 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3309 if (!consumer_data
.channel_ht
) {
3313 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3314 if (!consumer_data
.relayd_ht
) {
3318 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3319 if (!consumer_data
.stream_list_ht
) {
3323 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3324 if (!consumer_data
.stream_per_chan_id_ht
) {
3328 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3333 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3345 * Process the ADD_RELAYD command receive by a consumer.
3347 * This will create a relayd socket pair and add it to the relayd hash table.
3348 * The caller MUST acquire a RCU read side lock before calling it.
3350 int consumer_add_relayd_socket(uint64_t net_seq_idx
, int sock_type
,
3351 struct lttng_consumer_local_data
*ctx
, int sock
,
3352 struct pollfd
*consumer_sockpoll
,
3353 struct lttcomm_relayd_sock
*relayd_sock
, uint64_t sessiond_id
,
3354 uint64_t relayd_session_id
)
3356 int fd
= -1, ret
= -1, relayd_created
= 0;
3357 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3358 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3361 assert(relayd_sock
);
3363 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", net_seq_idx
);
3365 /* Get relayd reference if exists. */
3366 relayd
= consumer_find_relayd(net_seq_idx
);
3367 if (relayd
== NULL
) {
3368 assert(sock_type
== LTTNG_STREAM_CONTROL
);
3369 /* Not found. Allocate one. */
3370 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3371 if (relayd
== NULL
) {
3373 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3376 relayd
->sessiond_session_id
= sessiond_id
;
3381 * This code path MUST continue to the consumer send status message to
3382 * we can notify the session daemon and continue our work without
3383 * killing everything.
3387 * relayd key should never be found for control socket.
3389 assert(sock_type
!= LTTNG_STREAM_CONTROL
);
3392 /* First send a status message before receiving the fds. */
3393 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3395 /* Somehow, the session daemon is not responding anymore. */
3396 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3397 goto error_nosignal
;
3400 /* Poll on consumer socket. */
3401 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3403 /* Needing to exit in the middle of a command: error. */
3404 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3406 goto error_nosignal
;
3409 /* Get relayd socket from session daemon */
3410 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3411 if (ret
!= sizeof(fd
)) {
3413 fd
= -1; /* Just in case it gets set with an invalid value. */
3416 * Failing to receive FDs might indicate a major problem such as
3417 * reaching a fd limit during the receive where the kernel returns a
3418 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3419 * don't take any chances and stop everything.
3421 * XXX: Feature request #558 will fix that and avoid this possible
3422 * issue when reaching the fd limit.
3424 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3425 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3429 /* Copy socket information and received FD */
3430 switch (sock_type
) {
3431 case LTTNG_STREAM_CONTROL
:
3432 /* Copy received lttcomm socket */
3433 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3434 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3435 /* Handle create_sock error. */
3437 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3441 * Close the socket created internally by
3442 * lttcomm_create_sock, so we can replace it by the one
3443 * received from sessiond.
3445 if (close(relayd
->control_sock
.sock
.fd
)) {
3449 /* Assign new file descriptor */
3450 relayd
->control_sock
.sock
.fd
= fd
;
3451 fd
= -1; /* For error path */
3452 /* Assign version values. */
3453 relayd
->control_sock
.major
= relayd_sock
->major
;
3454 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3456 relayd
->relayd_session_id
= relayd_session_id
;
3459 case LTTNG_STREAM_DATA
:
3460 /* Copy received lttcomm socket */
3461 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3462 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3463 /* Handle create_sock error. */
3465 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3469 * Close the socket created internally by
3470 * lttcomm_create_sock, so we can replace it by the one
3471 * received from sessiond.
3473 if (close(relayd
->data_sock
.sock
.fd
)) {
3477 /* Assign new file descriptor */
3478 relayd
->data_sock
.sock
.fd
= fd
;
3479 fd
= -1; /* for eventual error paths */
3480 /* Assign version values. */
3481 relayd
->data_sock
.major
= relayd_sock
->major
;
3482 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3485 ERR("Unknown relayd socket type (%d)", sock_type
);
3487 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3491 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3492 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3493 relayd
->net_seq_idx
, fd
);
3495 /* We successfully added the socket. Send status back. */
3496 ret
= consumer_send_status_msg(sock
, ret_code
);
3498 /* Somehow, the session daemon is not responding anymore. */
3499 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3500 goto error_nosignal
;
3504 * Add relayd socket pair to consumer data hashtable. If object already
3505 * exists or on error, the function gracefully returns.
3513 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3514 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3518 /* Close received socket if valid. */
3521 PERROR("close received socket");
3525 if (relayd_created
) {
3533 * Try to lock the stream mutex.
3535 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
3537 static int stream_try_lock(struct lttng_consumer_stream
*stream
)
3544 * Try to lock the stream mutex. On failure, we know that the stream is
3545 * being used else where hence there is data still being extracted.
3547 ret
= pthread_mutex_trylock(&stream
->lock
);
3549 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
3561 * Search for a relayd associated to the session id and return the reference.
3563 * A rcu read side lock MUST be acquire before calling this function and locked
3564 * until the relayd object is no longer necessary.
3566 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3568 struct lttng_ht_iter iter
;
3569 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3571 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3572 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3575 * Check by sessiond id which is unique here where the relayd session
3576 * id might not be when having multiple relayd.
3578 if (relayd
->sessiond_session_id
== id
) {
3579 /* Found the relayd. There can be only one per id. */
3591 * Check if for a given session id there is still data needed to be extract
3594 * Return 1 if data is pending or else 0 meaning ready to be read.
3596 int consumer_data_pending(uint64_t id
)
3599 struct lttng_ht_iter iter
;
3600 struct lttng_ht
*ht
;
3601 struct lttng_consumer_stream
*stream
;
3602 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3603 int (*data_pending
)(struct lttng_consumer_stream
*);
3605 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3608 pthread_mutex_lock(&consumer_data
.lock
);
3610 switch (consumer_data
.type
) {
3611 case LTTNG_CONSUMER_KERNEL
:
3612 data_pending
= lttng_kconsumer_data_pending
;
3614 case LTTNG_CONSUMER32_UST
:
3615 case LTTNG_CONSUMER64_UST
:
3616 data_pending
= lttng_ustconsumer_data_pending
;
3619 ERR("Unknown consumer data type");
3623 /* Ease our life a bit */
3624 ht
= consumer_data
.stream_list_ht
;
3626 relayd
= find_relayd_by_session_id(id
);
3628 /* Send init command for data pending. */
3629 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3630 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3631 relayd
->relayd_session_id
);
3632 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3634 /* Communication error thus the relayd so no data pending. */
3635 goto data_not_pending
;
3639 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3640 ht
->hash_fct(&id
, lttng_ht_seed
),
3642 &iter
.iter
, stream
, node_session_id
.node
) {
3643 /* If this call fails, the stream is being used hence data pending. */
3644 ret
= stream_try_lock(stream
);
3650 * A removed node from the hash table indicates that the stream has
3651 * been deleted thus having a guarantee that the buffers are closed
3652 * on the consumer side. However, data can still be transmitted
3653 * over the network so don't skip the relayd check.
3655 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3657 /* Check the stream if there is data in the buffers. */
3658 ret
= data_pending(stream
);
3660 pthread_mutex_unlock(&stream
->lock
);
3667 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3668 if (stream
->metadata_flag
) {
3669 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3670 stream
->relayd_stream_id
);
3672 ret
= relayd_data_pending(&relayd
->control_sock
,
3673 stream
->relayd_stream_id
,
3674 stream
->next_net_seq_num
- 1);
3676 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3678 pthread_mutex_unlock(&stream
->lock
);
3682 pthread_mutex_unlock(&stream
->lock
);
3686 unsigned int is_data_inflight
= 0;
3688 /* Send init command for data pending. */
3689 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3690 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3691 relayd
->relayd_session_id
, &is_data_inflight
);
3692 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3694 goto data_not_pending
;
3696 if (is_data_inflight
) {
3702 * Finding _no_ node in the hash table and no inflight data means that the
3703 * stream(s) have been removed thus data is guaranteed to be available for
3704 * analysis from the trace files.
3708 /* Data is available to be read by a viewer. */
3709 pthread_mutex_unlock(&consumer_data
.lock
);
3714 /* Data is still being extracted from buffers. */
3715 pthread_mutex_unlock(&consumer_data
.lock
);
3721 * Send a ret code status message to the sessiond daemon.
3723 * Return the sendmsg() return value.
3725 int consumer_send_status_msg(int sock
, int ret_code
)
3727 struct lttcomm_consumer_status_msg msg
;
3729 memset(&msg
, 0, sizeof(msg
));
3730 msg
.ret_code
= ret_code
;
3732 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3736 * Send a channel status message to the sessiond daemon.
3738 * Return the sendmsg() return value.
3740 int consumer_send_status_channel(int sock
,
3741 struct lttng_consumer_channel
*channel
)
3743 struct lttcomm_consumer_status_channel msg
;
3747 memset(&msg
, 0, sizeof(msg
));
3749 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3751 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3752 msg
.key
= channel
->key
;
3753 msg
.stream_count
= channel
->streams
.count
;
3756 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3759 unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos
,
3760 unsigned long produced_pos
, uint64_t nb_packets_per_stream
,
3761 uint64_t max_sb_size
)
3763 unsigned long start_pos
;
3765 if (!nb_packets_per_stream
) {
3766 return consumed_pos
; /* Grab everything */
3768 start_pos
= produced_pos
- offset_align_floor(produced_pos
, max_sb_size
);
3769 start_pos
-= max_sb_size
* nb_packets_per_stream
;
3770 if ((long) (start_pos
- consumed_pos
) < 0) {
3771 return consumed_pos
; /* Grab everything */