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_file
= NULL
;
574 stream
->last_sequence_number
= -1ULL;
575 pthread_mutex_init(&stream
->lock
, NULL
);
576 pthread_mutex_init(&stream
->metadata_timer_lock
, NULL
);
578 /* If channel is the metadata, flag this stream as metadata. */
579 if (type
== CONSUMER_CHANNEL_TYPE_METADATA
) {
580 stream
->metadata_flag
= 1;
581 /* Metadata is flat out. */
582 strncpy(stream
->name
, DEFAULT_METADATA_NAME
, sizeof(stream
->name
));
583 /* Live rendez-vous point. */
584 pthread_cond_init(&stream
->metadata_rdv
, NULL
);
585 pthread_mutex_init(&stream
->metadata_rdv_lock
, NULL
);
587 /* Format stream name to <channel_name>_<cpu_number> */
588 ret
= snprintf(stream
->name
, sizeof(stream
->name
), "%s_%d",
591 PERROR("snprintf stream name");
596 /* Key is always the wait_fd for streams. */
597 lttng_ht_node_init_u64(&stream
->node
, stream
->key
);
599 /* Init node per channel id key */
600 lttng_ht_node_init_u64(&stream
->node_channel_id
, channel_key
);
602 /* Init session id node with the stream session id */
603 lttng_ht_node_init_u64(&stream
->node_session_id
, stream
->session_id
);
605 DBG3("Allocated stream %s (key %" PRIu64
", chan_key %" PRIu64
606 " relayd_id %" PRIu64
", session_id %" PRIu64
,
607 stream
->name
, stream
->key
, channel_key
,
608 stream
->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_file
) {
1629 lttng_index_file_put(stream
->index_file
);
1630 stream
->index_file
= lttng_index_file_create(stream
->chan
->pathname
,
1631 stream
->name
, stream
->uid
, stream
->gid
,
1632 stream
->chan
->tracefile_size
,
1633 stream
->tracefile_count_current
,
1634 CTF_INDEX_MAJOR
, CTF_INDEX_MINOR
);
1635 if (!stream
->index_file
) {
1640 /* Reset current size because we just perform a rotation. */
1641 stream
->tracefile_size_current
= 0;
1642 stream
->out_fd_offset
= 0;
1645 stream
->tracefile_size_current
+= len
;
1647 index
->offset
= htobe64(stream
->out_fd_offset
);
1652 * This call guarantee that len or less is returned. It's impossible to
1653 * receive a ret value that is bigger than len.
1655 ret
= lttng_write(outfd
, mmap_base
+ mmap_offset
, len
);
1656 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, len
);
1657 if (ret
< 0 || ((size_t) ret
!= len
)) {
1659 * Report error to caller if nothing was written else at least send the
1667 /* Socket operation failed. We consider the relayd dead */
1668 if (errno
== EPIPE
|| errno
== EINVAL
|| errno
== EBADF
) {
1670 * This is possible if the fd is closed on the other side
1671 * (outfd) or any write problem. It can be verbose a bit for a
1672 * normal execution if for instance the relayd is stopped
1673 * abruptly. This can happen so set this to a DBG statement.
1675 DBG("Consumer mmap write detected relayd hang up");
1677 /* Unhandled error, print it and stop function right now. */
1678 PERROR("Error in write mmap (ret %zd != len %lu)", ret
, len
);
1682 stream
->output_written
+= ret
;
1684 /* This call is useless on a socket so better save a syscall. */
1686 /* This won't block, but will start writeout asynchronously */
1687 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, len
,
1688 SYNC_FILE_RANGE_WRITE
);
1689 stream
->out_fd_offset
+= len
;
1690 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1695 * This is a special case that the relayd has closed its socket. Let's
1696 * cleanup the relayd object and all associated streams.
1698 if (relayd
&& relayd_hang_up
) {
1699 cleanup_relayd(relayd
, ctx
);
1703 /* Unlock only if ctrl socket used */
1704 if (relayd
&& stream
->metadata_flag
) {
1705 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1713 * Splice the data from the ring buffer to the tracefile.
1715 * It must be called with the stream lock held.
1717 * Returns the number of bytes spliced.
1719 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1720 struct lttng_consumer_local_data
*ctx
,
1721 struct lttng_consumer_stream
*stream
, unsigned long len
,
1722 unsigned long padding
,
1723 struct ctf_packet_index
*index
)
1725 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1727 off_t orig_offset
= stream
->out_fd_offset
;
1728 int fd
= stream
->wait_fd
;
1729 /* Default is on the disk */
1730 int outfd
= stream
->out_fd
;
1731 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1733 unsigned int relayd_hang_up
= 0;
1735 switch (consumer_data
.type
) {
1736 case LTTNG_CONSUMER_KERNEL
:
1738 case LTTNG_CONSUMER32_UST
:
1739 case LTTNG_CONSUMER64_UST
:
1740 /* Not supported for user space tracing */
1743 ERR("Unknown consumer_data type");
1747 /* RCU lock for the relayd pointer */
1750 /* Flag that the current stream if set for network streaming. */
1751 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1752 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1753 if (relayd
== NULL
) {
1758 splice_pipe
= stream
->splice_pipe
;
1760 /* Write metadata stream id before payload */
1762 unsigned long total_len
= len
;
1764 if (stream
->metadata_flag
) {
1766 * Lock the control socket for the complete duration of the function
1767 * since from this point on we will use the socket.
1769 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1771 if (stream
->reset_metadata_flag
) {
1772 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1773 stream
->relayd_stream_id
,
1774 stream
->metadata_version
);
1779 stream
->reset_metadata_flag
= 0;
1781 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
, relayd
,
1789 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1792 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1798 /* Use the returned socket. */
1801 /* No streaming, we have to set the len with the full padding */
1804 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1805 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1807 ERR("Reset metadata file");
1810 stream
->reset_metadata_flag
= 0;
1813 * Check if we need to change the tracefile before writing the packet.
1815 if (stream
->chan
->tracefile_size
> 0 &&
1816 (stream
->tracefile_size_current
+ len
) >
1817 stream
->chan
->tracefile_size
) {
1818 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1819 stream
->name
, stream
->chan
->tracefile_size
,
1820 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1821 stream
->out_fd
, &(stream
->tracefile_count_current
),
1825 ERR("Rotating output file");
1828 outfd
= stream
->out_fd
;
1830 if (stream
->index_file
) {
1831 lttng_index_file_put(stream
->index_file
);
1832 stream
->index_file
= lttng_index_file_create(stream
->chan
->pathname
,
1833 stream
->name
, stream
->uid
, stream
->gid
,
1834 stream
->chan
->tracefile_size
,
1835 stream
->tracefile_count_current
,
1836 CTF_INDEX_MAJOR
, CTF_INDEX_MINOR
);
1837 if (!stream
->index_file
) {
1842 /* Reset current size because we just perform a rotation. */
1843 stream
->tracefile_size_current
= 0;
1844 stream
->out_fd_offset
= 0;
1847 stream
->tracefile_size_current
+= len
;
1848 index
->offset
= htobe64(stream
->out_fd_offset
);
1852 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1853 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1854 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1855 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1856 DBG("splice chan to pipe, ret %zd", ret_splice
);
1857 if (ret_splice
< 0) {
1860 PERROR("Error in relay splice");
1864 /* Handle stream on the relayd if the output is on the network */
1865 if (relayd
&& stream
->metadata_flag
) {
1866 size_t metadata_payload_size
=
1867 sizeof(struct lttcomm_relayd_metadata_payload
);
1869 /* Update counter to fit the spliced data */
1870 ret_splice
+= metadata_payload_size
;
1871 len
+= metadata_payload_size
;
1873 * We do this so the return value can match the len passed as
1874 * argument to this function.
1876 written
-= metadata_payload_size
;
1879 /* Splice data out */
1880 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1881 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1882 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
1884 if (ret_splice
< 0) {
1889 } else if (ret_splice
> len
) {
1891 * We don't expect this code path to be executed but you never know
1892 * so this is an extra protection agains a buggy splice().
1895 written
+= ret_splice
;
1896 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice
,
1900 /* All good, update current len and continue. */
1904 /* This call is useless on a socket so better save a syscall. */
1906 /* This won't block, but will start writeout asynchronously */
1907 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1908 SYNC_FILE_RANGE_WRITE
);
1909 stream
->out_fd_offset
+= ret_splice
;
1911 stream
->output_written
+= ret_splice
;
1912 written
+= ret_splice
;
1915 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1921 * This is a special case that the relayd has closed its socket. Let's
1922 * cleanup the relayd object and all associated streams.
1924 if (relayd
&& relayd_hang_up
) {
1925 cleanup_relayd(relayd
, ctx
);
1926 /* Skip splice error so the consumer does not fail */
1931 /* send the appropriate error description to sessiond */
1934 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1937 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1940 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
1945 if (relayd
&& stream
->metadata_flag
) {
1946 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1954 * Take a snapshot for a specific fd
1956 * Returns 0 on success, < 0 on error
1958 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
1960 switch (consumer_data
.type
) {
1961 case LTTNG_CONSUMER_KERNEL
:
1962 return lttng_kconsumer_take_snapshot(stream
);
1963 case LTTNG_CONSUMER32_UST
:
1964 case LTTNG_CONSUMER64_UST
:
1965 return lttng_ustconsumer_take_snapshot(stream
);
1967 ERR("Unknown consumer_data type");
1974 * Get the produced position
1976 * Returns 0 on success, < 0 on error
1978 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
1981 switch (consumer_data
.type
) {
1982 case LTTNG_CONSUMER_KERNEL
:
1983 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
1984 case LTTNG_CONSUMER32_UST
:
1985 case LTTNG_CONSUMER64_UST
:
1986 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
1988 ERR("Unknown consumer_data type");
1994 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
1995 int sock
, struct pollfd
*consumer_sockpoll
)
1997 switch (consumer_data
.type
) {
1998 case LTTNG_CONSUMER_KERNEL
:
1999 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2000 case LTTNG_CONSUMER32_UST
:
2001 case LTTNG_CONSUMER64_UST
:
2002 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2004 ERR("Unknown consumer_data type");
2010 void lttng_consumer_close_all_metadata(void)
2012 switch (consumer_data
.type
) {
2013 case LTTNG_CONSUMER_KERNEL
:
2015 * The Kernel consumer has a different metadata scheme so we don't
2016 * close anything because the stream will be closed by the session
2020 case LTTNG_CONSUMER32_UST
:
2021 case LTTNG_CONSUMER64_UST
:
2023 * Close all metadata streams. The metadata hash table is passed and
2024 * this call iterates over it by closing all wakeup fd. This is safe
2025 * because at this point we are sure that the metadata producer is
2026 * either dead or blocked.
2028 lttng_ustconsumer_close_all_metadata(metadata_ht
);
2031 ERR("Unknown consumer_data type");
2037 * Clean up a metadata stream and free its memory.
2039 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
2040 struct lttng_ht
*ht
)
2042 struct lttng_consumer_channel
*free_chan
= NULL
;
2046 * This call should NEVER receive regular stream. It must always be
2047 * metadata stream and this is crucial for data structure synchronization.
2049 assert(stream
->metadata_flag
);
2051 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
2053 pthread_mutex_lock(&consumer_data
.lock
);
2054 pthread_mutex_lock(&stream
->chan
->lock
);
2055 pthread_mutex_lock(&stream
->lock
);
2057 /* Remove any reference to that stream. */
2058 consumer_stream_delete(stream
, ht
);
2060 /* Close down everything including the relayd if one. */
2061 consumer_stream_close(stream
);
2062 /* Destroy tracer buffers of the stream. */
2063 consumer_stream_destroy_buffers(stream
);
2065 /* Atomically decrement channel refcount since other threads can use it. */
2066 if (!uatomic_sub_return(&stream
->chan
->refcount
, 1)
2067 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
2068 /* Go for channel deletion! */
2069 free_chan
= stream
->chan
;
2073 * Nullify the stream reference so it is not used after deletion. The
2074 * channel lock MUST be acquired before being able to check for a NULL
2077 stream
->chan
->metadata_stream
= NULL
;
2079 pthread_mutex_unlock(&stream
->lock
);
2080 pthread_mutex_unlock(&stream
->chan
->lock
);
2081 pthread_mutex_unlock(&consumer_data
.lock
);
2084 consumer_del_channel(free_chan
);
2087 consumer_stream_free(stream
);
2091 * Action done with the metadata stream when adding it to the consumer internal
2092 * data structures to handle it.
2094 int consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
2096 struct lttng_ht
*ht
= metadata_ht
;
2098 struct lttng_ht_iter iter
;
2099 struct lttng_ht_node_u64
*node
;
2104 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2106 pthread_mutex_lock(&consumer_data
.lock
);
2107 pthread_mutex_lock(&stream
->chan
->lock
);
2108 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2109 pthread_mutex_lock(&stream
->lock
);
2112 * From here, refcounts are updated so be _careful_ when returning an error
2119 * Lookup the stream just to make sure it does not exist in our internal
2120 * state. This should NEVER happen.
2122 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2123 node
= lttng_ht_iter_get_node_u64(&iter
);
2127 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2128 * in terms of destroying the associated channel, because the action that
2129 * causes the count to become 0 also causes a stream to be added. The
2130 * channel deletion will thus be triggered by the following removal of this
2133 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2134 /* Increment refcount before decrementing nb_init_stream_left */
2136 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2139 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2141 lttng_ht_add_unique_u64(consumer_data
.stream_per_chan_id_ht
,
2142 &stream
->node_channel_id
);
2145 * Add stream to the stream_list_ht of the consumer data. No need to steal
2146 * the key since the HT does not use it and we allow to add redundant keys
2149 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2153 pthread_mutex_unlock(&stream
->lock
);
2154 pthread_mutex_unlock(&stream
->chan
->lock
);
2155 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2156 pthread_mutex_unlock(&consumer_data
.lock
);
2161 * Delete data stream that are flagged for deletion (endpoint_status).
2163 static void validate_endpoint_status_data_stream(void)
2165 struct lttng_ht_iter iter
;
2166 struct lttng_consumer_stream
*stream
;
2168 DBG("Consumer delete flagged data stream");
2171 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2172 /* Validate delete flag of the stream */
2173 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2176 /* Delete it right now */
2177 consumer_del_stream(stream
, data_ht
);
2183 * Delete metadata stream that are flagged for deletion (endpoint_status).
2185 static void validate_endpoint_status_metadata_stream(
2186 struct lttng_poll_event
*pollset
)
2188 struct lttng_ht_iter iter
;
2189 struct lttng_consumer_stream
*stream
;
2191 DBG("Consumer delete flagged metadata stream");
2196 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2197 /* Validate delete flag of the stream */
2198 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2202 * Remove from pollset so the metadata thread can continue without
2203 * blocking on a deleted stream.
2205 lttng_poll_del(pollset
, stream
->wait_fd
);
2207 /* Delete it right now */
2208 consumer_del_metadata_stream(stream
, metadata_ht
);
2214 * Thread polls on metadata file descriptor and write them on disk or on the
2217 void *consumer_thread_metadata_poll(void *data
)
2219 int ret
, i
, pollfd
, err
= -1;
2220 uint32_t revents
, nb_fd
;
2221 struct lttng_consumer_stream
*stream
= NULL
;
2222 struct lttng_ht_iter iter
;
2223 struct lttng_ht_node_u64
*node
;
2224 struct lttng_poll_event events
;
2225 struct lttng_consumer_local_data
*ctx
= data
;
2228 rcu_register_thread();
2230 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2232 if (testpoint(consumerd_thread_metadata
)) {
2233 goto error_testpoint
;
2236 health_code_update();
2238 DBG("Thread metadata poll started");
2240 /* Size is set to 1 for the consumer_metadata pipe */
2241 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2243 ERR("Poll set creation failed");
2247 ret
= lttng_poll_add(&events
,
2248 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2254 DBG("Metadata main loop started");
2258 health_code_update();
2259 health_poll_entry();
2260 DBG("Metadata poll wait");
2261 ret
= lttng_poll_wait(&events
, -1);
2262 DBG("Metadata poll return from wait with %d fd(s)",
2263 LTTNG_POLL_GETNB(&events
));
2265 DBG("Metadata event caught in thread");
2267 if (errno
== EINTR
) {
2268 ERR("Poll EINTR caught");
2271 if (LTTNG_POLL_GETNB(&events
) == 0) {
2272 err
= 0; /* All is OK */
2279 /* From here, the event is a metadata wait fd */
2280 for (i
= 0; i
< nb_fd
; i
++) {
2281 health_code_update();
2283 revents
= LTTNG_POLL_GETEV(&events
, i
);
2284 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2287 /* No activity for this FD (poll implementation). */
2291 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2292 if (revents
& LPOLLIN
) {
2295 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2296 &stream
, sizeof(stream
));
2297 if (pipe_len
< sizeof(stream
)) {
2299 PERROR("read metadata stream");
2302 * Remove the pipe from the poll set and continue the loop
2303 * since their might be data to consume.
2305 lttng_poll_del(&events
,
2306 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2307 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2311 /* A NULL stream means that the state has changed. */
2312 if (stream
== NULL
) {
2313 /* Check for deleted streams. */
2314 validate_endpoint_status_metadata_stream(&events
);
2318 DBG("Adding metadata stream %d to poll set",
2321 /* Add metadata stream to the global poll events list */
2322 lttng_poll_add(&events
, stream
->wait_fd
,
2323 LPOLLIN
| LPOLLPRI
| LPOLLHUP
);
2324 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2325 DBG("Metadata thread pipe hung up");
2327 * Remove the pipe from the poll set and continue the loop
2328 * since their might be data to consume.
2330 lttng_poll_del(&events
,
2331 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2332 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2335 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2339 /* Handle other stream */
2345 uint64_t tmp_id
= (uint64_t) pollfd
;
2347 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2349 node
= lttng_ht_iter_get_node_u64(&iter
);
2352 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2355 if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2356 /* Get the data out of the metadata file descriptor */
2357 DBG("Metadata available on fd %d", pollfd
);
2358 assert(stream
->wait_fd
== pollfd
);
2361 health_code_update();
2363 len
= ctx
->on_buffer_ready(stream
, ctx
);
2365 * We don't check the return value here since if we get
2366 * a negative len, it means an error occurred thus we
2367 * simply remove it from the poll set and free the
2372 /* It's ok to have an unavailable sub-buffer */
2373 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2374 /* Clean up stream from consumer and free it. */
2375 lttng_poll_del(&events
, stream
->wait_fd
);
2376 consumer_del_metadata_stream(stream
, metadata_ht
);
2378 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2379 DBG("Metadata fd %d is hup|err.", pollfd
);
2380 if (!stream
->hangup_flush_done
2381 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2382 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2383 DBG("Attempting to flush and consume the UST buffers");
2384 lttng_ustconsumer_on_stream_hangup(stream
);
2386 /* We just flushed the stream now read it. */
2388 health_code_update();
2390 len
= ctx
->on_buffer_ready(stream
, ctx
);
2392 * We don't check the return value here since if we get
2393 * a negative len, it means an error occurred thus we
2394 * simply remove it from the poll set and free the
2400 lttng_poll_del(&events
, stream
->wait_fd
);
2402 * This call update the channel states, closes file descriptors
2403 * and securely free the stream.
2405 consumer_del_metadata_stream(stream
, metadata_ht
);
2407 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2411 /* Release RCU lock for the stream looked up */
2419 DBG("Metadata poll thread exiting");
2421 lttng_poll_clean(&events
);
2426 ERR("Health error occurred in %s", __func__
);
2428 health_unregister(health_consumerd
);
2429 rcu_unregister_thread();
2434 * This thread polls the fds in the set to consume the data and write
2435 * it to tracefile if necessary.
2437 void *consumer_thread_data_poll(void *data
)
2439 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2440 struct pollfd
*pollfd
= NULL
;
2441 /* local view of the streams */
2442 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2443 /* local view of consumer_data.fds_count */
2445 struct lttng_consumer_local_data
*ctx
= data
;
2448 rcu_register_thread();
2450 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2452 if (testpoint(consumerd_thread_data
)) {
2453 goto error_testpoint
;
2456 health_code_update();
2458 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
*));
2459 if (local_stream
== NULL
) {
2460 PERROR("local_stream malloc");
2465 health_code_update();
2471 * the fds set has been updated, we need to update our
2472 * local array as well
2474 pthread_mutex_lock(&consumer_data
.lock
);
2475 if (consumer_data
.need_update
) {
2480 local_stream
= NULL
;
2483 * Allocate for all fds +1 for the consumer_data_pipe and +1 for
2486 pollfd
= zmalloc((consumer_data
.stream_count
+ 2) * sizeof(struct pollfd
));
2487 if (pollfd
== NULL
) {
2488 PERROR("pollfd malloc");
2489 pthread_mutex_unlock(&consumer_data
.lock
);
2493 local_stream
= zmalloc((consumer_data
.stream_count
+ 2) *
2494 sizeof(struct lttng_consumer_stream
*));
2495 if (local_stream
== NULL
) {
2496 PERROR("local_stream malloc");
2497 pthread_mutex_unlock(&consumer_data
.lock
);
2500 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2503 ERR("Error in allocating pollfd or local_outfds");
2504 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2505 pthread_mutex_unlock(&consumer_data
.lock
);
2509 consumer_data
.need_update
= 0;
2511 pthread_mutex_unlock(&consumer_data
.lock
);
2513 /* No FDs and consumer_quit, consumer_cleanup the thread */
2514 if (nb_fd
== 0 && consumer_quit
== 1) {
2515 err
= 0; /* All is OK */
2518 /* poll on the array of fds */
2520 DBG("polling on %d fd", nb_fd
+ 2);
2521 health_poll_entry();
2522 num_rdy
= poll(pollfd
, nb_fd
+ 2, -1);
2524 DBG("poll num_rdy : %d", num_rdy
);
2525 if (num_rdy
== -1) {
2527 * Restart interrupted system call.
2529 if (errno
== EINTR
) {
2532 PERROR("Poll error");
2533 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2535 } else if (num_rdy
== 0) {
2536 DBG("Polling thread timed out");
2541 * If the consumer_data_pipe triggered poll go directly to the
2542 * beginning of the loop to update the array. We want to prioritize
2543 * array update over low-priority reads.
2545 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2546 ssize_t pipe_readlen
;
2548 DBG("consumer_data_pipe wake up");
2549 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2550 &new_stream
, sizeof(new_stream
));
2551 if (pipe_readlen
< sizeof(new_stream
)) {
2552 PERROR("Consumer data pipe");
2553 /* Continue so we can at least handle the current stream(s). */
2558 * If the stream is NULL, just ignore it. It's also possible that
2559 * the sessiond poll thread changed the consumer_quit state and is
2560 * waking us up to test it.
2562 if (new_stream
== NULL
) {
2563 validate_endpoint_status_data_stream();
2567 /* Continue to update the local streams and handle prio ones */
2571 /* Handle wakeup pipe. */
2572 if (pollfd
[nb_fd
+ 1].revents
& (POLLIN
| POLLPRI
)) {
2574 ssize_t pipe_readlen
;
2576 pipe_readlen
= lttng_pipe_read(ctx
->consumer_wakeup_pipe
, &dummy
,
2578 if (pipe_readlen
< 0) {
2579 PERROR("Consumer data wakeup pipe");
2581 /* We've been awakened to handle stream(s). */
2582 ctx
->has_wakeup
= 0;
2585 /* Take care of high priority channels first. */
2586 for (i
= 0; i
< nb_fd
; i
++) {
2587 health_code_update();
2589 if (local_stream
[i
] == NULL
) {
2592 if (pollfd
[i
].revents
& POLLPRI
) {
2593 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2595 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2596 /* it's ok to have an unavailable sub-buffer */
2597 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2598 /* Clean the stream and free it. */
2599 consumer_del_stream(local_stream
[i
], data_ht
);
2600 local_stream
[i
] = NULL
;
2601 } else if (len
> 0) {
2602 local_stream
[i
]->data_read
= 1;
2608 * If we read high prio channel in this loop, try again
2609 * for more high prio data.
2615 /* Take care of low priority channels. */
2616 for (i
= 0; i
< nb_fd
; i
++) {
2617 health_code_update();
2619 if (local_stream
[i
] == NULL
) {
2622 if ((pollfd
[i
].revents
& POLLIN
) ||
2623 local_stream
[i
]->hangup_flush_done
||
2624 local_stream
[i
]->has_data
) {
2625 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2626 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2627 /* it's ok to have an unavailable sub-buffer */
2628 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2629 /* Clean the stream and free it. */
2630 consumer_del_stream(local_stream
[i
], data_ht
);
2631 local_stream
[i
] = NULL
;
2632 } else if (len
> 0) {
2633 local_stream
[i
]->data_read
= 1;
2638 /* Handle hangup and errors */
2639 for (i
= 0; i
< nb_fd
; i
++) {
2640 health_code_update();
2642 if (local_stream
[i
] == NULL
) {
2645 if (!local_stream
[i
]->hangup_flush_done
2646 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2647 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2648 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2649 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2651 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2652 /* Attempt read again, for the data we just flushed. */
2653 local_stream
[i
]->data_read
= 1;
2656 * If the poll flag is HUP/ERR/NVAL and we have
2657 * read no data in this pass, we can remove the
2658 * stream from its hash table.
2660 if ((pollfd
[i
].revents
& POLLHUP
)) {
2661 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2662 if (!local_stream
[i
]->data_read
) {
2663 consumer_del_stream(local_stream
[i
], data_ht
);
2664 local_stream
[i
] = NULL
;
2667 } else if (pollfd
[i
].revents
& POLLERR
) {
2668 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2669 if (!local_stream
[i
]->data_read
) {
2670 consumer_del_stream(local_stream
[i
], data_ht
);
2671 local_stream
[i
] = NULL
;
2674 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2675 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2676 if (!local_stream
[i
]->data_read
) {
2677 consumer_del_stream(local_stream
[i
], data_ht
);
2678 local_stream
[i
] = NULL
;
2682 if (local_stream
[i
] != NULL
) {
2683 local_stream
[i
]->data_read
= 0;
2690 DBG("polling thread exiting");
2695 * Close the write side of the pipe so epoll_wait() in
2696 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2697 * read side of the pipe. If we close them both, epoll_wait strangely does
2698 * not return and could create a endless wait period if the pipe is the
2699 * only tracked fd in the poll set. The thread will take care of closing
2702 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2707 ERR("Health error occurred in %s", __func__
);
2709 health_unregister(health_consumerd
);
2711 rcu_unregister_thread();
2716 * Close wake-up end of each stream belonging to the channel. This will
2717 * allow the poll() on the stream read-side to detect when the
2718 * write-side (application) finally closes them.
2721 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2723 struct lttng_ht
*ht
;
2724 struct lttng_consumer_stream
*stream
;
2725 struct lttng_ht_iter iter
;
2727 ht
= consumer_data
.stream_per_chan_id_ht
;
2730 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2731 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2732 ht
->match_fct
, &channel
->key
,
2733 &iter
.iter
, stream
, node_channel_id
.node
) {
2735 * Protect against teardown with mutex.
2737 pthread_mutex_lock(&stream
->lock
);
2738 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2741 switch (consumer_data
.type
) {
2742 case LTTNG_CONSUMER_KERNEL
:
2744 case LTTNG_CONSUMER32_UST
:
2745 case LTTNG_CONSUMER64_UST
:
2746 if (stream
->metadata_flag
) {
2747 /* Safe and protected by the stream lock. */
2748 lttng_ustconsumer_close_metadata(stream
->chan
);
2751 * Note: a mutex is taken internally within
2752 * liblttng-ust-ctl to protect timer wakeup_fd
2753 * use from concurrent close.
2755 lttng_ustconsumer_close_stream_wakeup(stream
);
2759 ERR("Unknown consumer_data type");
2763 pthread_mutex_unlock(&stream
->lock
);
2768 static void destroy_channel_ht(struct lttng_ht
*ht
)
2770 struct lttng_ht_iter iter
;
2771 struct lttng_consumer_channel
*channel
;
2779 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2780 ret
= lttng_ht_del(ht
, &iter
);
2785 lttng_ht_destroy(ht
);
2789 * This thread polls the channel fds to detect when they are being
2790 * closed. It closes all related streams if the channel is detected as
2791 * closed. It is currently only used as a shim layer for UST because the
2792 * consumerd needs to keep the per-stream wakeup end of pipes open for
2795 void *consumer_thread_channel_poll(void *data
)
2797 int ret
, i
, pollfd
, err
= -1;
2798 uint32_t revents
, nb_fd
;
2799 struct lttng_consumer_channel
*chan
= NULL
;
2800 struct lttng_ht_iter iter
;
2801 struct lttng_ht_node_u64
*node
;
2802 struct lttng_poll_event events
;
2803 struct lttng_consumer_local_data
*ctx
= data
;
2804 struct lttng_ht
*channel_ht
;
2806 rcu_register_thread();
2808 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
2810 if (testpoint(consumerd_thread_channel
)) {
2811 goto error_testpoint
;
2814 health_code_update();
2816 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2818 /* ENOMEM at this point. Better to bail out. */
2822 DBG("Thread channel poll started");
2824 /* Size is set to 1 for the consumer_channel pipe */
2825 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2827 ERR("Poll set creation failed");
2831 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2837 DBG("Channel main loop started");
2841 health_code_update();
2842 DBG("Channel poll wait");
2843 health_poll_entry();
2844 ret
= lttng_poll_wait(&events
, -1);
2845 DBG("Channel poll return from wait with %d fd(s)",
2846 LTTNG_POLL_GETNB(&events
));
2848 DBG("Channel event caught in thread");
2850 if (errno
== EINTR
) {
2851 ERR("Poll EINTR caught");
2854 if (LTTNG_POLL_GETNB(&events
) == 0) {
2855 err
= 0; /* All is OK */
2862 /* From here, the event is a channel wait fd */
2863 for (i
= 0; i
< nb_fd
; i
++) {
2864 health_code_update();
2866 revents
= LTTNG_POLL_GETEV(&events
, i
);
2867 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2870 /* No activity for this FD (poll implementation). */
2874 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2875 if (revents
& LPOLLIN
) {
2876 enum consumer_channel_action action
;
2879 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
2882 ERR("Error reading channel pipe");
2884 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2889 case CONSUMER_CHANNEL_ADD
:
2890 DBG("Adding channel %d to poll set",
2893 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
2896 lttng_ht_add_unique_u64(channel_ht
,
2897 &chan
->wait_fd_node
);
2899 /* Add channel to the global poll events list */
2900 lttng_poll_add(&events
, chan
->wait_fd
,
2901 LPOLLERR
| LPOLLHUP
);
2903 case CONSUMER_CHANNEL_DEL
:
2906 * This command should never be called if the channel
2907 * has streams monitored by either the data or metadata
2908 * thread. The consumer only notify this thread with a
2909 * channel del. command if it receives a destroy
2910 * channel command from the session daemon that send it
2911 * if a command prior to the GET_CHANNEL failed.
2915 chan
= consumer_find_channel(key
);
2918 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
2921 lttng_poll_del(&events
, chan
->wait_fd
);
2922 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
2923 ret
= lttng_ht_del(channel_ht
, &iter
);
2926 switch (consumer_data
.type
) {
2927 case LTTNG_CONSUMER_KERNEL
:
2929 case LTTNG_CONSUMER32_UST
:
2930 case LTTNG_CONSUMER64_UST
:
2931 health_code_update();
2932 /* Destroy streams that might have been left in the stream list. */
2933 clean_channel_stream_list(chan
);
2936 ERR("Unknown consumer_data type");
2941 * Release our own refcount. Force channel deletion even if
2942 * streams were not initialized.
2944 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
2945 consumer_del_channel(chan
);
2950 case CONSUMER_CHANNEL_QUIT
:
2952 * Remove the pipe from the poll set and continue the loop
2953 * since their might be data to consume.
2955 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2958 ERR("Unknown action");
2961 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2962 DBG("Channel thread pipe hung up");
2964 * Remove the pipe from the poll set and continue the loop
2965 * since their might be data to consume.
2967 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2970 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2974 /* Handle other stream */
2980 uint64_t tmp_id
= (uint64_t) pollfd
;
2982 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
2984 node
= lttng_ht_iter_get_node_u64(&iter
);
2987 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
2990 /* Check for error event */
2991 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2992 DBG("Channel fd %d is hup|err.", pollfd
);
2994 lttng_poll_del(&events
, chan
->wait_fd
);
2995 ret
= lttng_ht_del(channel_ht
, &iter
);
2999 * This will close the wait fd for each stream associated to
3000 * this channel AND monitored by the data/metadata thread thus
3001 * will be clean by the right thread.
3003 consumer_close_channel_streams(chan
);
3005 /* Release our own refcount */
3006 if (!uatomic_sub_return(&chan
->refcount
, 1)
3007 && !uatomic_read(&chan
->nb_init_stream_left
)) {
3008 consumer_del_channel(chan
);
3011 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3016 /* Release RCU lock for the channel looked up */
3024 lttng_poll_clean(&events
);
3026 destroy_channel_ht(channel_ht
);
3029 DBG("Channel poll thread exiting");
3032 ERR("Health error occurred in %s", __func__
);
3034 health_unregister(health_consumerd
);
3035 rcu_unregister_thread();
3039 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
3040 struct pollfd
*sockpoll
, int client_socket
)
3047 ret
= lttng_consumer_poll_socket(sockpoll
);
3051 DBG("Metadata connection on client_socket");
3053 /* Blocking call, waiting for transmission */
3054 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
3055 if (ctx
->consumer_metadata_socket
< 0) {
3056 WARN("On accept metadata");
3067 * This thread listens on the consumerd socket and receives the file
3068 * descriptors from the session daemon.
3070 void *consumer_thread_sessiond_poll(void *data
)
3072 int sock
= -1, client_socket
, ret
, err
= -1;
3074 * structure to poll for incoming data on communication socket avoids
3075 * making blocking sockets.
3077 struct pollfd consumer_sockpoll
[2];
3078 struct lttng_consumer_local_data
*ctx
= data
;
3080 rcu_register_thread();
3082 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
3084 if (testpoint(consumerd_thread_sessiond
)) {
3085 goto error_testpoint
;
3088 health_code_update();
3090 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
3091 unlink(ctx
->consumer_command_sock_path
);
3092 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
3093 if (client_socket
< 0) {
3094 ERR("Cannot create command socket");
3098 ret
= lttcomm_listen_unix_sock(client_socket
);
3103 DBG("Sending ready command to lttng-sessiond");
3104 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3105 /* return < 0 on error, but == 0 is not fatal */
3107 ERR("Error sending ready command to lttng-sessiond");
3111 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3112 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3113 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3114 consumer_sockpoll
[1].fd
= client_socket
;
3115 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3117 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3125 DBG("Connection on client_socket");
3127 /* Blocking call, waiting for transmission */
3128 sock
= lttcomm_accept_unix_sock(client_socket
);
3135 * Setup metadata socket which is the second socket connection on the
3136 * command unix socket.
3138 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3147 /* This socket is not useful anymore. */
3148 ret
= close(client_socket
);
3150 PERROR("close client_socket");
3154 /* update the polling structure to poll on the established socket */
3155 consumer_sockpoll
[1].fd
= sock
;
3156 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3159 health_code_update();
3161 health_poll_entry();
3162 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3171 DBG("Incoming command on sock");
3172 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3175 * This could simply be a session daemon quitting. Don't output
3178 DBG("Communication interrupted on command socket");
3182 if (consumer_quit
) {
3183 DBG("consumer_thread_receive_fds received quit from signal");
3184 err
= 0; /* All is OK */
3187 DBG("received command on sock");
3193 DBG("Consumer thread sessiond poll exiting");
3196 * Close metadata streams since the producer is the session daemon which
3199 * NOTE: for now, this only applies to the UST tracer.
3201 lttng_consumer_close_all_metadata();
3204 * when all fds have hung up, the polling thread
3210 * Notify the data poll thread to poll back again and test the
3211 * consumer_quit state that we just set so to quit gracefully.
3213 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3215 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3217 notify_health_quit_pipe(health_quit_pipe
);
3219 /* Cleaning up possibly open sockets. */
3223 PERROR("close sock sessiond poll");
3226 if (client_socket
>= 0) {
3227 ret
= close(client_socket
);
3229 PERROR("close client_socket sessiond poll");
3236 ERR("Health error occurred in %s", __func__
);
3238 health_unregister(health_consumerd
);
3240 rcu_unregister_thread();
3244 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3245 struct lttng_consumer_local_data
*ctx
)
3249 pthread_mutex_lock(&stream
->lock
);
3250 if (stream
->metadata_flag
) {
3251 pthread_mutex_lock(&stream
->metadata_rdv_lock
);
3254 switch (consumer_data
.type
) {
3255 case LTTNG_CONSUMER_KERNEL
:
3256 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
3258 case LTTNG_CONSUMER32_UST
:
3259 case LTTNG_CONSUMER64_UST
:
3260 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
3263 ERR("Unknown consumer_data type");
3269 if (stream
->metadata_flag
) {
3270 pthread_cond_broadcast(&stream
->metadata_rdv
);
3271 pthread_mutex_unlock(&stream
->metadata_rdv_lock
);
3273 pthread_mutex_unlock(&stream
->lock
);
3277 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3279 switch (consumer_data
.type
) {
3280 case LTTNG_CONSUMER_KERNEL
:
3281 return lttng_kconsumer_on_recv_stream(stream
);
3282 case LTTNG_CONSUMER32_UST
:
3283 case LTTNG_CONSUMER64_UST
:
3284 return lttng_ustconsumer_on_recv_stream(stream
);
3286 ERR("Unknown consumer_data type");
3293 * Allocate and set consumer data hash tables.
3295 int lttng_consumer_init(void)
3297 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3298 if (!consumer_data
.channel_ht
) {
3302 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3303 if (!consumer_data
.relayd_ht
) {
3307 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3308 if (!consumer_data
.stream_list_ht
) {
3312 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3313 if (!consumer_data
.stream_per_chan_id_ht
) {
3317 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3322 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3334 * Process the ADD_RELAYD command receive by a consumer.
3336 * This will create a relayd socket pair and add it to the relayd hash table.
3337 * The caller MUST acquire a RCU read side lock before calling it.
3339 int consumer_add_relayd_socket(uint64_t net_seq_idx
, int sock_type
,
3340 struct lttng_consumer_local_data
*ctx
, int sock
,
3341 struct pollfd
*consumer_sockpoll
,
3342 struct lttcomm_relayd_sock
*relayd_sock
, uint64_t sessiond_id
,
3343 uint64_t relayd_session_id
)
3345 int fd
= -1, ret
= -1, relayd_created
= 0;
3346 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3347 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3350 assert(relayd_sock
);
3352 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", net_seq_idx
);
3354 /* Get relayd reference if exists. */
3355 relayd
= consumer_find_relayd(net_seq_idx
);
3356 if (relayd
== NULL
) {
3357 assert(sock_type
== LTTNG_STREAM_CONTROL
);
3358 /* Not found. Allocate one. */
3359 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3360 if (relayd
== NULL
) {
3362 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3365 relayd
->sessiond_session_id
= sessiond_id
;
3370 * This code path MUST continue to the consumer send status message to
3371 * we can notify the session daemon and continue our work without
3372 * killing everything.
3376 * relayd key should never be found for control socket.
3378 assert(sock_type
!= LTTNG_STREAM_CONTROL
);
3381 /* First send a status message before receiving the fds. */
3382 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3384 /* Somehow, the session daemon is not responding anymore. */
3385 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3386 goto error_nosignal
;
3389 /* Poll on consumer socket. */
3390 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3392 /* Needing to exit in the middle of a command: error. */
3393 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3395 goto error_nosignal
;
3398 /* Get relayd socket from session daemon */
3399 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3400 if (ret
!= sizeof(fd
)) {
3402 fd
= -1; /* Just in case it gets set with an invalid value. */
3405 * Failing to receive FDs might indicate a major problem such as
3406 * reaching a fd limit during the receive where the kernel returns a
3407 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3408 * don't take any chances and stop everything.
3410 * XXX: Feature request #558 will fix that and avoid this possible
3411 * issue when reaching the fd limit.
3413 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3414 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3418 /* Copy socket information and received FD */
3419 switch (sock_type
) {
3420 case LTTNG_STREAM_CONTROL
:
3421 /* Copy received lttcomm socket */
3422 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3423 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3424 /* Handle create_sock error. */
3426 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3430 * Close the socket created internally by
3431 * lttcomm_create_sock, so we can replace it by the one
3432 * received from sessiond.
3434 if (close(relayd
->control_sock
.sock
.fd
)) {
3438 /* Assign new file descriptor */
3439 relayd
->control_sock
.sock
.fd
= fd
;
3440 fd
= -1; /* For error path */
3441 /* Assign version values. */
3442 relayd
->control_sock
.major
= relayd_sock
->major
;
3443 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3445 relayd
->relayd_session_id
= relayd_session_id
;
3448 case LTTNG_STREAM_DATA
:
3449 /* Copy received lttcomm socket */
3450 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3451 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3452 /* Handle create_sock error. */
3454 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3458 * Close the socket created internally by
3459 * lttcomm_create_sock, so we can replace it by the one
3460 * received from sessiond.
3462 if (close(relayd
->data_sock
.sock
.fd
)) {
3466 /* Assign new file descriptor */
3467 relayd
->data_sock
.sock
.fd
= fd
;
3468 fd
= -1; /* for eventual error paths */
3469 /* Assign version values. */
3470 relayd
->data_sock
.major
= relayd_sock
->major
;
3471 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3474 ERR("Unknown relayd socket type (%d)", sock_type
);
3476 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3480 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3481 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3482 relayd
->net_seq_idx
, fd
);
3484 /* We successfully added the socket. Send status back. */
3485 ret
= consumer_send_status_msg(sock
, ret_code
);
3487 /* Somehow, the session daemon is not responding anymore. */
3488 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3489 goto error_nosignal
;
3493 * Add relayd socket pair to consumer data hashtable. If object already
3494 * exists or on error, the function gracefully returns.
3502 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3503 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3507 /* Close received socket if valid. */
3510 PERROR("close received socket");
3514 if (relayd_created
) {
3522 * Try to lock the stream mutex.
3524 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
3526 static int stream_try_lock(struct lttng_consumer_stream
*stream
)
3533 * Try to lock the stream mutex. On failure, we know that the stream is
3534 * being used else where hence there is data still being extracted.
3536 ret
= pthread_mutex_trylock(&stream
->lock
);
3538 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
3550 * Search for a relayd associated to the session id and return the reference.
3552 * A rcu read side lock MUST be acquire before calling this function and locked
3553 * until the relayd object is no longer necessary.
3555 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3557 struct lttng_ht_iter iter
;
3558 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3560 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3561 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3564 * Check by sessiond id which is unique here where the relayd session
3565 * id might not be when having multiple relayd.
3567 if (relayd
->sessiond_session_id
== id
) {
3568 /* Found the relayd. There can be only one per id. */
3580 * Check if for a given session id there is still data needed to be extract
3583 * Return 1 if data is pending or else 0 meaning ready to be read.
3585 int consumer_data_pending(uint64_t id
)
3588 struct lttng_ht_iter iter
;
3589 struct lttng_ht
*ht
;
3590 struct lttng_consumer_stream
*stream
;
3591 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3592 int (*data_pending
)(struct lttng_consumer_stream
*);
3594 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3597 pthread_mutex_lock(&consumer_data
.lock
);
3599 switch (consumer_data
.type
) {
3600 case LTTNG_CONSUMER_KERNEL
:
3601 data_pending
= lttng_kconsumer_data_pending
;
3603 case LTTNG_CONSUMER32_UST
:
3604 case LTTNG_CONSUMER64_UST
:
3605 data_pending
= lttng_ustconsumer_data_pending
;
3608 ERR("Unknown consumer data type");
3612 /* Ease our life a bit */
3613 ht
= consumer_data
.stream_list_ht
;
3615 relayd
= find_relayd_by_session_id(id
);
3617 /* Send init command for data pending. */
3618 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3619 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3620 relayd
->relayd_session_id
);
3621 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3623 /* Communication error thus the relayd so no data pending. */
3624 goto data_not_pending
;
3628 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3629 ht
->hash_fct(&id
, lttng_ht_seed
),
3631 &iter
.iter
, stream
, node_session_id
.node
) {
3632 /* If this call fails, the stream is being used hence data pending. */
3633 ret
= stream_try_lock(stream
);
3639 * A removed node from the hash table indicates that the stream has
3640 * been deleted thus having a guarantee that the buffers are closed
3641 * on the consumer side. However, data can still be transmitted
3642 * over the network so don't skip the relayd check.
3644 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3646 /* Check the stream if there is data in the buffers. */
3647 ret
= data_pending(stream
);
3649 pthread_mutex_unlock(&stream
->lock
);
3656 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3657 if (stream
->metadata_flag
) {
3658 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3659 stream
->relayd_stream_id
);
3661 ret
= relayd_data_pending(&relayd
->control_sock
,
3662 stream
->relayd_stream_id
,
3663 stream
->next_net_seq_num
- 1);
3665 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3667 pthread_mutex_unlock(&stream
->lock
);
3671 pthread_mutex_unlock(&stream
->lock
);
3675 unsigned int is_data_inflight
= 0;
3677 /* Send init command for data pending. */
3678 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3679 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3680 relayd
->relayd_session_id
, &is_data_inflight
);
3681 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3683 goto data_not_pending
;
3685 if (is_data_inflight
) {
3691 * Finding _no_ node in the hash table and no inflight data means that the
3692 * stream(s) have been removed thus data is guaranteed to be available for
3693 * analysis from the trace files.
3697 /* Data is available to be read by a viewer. */
3698 pthread_mutex_unlock(&consumer_data
.lock
);
3703 /* Data is still being extracted from buffers. */
3704 pthread_mutex_unlock(&consumer_data
.lock
);
3710 * Send a ret code status message to the sessiond daemon.
3712 * Return the sendmsg() return value.
3714 int consumer_send_status_msg(int sock
, int ret_code
)
3716 struct lttcomm_consumer_status_msg msg
;
3718 memset(&msg
, 0, sizeof(msg
));
3719 msg
.ret_code
= ret_code
;
3721 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3725 * Send a channel status message to the sessiond daemon.
3727 * Return the sendmsg() return value.
3729 int consumer_send_status_channel(int sock
,
3730 struct lttng_consumer_channel
*channel
)
3732 struct lttcomm_consumer_status_channel msg
;
3736 memset(&msg
, 0, sizeof(msg
));
3738 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3740 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3741 msg
.key
= channel
->key
;
3742 msg
.stream_count
= channel
->streams
.count
;
3745 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3748 unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos
,
3749 unsigned long produced_pos
, uint64_t nb_packets_per_stream
,
3750 uint64_t max_sb_size
)
3752 unsigned long start_pos
;
3754 if (!nb_packets_per_stream
) {
3755 return consumed_pos
; /* Grab everything */
3757 start_pos
= produced_pos
- offset_align_floor(produced_pos
, max_sb_size
);
3758 start_pos
-= max_sb_size
* nb_packets_per_stream
;
3759 if ((long) (start_pos
- consumed_pos
) < 0) {
3760 return consumed_pos
; /* Grab everything */