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>
50 #include <common/consumer/consumer-metadata-cache.h>
52 struct lttng_consumer_global_data consumer_data
= {
55 .type
= LTTNG_CONSUMER_UNKNOWN
,
58 enum consumer_channel_action
{
61 CONSUMER_CHANNEL_QUIT
,
64 struct consumer_channel_msg
{
65 enum consumer_channel_action action
;
66 struct lttng_consumer_channel
*chan
; /* add */
67 uint64_t key
; /* del */
71 * Flag to inform the polling thread to quit when all fd hung up. Updated by
72 * the consumer_thread_receive_fds when it notices that all fds has hung up.
73 * Also updated by the signal handler (consumer_should_exit()). Read by the
76 volatile int consumer_quit
;
79 * Global hash table containing respectively metadata and data streams. The
80 * stream element in this ht should only be updated by the metadata poll thread
81 * for the metadata and the data poll thread for the data.
83 static struct lttng_ht
*metadata_ht
;
84 static struct lttng_ht
*data_ht
;
87 * Notify a thread lttng pipe to poll back again. This usually means that some
88 * global state has changed so we just send back the thread in a poll wait
91 static void notify_thread_lttng_pipe(struct lttng_pipe
*pipe
)
93 struct lttng_consumer_stream
*null_stream
= NULL
;
97 (void) lttng_pipe_write(pipe
, &null_stream
, sizeof(null_stream
));
100 static void notify_health_quit_pipe(int *pipe
)
104 ret
= lttng_write(pipe
[1], "4", 1);
106 PERROR("write consumer health quit");
110 static void notify_channel_pipe(struct lttng_consumer_local_data
*ctx
,
111 struct lttng_consumer_channel
*chan
,
113 enum consumer_channel_action action
)
115 struct consumer_channel_msg msg
;
118 memset(&msg
, 0, sizeof(msg
));
123 ret
= lttng_write(ctx
->consumer_channel_pipe
[1], &msg
, sizeof(msg
));
124 if (ret
< sizeof(msg
)) {
125 PERROR("notify_channel_pipe write error");
129 void notify_thread_del_channel(struct lttng_consumer_local_data
*ctx
,
132 notify_channel_pipe(ctx
, NULL
, key
, CONSUMER_CHANNEL_DEL
);
135 static int read_channel_pipe(struct lttng_consumer_local_data
*ctx
,
136 struct lttng_consumer_channel
**chan
,
138 enum consumer_channel_action
*action
)
140 struct consumer_channel_msg msg
;
143 ret
= lttng_read(ctx
->consumer_channel_pipe
[0], &msg
, sizeof(msg
));
144 if (ret
< sizeof(msg
)) {
148 *action
= msg
.action
;
156 * Cleanup the stream list of a channel. Those streams are not yet globally
159 static void clean_channel_stream_list(struct lttng_consumer_channel
*channel
)
161 struct lttng_consumer_stream
*stream
, *stmp
;
165 /* Delete streams that might have been left in the stream list. */
166 cds_list_for_each_entry_safe(stream
, stmp
, &channel
->streams
.head
,
168 cds_list_del(&stream
->send_node
);
170 * Once a stream is added to this list, the buffers were created so we
171 * have a guarantee that this call will succeed. Setting the monitor
172 * mode to 0 so we don't lock nor try to delete the stream from the
176 consumer_stream_destroy(stream
, NULL
);
181 * Find a stream. The consumer_data.lock must be locked during this
184 static struct lttng_consumer_stream
*find_stream(uint64_t key
,
187 struct lttng_ht_iter iter
;
188 struct lttng_ht_node_u64
*node
;
189 struct lttng_consumer_stream
*stream
= NULL
;
193 /* -1ULL keys are lookup failures */
194 if (key
== (uint64_t) -1ULL) {
200 lttng_ht_lookup(ht
, &key
, &iter
);
201 node
= lttng_ht_iter_get_node_u64(&iter
);
203 stream
= caa_container_of(node
, struct lttng_consumer_stream
, node
);
211 static void steal_stream_key(uint64_t key
, struct lttng_ht
*ht
)
213 struct lttng_consumer_stream
*stream
;
216 stream
= find_stream(key
, ht
);
218 stream
->key
= (uint64_t) -1ULL;
220 * We don't want the lookup to match, but we still need
221 * to iterate on this stream when iterating over the hash table. Just
222 * change the node key.
224 stream
->node
.key
= (uint64_t) -1ULL;
230 * Return a channel object for the given key.
232 * RCU read side lock MUST be acquired before calling this function and
233 * protects the channel ptr.
235 struct lttng_consumer_channel
*consumer_find_channel(uint64_t key
)
237 struct lttng_ht_iter iter
;
238 struct lttng_ht_node_u64
*node
;
239 struct lttng_consumer_channel
*channel
= NULL
;
241 /* -1ULL keys are lookup failures */
242 if (key
== (uint64_t) -1ULL) {
246 lttng_ht_lookup(consumer_data
.channel_ht
, &key
, &iter
);
247 node
= lttng_ht_iter_get_node_u64(&iter
);
249 channel
= caa_container_of(node
, struct lttng_consumer_channel
, node
);
256 * There is a possibility that the consumer does not have enough time between
257 * the close of the channel on the session daemon and the cleanup in here thus
258 * once we have a channel add with an existing key, we know for sure that this
259 * channel will eventually get cleaned up by all streams being closed.
261 * This function just nullifies the already existing channel key.
263 static void steal_channel_key(uint64_t key
)
265 struct lttng_consumer_channel
*channel
;
268 channel
= consumer_find_channel(key
);
270 channel
->key
= (uint64_t) -1ULL;
272 * We don't want the lookup to match, but we still need to iterate on
273 * this channel when iterating over the hash table. Just change the
276 channel
->node
.key
= (uint64_t) -1ULL;
281 static void free_channel_rcu(struct rcu_head
*head
)
283 struct lttng_ht_node_u64
*node
=
284 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
285 struct lttng_consumer_channel
*channel
=
286 caa_container_of(node
, struct lttng_consumer_channel
, node
);
288 switch (consumer_data
.type
) {
289 case LTTNG_CONSUMER_KERNEL
:
291 case LTTNG_CONSUMER32_UST
:
292 case LTTNG_CONSUMER64_UST
:
293 lttng_ustconsumer_free_channel(channel
);
296 ERR("Unknown consumer_data type");
303 * RCU protected relayd socket pair free.
305 static void free_relayd_rcu(struct rcu_head
*head
)
307 struct lttng_ht_node_u64
*node
=
308 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
309 struct consumer_relayd_sock_pair
*relayd
=
310 caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
313 * Close all sockets. This is done in the call RCU since we don't want the
314 * socket fds to be reassigned thus potentially creating bad state of the
317 * We do not have to lock the control socket mutex here since at this stage
318 * there is no one referencing to this relayd object.
320 (void) relayd_close(&relayd
->control_sock
);
321 (void) relayd_close(&relayd
->data_sock
);
327 * Destroy and free relayd socket pair object.
329 void consumer_destroy_relayd(struct consumer_relayd_sock_pair
*relayd
)
332 struct lttng_ht_iter iter
;
334 if (relayd
== NULL
) {
338 DBG("Consumer destroy and close relayd socket pair");
340 iter
.iter
.node
= &relayd
->node
.node
;
341 ret
= lttng_ht_del(consumer_data
.relayd_ht
, &iter
);
343 /* We assume the relayd is being or is destroyed */
347 /* RCU free() call */
348 call_rcu(&relayd
->node
.head
, free_relayd_rcu
);
352 * Remove a channel from the global list protected by a mutex. This function is
353 * also responsible for freeing its data structures.
355 void consumer_del_channel(struct lttng_consumer_channel
*channel
)
358 struct lttng_ht_iter iter
;
360 DBG("Consumer delete channel key %" PRIu64
, channel
->key
);
362 pthread_mutex_lock(&consumer_data
.lock
);
363 pthread_mutex_lock(&channel
->lock
);
365 /* Destroy streams that might have been left in the stream list. */
366 clean_channel_stream_list(channel
);
368 if (channel
->live_timer_enabled
== 1) {
369 consumer_timer_live_stop(channel
);
372 switch (consumer_data
.type
) {
373 case LTTNG_CONSUMER_KERNEL
:
375 case LTTNG_CONSUMER32_UST
:
376 case LTTNG_CONSUMER64_UST
:
377 lttng_ustconsumer_del_channel(channel
);
380 ERR("Unknown consumer_data type");
386 iter
.iter
.node
= &channel
->node
.node
;
387 ret
= lttng_ht_del(consumer_data
.channel_ht
, &iter
);
391 call_rcu(&channel
->node
.head
, free_channel_rcu
);
393 pthread_mutex_unlock(&channel
->lock
);
394 pthread_mutex_unlock(&consumer_data
.lock
);
398 * Iterate over the relayd hash table and destroy each element. Finally,
399 * destroy the whole hash table.
401 static void cleanup_relayd_ht(void)
403 struct lttng_ht_iter iter
;
404 struct consumer_relayd_sock_pair
*relayd
;
408 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
410 consumer_destroy_relayd(relayd
);
415 lttng_ht_destroy(consumer_data
.relayd_ht
);
419 * Update the end point status of all streams having the given network sequence
420 * index (relayd index).
422 * It's atomically set without having the stream mutex locked which is fine
423 * because we handle the write/read race with a pipe wakeup for each thread.
425 static void update_endpoint_status_by_netidx(uint64_t net_seq_idx
,
426 enum consumer_endpoint_status status
)
428 struct lttng_ht_iter iter
;
429 struct lttng_consumer_stream
*stream
;
431 DBG("Consumer set delete flag on stream by idx %" PRIu64
, net_seq_idx
);
435 /* Let's begin with metadata */
436 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
437 if (stream
->net_seq_idx
== net_seq_idx
) {
438 uatomic_set(&stream
->endpoint_status
, status
);
439 DBG("Delete flag set to metadata stream %d", stream
->wait_fd
);
443 /* Follow up by the data streams */
444 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
445 if (stream
->net_seq_idx
== net_seq_idx
) {
446 uatomic_set(&stream
->endpoint_status
, status
);
447 DBG("Delete flag set to data stream %d", stream
->wait_fd
);
454 * Cleanup a relayd object by flagging every associated streams for deletion,
455 * destroying the object meaning removing it from the relayd hash table,
456 * closing the sockets and freeing the memory in a RCU call.
458 * If a local data context is available, notify the threads that the streams'
459 * state have changed.
461 static void cleanup_relayd(struct consumer_relayd_sock_pair
*relayd
,
462 struct lttng_consumer_local_data
*ctx
)
468 DBG("Cleaning up relayd sockets");
470 /* Save the net sequence index before destroying the object */
471 netidx
= relayd
->net_seq_idx
;
474 * Delete the relayd from the relayd hash table, close the sockets and free
475 * the object in a RCU call.
477 consumer_destroy_relayd(relayd
);
479 /* Set inactive endpoint to all streams */
480 update_endpoint_status_by_netidx(netidx
, CONSUMER_ENDPOINT_INACTIVE
);
483 * With a local data context, notify the threads that the streams' state
484 * have changed. The write() action on the pipe acts as an "implicit"
485 * memory barrier ordering the updates of the end point status from the
486 * read of this status which happens AFTER receiving this notify.
489 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
490 notify_thread_lttng_pipe(ctx
->consumer_metadata_pipe
);
495 * Flag a relayd socket pair for destruction. Destroy it if the refcount
498 * RCU read side lock MUST be aquired before calling this function.
500 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair
*relayd
)
504 /* Set destroy flag for this object */
505 uatomic_set(&relayd
->destroy_flag
, 1);
507 /* Destroy the relayd if refcount is 0 */
508 if (uatomic_read(&relayd
->refcount
) == 0) {
509 consumer_destroy_relayd(relayd
);
514 * Completly destroy stream from every visiable data structure and the given
517 * One this call returns, the stream object is not longer usable nor visible.
519 void consumer_del_stream(struct lttng_consumer_stream
*stream
,
522 consumer_stream_destroy(stream
, ht
);
526 * XXX naming of del vs destroy is all mixed up.
528 void consumer_del_stream_for_data(struct lttng_consumer_stream
*stream
)
530 consumer_stream_destroy(stream
, data_ht
);
533 void consumer_del_stream_for_metadata(struct lttng_consumer_stream
*stream
)
535 consumer_stream_destroy(stream
, metadata_ht
);
538 struct lttng_consumer_stream
*consumer_allocate_stream(uint64_t channel_key
,
540 enum lttng_consumer_stream_state state
,
541 const char *channel_name
,
548 enum consumer_channel_type type
,
549 unsigned int monitor
)
552 struct lttng_consumer_stream
*stream
;
554 stream
= zmalloc(sizeof(*stream
));
555 if (stream
== NULL
) {
556 PERROR("malloc struct lttng_consumer_stream");
563 stream
->key
= stream_key
;
565 stream
->out_fd_offset
= 0;
566 stream
->output_written
= 0;
567 stream
->state
= state
;
570 stream
->net_seq_idx
= relayd_id
;
571 stream
->session_id
= session_id
;
572 stream
->monitor
= monitor
;
573 stream
->endpoint_status
= CONSUMER_ENDPOINT_ACTIVE
;
574 stream
->index_file
= NULL
;
575 stream
->last_sequence_number
= -1ULL;
576 pthread_mutex_init(&stream
->lock
, NULL
);
577 pthread_mutex_init(&stream
->metadata_timer_lock
, NULL
);
579 /* If channel is the metadata, flag this stream as metadata. */
580 if (type
== CONSUMER_CHANNEL_TYPE_METADATA
) {
581 stream
->metadata_flag
= 1;
582 /* Metadata is flat out. */
583 strncpy(stream
->name
, DEFAULT_METADATA_NAME
, sizeof(stream
->name
));
584 /* Live rendez-vous point. */
585 pthread_cond_init(&stream
->metadata_rdv
, NULL
);
586 pthread_mutex_init(&stream
->metadata_rdv_lock
, NULL
);
588 /* Format stream name to <channel_name>_<cpu_number> */
589 ret
= snprintf(stream
->name
, sizeof(stream
->name
), "%s_%d",
592 PERROR("snprintf stream name");
597 /* Key is always the wait_fd for streams. */
598 lttng_ht_node_init_u64(&stream
->node
, stream
->key
);
600 /* Init node per channel id key */
601 lttng_ht_node_init_u64(&stream
->node_channel_id
, channel_key
);
603 /* Init session id node with the stream session id */
604 lttng_ht_node_init_u64(&stream
->node_session_id
, stream
->session_id
);
606 DBG3("Allocated stream %s (key %" PRIu64
", chan_key %" PRIu64
607 " relayd_id %" PRIu64
", session_id %" PRIu64
,
608 stream
->name
, stream
->key
, channel_key
,
609 stream
->net_seq_idx
, stream
->session_id
);
625 * Add a stream to the global list protected by a mutex.
627 int consumer_add_data_stream(struct lttng_consumer_stream
*stream
)
629 struct lttng_ht
*ht
= data_ht
;
635 DBG3("Adding consumer stream %" PRIu64
, stream
->key
);
637 pthread_mutex_lock(&consumer_data
.lock
);
638 pthread_mutex_lock(&stream
->chan
->lock
);
639 pthread_mutex_lock(&stream
->chan
->timer_lock
);
640 pthread_mutex_lock(&stream
->lock
);
643 /* Steal stream identifier to avoid having streams with the same key */
644 steal_stream_key(stream
->key
, ht
);
646 lttng_ht_add_unique_u64(ht
, &stream
->node
);
648 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
649 &stream
->node_channel_id
);
652 * Add stream to the stream_list_ht of the consumer data. No need to steal
653 * the key since the HT does not use it and we allow to add redundant keys
656 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
659 * When nb_init_stream_left reaches 0, we don't need to trigger any action
660 * in terms of destroying the associated channel, because the action that
661 * causes the count to become 0 also causes a stream to be added. The
662 * channel deletion will thus be triggered by the following removal of this
665 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
666 /* Increment refcount before decrementing nb_init_stream_left */
668 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
671 /* Update consumer data once the node is inserted. */
672 consumer_data
.stream_count
++;
673 consumer_data
.need_update
= 1;
676 pthread_mutex_unlock(&stream
->lock
);
677 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
678 pthread_mutex_unlock(&stream
->chan
->lock
);
679 pthread_mutex_unlock(&consumer_data
.lock
);
684 void consumer_del_data_stream(struct lttng_consumer_stream
*stream
)
686 consumer_del_stream(stream
, data_ht
);
690 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
691 * be acquired before calling this.
693 static int add_relayd(struct consumer_relayd_sock_pair
*relayd
)
696 struct lttng_ht_node_u64
*node
;
697 struct lttng_ht_iter iter
;
701 lttng_ht_lookup(consumer_data
.relayd_ht
,
702 &relayd
->net_seq_idx
, &iter
);
703 node
= lttng_ht_iter_get_node_u64(&iter
);
707 lttng_ht_add_unique_u64(consumer_data
.relayd_ht
, &relayd
->node
);
714 * Allocate and return a consumer relayd socket.
716 static struct consumer_relayd_sock_pair
*consumer_allocate_relayd_sock_pair(
717 uint64_t net_seq_idx
)
719 struct consumer_relayd_sock_pair
*obj
= NULL
;
721 /* net sequence index of -1 is a failure */
722 if (net_seq_idx
== (uint64_t) -1ULL) {
726 obj
= zmalloc(sizeof(struct consumer_relayd_sock_pair
));
728 PERROR("zmalloc relayd sock");
732 obj
->net_seq_idx
= net_seq_idx
;
734 obj
->destroy_flag
= 0;
735 obj
->control_sock
.sock
.fd
= -1;
736 obj
->data_sock
.sock
.fd
= -1;
737 lttng_ht_node_init_u64(&obj
->node
, obj
->net_seq_idx
);
738 pthread_mutex_init(&obj
->ctrl_sock_mutex
, NULL
);
745 * Find a relayd socket pair in the global consumer data.
747 * Return the object if found else NULL.
748 * RCU read-side lock must be held across this call and while using the
751 struct consumer_relayd_sock_pair
*consumer_find_relayd(uint64_t key
)
753 struct lttng_ht_iter iter
;
754 struct lttng_ht_node_u64
*node
;
755 struct consumer_relayd_sock_pair
*relayd
= NULL
;
757 /* Negative keys are lookup failures */
758 if (key
== (uint64_t) -1ULL) {
762 lttng_ht_lookup(consumer_data
.relayd_ht
, &key
,
764 node
= lttng_ht_iter_get_node_u64(&iter
);
766 relayd
= caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
774 * Find a relayd and send the stream
776 * Returns 0 on success, < 0 on error
778 int consumer_send_relayd_stream(struct lttng_consumer_stream
*stream
,
782 struct consumer_relayd_sock_pair
*relayd
;
785 assert(stream
->net_seq_idx
!= -1ULL);
788 /* The stream is not metadata. Get relayd reference if exists. */
790 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
791 if (relayd
!= NULL
) {
792 /* Add stream on the relayd */
793 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
794 ret
= relayd_add_stream(&relayd
->control_sock
, stream
->name
,
795 path
, &stream
->relayd_stream_id
,
796 stream
->chan
->tracefile_size
, stream
->chan
->tracefile_count
);
797 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
802 uatomic_inc(&relayd
->refcount
);
803 stream
->sent_to_relayd
= 1;
805 ERR("Stream %" PRIu64
" relayd ID %" PRIu64
" unknown. Can't send it.",
806 stream
->key
, stream
->net_seq_idx
);
811 DBG("Stream %s with key %" PRIu64
" sent to relayd id %" PRIu64
,
812 stream
->name
, stream
->key
, stream
->net_seq_idx
);
820 * Find a relayd and send the streams sent message
822 * Returns 0 on success, < 0 on error
824 int consumer_send_relayd_streams_sent(uint64_t net_seq_idx
)
827 struct consumer_relayd_sock_pair
*relayd
;
829 assert(net_seq_idx
!= -1ULL);
831 /* The stream is not metadata. Get relayd reference if exists. */
833 relayd
= consumer_find_relayd(net_seq_idx
);
834 if (relayd
!= NULL
) {
835 /* Add stream on the relayd */
836 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
837 ret
= relayd_streams_sent(&relayd
->control_sock
);
838 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
843 ERR("Relayd ID %" PRIu64
" unknown. Can't send streams_sent.",
850 DBG("All streams sent relayd id %" PRIu64
, net_seq_idx
);
858 * Find a relayd and close the stream
860 void close_relayd_stream(struct lttng_consumer_stream
*stream
)
862 struct consumer_relayd_sock_pair
*relayd
;
864 /* The stream is not metadata. Get relayd reference if exists. */
866 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
868 consumer_stream_relayd_close(stream
, relayd
);
874 * Handle stream for relayd transmission if the stream applies for network
875 * streaming where the net sequence index is set.
877 * Return destination file descriptor or negative value on error.
879 static int write_relayd_stream_header(struct lttng_consumer_stream
*stream
,
880 size_t data_size
, unsigned long padding
,
881 struct consumer_relayd_sock_pair
*relayd
)
884 struct lttcomm_relayd_data_hdr data_hdr
;
890 /* Reset data header */
891 memset(&data_hdr
, 0, sizeof(data_hdr
));
893 if (stream
->metadata_flag
) {
894 /* Caller MUST acquire the relayd control socket lock */
895 ret
= relayd_send_metadata(&relayd
->control_sock
, data_size
);
900 /* Metadata are always sent on the control socket. */
901 outfd
= relayd
->control_sock
.sock
.fd
;
903 /* Set header with stream information */
904 data_hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
905 data_hdr
.data_size
= htobe32(data_size
);
906 data_hdr
.padding_size
= htobe32(padding
);
908 * Note that net_seq_num below is assigned with the *current* value of
909 * next_net_seq_num and only after that the next_net_seq_num will be
910 * increment. This is why when issuing a command on the relayd using
911 * this next value, 1 should always be substracted in order to compare
912 * the last seen sequence number on the relayd side to the last sent.
914 data_hdr
.net_seq_num
= htobe64(stream
->next_net_seq_num
);
915 /* Other fields are zeroed previously */
917 ret
= relayd_send_data_hdr(&relayd
->data_sock
, &data_hdr
,
923 ++stream
->next_net_seq_num
;
925 /* Set to go on data socket */
926 outfd
= relayd
->data_sock
.sock
.fd
;
934 * Allocate and return a new lttng_consumer_channel object using the given key
935 * to initialize the hash table node.
937 * On error, return NULL.
939 struct lttng_consumer_channel
*consumer_allocate_channel(uint64_t key
,
941 const char *pathname
,
946 enum lttng_event_output output
,
947 uint64_t tracefile_size
,
948 uint64_t tracefile_count
,
949 uint64_t session_id_per_pid
,
950 unsigned int monitor
,
951 unsigned int live_timer_interval
,
952 const char *root_shm_path
,
953 const char *shm_path
)
955 struct lttng_consumer_channel
*channel
;
957 channel
= zmalloc(sizeof(*channel
));
958 if (channel
== NULL
) {
959 PERROR("malloc struct lttng_consumer_channel");
964 channel
->refcount
= 0;
965 channel
->session_id
= session_id
;
966 channel
->session_id_per_pid
= session_id_per_pid
;
969 channel
->relayd_id
= relayd_id
;
970 channel
->tracefile_size
= tracefile_size
;
971 channel
->tracefile_count
= tracefile_count
;
972 channel
->monitor
= monitor
;
973 channel
->live_timer_interval
= live_timer_interval
;
974 pthread_mutex_init(&channel
->lock
, NULL
);
975 pthread_mutex_init(&channel
->timer_lock
, NULL
);
978 case LTTNG_EVENT_SPLICE
:
979 channel
->output
= CONSUMER_CHANNEL_SPLICE
;
981 case LTTNG_EVENT_MMAP
:
982 channel
->output
= CONSUMER_CHANNEL_MMAP
;
992 * In monitor mode, the streams associated with the channel will be put in
993 * a special list ONLY owned by this channel. So, the refcount is set to 1
994 * here meaning that the channel itself has streams that are referenced.
996 * On a channel deletion, once the channel is no longer visible, the
997 * refcount is decremented and checked for a zero value to delete it. With
998 * streams in no monitor mode, it will now be safe to destroy the channel.
1000 if (!channel
->monitor
) {
1001 channel
->refcount
= 1;
1004 strncpy(channel
->pathname
, pathname
, sizeof(channel
->pathname
));
1005 channel
->pathname
[sizeof(channel
->pathname
) - 1] = '\0';
1007 strncpy(channel
->name
, name
, sizeof(channel
->name
));
1008 channel
->name
[sizeof(channel
->name
) - 1] = '\0';
1010 if (root_shm_path
) {
1011 strncpy(channel
->root_shm_path
, root_shm_path
, sizeof(channel
->root_shm_path
));
1012 channel
->root_shm_path
[sizeof(channel
->root_shm_path
) - 1] = '\0';
1015 strncpy(channel
->shm_path
, shm_path
, sizeof(channel
->shm_path
));
1016 channel
->shm_path
[sizeof(channel
->shm_path
) - 1] = '\0';
1019 lttng_ht_node_init_u64(&channel
->node
, channel
->key
);
1021 channel
->wait_fd
= -1;
1023 CDS_INIT_LIST_HEAD(&channel
->streams
.head
);
1025 DBG("Allocated channel (key %" PRIu64
")", channel
->key
);
1032 * Add a channel to the global list protected by a mutex.
1034 * Always return 0 indicating success.
1036 int consumer_add_channel(struct lttng_consumer_channel
*channel
,
1037 struct lttng_consumer_local_data
*ctx
)
1039 pthread_mutex_lock(&consumer_data
.lock
);
1040 pthread_mutex_lock(&channel
->lock
);
1041 pthread_mutex_lock(&channel
->timer_lock
);
1044 * This gives us a guarantee that the channel we are about to add to the
1045 * channel hash table will be unique. See this function comment on the why
1046 * we need to steel the channel key at this stage.
1048 steal_channel_key(channel
->key
);
1051 lttng_ht_add_unique_u64(consumer_data
.channel_ht
, &channel
->node
);
1054 pthread_mutex_unlock(&channel
->timer_lock
);
1055 pthread_mutex_unlock(&channel
->lock
);
1056 pthread_mutex_unlock(&consumer_data
.lock
);
1058 if (channel
->wait_fd
!= -1 && channel
->type
== CONSUMER_CHANNEL_TYPE_DATA
) {
1059 notify_channel_pipe(ctx
, channel
, -1, CONSUMER_CHANNEL_ADD
);
1066 * Allocate the pollfd structure and the local view of the out fds to avoid
1067 * doing a lookup in the linked list and concurrency issues when writing is
1068 * needed. Called with consumer_data.lock held.
1070 * Returns the number of fds in the structures.
1072 static int update_poll_array(struct lttng_consumer_local_data
*ctx
,
1073 struct pollfd
**pollfd
, struct lttng_consumer_stream
**local_stream
,
1074 struct lttng_ht
*ht
)
1077 struct lttng_ht_iter iter
;
1078 struct lttng_consumer_stream
*stream
;
1083 assert(local_stream
);
1085 DBG("Updating poll fd array");
1087 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1089 * Only active streams with an active end point can be added to the
1090 * poll set and local stream storage of the thread.
1092 * There is a potential race here for endpoint_status to be updated
1093 * just after the check. However, this is OK since the stream(s) will
1094 * be deleted once the thread is notified that the end point state has
1095 * changed where this function will be called back again.
1097 if (stream
->state
!= LTTNG_CONSUMER_ACTIVE_STREAM
||
1098 stream
->endpoint_status
== CONSUMER_ENDPOINT_INACTIVE
) {
1102 * This clobbers way too much the debug output. Uncomment that if you
1103 * need it for debugging purposes.
1105 * DBG("Active FD %d", stream->wait_fd);
1107 (*pollfd
)[i
].fd
= stream
->wait_fd
;
1108 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1109 local_stream
[i
] = stream
;
1115 * Insert the consumer_data_pipe at the end of the array and don't
1116 * increment i so nb_fd is the number of real FD.
1118 (*pollfd
)[i
].fd
= lttng_pipe_get_readfd(ctx
->consumer_data_pipe
);
1119 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1121 (*pollfd
)[i
+ 1].fd
= lttng_pipe_get_readfd(ctx
->consumer_wakeup_pipe
);
1122 (*pollfd
)[i
+ 1].events
= POLLIN
| POLLPRI
;
1127 * Poll on the should_quit pipe and the command socket return -1 on
1128 * error, 1 if should exit, 0 if data is available on the command socket
1130 int lttng_consumer_poll_socket(struct pollfd
*consumer_sockpoll
)
1135 num_rdy
= poll(consumer_sockpoll
, 2, -1);
1136 if (num_rdy
== -1) {
1138 * Restart interrupted system call.
1140 if (errno
== EINTR
) {
1143 PERROR("Poll error");
1146 if (consumer_sockpoll
[0].revents
& (POLLIN
| POLLPRI
)) {
1147 DBG("consumer_should_quit wake up");
1154 * Set the error socket.
1156 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data
*ctx
,
1159 ctx
->consumer_error_socket
= sock
;
1163 * Set the command socket path.
1165 void lttng_consumer_set_command_sock_path(
1166 struct lttng_consumer_local_data
*ctx
, char *sock
)
1168 ctx
->consumer_command_sock_path
= sock
;
1172 * Send return code to the session daemon.
1173 * If the socket is not defined, we return 0, it is not a fatal error
1175 int lttng_consumer_send_error(struct lttng_consumer_local_data
*ctx
, int cmd
)
1177 if (ctx
->consumer_error_socket
> 0) {
1178 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
1179 sizeof(enum lttcomm_sessiond_command
));
1186 * Close all the tracefiles and stream fds and MUST be called when all
1187 * instances are destroyed i.e. when all threads were joined and are ended.
1189 void lttng_consumer_cleanup(void)
1191 struct lttng_ht_iter iter
;
1192 struct lttng_consumer_channel
*channel
;
1196 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
, channel
,
1198 consumer_del_channel(channel
);
1203 lttng_ht_destroy(consumer_data
.channel_ht
);
1205 cleanup_relayd_ht();
1207 lttng_ht_destroy(consumer_data
.stream_per_chan_id_ht
);
1210 * This HT contains streams that are freed by either the metadata thread or
1211 * the data thread so we do *nothing* on the hash table and simply destroy
1214 lttng_ht_destroy(consumer_data
.stream_list_ht
);
1218 * Called from signal handler.
1220 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
1225 ret
= lttng_write(ctx
->consumer_should_quit
[1], "4", 1);
1227 PERROR("write consumer quit");
1230 DBG("Consumer flag that it should quit");
1235 * Flush pending writes to trace output disk file.
1238 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1242 int outfd
= stream
->out_fd
;
1245 * This does a blocking write-and-wait on any page that belongs to the
1246 * subbuffer prior to the one we just wrote.
1247 * Don't care about error values, as these are just hints and ways to
1248 * limit the amount of page cache used.
1250 if (orig_offset
< stream
->max_sb_size
) {
1253 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1254 stream
->max_sb_size
,
1255 SYNC_FILE_RANGE_WAIT_BEFORE
1256 | SYNC_FILE_RANGE_WRITE
1257 | SYNC_FILE_RANGE_WAIT_AFTER
);
1259 * Give hints to the kernel about how we access the file:
1260 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1263 * We need to call fadvise again after the file grows because the
1264 * kernel does not seem to apply fadvise to non-existing parts of the
1267 * Call fadvise _after_ having waited for the page writeback to
1268 * complete because the dirty page writeback semantic is not well
1269 * defined. So it can be expected to lead to lower throughput in
1272 ret
= posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1273 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1274 if (ret
&& ret
!= -ENOSYS
) {
1276 PERROR("posix_fadvise on fd %i", outfd
);
1281 * Initialise the necessary environnement :
1282 * - create a new context
1283 * - create the poll_pipe
1284 * - create the should_quit pipe (for signal handler)
1285 * - create the thread pipe (for splice)
1287 * Takes a function pointer as argument, this function is called when data is
1288 * available on a buffer. This function is responsible to do the
1289 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1290 * buffer configuration and then kernctl_put_next_subbuf at the end.
1292 * Returns a pointer to the new context or NULL on error.
1294 struct lttng_consumer_local_data
*lttng_consumer_create(
1295 enum lttng_consumer_type type
,
1296 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1297 struct lttng_consumer_local_data
*ctx
),
1298 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1299 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1300 int (*update_stream
)(uint64_t stream_key
, uint32_t state
))
1303 struct lttng_consumer_local_data
*ctx
;
1305 assert(consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1306 consumer_data
.type
== type
);
1307 consumer_data
.type
= type
;
1309 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
1311 PERROR("allocating context");
1315 ctx
->consumer_error_socket
= -1;
1316 ctx
->consumer_metadata_socket
= -1;
1317 pthread_mutex_init(&ctx
->metadata_socket_lock
, NULL
);
1318 /* assign the callbacks */
1319 ctx
->on_buffer_ready
= buffer_ready
;
1320 ctx
->on_recv_channel
= recv_channel
;
1321 ctx
->on_recv_stream
= recv_stream
;
1322 ctx
->on_update_stream
= update_stream
;
1324 ctx
->consumer_data_pipe
= lttng_pipe_open(0);
1325 if (!ctx
->consumer_data_pipe
) {
1326 goto error_poll_pipe
;
1329 ctx
->consumer_wakeup_pipe
= lttng_pipe_open(0);
1330 if (!ctx
->consumer_wakeup_pipe
) {
1331 goto error_wakeup_pipe
;
1334 ret
= pipe(ctx
->consumer_should_quit
);
1336 PERROR("Error creating recv pipe");
1337 goto error_quit_pipe
;
1340 ret
= pipe(ctx
->consumer_channel_pipe
);
1342 PERROR("Error creating channel pipe");
1343 goto error_channel_pipe
;
1346 ctx
->consumer_metadata_pipe
= lttng_pipe_open(0);
1347 if (!ctx
->consumer_metadata_pipe
) {
1348 goto error_metadata_pipe
;
1353 error_metadata_pipe
:
1354 utils_close_pipe(ctx
->consumer_channel_pipe
);
1356 utils_close_pipe(ctx
->consumer_should_quit
);
1358 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1360 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1368 * Iterate over all streams of the hashtable and free them properly.
1370 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1372 struct lttng_ht_iter iter
;
1373 struct lttng_consumer_stream
*stream
;
1380 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1382 * Ignore return value since we are currently cleaning up so any error
1385 (void) consumer_del_stream(stream
, ht
);
1389 lttng_ht_destroy(ht
);
1393 * Iterate over all streams of the metadata hashtable and free them
1396 static void destroy_metadata_stream_ht(struct lttng_ht
*ht
)
1398 struct lttng_ht_iter iter
;
1399 struct lttng_consumer_stream
*stream
;
1406 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1408 * Ignore return value since we are currently cleaning up so any error
1411 (void) consumer_del_metadata_stream(stream
, ht
);
1415 lttng_ht_destroy(ht
);
1419 * Close all fds associated with the instance and free the context.
1421 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1425 DBG("Consumer destroying it. Closing everything.");
1431 destroy_data_stream_ht(data_ht
);
1432 destroy_metadata_stream_ht(metadata_ht
);
1434 ret
= close(ctx
->consumer_error_socket
);
1438 ret
= close(ctx
->consumer_metadata_socket
);
1442 utils_close_pipe(ctx
->consumer_channel_pipe
);
1443 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1444 lttng_pipe_destroy(ctx
->consumer_metadata_pipe
);
1445 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1446 utils_close_pipe(ctx
->consumer_should_quit
);
1448 unlink(ctx
->consumer_command_sock_path
);
1453 * Write the metadata stream id on the specified file descriptor.
1455 static int write_relayd_metadata_id(int fd
,
1456 struct lttng_consumer_stream
*stream
,
1457 struct consumer_relayd_sock_pair
*relayd
, unsigned long padding
)
1460 struct lttcomm_relayd_metadata_payload hdr
;
1462 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1463 hdr
.padding_size
= htobe32(padding
);
1464 ret
= lttng_write(fd
, (void *) &hdr
, sizeof(hdr
));
1465 if (ret
< sizeof(hdr
)) {
1467 * This error means that the fd's end is closed so ignore the PERROR
1468 * not to clubber the error output since this can happen in a normal
1471 if (errno
!= EPIPE
) {
1472 PERROR("write metadata stream id");
1474 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1476 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1477 * handle writting the missing part so report that as an error and
1478 * don't lie to the caller.
1483 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1484 stream
->relayd_stream_id
, padding
);
1491 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1492 * core function for writing trace buffers to either the local filesystem or
1495 * It must be called with the stream lock held.
1497 * Careful review MUST be put if any changes occur!
1499 * Returns the number of bytes written
1501 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1502 struct lttng_consumer_local_data
*ctx
,
1503 struct lttng_consumer_stream
*stream
, unsigned long len
,
1504 unsigned long padding
,
1505 struct ctf_packet_index
*index
)
1507 unsigned long mmap_offset
;
1510 off_t orig_offset
= stream
->out_fd_offset
;
1511 /* Default is on the disk */
1512 int outfd
= stream
->out_fd
;
1513 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1514 unsigned int relayd_hang_up
= 0;
1516 /* RCU lock for the relayd pointer */
1519 /* Flag that the current stream if set for network streaming. */
1520 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1521 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1522 if (relayd
== NULL
) {
1528 /* get the offset inside the fd to mmap */
1529 switch (consumer_data
.type
) {
1530 case LTTNG_CONSUMER_KERNEL
:
1531 mmap_base
= stream
->mmap_base
;
1532 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
);
2056 if (stream
->chan
->metadata_cache
) {
2057 /* Only applicable to userspace consumers. */
2058 pthread_mutex_lock(&stream
->chan
->metadata_cache
->lock
);
2061 /* Remove any reference to that stream. */
2062 consumer_stream_delete(stream
, ht
);
2064 /* Close down everything including the relayd if one. */
2065 consumer_stream_close(stream
);
2066 /* Destroy tracer buffers of the stream. */
2067 consumer_stream_destroy_buffers(stream
);
2069 /* Atomically decrement channel refcount since other threads can use it. */
2070 if (!uatomic_sub_return(&stream
->chan
->refcount
, 1)
2071 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
2072 /* Go for channel deletion! */
2073 free_chan
= stream
->chan
;
2077 * Nullify the stream reference so it is not used after deletion. The
2078 * channel lock MUST be acquired before being able to check for a NULL
2081 stream
->chan
->metadata_stream
= NULL
;
2083 if (stream
->chan
->metadata_cache
) {
2084 pthread_mutex_unlock(&stream
->chan
->metadata_cache
->lock
);
2086 pthread_mutex_unlock(&stream
->lock
);
2087 pthread_mutex_unlock(&stream
->chan
->lock
);
2088 pthread_mutex_unlock(&consumer_data
.lock
);
2091 consumer_del_channel(free_chan
);
2094 consumer_stream_free(stream
);
2098 * Action done with the metadata stream when adding it to the consumer internal
2099 * data structures to handle it.
2101 int consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
2103 struct lttng_ht
*ht
= metadata_ht
;
2105 struct lttng_ht_iter iter
;
2106 struct lttng_ht_node_u64
*node
;
2111 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2113 pthread_mutex_lock(&consumer_data
.lock
);
2114 pthread_mutex_lock(&stream
->chan
->lock
);
2115 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2116 pthread_mutex_lock(&stream
->lock
);
2119 * From here, refcounts are updated so be _careful_ when returning an error
2126 * Lookup the stream just to make sure it does not exist in our internal
2127 * state. This should NEVER happen.
2129 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2130 node
= lttng_ht_iter_get_node_u64(&iter
);
2134 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2135 * in terms of destroying the associated channel, because the action that
2136 * causes the count to become 0 also causes a stream to be added. The
2137 * channel deletion will thus be triggered by the following removal of this
2140 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2141 /* Increment refcount before decrementing nb_init_stream_left */
2143 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2146 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2148 lttng_ht_add_unique_u64(consumer_data
.stream_per_chan_id_ht
,
2149 &stream
->node_channel_id
);
2152 * Add stream to the stream_list_ht of the consumer data. No need to steal
2153 * the key since the HT does not use it and we allow to add redundant keys
2156 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2160 pthread_mutex_unlock(&stream
->lock
);
2161 pthread_mutex_unlock(&stream
->chan
->lock
);
2162 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2163 pthread_mutex_unlock(&consumer_data
.lock
);
2168 * Delete data stream that are flagged for deletion (endpoint_status).
2170 static void validate_endpoint_status_data_stream(void)
2172 struct lttng_ht_iter iter
;
2173 struct lttng_consumer_stream
*stream
;
2175 DBG("Consumer delete flagged data stream");
2178 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2179 /* Validate delete flag of the stream */
2180 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2183 /* Delete it right now */
2184 consumer_del_stream(stream
, data_ht
);
2190 * Delete metadata stream that are flagged for deletion (endpoint_status).
2192 static void validate_endpoint_status_metadata_stream(
2193 struct lttng_poll_event
*pollset
)
2195 struct lttng_ht_iter iter
;
2196 struct lttng_consumer_stream
*stream
;
2198 DBG("Consumer delete flagged metadata stream");
2203 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2204 /* Validate delete flag of the stream */
2205 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2209 * Remove from pollset so the metadata thread can continue without
2210 * blocking on a deleted stream.
2212 lttng_poll_del(pollset
, stream
->wait_fd
);
2214 /* Delete it right now */
2215 consumer_del_metadata_stream(stream
, metadata_ht
);
2221 * Thread polls on metadata file descriptor and write them on disk or on the
2224 void *consumer_thread_metadata_poll(void *data
)
2226 int ret
, i
, pollfd
, err
= -1;
2227 uint32_t revents
, nb_fd
;
2228 struct lttng_consumer_stream
*stream
= NULL
;
2229 struct lttng_ht_iter iter
;
2230 struct lttng_ht_node_u64
*node
;
2231 struct lttng_poll_event events
;
2232 struct lttng_consumer_local_data
*ctx
= data
;
2235 rcu_register_thread();
2237 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2239 if (testpoint(consumerd_thread_metadata
)) {
2240 goto error_testpoint
;
2243 health_code_update();
2245 DBG("Thread metadata poll started");
2247 /* Size is set to 1 for the consumer_metadata pipe */
2248 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2250 ERR("Poll set creation failed");
2254 ret
= lttng_poll_add(&events
,
2255 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2261 DBG("Metadata main loop started");
2265 health_code_update();
2266 health_poll_entry();
2267 DBG("Metadata poll wait");
2268 ret
= lttng_poll_wait(&events
, -1);
2269 DBG("Metadata poll return from wait with %d fd(s)",
2270 LTTNG_POLL_GETNB(&events
));
2272 DBG("Metadata event caught in thread");
2274 if (errno
== EINTR
) {
2275 ERR("Poll EINTR caught");
2278 if (LTTNG_POLL_GETNB(&events
) == 0) {
2279 err
= 0; /* All is OK */
2286 /* From here, the event is a metadata wait fd */
2287 for (i
= 0; i
< nb_fd
; i
++) {
2288 health_code_update();
2290 revents
= LTTNG_POLL_GETEV(&events
, i
);
2291 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2294 /* No activity for this FD (poll implementation). */
2298 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2299 if (revents
& LPOLLIN
) {
2302 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2303 &stream
, sizeof(stream
));
2304 if (pipe_len
< sizeof(stream
)) {
2306 PERROR("read metadata stream");
2309 * Remove the pipe from the poll set and continue the loop
2310 * since their might be data to consume.
2312 lttng_poll_del(&events
,
2313 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2314 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2318 /* A NULL stream means that the state has changed. */
2319 if (stream
== NULL
) {
2320 /* Check for deleted streams. */
2321 validate_endpoint_status_metadata_stream(&events
);
2325 DBG("Adding metadata stream %d to poll set",
2328 /* Add metadata stream to the global poll events list */
2329 lttng_poll_add(&events
, stream
->wait_fd
,
2330 LPOLLIN
| LPOLLPRI
| LPOLLHUP
);
2331 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2332 DBG("Metadata thread pipe hung up");
2334 * Remove the pipe from the poll set and continue the loop
2335 * since their might be data to consume.
2337 lttng_poll_del(&events
,
2338 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2339 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2342 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2346 /* Handle other stream */
2352 uint64_t tmp_id
= (uint64_t) pollfd
;
2354 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2356 node
= lttng_ht_iter_get_node_u64(&iter
);
2359 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2362 if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2363 /* Get the data out of the metadata file descriptor */
2364 DBG("Metadata available on fd %d", pollfd
);
2365 assert(stream
->wait_fd
== pollfd
);
2368 health_code_update();
2370 len
= ctx
->on_buffer_ready(stream
, ctx
);
2372 * We don't check the return value here since if we get
2373 * a negative len, it means an error occurred thus we
2374 * simply remove it from the poll set and free the
2379 /* It's ok to have an unavailable sub-buffer */
2380 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2381 /* Clean up stream from consumer and free it. */
2382 lttng_poll_del(&events
, stream
->wait_fd
);
2383 consumer_del_metadata_stream(stream
, metadata_ht
);
2385 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2386 DBG("Metadata fd %d is hup|err.", pollfd
);
2387 if (!stream
->hangup_flush_done
2388 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2389 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2390 DBG("Attempting to flush and consume the UST buffers");
2391 lttng_ustconsumer_on_stream_hangup(stream
);
2393 /* We just flushed the stream now read it. */
2395 health_code_update();
2397 len
= ctx
->on_buffer_ready(stream
, ctx
);
2399 * We don't check the return value here since if we get
2400 * a negative len, it means an error occurred thus we
2401 * simply remove it from the poll set and free the
2407 lttng_poll_del(&events
, stream
->wait_fd
);
2409 * This call update the channel states, closes file descriptors
2410 * and securely free the stream.
2412 consumer_del_metadata_stream(stream
, metadata_ht
);
2414 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2418 /* Release RCU lock for the stream looked up */
2426 DBG("Metadata poll thread exiting");
2428 lttng_poll_clean(&events
);
2433 ERR("Health error occurred in %s", __func__
);
2435 health_unregister(health_consumerd
);
2436 rcu_unregister_thread();
2441 * This thread polls the fds in the set to consume the data and write
2442 * it to tracefile if necessary.
2444 void *consumer_thread_data_poll(void *data
)
2446 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2447 struct pollfd
*pollfd
= NULL
;
2448 /* local view of the streams */
2449 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2450 /* local view of consumer_data.fds_count */
2452 struct lttng_consumer_local_data
*ctx
= data
;
2455 rcu_register_thread();
2457 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2459 if (testpoint(consumerd_thread_data
)) {
2460 goto error_testpoint
;
2463 health_code_update();
2465 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
*));
2466 if (local_stream
== NULL
) {
2467 PERROR("local_stream malloc");
2472 health_code_update();
2478 * the fds set has been updated, we need to update our
2479 * local array as well
2481 pthread_mutex_lock(&consumer_data
.lock
);
2482 if (consumer_data
.need_update
) {
2487 local_stream
= NULL
;
2490 * Allocate for all fds +1 for the consumer_data_pipe and +1 for
2493 pollfd
= zmalloc((consumer_data
.stream_count
+ 2) * sizeof(struct pollfd
));
2494 if (pollfd
== NULL
) {
2495 PERROR("pollfd malloc");
2496 pthread_mutex_unlock(&consumer_data
.lock
);
2500 local_stream
= zmalloc((consumer_data
.stream_count
+ 2) *
2501 sizeof(struct lttng_consumer_stream
*));
2502 if (local_stream
== NULL
) {
2503 PERROR("local_stream malloc");
2504 pthread_mutex_unlock(&consumer_data
.lock
);
2507 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2510 ERR("Error in allocating pollfd or local_outfds");
2511 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2512 pthread_mutex_unlock(&consumer_data
.lock
);
2516 consumer_data
.need_update
= 0;
2518 pthread_mutex_unlock(&consumer_data
.lock
);
2520 /* No FDs and consumer_quit, consumer_cleanup the thread */
2521 if (nb_fd
== 0 && consumer_quit
== 1) {
2522 err
= 0; /* All is OK */
2525 /* poll on the array of fds */
2527 DBG("polling on %d fd", nb_fd
+ 2);
2528 health_poll_entry();
2529 num_rdy
= poll(pollfd
, nb_fd
+ 2, -1);
2531 DBG("poll num_rdy : %d", num_rdy
);
2532 if (num_rdy
== -1) {
2534 * Restart interrupted system call.
2536 if (errno
== EINTR
) {
2539 PERROR("Poll error");
2540 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2542 } else if (num_rdy
== 0) {
2543 DBG("Polling thread timed out");
2548 * If the consumer_data_pipe triggered poll go directly to the
2549 * beginning of the loop to update the array. We want to prioritize
2550 * array update over low-priority reads.
2552 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2553 ssize_t pipe_readlen
;
2555 DBG("consumer_data_pipe wake up");
2556 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2557 &new_stream
, sizeof(new_stream
));
2558 if (pipe_readlen
< sizeof(new_stream
)) {
2559 PERROR("Consumer data pipe");
2560 /* Continue so we can at least handle the current stream(s). */
2565 * If the stream is NULL, just ignore it. It's also possible that
2566 * the sessiond poll thread changed the consumer_quit state and is
2567 * waking us up to test it.
2569 if (new_stream
== NULL
) {
2570 validate_endpoint_status_data_stream();
2574 /* Continue to update the local streams and handle prio ones */
2578 /* Handle wakeup pipe. */
2579 if (pollfd
[nb_fd
+ 1].revents
& (POLLIN
| POLLPRI
)) {
2581 ssize_t pipe_readlen
;
2583 pipe_readlen
= lttng_pipe_read(ctx
->consumer_wakeup_pipe
, &dummy
,
2585 if (pipe_readlen
< 0) {
2586 PERROR("Consumer data wakeup pipe");
2588 /* We've been awakened to handle stream(s). */
2589 ctx
->has_wakeup
= 0;
2592 /* Take care of high priority channels first. */
2593 for (i
= 0; i
< nb_fd
; i
++) {
2594 health_code_update();
2596 if (local_stream
[i
] == NULL
) {
2599 if (pollfd
[i
].revents
& POLLPRI
) {
2600 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2602 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2603 /* it's ok to have an unavailable sub-buffer */
2604 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2605 /* Clean the stream and free it. */
2606 consumer_del_stream(local_stream
[i
], data_ht
);
2607 local_stream
[i
] = NULL
;
2608 } else if (len
> 0) {
2609 local_stream
[i
]->data_read
= 1;
2615 * If we read high prio channel in this loop, try again
2616 * for more high prio data.
2622 /* Take care of low priority channels. */
2623 for (i
= 0; i
< nb_fd
; i
++) {
2624 health_code_update();
2626 if (local_stream
[i
] == NULL
) {
2629 if ((pollfd
[i
].revents
& POLLIN
) ||
2630 local_stream
[i
]->hangup_flush_done
||
2631 local_stream
[i
]->has_data
) {
2632 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2633 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2634 /* it's ok to have an unavailable sub-buffer */
2635 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2636 /* Clean the stream and free it. */
2637 consumer_del_stream(local_stream
[i
], data_ht
);
2638 local_stream
[i
] = NULL
;
2639 } else if (len
> 0) {
2640 local_stream
[i
]->data_read
= 1;
2645 /* Handle hangup and errors */
2646 for (i
= 0; i
< nb_fd
; i
++) {
2647 health_code_update();
2649 if (local_stream
[i
] == NULL
) {
2652 if (!local_stream
[i
]->hangup_flush_done
2653 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2654 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2655 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2656 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2658 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2659 /* Attempt read again, for the data we just flushed. */
2660 local_stream
[i
]->data_read
= 1;
2663 * If the poll flag is HUP/ERR/NVAL and we have
2664 * read no data in this pass, we can remove the
2665 * stream from its hash table.
2667 if ((pollfd
[i
].revents
& POLLHUP
)) {
2668 DBG("Polling fd %d tells it has hung up.", 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
& POLLERR
) {
2675 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2676 if (!local_stream
[i
]->data_read
) {
2677 consumer_del_stream(local_stream
[i
], data_ht
);
2678 local_stream
[i
] = NULL
;
2681 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2682 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2683 if (!local_stream
[i
]->data_read
) {
2684 consumer_del_stream(local_stream
[i
], data_ht
);
2685 local_stream
[i
] = NULL
;
2689 if (local_stream
[i
] != NULL
) {
2690 local_stream
[i
]->data_read
= 0;
2697 DBG("polling thread exiting");
2702 * Close the write side of the pipe so epoll_wait() in
2703 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2704 * read side of the pipe. If we close them both, epoll_wait strangely does
2705 * not return and could create a endless wait period if the pipe is the
2706 * only tracked fd in the poll set. The thread will take care of closing
2709 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2714 ERR("Health error occurred in %s", __func__
);
2716 health_unregister(health_consumerd
);
2718 rcu_unregister_thread();
2723 * Close wake-up end of each stream belonging to the channel. This will
2724 * allow the poll() on the stream read-side to detect when the
2725 * write-side (application) finally closes them.
2728 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2730 struct lttng_ht
*ht
;
2731 struct lttng_consumer_stream
*stream
;
2732 struct lttng_ht_iter iter
;
2734 ht
= consumer_data
.stream_per_chan_id_ht
;
2737 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2738 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2739 ht
->match_fct
, &channel
->key
,
2740 &iter
.iter
, stream
, node_channel_id
.node
) {
2742 * Protect against teardown with mutex.
2744 pthread_mutex_lock(&stream
->lock
);
2745 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2748 switch (consumer_data
.type
) {
2749 case LTTNG_CONSUMER_KERNEL
:
2751 case LTTNG_CONSUMER32_UST
:
2752 case LTTNG_CONSUMER64_UST
:
2753 if (stream
->metadata_flag
) {
2754 /* Safe and protected by the stream lock. */
2755 lttng_ustconsumer_close_metadata(stream
->chan
);
2758 * Note: a mutex is taken internally within
2759 * liblttng-ust-ctl to protect timer wakeup_fd
2760 * use from concurrent close.
2762 lttng_ustconsumer_close_stream_wakeup(stream
);
2766 ERR("Unknown consumer_data type");
2770 pthread_mutex_unlock(&stream
->lock
);
2775 static void destroy_channel_ht(struct lttng_ht
*ht
)
2777 struct lttng_ht_iter iter
;
2778 struct lttng_consumer_channel
*channel
;
2786 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2787 ret
= lttng_ht_del(ht
, &iter
);
2792 lttng_ht_destroy(ht
);
2796 * This thread polls the channel fds to detect when they are being
2797 * closed. It closes all related streams if the channel is detected as
2798 * closed. It is currently only used as a shim layer for UST because the
2799 * consumerd needs to keep the per-stream wakeup end of pipes open for
2802 void *consumer_thread_channel_poll(void *data
)
2804 int ret
, i
, pollfd
, err
= -1;
2805 uint32_t revents
, nb_fd
;
2806 struct lttng_consumer_channel
*chan
= NULL
;
2807 struct lttng_ht_iter iter
;
2808 struct lttng_ht_node_u64
*node
;
2809 struct lttng_poll_event events
;
2810 struct lttng_consumer_local_data
*ctx
= data
;
2811 struct lttng_ht
*channel_ht
;
2813 rcu_register_thread();
2815 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
2817 if (testpoint(consumerd_thread_channel
)) {
2818 goto error_testpoint
;
2821 health_code_update();
2823 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2825 /* ENOMEM at this point. Better to bail out. */
2829 DBG("Thread channel poll started");
2831 /* Size is set to 1 for the consumer_channel pipe */
2832 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2834 ERR("Poll set creation failed");
2838 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2844 DBG("Channel main loop started");
2848 health_code_update();
2849 DBG("Channel poll wait");
2850 health_poll_entry();
2851 ret
= lttng_poll_wait(&events
, -1);
2852 DBG("Channel poll return from wait with %d fd(s)",
2853 LTTNG_POLL_GETNB(&events
));
2855 DBG("Channel event caught in thread");
2857 if (errno
== EINTR
) {
2858 ERR("Poll EINTR caught");
2861 if (LTTNG_POLL_GETNB(&events
) == 0) {
2862 err
= 0; /* All is OK */
2869 /* From here, the event is a channel wait fd */
2870 for (i
= 0; i
< nb_fd
; i
++) {
2871 health_code_update();
2873 revents
= LTTNG_POLL_GETEV(&events
, i
);
2874 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2877 /* No activity for this FD (poll implementation). */
2881 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2882 if (revents
& LPOLLIN
) {
2883 enum consumer_channel_action action
;
2886 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
2889 ERR("Error reading channel pipe");
2891 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2896 case CONSUMER_CHANNEL_ADD
:
2897 DBG("Adding channel %d to poll set",
2900 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
2903 lttng_ht_add_unique_u64(channel_ht
,
2904 &chan
->wait_fd_node
);
2906 /* Add channel to the global poll events list */
2907 lttng_poll_add(&events
, chan
->wait_fd
,
2908 LPOLLERR
| LPOLLHUP
);
2910 case CONSUMER_CHANNEL_DEL
:
2913 * This command should never be called if the channel
2914 * has streams monitored by either the data or metadata
2915 * thread. The consumer only notify this thread with a
2916 * channel del. command if it receives a destroy
2917 * channel command from the session daemon that send it
2918 * if a command prior to the GET_CHANNEL failed.
2922 chan
= consumer_find_channel(key
);
2925 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
2928 lttng_poll_del(&events
, chan
->wait_fd
);
2929 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
2930 ret
= lttng_ht_del(channel_ht
, &iter
);
2933 switch (consumer_data
.type
) {
2934 case LTTNG_CONSUMER_KERNEL
:
2936 case LTTNG_CONSUMER32_UST
:
2937 case LTTNG_CONSUMER64_UST
:
2938 health_code_update();
2939 /* Destroy streams that might have been left in the stream list. */
2940 clean_channel_stream_list(chan
);
2943 ERR("Unknown consumer_data type");
2948 * Release our own refcount. Force channel deletion even if
2949 * streams were not initialized.
2951 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
2952 consumer_del_channel(chan
);
2957 case CONSUMER_CHANNEL_QUIT
:
2959 * Remove the pipe from the poll set and continue the loop
2960 * since their might be data to consume.
2962 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2965 ERR("Unknown action");
2968 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2969 DBG("Channel thread pipe hung up");
2971 * Remove the pipe from the poll set and continue the loop
2972 * since their might be data to consume.
2974 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2977 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2981 /* Handle other stream */
2987 uint64_t tmp_id
= (uint64_t) pollfd
;
2989 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
2991 node
= lttng_ht_iter_get_node_u64(&iter
);
2994 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
2997 /* Check for error event */
2998 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2999 DBG("Channel fd %d is hup|err.", pollfd
);
3001 lttng_poll_del(&events
, chan
->wait_fd
);
3002 ret
= lttng_ht_del(channel_ht
, &iter
);
3006 * This will close the wait fd for each stream associated to
3007 * this channel AND monitored by the data/metadata thread thus
3008 * will be clean by the right thread.
3010 consumer_close_channel_streams(chan
);
3012 /* Release our own refcount */
3013 if (!uatomic_sub_return(&chan
->refcount
, 1)
3014 && !uatomic_read(&chan
->nb_init_stream_left
)) {
3015 consumer_del_channel(chan
);
3018 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3023 /* Release RCU lock for the channel looked up */
3031 lttng_poll_clean(&events
);
3033 destroy_channel_ht(channel_ht
);
3036 DBG("Channel poll thread exiting");
3039 ERR("Health error occurred in %s", __func__
);
3041 health_unregister(health_consumerd
);
3042 rcu_unregister_thread();
3046 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
3047 struct pollfd
*sockpoll
, int client_socket
)
3054 ret
= lttng_consumer_poll_socket(sockpoll
);
3058 DBG("Metadata connection on client_socket");
3060 /* Blocking call, waiting for transmission */
3061 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
3062 if (ctx
->consumer_metadata_socket
< 0) {
3063 WARN("On accept metadata");
3074 * This thread listens on the consumerd socket and receives the file
3075 * descriptors from the session daemon.
3077 void *consumer_thread_sessiond_poll(void *data
)
3079 int sock
= -1, client_socket
, ret
, err
= -1;
3081 * structure to poll for incoming data on communication socket avoids
3082 * making blocking sockets.
3084 struct pollfd consumer_sockpoll
[2];
3085 struct lttng_consumer_local_data
*ctx
= data
;
3087 rcu_register_thread();
3089 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
3091 if (testpoint(consumerd_thread_sessiond
)) {
3092 goto error_testpoint
;
3095 health_code_update();
3097 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
3098 unlink(ctx
->consumer_command_sock_path
);
3099 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
3100 if (client_socket
< 0) {
3101 ERR("Cannot create command socket");
3105 ret
= lttcomm_listen_unix_sock(client_socket
);
3110 DBG("Sending ready command to lttng-sessiond");
3111 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3112 /* return < 0 on error, but == 0 is not fatal */
3114 ERR("Error sending ready command to lttng-sessiond");
3118 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3119 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3120 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3121 consumer_sockpoll
[1].fd
= client_socket
;
3122 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3124 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3132 DBG("Connection on client_socket");
3134 /* Blocking call, waiting for transmission */
3135 sock
= lttcomm_accept_unix_sock(client_socket
);
3142 * Setup metadata socket which is the second socket connection on the
3143 * command unix socket.
3145 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3154 /* This socket is not useful anymore. */
3155 ret
= close(client_socket
);
3157 PERROR("close client_socket");
3161 /* update the polling structure to poll on the established socket */
3162 consumer_sockpoll
[1].fd
= sock
;
3163 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3166 health_code_update();
3168 health_poll_entry();
3169 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3178 DBG("Incoming command on sock");
3179 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3182 * This could simply be a session daemon quitting. Don't output
3185 DBG("Communication interrupted on command socket");
3189 if (consumer_quit
) {
3190 DBG("consumer_thread_receive_fds received quit from signal");
3191 err
= 0; /* All is OK */
3194 DBG("received command on sock");
3200 DBG("Consumer thread sessiond poll exiting");
3203 * Close metadata streams since the producer is the session daemon which
3206 * NOTE: for now, this only applies to the UST tracer.
3208 lttng_consumer_close_all_metadata();
3211 * when all fds have hung up, the polling thread
3217 * Notify the data poll thread to poll back again and test the
3218 * consumer_quit state that we just set so to quit gracefully.
3220 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3222 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3224 notify_health_quit_pipe(health_quit_pipe
);
3226 /* Cleaning up possibly open sockets. */
3230 PERROR("close sock sessiond poll");
3233 if (client_socket
>= 0) {
3234 ret
= close(client_socket
);
3236 PERROR("close client_socket sessiond poll");
3243 ERR("Health error occurred in %s", __func__
);
3245 health_unregister(health_consumerd
);
3247 rcu_unregister_thread();
3251 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3252 struct lttng_consumer_local_data
*ctx
)
3256 pthread_mutex_lock(&stream
->lock
);
3257 if (stream
->metadata_flag
) {
3258 pthread_mutex_lock(&stream
->metadata_rdv_lock
);
3261 switch (consumer_data
.type
) {
3262 case LTTNG_CONSUMER_KERNEL
:
3263 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
3265 case LTTNG_CONSUMER32_UST
:
3266 case LTTNG_CONSUMER64_UST
:
3267 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
3270 ERR("Unknown consumer_data type");
3276 if (stream
->metadata_flag
) {
3277 pthread_cond_broadcast(&stream
->metadata_rdv
);
3278 pthread_mutex_unlock(&stream
->metadata_rdv_lock
);
3280 pthread_mutex_unlock(&stream
->lock
);
3284 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3286 switch (consumer_data
.type
) {
3287 case LTTNG_CONSUMER_KERNEL
:
3288 return lttng_kconsumer_on_recv_stream(stream
);
3289 case LTTNG_CONSUMER32_UST
:
3290 case LTTNG_CONSUMER64_UST
:
3291 return lttng_ustconsumer_on_recv_stream(stream
);
3293 ERR("Unknown consumer_data type");
3300 * Allocate and set consumer data hash tables.
3302 int lttng_consumer_init(void)
3304 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3305 if (!consumer_data
.channel_ht
) {
3309 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3310 if (!consumer_data
.relayd_ht
) {
3314 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3315 if (!consumer_data
.stream_list_ht
) {
3319 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3320 if (!consumer_data
.stream_per_chan_id_ht
) {
3324 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3329 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3341 * Process the ADD_RELAYD command receive by a consumer.
3343 * This will create a relayd socket pair and add it to the relayd hash table.
3344 * The caller MUST acquire a RCU read side lock before calling it.
3346 int consumer_add_relayd_socket(uint64_t net_seq_idx
, int sock_type
,
3347 struct lttng_consumer_local_data
*ctx
, int sock
,
3348 struct pollfd
*consumer_sockpoll
,
3349 struct lttcomm_relayd_sock
*relayd_sock
, uint64_t sessiond_id
,
3350 uint64_t relayd_session_id
)
3352 int fd
= -1, ret
= -1, relayd_created
= 0;
3353 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3354 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3357 assert(relayd_sock
);
3359 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", net_seq_idx
);
3361 /* Get relayd reference if exists. */
3362 relayd
= consumer_find_relayd(net_seq_idx
);
3363 if (relayd
== NULL
) {
3364 assert(sock_type
== LTTNG_STREAM_CONTROL
);
3365 /* Not found. Allocate one. */
3366 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3367 if (relayd
== NULL
) {
3369 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3372 relayd
->sessiond_session_id
= sessiond_id
;
3377 * This code path MUST continue to the consumer send status message to
3378 * we can notify the session daemon and continue our work without
3379 * killing everything.
3383 * relayd key should never be found for control socket.
3385 assert(sock_type
!= LTTNG_STREAM_CONTROL
);
3388 /* First send a status message before receiving the fds. */
3389 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3391 /* Somehow, the session daemon is not responding anymore. */
3392 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3393 goto error_nosignal
;
3396 /* Poll on consumer socket. */
3397 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3399 /* Needing to exit in the middle of a command: error. */
3400 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3402 goto error_nosignal
;
3405 /* Get relayd socket from session daemon */
3406 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3407 if (ret
!= sizeof(fd
)) {
3409 fd
= -1; /* Just in case it gets set with an invalid value. */
3412 * Failing to receive FDs might indicate a major problem such as
3413 * reaching a fd limit during the receive where the kernel returns a
3414 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3415 * don't take any chances and stop everything.
3417 * XXX: Feature request #558 will fix that and avoid this possible
3418 * issue when reaching the fd limit.
3420 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3421 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3425 /* Copy socket information and received FD */
3426 switch (sock_type
) {
3427 case LTTNG_STREAM_CONTROL
:
3428 /* Copy received lttcomm socket */
3429 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3430 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3431 /* Handle create_sock error. */
3433 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3437 * Close the socket created internally by
3438 * lttcomm_create_sock, so we can replace it by the one
3439 * received from sessiond.
3441 if (close(relayd
->control_sock
.sock
.fd
)) {
3445 /* Assign new file descriptor */
3446 relayd
->control_sock
.sock
.fd
= fd
;
3447 fd
= -1; /* For error path */
3448 /* Assign version values. */
3449 relayd
->control_sock
.major
= relayd_sock
->major
;
3450 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3452 relayd
->relayd_session_id
= relayd_session_id
;
3455 case LTTNG_STREAM_DATA
:
3456 /* Copy received lttcomm socket */
3457 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3458 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3459 /* Handle create_sock error. */
3461 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3465 * Close the socket created internally by
3466 * lttcomm_create_sock, so we can replace it by the one
3467 * received from sessiond.
3469 if (close(relayd
->data_sock
.sock
.fd
)) {
3473 /* Assign new file descriptor */
3474 relayd
->data_sock
.sock
.fd
= fd
;
3475 fd
= -1; /* for eventual error paths */
3476 /* Assign version values. */
3477 relayd
->data_sock
.major
= relayd_sock
->major
;
3478 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3481 ERR("Unknown relayd socket type (%d)", sock_type
);
3483 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3487 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3488 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3489 relayd
->net_seq_idx
, fd
);
3491 /* We successfully added the socket. Send status back. */
3492 ret
= consumer_send_status_msg(sock
, ret_code
);
3494 /* Somehow, the session daemon is not responding anymore. */
3495 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3496 goto error_nosignal
;
3500 * Add relayd socket pair to consumer data hashtable. If object already
3501 * exists or on error, the function gracefully returns.
3509 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3510 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3514 /* Close received socket if valid. */
3517 PERROR("close received socket");
3521 if (relayd_created
) {
3529 * Try to lock the stream mutex.
3531 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
3533 static int stream_try_lock(struct lttng_consumer_stream
*stream
)
3540 * Try to lock the stream mutex. On failure, we know that the stream is
3541 * being used else where hence there is data still being extracted.
3543 ret
= pthread_mutex_trylock(&stream
->lock
);
3545 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
3557 * Search for a relayd associated to the session id and return the reference.
3559 * A rcu read side lock MUST be acquire before calling this function and locked
3560 * until the relayd object is no longer necessary.
3562 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3564 struct lttng_ht_iter iter
;
3565 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3567 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3568 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3571 * Check by sessiond id which is unique here where the relayd session
3572 * id might not be when having multiple relayd.
3574 if (relayd
->sessiond_session_id
== id
) {
3575 /* Found the relayd. There can be only one per id. */
3587 * Check if for a given session id there is still data needed to be extract
3590 * Return 1 if data is pending or else 0 meaning ready to be read.
3592 int consumer_data_pending(uint64_t id
)
3595 struct lttng_ht_iter iter
;
3596 struct lttng_ht
*ht
;
3597 struct lttng_consumer_stream
*stream
;
3598 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3599 int (*data_pending
)(struct lttng_consumer_stream
*);
3601 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3604 pthread_mutex_lock(&consumer_data
.lock
);
3606 switch (consumer_data
.type
) {
3607 case LTTNG_CONSUMER_KERNEL
:
3608 data_pending
= lttng_kconsumer_data_pending
;
3610 case LTTNG_CONSUMER32_UST
:
3611 case LTTNG_CONSUMER64_UST
:
3612 data_pending
= lttng_ustconsumer_data_pending
;
3615 ERR("Unknown consumer data type");
3619 /* Ease our life a bit */
3620 ht
= consumer_data
.stream_list_ht
;
3622 relayd
= find_relayd_by_session_id(id
);
3624 /* Send init command for data pending. */
3625 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3626 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3627 relayd
->relayd_session_id
);
3628 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3630 /* Communication error thus the relayd so no data pending. */
3631 goto data_not_pending
;
3635 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3636 ht
->hash_fct(&id
, lttng_ht_seed
),
3638 &iter
.iter
, stream
, node_session_id
.node
) {
3639 /* If this call fails, the stream is being used hence data pending. */
3640 ret
= stream_try_lock(stream
);
3646 * A removed node from the hash table indicates that the stream has
3647 * been deleted thus having a guarantee that the buffers are closed
3648 * on the consumer side. However, data can still be transmitted
3649 * over the network so don't skip the relayd check.
3651 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3653 /* Check the stream if there is data in the buffers. */
3654 ret
= data_pending(stream
);
3656 pthread_mutex_unlock(&stream
->lock
);
3663 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3664 if (stream
->metadata_flag
) {
3665 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3666 stream
->relayd_stream_id
);
3668 ret
= relayd_data_pending(&relayd
->control_sock
,
3669 stream
->relayd_stream_id
,
3670 stream
->next_net_seq_num
- 1);
3672 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3674 pthread_mutex_unlock(&stream
->lock
);
3678 pthread_mutex_unlock(&stream
->lock
);
3682 unsigned int is_data_inflight
= 0;
3684 /* Send init command for data pending. */
3685 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3686 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3687 relayd
->relayd_session_id
, &is_data_inflight
);
3688 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3690 goto data_not_pending
;
3692 if (is_data_inflight
) {
3698 * Finding _no_ node in the hash table and no inflight data means that the
3699 * stream(s) have been removed thus data is guaranteed to be available for
3700 * analysis from the trace files.
3704 /* Data is available to be read by a viewer. */
3705 pthread_mutex_unlock(&consumer_data
.lock
);
3710 /* Data is still being extracted from buffers. */
3711 pthread_mutex_unlock(&consumer_data
.lock
);
3717 * Send a ret code status message to the sessiond daemon.
3719 * Return the sendmsg() return value.
3721 int consumer_send_status_msg(int sock
, int ret_code
)
3723 struct lttcomm_consumer_status_msg msg
;
3725 memset(&msg
, 0, sizeof(msg
));
3726 msg
.ret_code
= ret_code
;
3728 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3732 * Send a channel status message to the sessiond daemon.
3734 * Return the sendmsg() return value.
3736 int consumer_send_status_channel(int sock
,
3737 struct lttng_consumer_channel
*channel
)
3739 struct lttcomm_consumer_status_channel msg
;
3743 memset(&msg
, 0, sizeof(msg
));
3745 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3747 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3748 msg
.key
= channel
->key
;
3749 msg
.stream_count
= channel
->streams
.count
;
3752 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3755 unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos
,
3756 unsigned long produced_pos
, uint64_t nb_packets_per_stream
,
3757 uint64_t max_sb_size
)
3759 unsigned long start_pos
;
3761 if (!nb_packets_per_stream
) {
3762 return consumed_pos
; /* Grab everything */
3764 start_pos
= produced_pos
- offset_align_floor(produced_pos
, max_sb_size
);
3765 start_pos
-= max_sb_size
* nb_packets_per_stream
;
3766 if ((long) (start_pos
- consumed_pos
) < 0) {
3767 return consumed_pos
; /* Grab everything */