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");
1232 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1236 int outfd
= stream
->out_fd
;
1239 * This does a blocking write-and-wait on any page that belongs to the
1240 * subbuffer prior to the one we just wrote.
1241 * Don't care about error values, as these are just hints and ways to
1242 * limit the amount of page cache used.
1244 if (orig_offset
< stream
->max_sb_size
) {
1247 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1248 stream
->max_sb_size
,
1249 SYNC_FILE_RANGE_WAIT_BEFORE
1250 | SYNC_FILE_RANGE_WRITE
1251 | SYNC_FILE_RANGE_WAIT_AFTER
);
1253 * Give hints to the kernel about how we access the file:
1254 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1257 * We need to call fadvise again after the file grows because the
1258 * kernel does not seem to apply fadvise to non-existing parts of the
1261 * Call fadvise _after_ having waited for the page writeback to
1262 * complete because the dirty page writeback semantic is not well
1263 * defined. So it can be expected to lead to lower throughput in
1266 ret
= posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1267 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1269 WARN("posix_fadvise() error (%i)", ret
);
1274 * Initialise the necessary environnement :
1275 * - create a new context
1276 * - create the poll_pipe
1277 * - create the should_quit pipe (for signal handler)
1278 * - create the thread pipe (for splice)
1280 * Takes a function pointer as argument, this function is called when data is
1281 * available on a buffer. This function is responsible to do the
1282 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1283 * buffer configuration and then kernctl_put_next_subbuf at the end.
1285 * Returns a pointer to the new context or NULL on error.
1287 struct lttng_consumer_local_data
*lttng_consumer_create(
1288 enum lttng_consumer_type type
,
1289 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1290 struct lttng_consumer_local_data
*ctx
),
1291 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1292 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1293 int (*update_stream
)(uint64_t stream_key
, uint32_t state
))
1296 struct lttng_consumer_local_data
*ctx
;
1298 assert(consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1299 consumer_data
.type
== type
);
1300 consumer_data
.type
= type
;
1302 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
1304 PERROR("allocating context");
1308 ctx
->consumer_error_socket
= -1;
1309 ctx
->consumer_metadata_socket
= -1;
1310 pthread_mutex_init(&ctx
->metadata_socket_lock
, NULL
);
1311 /* assign the callbacks */
1312 ctx
->on_buffer_ready
= buffer_ready
;
1313 ctx
->on_recv_channel
= recv_channel
;
1314 ctx
->on_recv_stream
= recv_stream
;
1315 ctx
->on_update_stream
= update_stream
;
1317 ctx
->consumer_data_pipe
= lttng_pipe_open(0);
1318 if (!ctx
->consumer_data_pipe
) {
1319 goto error_poll_pipe
;
1322 ctx
->consumer_wakeup_pipe
= lttng_pipe_open(0);
1323 if (!ctx
->consumer_wakeup_pipe
) {
1324 goto error_wakeup_pipe
;
1327 ret
= pipe(ctx
->consumer_should_quit
);
1329 PERROR("Error creating recv pipe");
1330 goto error_quit_pipe
;
1333 ret
= pipe(ctx
->consumer_channel_pipe
);
1335 PERROR("Error creating channel pipe");
1336 goto error_channel_pipe
;
1339 ctx
->consumer_metadata_pipe
= lttng_pipe_open(0);
1340 if (!ctx
->consumer_metadata_pipe
) {
1341 goto error_metadata_pipe
;
1346 error_metadata_pipe
:
1347 utils_close_pipe(ctx
->consumer_channel_pipe
);
1349 utils_close_pipe(ctx
->consumer_should_quit
);
1351 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1353 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1361 * Iterate over all streams of the hashtable and free them properly.
1363 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1365 struct lttng_ht_iter iter
;
1366 struct lttng_consumer_stream
*stream
;
1373 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1375 * Ignore return value since we are currently cleaning up so any error
1378 (void) consumer_del_stream(stream
, ht
);
1382 lttng_ht_destroy(ht
);
1386 * Iterate over all streams of the metadata hashtable and free them
1389 static void destroy_metadata_stream_ht(struct lttng_ht
*ht
)
1391 struct lttng_ht_iter iter
;
1392 struct lttng_consumer_stream
*stream
;
1399 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1401 * Ignore return value since we are currently cleaning up so any error
1404 (void) consumer_del_metadata_stream(stream
, ht
);
1408 lttng_ht_destroy(ht
);
1412 * Close all fds associated with the instance and free the context.
1414 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1418 DBG("Consumer destroying it. Closing everything.");
1424 destroy_data_stream_ht(data_ht
);
1425 destroy_metadata_stream_ht(metadata_ht
);
1427 ret
= close(ctx
->consumer_error_socket
);
1431 ret
= close(ctx
->consumer_metadata_socket
);
1435 utils_close_pipe(ctx
->consumer_channel_pipe
);
1436 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1437 lttng_pipe_destroy(ctx
->consumer_metadata_pipe
);
1438 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1439 utils_close_pipe(ctx
->consumer_should_quit
);
1441 unlink(ctx
->consumer_command_sock_path
);
1446 * Write the metadata stream id on the specified file descriptor.
1448 static int write_relayd_metadata_id(int fd
,
1449 struct lttng_consumer_stream
*stream
,
1450 struct consumer_relayd_sock_pair
*relayd
, unsigned long padding
)
1453 struct lttcomm_relayd_metadata_payload hdr
;
1455 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1456 hdr
.padding_size
= htobe32(padding
);
1457 ret
= lttng_write(fd
, (void *) &hdr
, sizeof(hdr
));
1458 if (ret
< sizeof(hdr
)) {
1460 * This error means that the fd's end is closed so ignore the PERROR
1461 * not to clubber the error output since this can happen in a normal
1464 if (errno
!= EPIPE
) {
1465 PERROR("write metadata stream id");
1467 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1469 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1470 * handle writting the missing part so report that as an error and
1471 * don't lie to the caller.
1476 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1477 stream
->relayd_stream_id
, padding
);
1484 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1485 * core function for writing trace buffers to either the local filesystem or
1488 * It must be called with the stream lock held.
1490 * Careful review MUST be put if any changes occur!
1492 * Returns the number of bytes written
1494 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1495 struct lttng_consumer_local_data
*ctx
,
1496 struct lttng_consumer_stream
*stream
, unsigned long len
,
1497 unsigned long padding
,
1498 struct ctf_packet_index
*index
)
1500 unsigned long mmap_offset
;
1503 off_t orig_offset
= stream
->out_fd_offset
;
1504 /* Default is on the disk */
1505 int outfd
= stream
->out_fd
;
1506 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1507 unsigned int relayd_hang_up
= 0;
1509 /* RCU lock for the relayd pointer */
1512 /* Flag that the current stream if set for network streaming. */
1513 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1514 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1515 if (relayd
== NULL
) {
1521 /* get the offset inside the fd to mmap */
1522 switch (consumer_data
.type
) {
1523 case LTTNG_CONSUMER_KERNEL
:
1524 mmap_base
= stream
->mmap_base
;
1525 ret
= kernctl_get_mmap_read_offset(stream
->wait_fd
, &mmap_offset
);
1528 PERROR("tracer ctl get_mmap_read_offset");
1532 case LTTNG_CONSUMER32_UST
:
1533 case LTTNG_CONSUMER64_UST
:
1534 mmap_base
= lttng_ustctl_get_mmap_base(stream
);
1536 ERR("read mmap get mmap base for stream %s", stream
->name
);
1540 ret
= lttng_ustctl_get_mmap_read_offset(stream
, &mmap_offset
);
1542 PERROR("tracer ctl get_mmap_read_offset");
1548 ERR("Unknown consumer_data type");
1552 /* Handle stream on the relayd if the output is on the network */
1554 unsigned long netlen
= len
;
1557 * Lock the control socket for the complete duration of the function
1558 * since from this point on we will use the socket.
1560 if (stream
->metadata_flag
) {
1561 /* Metadata requires the control socket. */
1562 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1563 if (stream
->reset_metadata_flag
) {
1564 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1565 stream
->relayd_stream_id
,
1566 stream
->metadata_version
);
1571 stream
->reset_metadata_flag
= 0;
1573 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1576 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1581 /* Use the returned socket. */
1584 /* Write metadata stream id before payload */
1585 if (stream
->metadata_flag
) {
1586 ret
= write_relayd_metadata_id(outfd
, stream
, relayd
, padding
);
1593 /* No streaming, we have to set the len with the full padding */
1596 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1597 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1599 ERR("Reset metadata file");
1602 stream
->reset_metadata_flag
= 0;
1606 * Check if we need to change the tracefile before writing the packet.
1608 if (stream
->chan
->tracefile_size
> 0 &&
1609 (stream
->tracefile_size_current
+ len
) >
1610 stream
->chan
->tracefile_size
) {
1611 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1612 stream
->name
, stream
->chan
->tracefile_size
,
1613 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1614 stream
->out_fd
, &(stream
->tracefile_count_current
),
1617 ERR("Rotating output file");
1620 outfd
= stream
->out_fd
;
1622 if (stream
->index_fd
>= 0) {
1623 ret
= close(stream
->index_fd
);
1625 PERROR("Closing index");
1628 stream
->index_fd
= -1;
1629 ret
= index_create_file(stream
->chan
->pathname
,
1630 stream
->name
, stream
->uid
, stream
->gid
,
1631 stream
->chan
->tracefile_size
,
1632 stream
->tracefile_count_current
);
1636 stream
->index_fd
= ret
;
1639 /* Reset current size because we just perform a rotation. */
1640 stream
->tracefile_size_current
= 0;
1641 stream
->out_fd_offset
= 0;
1644 stream
->tracefile_size_current
+= len
;
1646 index
->offset
= htobe64(stream
->out_fd_offset
);
1651 * This call guarantee that len or less is returned. It's impossible to
1652 * receive a ret value that is bigger than len.
1654 ret
= lttng_write(outfd
, mmap_base
+ mmap_offset
, len
);
1655 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, len
);
1656 if (ret
< 0 || ((size_t) ret
!= len
)) {
1658 * Report error to caller if nothing was written else at least send the
1666 /* Socket operation failed. We consider the relayd dead */
1667 if (errno
== EPIPE
|| errno
== EINVAL
|| errno
== EBADF
) {
1669 * This is possible if the fd is closed on the other side
1670 * (outfd) or any write problem. It can be verbose a bit for a
1671 * normal execution if for instance the relayd is stopped
1672 * abruptly. This can happen so set this to a DBG statement.
1674 DBG("Consumer mmap write detected relayd hang up");
1676 /* Unhandled error, print it and stop function right now. */
1677 PERROR("Error in write mmap (ret %zd != len %lu)", ret
, len
);
1681 stream
->output_written
+= ret
;
1683 /* This call is useless on a socket so better save a syscall. */
1685 /* This won't block, but will start writeout asynchronously */
1686 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, len
,
1687 SYNC_FILE_RANGE_WRITE
);
1688 stream
->out_fd_offset
+= len
;
1690 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1694 * This is a special case that the relayd has closed its socket. Let's
1695 * cleanup the relayd object and all associated streams.
1697 if (relayd
&& relayd_hang_up
) {
1698 cleanup_relayd(relayd
, ctx
);
1702 /* Unlock only if ctrl socket used */
1703 if (relayd
&& stream
->metadata_flag
) {
1704 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1712 * Splice the data from the ring buffer to the tracefile.
1714 * It must be called with the stream lock held.
1716 * Returns the number of bytes spliced.
1718 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1719 struct lttng_consumer_local_data
*ctx
,
1720 struct lttng_consumer_stream
*stream
, unsigned long len
,
1721 unsigned long padding
,
1722 struct ctf_packet_index
*index
)
1724 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1726 off_t orig_offset
= stream
->out_fd_offset
;
1727 int fd
= stream
->wait_fd
;
1728 /* Default is on the disk */
1729 int outfd
= stream
->out_fd
;
1730 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1732 unsigned int relayd_hang_up
= 0;
1734 switch (consumer_data
.type
) {
1735 case LTTNG_CONSUMER_KERNEL
:
1737 case LTTNG_CONSUMER32_UST
:
1738 case LTTNG_CONSUMER64_UST
:
1739 /* Not supported for user space tracing */
1742 ERR("Unknown consumer_data type");
1746 /* RCU lock for the relayd pointer */
1749 /* Flag that the current stream if set for network streaming. */
1750 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1751 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1752 if (relayd
== NULL
) {
1757 splice_pipe
= stream
->splice_pipe
;
1759 /* Write metadata stream id before payload */
1761 unsigned long total_len
= len
;
1763 if (stream
->metadata_flag
) {
1765 * Lock the control socket for the complete duration of the function
1766 * since from this point on we will use the socket.
1768 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1770 if (stream
->reset_metadata_flag
) {
1771 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1772 stream
->relayd_stream_id
,
1773 stream
->metadata_version
);
1778 stream
->reset_metadata_flag
= 0;
1780 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
, relayd
,
1788 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1791 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1797 /* Use the returned socket. */
1800 /* No streaming, we have to set the len with the full padding */
1803 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1804 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1806 ERR("Reset metadata file");
1809 stream
->reset_metadata_flag
= 0;
1812 * Check if we need to change the tracefile before writing the packet.
1814 if (stream
->chan
->tracefile_size
> 0 &&
1815 (stream
->tracefile_size_current
+ len
) >
1816 stream
->chan
->tracefile_size
) {
1817 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1818 stream
->name
, stream
->chan
->tracefile_size
,
1819 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1820 stream
->out_fd
, &(stream
->tracefile_count_current
),
1824 ERR("Rotating output file");
1827 outfd
= stream
->out_fd
;
1829 if (stream
->index_fd
>= 0) {
1830 ret
= close(stream
->index_fd
);
1832 PERROR("Closing index");
1835 stream
->index_fd
= -1;
1836 ret
= index_create_file(stream
->chan
->pathname
,
1837 stream
->name
, stream
->uid
, stream
->gid
,
1838 stream
->chan
->tracefile_size
,
1839 stream
->tracefile_count_current
);
1844 stream
->index_fd
= ret
;
1847 /* Reset current size because we just perform a rotation. */
1848 stream
->tracefile_size_current
= 0;
1849 stream
->out_fd_offset
= 0;
1852 stream
->tracefile_size_current
+= len
;
1853 index
->offset
= htobe64(stream
->out_fd_offset
);
1857 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1858 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1859 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1860 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1861 DBG("splice chan to pipe, ret %zd", ret_splice
);
1862 if (ret_splice
< 0) {
1865 PERROR("Error in relay splice");
1869 /* Handle stream on the relayd if the output is on the network */
1870 if (relayd
&& stream
->metadata_flag
) {
1871 size_t metadata_payload_size
=
1872 sizeof(struct lttcomm_relayd_metadata_payload
);
1874 /* Update counter to fit the spliced data */
1875 ret_splice
+= metadata_payload_size
;
1876 len
+= metadata_payload_size
;
1878 * We do this so the return value can match the len passed as
1879 * argument to this function.
1881 written
-= metadata_payload_size
;
1884 /* Splice data out */
1885 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1886 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1887 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
1889 if (ret_splice
< 0) {
1894 } else if (ret_splice
> len
) {
1896 * We don't expect this code path to be executed but you never know
1897 * so this is an extra protection agains a buggy splice().
1900 written
+= ret_splice
;
1901 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice
,
1905 /* All good, update current len and continue. */
1909 /* This call is useless on a socket so better save a syscall. */
1911 /* This won't block, but will start writeout asynchronously */
1912 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1913 SYNC_FILE_RANGE_WRITE
);
1914 stream
->out_fd_offset
+= ret_splice
;
1916 stream
->output_written
+= ret_splice
;
1917 written
+= ret_splice
;
1919 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1924 * This is a special case that the relayd has closed its socket. Let's
1925 * cleanup the relayd object and all associated streams.
1927 if (relayd
&& relayd_hang_up
) {
1928 cleanup_relayd(relayd
, ctx
);
1929 /* Skip splice error so the consumer does not fail */
1934 /* send the appropriate error description to sessiond */
1937 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1940 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1943 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
1948 if (relayd
&& stream
->metadata_flag
) {
1949 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1957 * Take a snapshot for a specific fd
1959 * Returns 0 on success, < 0 on error
1961 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
1963 switch (consumer_data
.type
) {
1964 case LTTNG_CONSUMER_KERNEL
:
1965 return lttng_kconsumer_take_snapshot(stream
);
1966 case LTTNG_CONSUMER32_UST
:
1967 case LTTNG_CONSUMER64_UST
:
1968 return lttng_ustconsumer_take_snapshot(stream
);
1970 ERR("Unknown consumer_data type");
1977 * Get the produced position
1979 * Returns 0 on success, < 0 on error
1981 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
1984 switch (consumer_data
.type
) {
1985 case LTTNG_CONSUMER_KERNEL
:
1986 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
1987 case LTTNG_CONSUMER32_UST
:
1988 case LTTNG_CONSUMER64_UST
:
1989 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
1991 ERR("Unknown consumer_data type");
1997 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
1998 int sock
, struct pollfd
*consumer_sockpoll
)
2000 switch (consumer_data
.type
) {
2001 case LTTNG_CONSUMER_KERNEL
:
2002 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2003 case LTTNG_CONSUMER32_UST
:
2004 case LTTNG_CONSUMER64_UST
:
2005 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2007 ERR("Unknown consumer_data type");
2013 void lttng_consumer_close_all_metadata(void)
2015 switch (consumer_data
.type
) {
2016 case LTTNG_CONSUMER_KERNEL
:
2018 * The Kernel consumer has a different metadata scheme so we don't
2019 * close anything because the stream will be closed by the session
2023 case LTTNG_CONSUMER32_UST
:
2024 case LTTNG_CONSUMER64_UST
:
2026 * Close all metadata streams. The metadata hash table is passed and
2027 * this call iterates over it by closing all wakeup fd. This is safe
2028 * because at this point we are sure that the metadata producer is
2029 * either dead or blocked.
2031 lttng_ustconsumer_close_all_metadata(metadata_ht
);
2034 ERR("Unknown consumer_data type");
2040 * Clean up a metadata stream and free its memory.
2042 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
2043 struct lttng_ht
*ht
)
2045 struct lttng_consumer_channel
*free_chan
= NULL
;
2049 * This call should NEVER receive regular stream. It must always be
2050 * metadata stream and this is crucial for data structure synchronization.
2052 assert(stream
->metadata_flag
);
2054 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
2056 pthread_mutex_lock(&consumer_data
.lock
);
2057 pthread_mutex_lock(&stream
->chan
->lock
);
2058 pthread_mutex_lock(&stream
->lock
);
2060 /* Remove any reference to that stream. */
2061 consumer_stream_delete(stream
, ht
);
2063 /* Close down everything including the relayd if one. */
2064 consumer_stream_close(stream
);
2065 /* Destroy tracer buffers of the stream. */
2066 consumer_stream_destroy_buffers(stream
);
2068 /* Atomically decrement channel refcount since other threads can use it. */
2069 if (!uatomic_sub_return(&stream
->chan
->refcount
, 1)
2070 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
2071 /* Go for channel deletion! */
2072 free_chan
= stream
->chan
;
2076 * Nullify the stream reference so it is not used after deletion. The
2077 * channel lock MUST be acquired before being able to check for a NULL
2080 stream
->chan
->metadata_stream
= NULL
;
2082 pthread_mutex_unlock(&stream
->lock
);
2083 pthread_mutex_unlock(&stream
->chan
->lock
);
2084 pthread_mutex_unlock(&consumer_data
.lock
);
2087 consumer_del_channel(free_chan
);
2090 consumer_stream_free(stream
);
2094 * Action done with the metadata stream when adding it to the consumer internal
2095 * data structures to handle it.
2097 int consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
2099 struct lttng_ht
*ht
= metadata_ht
;
2101 struct lttng_ht_iter iter
;
2102 struct lttng_ht_node_u64
*node
;
2107 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2109 pthread_mutex_lock(&consumer_data
.lock
);
2110 pthread_mutex_lock(&stream
->chan
->lock
);
2111 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2112 pthread_mutex_lock(&stream
->lock
);
2115 * From here, refcounts are updated so be _careful_ when returning an error
2122 * Lookup the stream just to make sure it does not exist in our internal
2123 * state. This should NEVER happen.
2125 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2126 node
= lttng_ht_iter_get_node_u64(&iter
);
2130 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2131 * in terms of destroying the associated channel, because the action that
2132 * causes the count to become 0 also causes a stream to be added. The
2133 * channel deletion will thus be triggered by the following removal of this
2136 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2137 /* Increment refcount before decrementing nb_init_stream_left */
2139 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2142 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2144 lttng_ht_add_unique_u64(consumer_data
.stream_per_chan_id_ht
,
2145 &stream
->node_channel_id
);
2148 * Add stream to the stream_list_ht of the consumer data. No need to steal
2149 * the key since the HT does not use it and we allow to add redundant keys
2152 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2156 pthread_mutex_unlock(&stream
->lock
);
2157 pthread_mutex_unlock(&stream
->chan
->lock
);
2158 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2159 pthread_mutex_unlock(&consumer_data
.lock
);
2164 * Delete data stream that are flagged for deletion (endpoint_status).
2166 static void validate_endpoint_status_data_stream(void)
2168 struct lttng_ht_iter iter
;
2169 struct lttng_consumer_stream
*stream
;
2171 DBG("Consumer delete flagged data stream");
2174 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2175 /* Validate delete flag of the stream */
2176 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2179 /* Delete it right now */
2180 consumer_del_stream(stream
, data_ht
);
2186 * Delete metadata stream that are flagged for deletion (endpoint_status).
2188 static void validate_endpoint_status_metadata_stream(
2189 struct lttng_poll_event
*pollset
)
2191 struct lttng_ht_iter iter
;
2192 struct lttng_consumer_stream
*stream
;
2194 DBG("Consumer delete flagged metadata stream");
2199 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2200 /* Validate delete flag of the stream */
2201 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2205 * Remove from pollset so the metadata thread can continue without
2206 * blocking on a deleted stream.
2208 lttng_poll_del(pollset
, stream
->wait_fd
);
2210 /* Delete it right now */
2211 consumer_del_metadata_stream(stream
, metadata_ht
);
2217 * Thread polls on metadata file descriptor and write them on disk or on the
2220 void *consumer_thread_metadata_poll(void *data
)
2222 int ret
, i
, pollfd
, err
= -1;
2223 uint32_t revents
, nb_fd
;
2224 struct lttng_consumer_stream
*stream
= NULL
;
2225 struct lttng_ht_iter iter
;
2226 struct lttng_ht_node_u64
*node
;
2227 struct lttng_poll_event events
;
2228 struct lttng_consumer_local_data
*ctx
= data
;
2231 rcu_register_thread();
2233 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2235 if (testpoint(consumerd_thread_metadata
)) {
2236 goto error_testpoint
;
2239 health_code_update();
2241 DBG("Thread metadata poll started");
2243 /* Size is set to 1 for the consumer_metadata pipe */
2244 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2246 ERR("Poll set creation failed");
2250 ret
= lttng_poll_add(&events
,
2251 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2257 DBG("Metadata main loop started");
2261 health_code_update();
2262 health_poll_entry();
2263 DBG("Metadata poll wait");
2264 ret
= lttng_poll_wait(&events
, -1);
2265 DBG("Metadata poll return from wait with %d fd(s)",
2266 LTTNG_POLL_GETNB(&events
));
2268 DBG("Metadata event caught in thread");
2270 if (errno
== EINTR
) {
2271 ERR("Poll EINTR caught");
2274 if (LTTNG_POLL_GETNB(&events
) == 0) {
2275 err
= 0; /* All is OK */
2282 /* From here, the event is a metadata wait fd */
2283 for (i
= 0; i
< nb_fd
; i
++) {
2284 health_code_update();
2286 revents
= LTTNG_POLL_GETEV(&events
, i
);
2287 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2290 /* No activity for this FD (poll implementation). */
2294 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2295 if (revents
& LPOLLIN
) {
2298 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2299 &stream
, sizeof(stream
));
2300 if (pipe_len
< sizeof(stream
)) {
2302 PERROR("read metadata stream");
2305 * Remove the pipe from the poll set and continue the loop
2306 * since their might be data to consume.
2308 lttng_poll_del(&events
,
2309 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2310 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2314 /* A NULL stream means that the state has changed. */
2315 if (stream
== NULL
) {
2316 /* Check for deleted streams. */
2317 validate_endpoint_status_metadata_stream(&events
);
2321 DBG("Adding metadata stream %d to poll set",
2324 /* Add metadata stream to the global poll events list */
2325 lttng_poll_add(&events
, stream
->wait_fd
,
2326 LPOLLIN
| LPOLLPRI
| LPOLLHUP
);
2327 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2328 DBG("Metadata thread pipe hung up");
2330 * Remove the pipe from the poll set and continue the loop
2331 * since their might be data to consume.
2333 lttng_poll_del(&events
,
2334 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2335 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2338 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2342 /* Handle other stream */
2348 uint64_t tmp_id
= (uint64_t) pollfd
;
2350 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2352 node
= lttng_ht_iter_get_node_u64(&iter
);
2355 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2358 if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2359 /* Get the data out of the metadata file descriptor */
2360 DBG("Metadata available on fd %d", pollfd
);
2361 assert(stream
->wait_fd
== pollfd
);
2364 health_code_update();
2366 len
= ctx
->on_buffer_ready(stream
, ctx
);
2368 * We don't check the return value here since if we get
2369 * a negative len, it means an error occured thus we
2370 * simply remove it from the poll set and free the
2375 /* It's ok to have an unavailable sub-buffer */
2376 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2377 /* Clean up stream from consumer and free it. */
2378 lttng_poll_del(&events
, stream
->wait_fd
);
2379 consumer_del_metadata_stream(stream
, metadata_ht
);
2381 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2382 DBG("Metadata fd %d is hup|err.", pollfd
);
2383 if (!stream
->hangup_flush_done
2384 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2385 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2386 DBG("Attempting to flush and consume the UST buffers");
2387 lttng_ustconsumer_on_stream_hangup(stream
);
2389 /* We just flushed the stream now read it. */
2391 health_code_update();
2393 len
= ctx
->on_buffer_ready(stream
, ctx
);
2395 * We don't check the return value here since if we get
2396 * a negative len, it means an error occured thus we
2397 * simply remove it from the poll set and free the
2403 lttng_poll_del(&events
, stream
->wait_fd
);
2405 * This call update the channel states, closes file descriptors
2406 * and securely free the stream.
2408 consumer_del_metadata_stream(stream
, metadata_ht
);
2410 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2414 /* Release RCU lock for the stream looked up */
2422 DBG("Metadata poll thread exiting");
2424 lttng_poll_clean(&events
);
2429 ERR("Health error occurred in %s", __func__
);
2431 health_unregister(health_consumerd
);
2432 rcu_unregister_thread();
2437 * This thread polls the fds in the set to consume the data and write
2438 * it to tracefile if necessary.
2440 void *consumer_thread_data_poll(void *data
)
2442 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2443 struct pollfd
*pollfd
= NULL
;
2444 /* local view of the streams */
2445 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2446 /* local view of consumer_data.fds_count */
2448 struct lttng_consumer_local_data
*ctx
= data
;
2451 rcu_register_thread();
2453 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2455 if (testpoint(consumerd_thread_data
)) {
2456 goto error_testpoint
;
2459 health_code_update();
2461 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
*));
2462 if (local_stream
== NULL
) {
2463 PERROR("local_stream malloc");
2468 health_code_update();
2474 * the fds set has been updated, we need to update our
2475 * local array as well
2477 pthread_mutex_lock(&consumer_data
.lock
);
2478 if (consumer_data
.need_update
) {
2483 local_stream
= NULL
;
2486 * Allocate for all fds +1 for the consumer_data_pipe and +1 for
2489 pollfd
= zmalloc((consumer_data
.stream_count
+ 2) * sizeof(struct pollfd
));
2490 if (pollfd
== NULL
) {
2491 PERROR("pollfd malloc");
2492 pthread_mutex_unlock(&consumer_data
.lock
);
2496 local_stream
= zmalloc((consumer_data
.stream_count
+ 2) *
2497 sizeof(struct lttng_consumer_stream
*));
2498 if (local_stream
== NULL
) {
2499 PERROR("local_stream malloc");
2500 pthread_mutex_unlock(&consumer_data
.lock
);
2503 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2506 ERR("Error in allocating pollfd or local_outfds");
2507 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2508 pthread_mutex_unlock(&consumer_data
.lock
);
2512 consumer_data
.need_update
= 0;
2514 pthread_mutex_unlock(&consumer_data
.lock
);
2516 /* No FDs and consumer_quit, consumer_cleanup the thread */
2517 if (nb_fd
== 0 && consumer_quit
== 1) {
2518 err
= 0; /* All is OK */
2521 /* poll on the array of fds */
2523 DBG("polling on %d fd", nb_fd
+ 2);
2524 health_poll_entry();
2525 num_rdy
= poll(pollfd
, nb_fd
+ 2, -1);
2527 DBG("poll num_rdy : %d", num_rdy
);
2528 if (num_rdy
== -1) {
2530 * Restart interrupted system call.
2532 if (errno
== EINTR
) {
2535 PERROR("Poll error");
2536 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2538 } else if (num_rdy
== 0) {
2539 DBG("Polling thread timed out");
2544 * If the consumer_data_pipe triggered poll go directly to the
2545 * beginning of the loop to update the array. We want to prioritize
2546 * array update over low-priority reads.
2548 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2549 ssize_t pipe_readlen
;
2551 DBG("consumer_data_pipe wake up");
2552 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2553 &new_stream
, sizeof(new_stream
));
2554 if (pipe_readlen
< sizeof(new_stream
)) {
2555 PERROR("Consumer data pipe");
2556 /* Continue so we can at least handle the current stream(s). */
2561 * If the stream is NULL, just ignore it. It's also possible that
2562 * the sessiond poll thread changed the consumer_quit state and is
2563 * waking us up to test it.
2565 if (new_stream
== NULL
) {
2566 validate_endpoint_status_data_stream();
2570 /* Continue to update the local streams and handle prio ones */
2574 /* Handle wakeup pipe. */
2575 if (pollfd
[nb_fd
+ 1].revents
& (POLLIN
| POLLPRI
)) {
2577 ssize_t pipe_readlen
;
2579 pipe_readlen
= lttng_pipe_read(ctx
->consumer_wakeup_pipe
, &dummy
,
2581 if (pipe_readlen
< 0) {
2582 PERROR("Consumer data wakeup pipe");
2584 /* We've been awakened to handle stream(s). */
2585 ctx
->has_wakeup
= 0;
2588 /* Take care of high priority channels first. */
2589 for (i
= 0; i
< nb_fd
; i
++) {
2590 health_code_update();
2592 if (local_stream
[i
] == NULL
) {
2595 if (pollfd
[i
].revents
& POLLPRI
) {
2596 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2598 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2599 /* it's ok to have an unavailable sub-buffer */
2600 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2601 /* Clean the stream and free it. */
2602 consumer_del_stream(local_stream
[i
], data_ht
);
2603 local_stream
[i
] = NULL
;
2604 } else if (len
> 0) {
2605 local_stream
[i
]->data_read
= 1;
2611 * If we read high prio channel in this loop, try again
2612 * for more high prio data.
2618 /* Take care of low priority channels. */
2619 for (i
= 0; i
< nb_fd
; i
++) {
2620 health_code_update();
2622 if (local_stream
[i
] == NULL
) {
2625 if ((pollfd
[i
].revents
& POLLIN
) ||
2626 local_stream
[i
]->hangup_flush_done
||
2627 local_stream
[i
]->has_data
) {
2628 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2629 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2630 /* it's ok to have an unavailable sub-buffer */
2631 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2632 /* Clean the stream and free it. */
2633 consumer_del_stream(local_stream
[i
], data_ht
);
2634 local_stream
[i
] = NULL
;
2635 } else if (len
> 0) {
2636 local_stream
[i
]->data_read
= 1;
2641 /* Handle hangup and errors */
2642 for (i
= 0; i
< nb_fd
; i
++) {
2643 health_code_update();
2645 if (local_stream
[i
] == NULL
) {
2648 if (!local_stream
[i
]->hangup_flush_done
2649 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2650 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2651 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2652 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2654 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2655 /* Attempt read again, for the data we just flushed. */
2656 local_stream
[i
]->data_read
= 1;
2659 * If the poll flag is HUP/ERR/NVAL and we have
2660 * read no data in this pass, we can remove the
2661 * stream from its hash table.
2663 if ((pollfd
[i
].revents
& POLLHUP
)) {
2664 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2665 if (!local_stream
[i
]->data_read
) {
2666 consumer_del_stream(local_stream
[i
], data_ht
);
2667 local_stream
[i
] = NULL
;
2670 } else if (pollfd
[i
].revents
& POLLERR
) {
2671 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2672 if (!local_stream
[i
]->data_read
) {
2673 consumer_del_stream(local_stream
[i
], data_ht
);
2674 local_stream
[i
] = NULL
;
2677 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2678 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2679 if (!local_stream
[i
]->data_read
) {
2680 consumer_del_stream(local_stream
[i
], data_ht
);
2681 local_stream
[i
] = NULL
;
2685 if (local_stream
[i
] != NULL
) {
2686 local_stream
[i
]->data_read
= 0;
2693 DBG("polling thread exiting");
2698 * Close the write side of the pipe so epoll_wait() in
2699 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2700 * read side of the pipe. If we close them both, epoll_wait strangely does
2701 * not return and could create a endless wait period if the pipe is the
2702 * only tracked fd in the poll set. The thread will take care of closing
2705 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2710 ERR("Health error occurred in %s", __func__
);
2712 health_unregister(health_consumerd
);
2714 rcu_unregister_thread();
2719 * Close wake-up end of each stream belonging to the channel. This will
2720 * allow the poll() on the stream read-side to detect when the
2721 * write-side (application) finally closes them.
2724 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2726 struct lttng_ht
*ht
;
2727 struct lttng_consumer_stream
*stream
;
2728 struct lttng_ht_iter iter
;
2730 ht
= consumer_data
.stream_per_chan_id_ht
;
2733 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2734 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2735 ht
->match_fct
, &channel
->key
,
2736 &iter
.iter
, stream
, node_channel_id
.node
) {
2738 * Protect against teardown with mutex.
2740 pthread_mutex_lock(&stream
->lock
);
2741 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2744 switch (consumer_data
.type
) {
2745 case LTTNG_CONSUMER_KERNEL
:
2747 case LTTNG_CONSUMER32_UST
:
2748 case LTTNG_CONSUMER64_UST
:
2749 if (stream
->metadata_flag
) {
2750 /* Safe and protected by the stream lock. */
2751 lttng_ustconsumer_close_metadata(stream
->chan
);
2754 * Note: a mutex is taken internally within
2755 * liblttng-ust-ctl to protect timer wakeup_fd
2756 * use from concurrent close.
2758 lttng_ustconsumer_close_stream_wakeup(stream
);
2762 ERR("Unknown consumer_data type");
2766 pthread_mutex_unlock(&stream
->lock
);
2771 static void destroy_channel_ht(struct lttng_ht
*ht
)
2773 struct lttng_ht_iter iter
;
2774 struct lttng_consumer_channel
*channel
;
2782 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2783 ret
= lttng_ht_del(ht
, &iter
);
2788 lttng_ht_destroy(ht
);
2792 * This thread polls the channel fds to detect when they are being
2793 * closed. It closes all related streams if the channel is detected as
2794 * closed. It is currently only used as a shim layer for UST because the
2795 * consumerd needs to keep the per-stream wakeup end of pipes open for
2798 void *consumer_thread_channel_poll(void *data
)
2800 int ret
, i
, pollfd
, err
= -1;
2801 uint32_t revents
, nb_fd
;
2802 struct lttng_consumer_channel
*chan
= NULL
;
2803 struct lttng_ht_iter iter
;
2804 struct lttng_ht_node_u64
*node
;
2805 struct lttng_poll_event events
;
2806 struct lttng_consumer_local_data
*ctx
= data
;
2807 struct lttng_ht
*channel_ht
;
2809 rcu_register_thread();
2811 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
2813 if (testpoint(consumerd_thread_channel
)) {
2814 goto error_testpoint
;
2817 health_code_update();
2819 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2821 /* ENOMEM at this point. Better to bail out. */
2825 DBG("Thread channel poll started");
2827 /* Size is set to 1 for the consumer_channel pipe */
2828 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2830 ERR("Poll set creation failed");
2834 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2840 DBG("Channel main loop started");
2844 health_code_update();
2845 DBG("Channel poll wait");
2846 health_poll_entry();
2847 ret
= lttng_poll_wait(&events
, -1);
2848 DBG("Channel poll return from wait with %d fd(s)",
2849 LTTNG_POLL_GETNB(&events
));
2851 DBG("Channel event caught in thread");
2853 if (errno
== EINTR
) {
2854 ERR("Poll EINTR caught");
2857 if (LTTNG_POLL_GETNB(&events
) == 0) {
2858 err
= 0; /* All is OK */
2865 /* From here, the event is a channel wait fd */
2866 for (i
= 0; i
< nb_fd
; i
++) {
2867 health_code_update();
2869 revents
= LTTNG_POLL_GETEV(&events
, i
);
2870 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2873 /* No activity for this FD (poll implementation). */
2877 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2878 if (revents
& LPOLLIN
) {
2879 enum consumer_channel_action action
;
2882 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
2885 ERR("Error reading channel pipe");
2887 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2892 case CONSUMER_CHANNEL_ADD
:
2893 DBG("Adding channel %d to poll set",
2896 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
2899 lttng_ht_add_unique_u64(channel_ht
,
2900 &chan
->wait_fd_node
);
2902 /* Add channel to the global poll events list */
2903 lttng_poll_add(&events
, chan
->wait_fd
,
2904 LPOLLERR
| LPOLLHUP
);
2906 case CONSUMER_CHANNEL_DEL
:
2909 * This command should never be called if the channel
2910 * has streams monitored by either the data or metadata
2911 * thread. The consumer only notify this thread with a
2912 * channel del. command if it receives a destroy
2913 * channel command from the session daemon that send it
2914 * if a command prior to the GET_CHANNEL failed.
2918 chan
= consumer_find_channel(key
);
2921 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
2924 lttng_poll_del(&events
, chan
->wait_fd
);
2925 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
2926 ret
= lttng_ht_del(channel_ht
, &iter
);
2929 switch (consumer_data
.type
) {
2930 case LTTNG_CONSUMER_KERNEL
:
2932 case LTTNG_CONSUMER32_UST
:
2933 case LTTNG_CONSUMER64_UST
:
2934 health_code_update();
2935 /* Destroy streams that might have been left in the stream list. */
2936 clean_channel_stream_list(chan
);
2939 ERR("Unknown consumer_data type");
2944 * Release our own refcount. Force channel deletion even if
2945 * streams were not initialized.
2947 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
2948 consumer_del_channel(chan
);
2953 case CONSUMER_CHANNEL_QUIT
:
2955 * Remove the pipe from the poll set and continue the loop
2956 * since their might be data to consume.
2958 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2961 ERR("Unknown action");
2964 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2965 DBG("Channel thread pipe hung up");
2967 * Remove the pipe from the poll set and continue the loop
2968 * since their might be data to consume.
2970 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2973 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2977 /* Handle other stream */
2983 uint64_t tmp_id
= (uint64_t) pollfd
;
2985 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
2987 node
= lttng_ht_iter_get_node_u64(&iter
);
2990 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
2993 /* Check for error event */
2994 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2995 DBG("Channel fd %d is hup|err.", pollfd
);
2997 lttng_poll_del(&events
, chan
->wait_fd
);
2998 ret
= lttng_ht_del(channel_ht
, &iter
);
3002 * This will close the wait fd for each stream associated to
3003 * this channel AND monitored by the data/metadata thread thus
3004 * will be clean by the right thread.
3006 consumer_close_channel_streams(chan
);
3008 /* Release our own refcount */
3009 if (!uatomic_sub_return(&chan
->refcount
, 1)
3010 && !uatomic_read(&chan
->nb_init_stream_left
)) {
3011 consumer_del_channel(chan
);
3014 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3019 /* Release RCU lock for the channel looked up */
3027 lttng_poll_clean(&events
);
3029 destroy_channel_ht(channel_ht
);
3032 DBG("Channel poll thread exiting");
3035 ERR("Health error occurred in %s", __func__
);
3037 health_unregister(health_consumerd
);
3038 rcu_unregister_thread();
3042 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
3043 struct pollfd
*sockpoll
, int client_socket
)
3050 ret
= lttng_consumer_poll_socket(sockpoll
);
3054 DBG("Metadata connection on client_socket");
3056 /* Blocking call, waiting for transmission */
3057 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
3058 if (ctx
->consumer_metadata_socket
< 0) {
3059 WARN("On accept metadata");
3070 * This thread listens on the consumerd socket and receives the file
3071 * descriptors from the session daemon.
3073 void *consumer_thread_sessiond_poll(void *data
)
3075 int sock
= -1, client_socket
, ret
, err
= -1;
3077 * structure to poll for incoming data on communication socket avoids
3078 * making blocking sockets.
3080 struct pollfd consumer_sockpoll
[2];
3081 struct lttng_consumer_local_data
*ctx
= data
;
3083 rcu_register_thread();
3085 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
3087 if (testpoint(consumerd_thread_sessiond
)) {
3088 goto error_testpoint
;
3091 health_code_update();
3093 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
3094 unlink(ctx
->consumer_command_sock_path
);
3095 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
3096 if (client_socket
< 0) {
3097 ERR("Cannot create command socket");
3101 ret
= lttcomm_listen_unix_sock(client_socket
);
3106 DBG("Sending ready command to lttng-sessiond");
3107 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3108 /* return < 0 on error, but == 0 is not fatal */
3110 ERR("Error sending ready command to lttng-sessiond");
3114 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3115 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3116 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3117 consumer_sockpoll
[1].fd
= client_socket
;
3118 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3120 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3128 DBG("Connection on client_socket");
3130 /* Blocking call, waiting for transmission */
3131 sock
= lttcomm_accept_unix_sock(client_socket
);
3138 * Setup metadata socket which is the second socket connection on the
3139 * command unix socket.
3141 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3150 /* This socket is not useful anymore. */
3151 ret
= close(client_socket
);
3153 PERROR("close client_socket");
3157 /* update the polling structure to poll on the established socket */
3158 consumer_sockpoll
[1].fd
= sock
;
3159 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3162 health_code_update();
3164 health_poll_entry();
3165 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3174 DBG("Incoming command on sock");
3175 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3178 * This could simply be a session daemon quitting. Don't output
3181 DBG("Communication interrupted on command socket");
3185 if (consumer_quit
) {
3186 DBG("consumer_thread_receive_fds received quit from signal");
3187 err
= 0; /* All is OK */
3190 DBG("received command on sock");
3196 DBG("Consumer thread sessiond poll exiting");
3199 * Close metadata streams since the producer is the session daemon which
3202 * NOTE: for now, this only applies to the UST tracer.
3204 lttng_consumer_close_all_metadata();
3207 * when all fds have hung up, the polling thread
3213 * Notify the data poll thread to poll back again and test the
3214 * consumer_quit state that we just set so to quit gracefully.
3216 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3218 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3220 notify_health_quit_pipe(health_quit_pipe
);
3222 /* Cleaning up possibly open sockets. */
3226 PERROR("close sock sessiond poll");
3229 if (client_socket
>= 0) {
3230 ret
= close(client_socket
);
3232 PERROR("close client_socket sessiond poll");
3239 ERR("Health error occurred in %s", __func__
);
3241 health_unregister(health_consumerd
);
3243 rcu_unregister_thread();
3247 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3248 struct lttng_consumer_local_data
*ctx
)
3252 pthread_mutex_lock(&stream
->lock
);
3253 if (stream
->metadata_flag
) {
3254 pthread_mutex_lock(&stream
->metadata_rdv_lock
);
3257 switch (consumer_data
.type
) {
3258 case LTTNG_CONSUMER_KERNEL
:
3259 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
3261 case LTTNG_CONSUMER32_UST
:
3262 case LTTNG_CONSUMER64_UST
:
3263 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
3266 ERR("Unknown consumer_data type");
3272 if (stream
->metadata_flag
) {
3273 pthread_cond_broadcast(&stream
->metadata_rdv
);
3274 pthread_mutex_unlock(&stream
->metadata_rdv_lock
);
3276 pthread_mutex_unlock(&stream
->lock
);
3280 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3282 switch (consumer_data
.type
) {
3283 case LTTNG_CONSUMER_KERNEL
:
3284 return lttng_kconsumer_on_recv_stream(stream
);
3285 case LTTNG_CONSUMER32_UST
:
3286 case LTTNG_CONSUMER64_UST
:
3287 return lttng_ustconsumer_on_recv_stream(stream
);
3289 ERR("Unknown consumer_data type");
3296 * Allocate and set consumer data hash tables.
3298 int lttng_consumer_init(void)
3300 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3301 if (!consumer_data
.channel_ht
) {
3305 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3306 if (!consumer_data
.relayd_ht
) {
3310 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3311 if (!consumer_data
.stream_list_ht
) {
3315 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3316 if (!consumer_data
.stream_per_chan_id_ht
) {
3320 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3325 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3337 * Process the ADD_RELAYD command receive by a consumer.
3339 * This will create a relayd socket pair and add it to the relayd hash table.
3340 * The caller MUST acquire a RCU read side lock before calling it.
3342 int consumer_add_relayd_socket(uint64_t net_seq_idx
, int sock_type
,
3343 struct lttng_consumer_local_data
*ctx
, int sock
,
3344 struct pollfd
*consumer_sockpoll
,
3345 struct lttcomm_relayd_sock
*relayd_sock
, uint64_t sessiond_id
,
3346 uint64_t relayd_session_id
)
3348 int fd
= -1, ret
= -1, relayd_created
= 0;
3349 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3350 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3353 assert(relayd_sock
);
3355 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", net_seq_idx
);
3357 /* Get relayd reference if exists. */
3358 relayd
= consumer_find_relayd(net_seq_idx
);
3359 if (relayd
== NULL
) {
3360 assert(sock_type
== LTTNG_STREAM_CONTROL
);
3361 /* Not found. Allocate one. */
3362 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3363 if (relayd
== NULL
) {
3365 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3368 relayd
->sessiond_session_id
= sessiond_id
;
3373 * This code path MUST continue to the consumer send status message to
3374 * we can notify the session daemon and continue our work without
3375 * killing everything.
3379 * relayd key should never be found for control socket.
3381 assert(sock_type
!= LTTNG_STREAM_CONTROL
);
3384 /* First send a status message before receiving the fds. */
3385 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3387 /* Somehow, the session daemon is not responding anymore. */
3388 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3389 goto error_nosignal
;
3392 /* Poll on consumer socket. */
3393 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3395 /* Needing to exit in the middle of a command: error. */
3396 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3398 goto error_nosignal
;
3401 /* Get relayd socket from session daemon */
3402 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3403 if (ret
!= sizeof(fd
)) {
3405 fd
= -1; /* Just in case it gets set with an invalid value. */
3408 * Failing to receive FDs might indicate a major problem such as
3409 * reaching a fd limit during the receive where the kernel returns a
3410 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3411 * don't take any chances and stop everything.
3413 * XXX: Feature request #558 will fix that and avoid this possible
3414 * issue when reaching the fd limit.
3416 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3417 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3421 /* Copy socket information and received FD */
3422 switch (sock_type
) {
3423 case LTTNG_STREAM_CONTROL
:
3424 /* Copy received lttcomm socket */
3425 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3426 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3427 /* Handle create_sock error. */
3429 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3433 * Close the socket created internally by
3434 * lttcomm_create_sock, so we can replace it by the one
3435 * received from sessiond.
3437 if (close(relayd
->control_sock
.sock
.fd
)) {
3441 /* Assign new file descriptor */
3442 relayd
->control_sock
.sock
.fd
= fd
;
3443 fd
= -1; /* For error path */
3444 /* Assign version values. */
3445 relayd
->control_sock
.major
= relayd_sock
->major
;
3446 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3448 relayd
->relayd_session_id
= relayd_session_id
;
3451 case LTTNG_STREAM_DATA
:
3452 /* Copy received lttcomm socket */
3453 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3454 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3455 /* Handle create_sock error. */
3457 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3461 * Close the socket created internally by
3462 * lttcomm_create_sock, so we can replace it by the one
3463 * received from sessiond.
3465 if (close(relayd
->data_sock
.sock
.fd
)) {
3469 /* Assign new file descriptor */
3470 relayd
->data_sock
.sock
.fd
= fd
;
3471 fd
= -1; /* for eventual error paths */
3472 /* Assign version values. */
3473 relayd
->data_sock
.major
= relayd_sock
->major
;
3474 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3477 ERR("Unknown relayd socket type (%d)", sock_type
);
3479 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3483 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3484 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3485 relayd
->net_seq_idx
, fd
);
3487 /* We successfully added the socket. Send status back. */
3488 ret
= consumer_send_status_msg(sock
, ret_code
);
3490 /* Somehow, the session daemon is not responding anymore. */
3491 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3492 goto error_nosignal
;
3496 * Add relayd socket pair to consumer data hashtable. If object already
3497 * exists or on error, the function gracefully returns.
3505 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3506 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3510 /* Close received socket if valid. */
3513 PERROR("close received socket");
3517 if (relayd_created
) {
3525 * Try to lock the stream mutex.
3527 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
3529 static int stream_try_lock(struct lttng_consumer_stream
*stream
)
3536 * Try to lock the stream mutex. On failure, we know that the stream is
3537 * being used else where hence there is data still being extracted.
3539 ret
= pthread_mutex_trylock(&stream
->lock
);
3541 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
3553 * Search for a relayd associated to the session id and return the reference.
3555 * A rcu read side lock MUST be acquire before calling this function and locked
3556 * until the relayd object is no longer necessary.
3558 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3560 struct lttng_ht_iter iter
;
3561 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3563 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3564 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3567 * Check by sessiond id which is unique here where the relayd session
3568 * id might not be when having multiple relayd.
3570 if (relayd
->sessiond_session_id
== id
) {
3571 /* Found the relayd. There can be only one per id. */
3583 * Check if for a given session id there is still data needed to be extract
3586 * Return 1 if data is pending or else 0 meaning ready to be read.
3588 int consumer_data_pending(uint64_t id
)
3591 struct lttng_ht_iter iter
;
3592 struct lttng_ht
*ht
;
3593 struct lttng_consumer_stream
*stream
;
3594 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3595 int (*data_pending
)(struct lttng_consumer_stream
*);
3597 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3600 pthread_mutex_lock(&consumer_data
.lock
);
3602 switch (consumer_data
.type
) {
3603 case LTTNG_CONSUMER_KERNEL
:
3604 data_pending
= lttng_kconsumer_data_pending
;
3606 case LTTNG_CONSUMER32_UST
:
3607 case LTTNG_CONSUMER64_UST
:
3608 data_pending
= lttng_ustconsumer_data_pending
;
3611 ERR("Unknown consumer data type");
3615 /* Ease our life a bit */
3616 ht
= consumer_data
.stream_list_ht
;
3618 relayd
= find_relayd_by_session_id(id
);
3620 /* Send init command for data pending. */
3621 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3622 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3623 relayd
->relayd_session_id
);
3624 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3626 /* Communication error thus the relayd so no data pending. */
3627 goto data_not_pending
;
3631 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3632 ht
->hash_fct(&id
, lttng_ht_seed
),
3634 &iter
.iter
, stream
, node_session_id
.node
) {
3635 /* If this call fails, the stream is being used hence data pending. */
3636 ret
= stream_try_lock(stream
);
3642 * A removed node from the hash table indicates that the stream has
3643 * been deleted thus having a guarantee that the buffers are closed
3644 * on the consumer side. However, data can still be transmitted
3645 * over the network so don't skip the relayd check.
3647 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3649 /* Check the stream if there is data in the buffers. */
3650 ret
= data_pending(stream
);
3652 pthread_mutex_unlock(&stream
->lock
);
3659 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3660 if (stream
->metadata_flag
) {
3661 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3662 stream
->relayd_stream_id
);
3664 ret
= relayd_data_pending(&relayd
->control_sock
,
3665 stream
->relayd_stream_id
,
3666 stream
->next_net_seq_num
- 1);
3668 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3670 pthread_mutex_unlock(&stream
->lock
);
3674 pthread_mutex_unlock(&stream
->lock
);
3678 unsigned int is_data_inflight
= 0;
3680 /* Send init command for data pending. */
3681 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3682 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3683 relayd
->relayd_session_id
, &is_data_inflight
);
3684 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3686 goto data_not_pending
;
3688 if (is_data_inflight
) {
3694 * Finding _no_ node in the hash table and no inflight data means that the
3695 * stream(s) have been removed thus data is guaranteed to be available for
3696 * analysis from the trace files.
3700 /* Data is available to be read by a viewer. */
3701 pthread_mutex_unlock(&consumer_data
.lock
);
3706 /* Data is still being extracted from buffers. */
3707 pthread_mutex_unlock(&consumer_data
.lock
);
3713 * Send a ret code status message to the sessiond daemon.
3715 * Return the sendmsg() return value.
3717 int consumer_send_status_msg(int sock
, int ret_code
)
3719 struct lttcomm_consumer_status_msg msg
;
3721 memset(&msg
, 0, sizeof(msg
));
3722 msg
.ret_code
= ret_code
;
3724 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3728 * Send a channel status message to the sessiond daemon.
3730 * Return the sendmsg() return value.
3732 int consumer_send_status_channel(int sock
,
3733 struct lttng_consumer_channel
*channel
)
3735 struct lttcomm_consumer_status_channel msg
;
3739 memset(&msg
, 0, sizeof(msg
));
3741 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3743 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3744 msg
.key
= channel
->key
;
3745 msg
.stream_count
= channel
->streams
.count
;
3748 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3751 unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos
,
3752 unsigned long produced_pos
, uint64_t nb_packets_per_stream
,
3753 uint64_t max_sb_size
)
3755 unsigned long start_pos
;
3757 if (!nb_packets_per_stream
) {
3758 return consumed_pos
; /* Grab everything */
3760 start_pos
= produced_pos
- offset_align_floor(produced_pos
, max_sb_size
);
3761 start_pos
-= max_sb_size
* nb_packets_per_stream
;
3762 if ((long) (start_pos
- consumed_pos
) < 0) {
3763 return consumed_pos
; /* Grab everything */