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
, int *nb_inactive_fd
)
1077 struct lttng_ht_iter iter
;
1078 struct lttng_consumer_stream
*stream
;
1083 assert(local_stream
);
1085 DBG("Updating poll fd array");
1086 *nb_inactive_fd
= 0;
1088 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1090 * Only active streams with an active end point can be added to the
1091 * poll set and local stream storage of the thread.
1093 * There is a potential race here for endpoint_status to be updated
1094 * just after the check. However, this is OK since the stream(s) will
1095 * be deleted once the thread is notified that the end point state has
1096 * changed where this function will be called back again.
1098 * We track the number of inactive FDs because they still need to be
1099 * closed by the polling thread after a wakeup on the data_pipe or
1102 if (stream
->state
!= LTTNG_CONSUMER_ACTIVE_STREAM
||
1103 stream
->endpoint_status
== CONSUMER_ENDPOINT_INACTIVE
) {
1104 (*nb_inactive_fd
)++;
1108 * This clobbers way too much the debug output. Uncomment that if you
1109 * need it for debugging purposes.
1111 * DBG("Active FD %d", stream->wait_fd);
1113 (*pollfd
)[i
].fd
= stream
->wait_fd
;
1114 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1115 local_stream
[i
] = stream
;
1121 * Insert the consumer_data_pipe at the end of the array and don't
1122 * increment i so nb_fd is the number of real FD.
1124 (*pollfd
)[i
].fd
= lttng_pipe_get_readfd(ctx
->consumer_data_pipe
);
1125 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1127 (*pollfd
)[i
+ 1].fd
= lttng_pipe_get_readfd(ctx
->consumer_wakeup_pipe
);
1128 (*pollfd
)[i
+ 1].events
= POLLIN
| POLLPRI
;
1133 * Poll on the should_quit pipe and the command socket return -1 on
1134 * error, 1 if should exit, 0 if data is available on the command socket
1136 int lttng_consumer_poll_socket(struct pollfd
*consumer_sockpoll
)
1141 num_rdy
= poll(consumer_sockpoll
, 2, -1);
1142 if (num_rdy
== -1) {
1144 * Restart interrupted system call.
1146 if (errno
== EINTR
) {
1149 PERROR("Poll error");
1152 if (consumer_sockpoll
[0].revents
& (POLLIN
| POLLPRI
)) {
1153 DBG("consumer_should_quit wake up");
1160 * Set the error socket.
1162 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data
*ctx
,
1165 ctx
->consumer_error_socket
= sock
;
1169 * Set the command socket path.
1171 void lttng_consumer_set_command_sock_path(
1172 struct lttng_consumer_local_data
*ctx
, char *sock
)
1174 ctx
->consumer_command_sock_path
= sock
;
1178 * Send return code to the session daemon.
1179 * If the socket is not defined, we return 0, it is not a fatal error
1181 int lttng_consumer_send_error(struct lttng_consumer_local_data
*ctx
, int cmd
)
1183 if (ctx
->consumer_error_socket
> 0) {
1184 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
1185 sizeof(enum lttcomm_sessiond_command
));
1192 * Close all the tracefiles and stream fds and MUST be called when all
1193 * instances are destroyed i.e. when all threads were joined and are ended.
1195 void lttng_consumer_cleanup(void)
1197 struct lttng_ht_iter iter
;
1198 struct lttng_consumer_channel
*channel
;
1202 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
, channel
,
1204 consumer_del_channel(channel
);
1209 lttng_ht_destroy(consumer_data
.channel_ht
);
1211 cleanup_relayd_ht();
1213 lttng_ht_destroy(consumer_data
.stream_per_chan_id_ht
);
1216 * This HT contains streams that are freed by either the metadata thread or
1217 * the data thread so we do *nothing* on the hash table and simply destroy
1220 lttng_ht_destroy(consumer_data
.stream_list_ht
);
1224 * Called from signal handler.
1226 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
1231 ret
= lttng_write(ctx
->consumer_should_quit
[1], "4", 1);
1233 PERROR("write consumer quit");
1236 DBG("Consumer flag that it should quit");
1241 * Flush pending writes to trace output disk file.
1244 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1248 int outfd
= stream
->out_fd
;
1251 * This does a blocking write-and-wait on any page that belongs to the
1252 * subbuffer prior to the one we just wrote.
1253 * Don't care about error values, as these are just hints and ways to
1254 * limit the amount of page cache used.
1256 if (orig_offset
< stream
->max_sb_size
) {
1259 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1260 stream
->max_sb_size
,
1261 SYNC_FILE_RANGE_WAIT_BEFORE
1262 | SYNC_FILE_RANGE_WRITE
1263 | SYNC_FILE_RANGE_WAIT_AFTER
);
1265 * Give hints to the kernel about how we access the file:
1266 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1269 * We need to call fadvise again after the file grows because the
1270 * kernel does not seem to apply fadvise to non-existing parts of the
1273 * Call fadvise _after_ having waited for the page writeback to
1274 * complete because the dirty page writeback semantic is not well
1275 * defined. So it can be expected to lead to lower throughput in
1278 ret
= posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1279 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1280 if (ret
&& ret
!= -ENOSYS
) {
1282 PERROR("posix_fadvise on fd %i", outfd
);
1287 * Initialise the necessary environnement :
1288 * - create a new context
1289 * - create the poll_pipe
1290 * - create the should_quit pipe (for signal handler)
1291 * - create the thread pipe (for splice)
1293 * Takes a function pointer as argument, this function is called when data is
1294 * available on a buffer. This function is responsible to do the
1295 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1296 * buffer configuration and then kernctl_put_next_subbuf at the end.
1298 * Returns a pointer to the new context or NULL on error.
1300 struct lttng_consumer_local_data
*lttng_consumer_create(
1301 enum lttng_consumer_type type
,
1302 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1303 struct lttng_consumer_local_data
*ctx
),
1304 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1305 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1306 int (*update_stream
)(uint64_t stream_key
, uint32_t state
))
1309 struct lttng_consumer_local_data
*ctx
;
1311 assert(consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1312 consumer_data
.type
== type
);
1313 consumer_data
.type
= type
;
1315 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
1317 PERROR("allocating context");
1321 ctx
->consumer_error_socket
= -1;
1322 ctx
->consumer_metadata_socket
= -1;
1323 pthread_mutex_init(&ctx
->metadata_socket_lock
, NULL
);
1324 /* assign the callbacks */
1325 ctx
->on_buffer_ready
= buffer_ready
;
1326 ctx
->on_recv_channel
= recv_channel
;
1327 ctx
->on_recv_stream
= recv_stream
;
1328 ctx
->on_update_stream
= update_stream
;
1330 ctx
->consumer_data_pipe
= lttng_pipe_open(0);
1331 if (!ctx
->consumer_data_pipe
) {
1332 goto error_poll_pipe
;
1335 ctx
->consumer_wakeup_pipe
= lttng_pipe_open(0);
1336 if (!ctx
->consumer_wakeup_pipe
) {
1337 goto error_wakeup_pipe
;
1340 ret
= pipe(ctx
->consumer_should_quit
);
1342 PERROR("Error creating recv pipe");
1343 goto error_quit_pipe
;
1346 ret
= pipe(ctx
->consumer_channel_pipe
);
1348 PERROR("Error creating channel pipe");
1349 goto error_channel_pipe
;
1352 ctx
->consumer_metadata_pipe
= lttng_pipe_open(0);
1353 if (!ctx
->consumer_metadata_pipe
) {
1354 goto error_metadata_pipe
;
1359 error_metadata_pipe
:
1360 utils_close_pipe(ctx
->consumer_channel_pipe
);
1362 utils_close_pipe(ctx
->consumer_should_quit
);
1364 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1366 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1374 * Iterate over all streams of the hashtable and free them properly.
1376 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1378 struct lttng_ht_iter iter
;
1379 struct lttng_consumer_stream
*stream
;
1386 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1388 * Ignore return value since we are currently cleaning up so any error
1391 (void) consumer_del_stream(stream
, ht
);
1395 lttng_ht_destroy(ht
);
1399 * Iterate over all streams of the metadata hashtable and free them
1402 static void destroy_metadata_stream_ht(struct lttng_ht
*ht
)
1404 struct lttng_ht_iter iter
;
1405 struct lttng_consumer_stream
*stream
;
1412 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1414 * Ignore return value since we are currently cleaning up so any error
1417 (void) consumer_del_metadata_stream(stream
, ht
);
1421 lttng_ht_destroy(ht
);
1425 * Close all fds associated with the instance and free the context.
1427 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1431 DBG("Consumer destroying it. Closing everything.");
1437 destroy_data_stream_ht(data_ht
);
1438 destroy_metadata_stream_ht(metadata_ht
);
1440 ret
= close(ctx
->consumer_error_socket
);
1444 ret
= close(ctx
->consumer_metadata_socket
);
1448 utils_close_pipe(ctx
->consumer_channel_pipe
);
1449 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1450 lttng_pipe_destroy(ctx
->consumer_metadata_pipe
);
1451 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1452 utils_close_pipe(ctx
->consumer_should_quit
);
1454 unlink(ctx
->consumer_command_sock_path
);
1459 * Write the metadata stream id on the specified file descriptor.
1461 static int write_relayd_metadata_id(int fd
,
1462 struct lttng_consumer_stream
*stream
,
1463 struct consumer_relayd_sock_pair
*relayd
, unsigned long padding
)
1466 struct lttcomm_relayd_metadata_payload hdr
;
1468 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1469 hdr
.padding_size
= htobe32(padding
);
1470 ret
= lttng_write(fd
, (void *) &hdr
, sizeof(hdr
));
1471 if (ret
< sizeof(hdr
)) {
1473 * This error means that the fd's end is closed so ignore the PERROR
1474 * not to clubber the error output since this can happen in a normal
1477 if (errno
!= EPIPE
) {
1478 PERROR("write metadata stream id");
1480 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1482 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1483 * handle writting the missing part so report that as an error and
1484 * don't lie to the caller.
1489 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1490 stream
->relayd_stream_id
, padding
);
1497 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1498 * core function for writing trace buffers to either the local filesystem or
1501 * It must be called with the stream lock held.
1503 * Careful review MUST be put if any changes occur!
1505 * Returns the number of bytes written
1507 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1508 struct lttng_consumer_local_data
*ctx
,
1509 struct lttng_consumer_stream
*stream
, unsigned long len
,
1510 unsigned long padding
,
1511 struct ctf_packet_index
*index
)
1513 unsigned long mmap_offset
;
1516 off_t orig_offset
= stream
->out_fd_offset
;
1517 /* Default is on the disk */
1518 int outfd
= stream
->out_fd
;
1519 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1520 unsigned int relayd_hang_up
= 0;
1522 /* RCU lock for the relayd pointer */
1525 /* Flag that the current stream if set for network streaming. */
1526 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1527 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1528 if (relayd
== NULL
) {
1534 /* get the offset inside the fd to mmap */
1535 switch (consumer_data
.type
) {
1536 case LTTNG_CONSUMER_KERNEL
:
1537 mmap_base
= stream
->mmap_base
;
1538 ret
= kernctl_get_mmap_read_offset(stream
->wait_fd
, &mmap_offset
);
1540 PERROR("tracer ctl get_mmap_read_offset");
1544 case LTTNG_CONSUMER32_UST
:
1545 case LTTNG_CONSUMER64_UST
:
1546 mmap_base
= lttng_ustctl_get_mmap_base(stream
);
1548 ERR("read mmap get mmap base for stream %s", stream
->name
);
1552 ret
= lttng_ustctl_get_mmap_read_offset(stream
, &mmap_offset
);
1554 PERROR("tracer ctl get_mmap_read_offset");
1560 ERR("Unknown consumer_data type");
1564 /* Handle stream on the relayd if the output is on the network */
1566 unsigned long netlen
= len
;
1569 * Lock the control socket for the complete duration of the function
1570 * since from this point on we will use the socket.
1572 if (stream
->metadata_flag
) {
1573 /* Metadata requires the control socket. */
1574 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1575 if (stream
->reset_metadata_flag
) {
1576 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1577 stream
->relayd_stream_id
,
1578 stream
->metadata_version
);
1583 stream
->reset_metadata_flag
= 0;
1585 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1588 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1593 /* Use the returned socket. */
1596 /* Write metadata stream id before payload */
1597 if (stream
->metadata_flag
) {
1598 ret
= write_relayd_metadata_id(outfd
, stream
, relayd
, padding
);
1605 /* No streaming, we have to set the len with the full padding */
1608 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1609 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1611 ERR("Reset metadata file");
1614 stream
->reset_metadata_flag
= 0;
1618 * Check if we need to change the tracefile before writing the packet.
1620 if (stream
->chan
->tracefile_size
> 0 &&
1621 (stream
->tracefile_size_current
+ len
) >
1622 stream
->chan
->tracefile_size
) {
1623 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1624 stream
->name
, stream
->chan
->tracefile_size
,
1625 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1626 stream
->out_fd
, &(stream
->tracefile_count_current
),
1629 ERR("Rotating output file");
1632 outfd
= stream
->out_fd
;
1634 if (stream
->index_file
) {
1635 lttng_index_file_put(stream
->index_file
);
1636 stream
->index_file
= lttng_index_file_create(stream
->chan
->pathname
,
1637 stream
->name
, stream
->uid
, stream
->gid
,
1638 stream
->chan
->tracefile_size
,
1639 stream
->tracefile_count_current
,
1640 CTF_INDEX_MAJOR
, CTF_INDEX_MINOR
);
1641 if (!stream
->index_file
) {
1646 /* Reset current size because we just perform a rotation. */
1647 stream
->tracefile_size_current
= 0;
1648 stream
->out_fd_offset
= 0;
1651 stream
->tracefile_size_current
+= len
;
1653 index
->offset
= htobe64(stream
->out_fd_offset
);
1658 * This call guarantee that len or less is returned. It's impossible to
1659 * receive a ret value that is bigger than len.
1661 ret
= lttng_write(outfd
, mmap_base
+ mmap_offset
, len
);
1662 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, len
);
1663 if (ret
< 0 || ((size_t) ret
!= len
)) {
1665 * Report error to caller if nothing was written else at least send the
1673 /* Socket operation failed. We consider the relayd dead */
1674 if (errno
== EPIPE
|| errno
== EINVAL
|| errno
== EBADF
) {
1676 * This is possible if the fd is closed on the other side
1677 * (outfd) or any write problem. It can be verbose a bit for a
1678 * normal execution if for instance the relayd is stopped
1679 * abruptly. This can happen so set this to a DBG statement.
1681 DBG("Consumer mmap write detected relayd hang up");
1683 /* Unhandled error, print it and stop function right now. */
1684 PERROR("Error in write mmap (ret %zd != len %lu)", ret
, len
);
1688 stream
->output_written
+= ret
;
1690 /* This call is useless on a socket so better save a syscall. */
1692 /* This won't block, but will start writeout asynchronously */
1693 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, len
,
1694 SYNC_FILE_RANGE_WRITE
);
1695 stream
->out_fd_offset
+= len
;
1696 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1701 * This is a special case that the relayd has closed its socket. Let's
1702 * cleanup the relayd object and all associated streams.
1704 if (relayd
&& relayd_hang_up
) {
1705 cleanup_relayd(relayd
, ctx
);
1709 /* Unlock only if ctrl socket used */
1710 if (relayd
&& stream
->metadata_flag
) {
1711 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1719 * Splice the data from the ring buffer to the tracefile.
1721 * It must be called with the stream lock held.
1723 * Returns the number of bytes spliced.
1725 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1726 struct lttng_consumer_local_data
*ctx
,
1727 struct lttng_consumer_stream
*stream
, unsigned long len
,
1728 unsigned long padding
,
1729 struct ctf_packet_index
*index
)
1731 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1733 off_t orig_offset
= stream
->out_fd_offset
;
1734 int fd
= stream
->wait_fd
;
1735 /* Default is on the disk */
1736 int outfd
= stream
->out_fd
;
1737 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1739 unsigned int relayd_hang_up
= 0;
1741 switch (consumer_data
.type
) {
1742 case LTTNG_CONSUMER_KERNEL
:
1744 case LTTNG_CONSUMER32_UST
:
1745 case LTTNG_CONSUMER64_UST
:
1746 /* Not supported for user space tracing */
1749 ERR("Unknown consumer_data type");
1753 /* RCU lock for the relayd pointer */
1756 /* Flag that the current stream if set for network streaming. */
1757 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1758 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1759 if (relayd
== NULL
) {
1764 splice_pipe
= stream
->splice_pipe
;
1766 /* Write metadata stream id before payload */
1768 unsigned long total_len
= len
;
1770 if (stream
->metadata_flag
) {
1772 * Lock the control socket for the complete duration of the function
1773 * since from this point on we will use the socket.
1775 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1777 if (stream
->reset_metadata_flag
) {
1778 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1779 stream
->relayd_stream_id
,
1780 stream
->metadata_version
);
1785 stream
->reset_metadata_flag
= 0;
1787 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
, relayd
,
1795 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1798 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1804 /* Use the returned socket. */
1807 /* No streaming, we have to set the len with the full padding */
1810 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1811 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1813 ERR("Reset metadata file");
1816 stream
->reset_metadata_flag
= 0;
1819 * Check if we need to change the tracefile before writing the packet.
1821 if (stream
->chan
->tracefile_size
> 0 &&
1822 (stream
->tracefile_size_current
+ len
) >
1823 stream
->chan
->tracefile_size
) {
1824 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1825 stream
->name
, stream
->chan
->tracefile_size
,
1826 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1827 stream
->out_fd
, &(stream
->tracefile_count_current
),
1831 ERR("Rotating output file");
1834 outfd
= stream
->out_fd
;
1836 if (stream
->index_file
) {
1837 lttng_index_file_put(stream
->index_file
);
1838 stream
->index_file
= lttng_index_file_create(stream
->chan
->pathname
,
1839 stream
->name
, stream
->uid
, stream
->gid
,
1840 stream
->chan
->tracefile_size
,
1841 stream
->tracefile_count_current
,
1842 CTF_INDEX_MAJOR
, CTF_INDEX_MINOR
);
1843 if (!stream
->index_file
) {
1848 /* Reset current size because we just perform a rotation. */
1849 stream
->tracefile_size_current
= 0;
1850 stream
->out_fd_offset
= 0;
1853 stream
->tracefile_size_current
+= len
;
1854 index
->offset
= htobe64(stream
->out_fd_offset
);
1858 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1859 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1860 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1861 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1862 DBG("splice chan to pipe, ret %zd", ret_splice
);
1863 if (ret_splice
< 0) {
1866 PERROR("Error in relay splice");
1870 /* Handle stream on the relayd if the output is on the network */
1871 if (relayd
&& stream
->metadata_flag
) {
1872 size_t metadata_payload_size
=
1873 sizeof(struct lttcomm_relayd_metadata_payload
);
1875 /* Update counter to fit the spliced data */
1876 ret_splice
+= metadata_payload_size
;
1877 len
+= metadata_payload_size
;
1879 * We do this so the return value can match the len passed as
1880 * argument to this function.
1882 written
-= metadata_payload_size
;
1885 /* Splice data out */
1886 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1887 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1888 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
1890 if (ret_splice
< 0) {
1895 } else if (ret_splice
> len
) {
1897 * We don't expect this code path to be executed but you never know
1898 * so this is an extra protection agains a buggy splice().
1901 written
+= ret_splice
;
1902 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice
,
1906 /* All good, update current len and continue. */
1910 /* This call is useless on a socket so better save a syscall. */
1912 /* This won't block, but will start writeout asynchronously */
1913 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1914 SYNC_FILE_RANGE_WRITE
);
1915 stream
->out_fd_offset
+= ret_splice
;
1917 stream
->output_written
+= ret_splice
;
1918 written
+= ret_splice
;
1921 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1927 * This is a special case that the relayd has closed its socket. Let's
1928 * cleanup the relayd object and all associated streams.
1930 if (relayd
&& relayd_hang_up
) {
1931 cleanup_relayd(relayd
, ctx
);
1932 /* Skip splice error so the consumer does not fail */
1937 /* send the appropriate error description to sessiond */
1940 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1943 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1946 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
1951 if (relayd
&& stream
->metadata_flag
) {
1952 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1960 * Take a snapshot for a specific fd
1962 * Returns 0 on success, < 0 on error
1964 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
1966 switch (consumer_data
.type
) {
1967 case LTTNG_CONSUMER_KERNEL
:
1968 return lttng_kconsumer_take_snapshot(stream
);
1969 case LTTNG_CONSUMER32_UST
:
1970 case LTTNG_CONSUMER64_UST
:
1971 return lttng_ustconsumer_take_snapshot(stream
);
1973 ERR("Unknown consumer_data type");
1980 * Get the produced position
1982 * Returns 0 on success, < 0 on error
1984 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
1987 switch (consumer_data
.type
) {
1988 case LTTNG_CONSUMER_KERNEL
:
1989 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
1990 case LTTNG_CONSUMER32_UST
:
1991 case LTTNG_CONSUMER64_UST
:
1992 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
1994 ERR("Unknown consumer_data type");
2000 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
2001 int sock
, struct pollfd
*consumer_sockpoll
)
2003 switch (consumer_data
.type
) {
2004 case LTTNG_CONSUMER_KERNEL
:
2005 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2006 case LTTNG_CONSUMER32_UST
:
2007 case LTTNG_CONSUMER64_UST
:
2008 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2010 ERR("Unknown consumer_data type");
2016 void lttng_consumer_close_all_metadata(void)
2018 switch (consumer_data
.type
) {
2019 case LTTNG_CONSUMER_KERNEL
:
2021 * The Kernel consumer has a different metadata scheme so we don't
2022 * close anything because the stream will be closed by the session
2026 case LTTNG_CONSUMER32_UST
:
2027 case LTTNG_CONSUMER64_UST
:
2029 * Close all metadata streams. The metadata hash table is passed and
2030 * this call iterates over it by closing all wakeup fd. This is safe
2031 * because at this point we are sure that the metadata producer is
2032 * either dead or blocked.
2034 lttng_ustconsumer_close_all_metadata(metadata_ht
);
2037 ERR("Unknown consumer_data type");
2043 * Clean up a metadata stream and free its memory.
2045 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
2046 struct lttng_ht
*ht
)
2048 struct lttng_consumer_channel
*free_chan
= NULL
;
2052 * This call should NEVER receive regular stream. It must always be
2053 * metadata stream and this is crucial for data structure synchronization.
2055 assert(stream
->metadata_flag
);
2057 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
2059 pthread_mutex_lock(&consumer_data
.lock
);
2060 pthread_mutex_lock(&stream
->chan
->lock
);
2061 pthread_mutex_lock(&stream
->lock
);
2062 if (stream
->chan
->metadata_cache
) {
2063 /* Only applicable to userspace consumers. */
2064 pthread_mutex_lock(&stream
->chan
->metadata_cache
->lock
);
2067 /* Remove any reference to that stream. */
2068 consumer_stream_delete(stream
, ht
);
2070 /* Close down everything including the relayd if one. */
2071 consumer_stream_close(stream
);
2072 /* Destroy tracer buffers of the stream. */
2073 consumer_stream_destroy_buffers(stream
);
2075 /* Atomically decrement channel refcount since other threads can use it. */
2076 if (!uatomic_sub_return(&stream
->chan
->refcount
, 1)
2077 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
2078 /* Go for channel deletion! */
2079 free_chan
= stream
->chan
;
2083 * Nullify the stream reference so it is not used after deletion. The
2084 * channel lock MUST be acquired before being able to check for a NULL
2087 stream
->chan
->metadata_stream
= NULL
;
2089 if (stream
->chan
->metadata_cache
) {
2090 pthread_mutex_unlock(&stream
->chan
->metadata_cache
->lock
);
2092 pthread_mutex_unlock(&stream
->lock
);
2093 pthread_mutex_unlock(&stream
->chan
->lock
);
2094 pthread_mutex_unlock(&consumer_data
.lock
);
2097 consumer_del_channel(free_chan
);
2100 consumer_stream_free(stream
);
2104 * Action done with the metadata stream when adding it to the consumer internal
2105 * data structures to handle it.
2107 int consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
2109 struct lttng_ht
*ht
= metadata_ht
;
2111 struct lttng_ht_iter iter
;
2112 struct lttng_ht_node_u64
*node
;
2117 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2119 pthread_mutex_lock(&consumer_data
.lock
);
2120 pthread_mutex_lock(&stream
->chan
->lock
);
2121 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2122 pthread_mutex_lock(&stream
->lock
);
2125 * From here, refcounts are updated so be _careful_ when returning an error
2132 * Lookup the stream just to make sure it does not exist in our internal
2133 * state. This should NEVER happen.
2135 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2136 node
= lttng_ht_iter_get_node_u64(&iter
);
2140 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2141 * in terms of destroying the associated channel, because the action that
2142 * causes the count to become 0 also causes a stream to be added. The
2143 * channel deletion will thus be triggered by the following removal of this
2146 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2147 /* Increment refcount before decrementing nb_init_stream_left */
2149 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2152 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2154 lttng_ht_add_unique_u64(consumer_data
.stream_per_chan_id_ht
,
2155 &stream
->node_channel_id
);
2158 * Add stream to the stream_list_ht of the consumer data. No need to steal
2159 * the key since the HT does not use it and we allow to add redundant keys
2162 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2166 pthread_mutex_unlock(&stream
->lock
);
2167 pthread_mutex_unlock(&stream
->chan
->lock
);
2168 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2169 pthread_mutex_unlock(&consumer_data
.lock
);
2174 * Delete data stream that are flagged for deletion (endpoint_status).
2176 static void validate_endpoint_status_data_stream(void)
2178 struct lttng_ht_iter iter
;
2179 struct lttng_consumer_stream
*stream
;
2181 DBG("Consumer delete flagged data stream");
2184 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2185 /* Validate delete flag of the stream */
2186 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2189 /* Delete it right now */
2190 consumer_del_stream(stream
, data_ht
);
2196 * Delete metadata stream that are flagged for deletion (endpoint_status).
2198 static void validate_endpoint_status_metadata_stream(
2199 struct lttng_poll_event
*pollset
)
2201 struct lttng_ht_iter iter
;
2202 struct lttng_consumer_stream
*stream
;
2204 DBG("Consumer delete flagged metadata stream");
2209 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2210 /* Validate delete flag of the stream */
2211 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2215 * Remove from pollset so the metadata thread can continue without
2216 * blocking on a deleted stream.
2218 lttng_poll_del(pollset
, stream
->wait_fd
);
2220 /* Delete it right now */
2221 consumer_del_metadata_stream(stream
, metadata_ht
);
2227 * Thread polls on metadata file descriptor and write them on disk or on the
2230 void *consumer_thread_metadata_poll(void *data
)
2232 int ret
, i
, pollfd
, err
= -1;
2233 uint32_t revents
, nb_fd
;
2234 struct lttng_consumer_stream
*stream
= NULL
;
2235 struct lttng_ht_iter iter
;
2236 struct lttng_ht_node_u64
*node
;
2237 struct lttng_poll_event events
;
2238 struct lttng_consumer_local_data
*ctx
= data
;
2241 rcu_register_thread();
2243 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2245 if (testpoint(consumerd_thread_metadata
)) {
2246 goto error_testpoint
;
2249 health_code_update();
2251 DBG("Thread metadata poll started");
2253 /* Size is set to 1 for the consumer_metadata pipe */
2254 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2256 ERR("Poll set creation failed");
2260 ret
= lttng_poll_add(&events
,
2261 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2267 DBG("Metadata main loop started");
2271 health_code_update();
2272 health_poll_entry();
2273 DBG("Metadata poll wait");
2274 ret
= lttng_poll_wait(&events
, -1);
2275 DBG("Metadata poll return from wait with %d fd(s)",
2276 LTTNG_POLL_GETNB(&events
));
2278 DBG("Metadata event caught in thread");
2280 if (errno
== EINTR
) {
2281 ERR("Poll EINTR caught");
2284 if (LTTNG_POLL_GETNB(&events
) == 0) {
2285 err
= 0; /* All is OK */
2292 /* From here, the event is a metadata wait fd */
2293 for (i
= 0; i
< nb_fd
; i
++) {
2294 health_code_update();
2296 revents
= LTTNG_POLL_GETEV(&events
, i
);
2297 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2300 /* No activity for this FD (poll implementation). */
2304 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2305 if (revents
& LPOLLIN
) {
2308 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2309 &stream
, sizeof(stream
));
2310 if (pipe_len
< sizeof(stream
)) {
2312 PERROR("read metadata stream");
2315 * Remove the pipe from the poll set and continue the loop
2316 * since their might be data to consume.
2318 lttng_poll_del(&events
,
2319 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2320 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2324 /* A NULL stream means that the state has changed. */
2325 if (stream
== NULL
) {
2326 /* Check for deleted streams. */
2327 validate_endpoint_status_metadata_stream(&events
);
2331 DBG("Adding metadata stream %d to poll set",
2334 /* Add metadata stream to the global poll events list */
2335 lttng_poll_add(&events
, stream
->wait_fd
,
2336 LPOLLIN
| LPOLLPRI
| LPOLLHUP
);
2337 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2338 DBG("Metadata thread pipe hung up");
2340 * Remove the pipe from the poll set and continue the loop
2341 * since their might be data to consume.
2343 lttng_poll_del(&events
,
2344 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2345 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2348 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2352 /* Handle other stream */
2358 uint64_t tmp_id
= (uint64_t) pollfd
;
2360 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2362 node
= lttng_ht_iter_get_node_u64(&iter
);
2365 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2368 if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2369 /* Get the data out of the metadata file descriptor */
2370 DBG("Metadata available on fd %d", pollfd
);
2371 assert(stream
->wait_fd
== pollfd
);
2374 health_code_update();
2376 len
= ctx
->on_buffer_ready(stream
, ctx
);
2378 * We don't check the return value here since if we get
2379 * a negative len, it means an error occurred thus we
2380 * simply remove it from the poll set and free the
2385 /* It's ok to have an unavailable sub-buffer */
2386 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2387 /* Clean up stream from consumer and free it. */
2388 lttng_poll_del(&events
, stream
->wait_fd
);
2389 consumer_del_metadata_stream(stream
, metadata_ht
);
2391 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2392 DBG("Metadata fd %d is hup|err.", pollfd
);
2393 if (!stream
->hangup_flush_done
2394 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2395 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2396 DBG("Attempting to flush and consume the UST buffers");
2397 lttng_ustconsumer_on_stream_hangup(stream
);
2399 /* We just flushed the stream now read it. */
2401 health_code_update();
2403 len
= ctx
->on_buffer_ready(stream
, ctx
);
2405 * We don't check the return value here since if we get
2406 * a negative len, it means an error occurred thus we
2407 * simply remove it from the poll set and free the
2413 lttng_poll_del(&events
, stream
->wait_fd
);
2415 * This call update the channel states, closes file descriptors
2416 * and securely free the stream.
2418 consumer_del_metadata_stream(stream
, metadata_ht
);
2420 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2424 /* Release RCU lock for the stream looked up */
2432 DBG("Metadata poll thread exiting");
2434 lttng_poll_clean(&events
);
2439 ERR("Health error occurred in %s", __func__
);
2441 health_unregister(health_consumerd
);
2442 rcu_unregister_thread();
2447 * This thread polls the fds in the set to consume the data and write
2448 * it to tracefile if necessary.
2450 void *consumer_thread_data_poll(void *data
)
2452 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2453 struct pollfd
*pollfd
= NULL
;
2454 /* local view of the streams */
2455 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2456 /* local view of consumer_data.fds_count */
2458 /* Number of FDs with CONSUMER_ENDPOINT_INACTIVE but still open. */
2459 int nb_inactive_fd
= 0;
2460 struct lttng_consumer_local_data
*ctx
= data
;
2463 rcu_register_thread();
2465 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2467 if (testpoint(consumerd_thread_data
)) {
2468 goto error_testpoint
;
2471 health_code_update();
2473 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
*));
2474 if (local_stream
== NULL
) {
2475 PERROR("local_stream malloc");
2480 health_code_update();
2486 * the fds set has been updated, we need to update our
2487 * local array as well
2489 pthread_mutex_lock(&consumer_data
.lock
);
2490 if (consumer_data
.need_update
) {
2495 local_stream
= NULL
;
2498 * Allocate for all fds +1 for the consumer_data_pipe and +1 for
2501 pollfd
= zmalloc((consumer_data
.stream_count
+ 2) * sizeof(struct pollfd
));
2502 if (pollfd
== NULL
) {
2503 PERROR("pollfd malloc");
2504 pthread_mutex_unlock(&consumer_data
.lock
);
2508 local_stream
= zmalloc((consumer_data
.stream_count
+ 2) *
2509 sizeof(struct lttng_consumer_stream
*));
2510 if (local_stream
== NULL
) {
2511 PERROR("local_stream malloc");
2512 pthread_mutex_unlock(&consumer_data
.lock
);
2515 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2516 data_ht
, &nb_inactive_fd
);
2518 ERR("Error in allocating pollfd or local_outfds");
2519 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2520 pthread_mutex_unlock(&consumer_data
.lock
);
2524 consumer_data
.need_update
= 0;
2526 pthread_mutex_unlock(&consumer_data
.lock
);
2528 /* No FDs and consumer_quit, consumer_cleanup the thread */
2529 if (nb_fd
== 0 && consumer_quit
== 1 && nb_inactive_fd
== 0) {
2530 err
= 0; /* All is OK */
2533 /* poll on the array of fds */
2535 DBG("polling on %d fd", nb_fd
+ 2);
2536 health_poll_entry();
2537 num_rdy
= poll(pollfd
, nb_fd
+ 2, -1);
2539 DBG("poll num_rdy : %d", num_rdy
);
2540 if (num_rdy
== -1) {
2542 * Restart interrupted system call.
2544 if (errno
== EINTR
) {
2547 PERROR("Poll error");
2548 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2550 } else if (num_rdy
== 0) {
2551 DBG("Polling thread timed out");
2556 * If the consumer_data_pipe triggered poll go directly to the
2557 * beginning of the loop to update the array. We want to prioritize
2558 * array update over low-priority reads.
2560 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2561 ssize_t pipe_readlen
;
2563 DBG("consumer_data_pipe wake up");
2564 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2565 &new_stream
, sizeof(new_stream
));
2566 if (pipe_readlen
< sizeof(new_stream
)) {
2567 PERROR("Consumer data pipe");
2568 /* Continue so we can at least handle the current stream(s). */
2573 * If the stream is NULL, just ignore it. It's also possible that
2574 * the sessiond poll thread changed the consumer_quit state and is
2575 * waking us up to test it.
2577 if (new_stream
== NULL
) {
2578 validate_endpoint_status_data_stream();
2582 /* Continue to update the local streams and handle prio ones */
2586 /* Handle wakeup pipe. */
2587 if (pollfd
[nb_fd
+ 1].revents
& (POLLIN
| POLLPRI
)) {
2589 ssize_t pipe_readlen
;
2591 pipe_readlen
= lttng_pipe_read(ctx
->consumer_wakeup_pipe
, &dummy
,
2593 if (pipe_readlen
< 0) {
2594 PERROR("Consumer data wakeup pipe");
2596 /* We've been awakened to handle stream(s). */
2597 ctx
->has_wakeup
= 0;
2600 /* Take care of high priority channels first. */
2601 for (i
= 0; i
< nb_fd
; i
++) {
2602 health_code_update();
2604 if (local_stream
[i
] == NULL
) {
2607 if (pollfd
[i
].revents
& POLLPRI
) {
2608 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2610 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2611 /* it's ok to have an unavailable sub-buffer */
2612 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2613 /* Clean the stream and free it. */
2614 consumer_del_stream(local_stream
[i
], data_ht
);
2615 local_stream
[i
] = NULL
;
2616 } else if (len
> 0) {
2617 local_stream
[i
]->data_read
= 1;
2623 * If we read high prio channel in this loop, try again
2624 * for more high prio data.
2630 /* Take care of low priority channels. */
2631 for (i
= 0; i
< nb_fd
; i
++) {
2632 health_code_update();
2634 if (local_stream
[i
] == NULL
) {
2637 if ((pollfd
[i
].revents
& POLLIN
) ||
2638 local_stream
[i
]->hangup_flush_done
||
2639 local_stream
[i
]->has_data
) {
2640 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2641 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2642 /* it's ok to have an unavailable sub-buffer */
2643 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2644 /* Clean the stream and free it. */
2645 consumer_del_stream(local_stream
[i
], data_ht
);
2646 local_stream
[i
] = NULL
;
2647 } else if (len
> 0) {
2648 local_stream
[i
]->data_read
= 1;
2653 /* Handle hangup and errors */
2654 for (i
= 0; i
< nb_fd
; i
++) {
2655 health_code_update();
2657 if (local_stream
[i
] == NULL
) {
2660 if (!local_stream
[i
]->hangup_flush_done
2661 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2662 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2663 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2664 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2666 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2667 /* Attempt read again, for the data we just flushed. */
2668 local_stream
[i
]->data_read
= 1;
2671 * If the poll flag is HUP/ERR/NVAL and we have
2672 * read no data in this pass, we can remove the
2673 * stream from its hash table.
2675 if ((pollfd
[i
].revents
& POLLHUP
)) {
2676 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2677 if (!local_stream
[i
]->data_read
) {
2678 consumer_del_stream(local_stream
[i
], data_ht
);
2679 local_stream
[i
] = NULL
;
2682 } else if (pollfd
[i
].revents
& POLLERR
) {
2683 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2684 if (!local_stream
[i
]->data_read
) {
2685 consumer_del_stream(local_stream
[i
], data_ht
);
2686 local_stream
[i
] = NULL
;
2689 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2690 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2691 if (!local_stream
[i
]->data_read
) {
2692 consumer_del_stream(local_stream
[i
], data_ht
);
2693 local_stream
[i
] = NULL
;
2697 if (local_stream
[i
] != NULL
) {
2698 local_stream
[i
]->data_read
= 0;
2705 DBG("polling thread exiting");
2710 * Close the write side of the pipe so epoll_wait() in
2711 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2712 * read side of the pipe. If we close them both, epoll_wait strangely does
2713 * not return and could create a endless wait period if the pipe is the
2714 * only tracked fd in the poll set. The thread will take care of closing
2717 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2722 ERR("Health error occurred in %s", __func__
);
2724 health_unregister(health_consumerd
);
2726 rcu_unregister_thread();
2731 * Close wake-up end of each stream belonging to the channel. This will
2732 * allow the poll() on the stream read-side to detect when the
2733 * write-side (application) finally closes them.
2736 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2738 struct lttng_ht
*ht
;
2739 struct lttng_consumer_stream
*stream
;
2740 struct lttng_ht_iter iter
;
2742 ht
= consumer_data
.stream_per_chan_id_ht
;
2745 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2746 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2747 ht
->match_fct
, &channel
->key
,
2748 &iter
.iter
, stream
, node_channel_id
.node
) {
2750 * Protect against teardown with mutex.
2752 pthread_mutex_lock(&stream
->lock
);
2753 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2756 switch (consumer_data
.type
) {
2757 case LTTNG_CONSUMER_KERNEL
:
2759 case LTTNG_CONSUMER32_UST
:
2760 case LTTNG_CONSUMER64_UST
:
2761 if (stream
->metadata_flag
) {
2762 /* Safe and protected by the stream lock. */
2763 lttng_ustconsumer_close_metadata(stream
->chan
);
2766 * Note: a mutex is taken internally within
2767 * liblttng-ust-ctl to protect timer wakeup_fd
2768 * use from concurrent close.
2770 lttng_ustconsumer_close_stream_wakeup(stream
);
2774 ERR("Unknown consumer_data type");
2778 pthread_mutex_unlock(&stream
->lock
);
2783 static void destroy_channel_ht(struct lttng_ht
*ht
)
2785 struct lttng_ht_iter iter
;
2786 struct lttng_consumer_channel
*channel
;
2794 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2795 ret
= lttng_ht_del(ht
, &iter
);
2800 lttng_ht_destroy(ht
);
2804 * This thread polls the channel fds to detect when they are being
2805 * closed. It closes all related streams if the channel is detected as
2806 * closed. It is currently only used as a shim layer for UST because the
2807 * consumerd needs to keep the per-stream wakeup end of pipes open for
2810 void *consumer_thread_channel_poll(void *data
)
2812 int ret
, i
, pollfd
, err
= -1;
2813 uint32_t revents
, nb_fd
;
2814 struct lttng_consumer_channel
*chan
= NULL
;
2815 struct lttng_ht_iter iter
;
2816 struct lttng_ht_node_u64
*node
;
2817 struct lttng_poll_event events
;
2818 struct lttng_consumer_local_data
*ctx
= data
;
2819 struct lttng_ht
*channel_ht
;
2821 rcu_register_thread();
2823 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
2825 if (testpoint(consumerd_thread_channel
)) {
2826 goto error_testpoint
;
2829 health_code_update();
2831 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2833 /* ENOMEM at this point. Better to bail out. */
2837 DBG("Thread channel poll started");
2839 /* Size is set to 1 for the consumer_channel pipe */
2840 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2842 ERR("Poll set creation failed");
2846 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2852 DBG("Channel main loop started");
2856 health_code_update();
2857 DBG("Channel poll wait");
2858 health_poll_entry();
2859 ret
= lttng_poll_wait(&events
, -1);
2860 DBG("Channel poll return from wait with %d fd(s)",
2861 LTTNG_POLL_GETNB(&events
));
2863 DBG("Channel event caught in thread");
2865 if (errno
== EINTR
) {
2866 ERR("Poll EINTR caught");
2869 if (LTTNG_POLL_GETNB(&events
) == 0) {
2870 err
= 0; /* All is OK */
2877 /* From here, the event is a channel wait fd */
2878 for (i
= 0; i
< nb_fd
; i
++) {
2879 health_code_update();
2881 revents
= LTTNG_POLL_GETEV(&events
, i
);
2882 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2885 /* No activity for this FD (poll implementation). */
2889 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2890 if (revents
& LPOLLIN
) {
2891 enum consumer_channel_action action
;
2894 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
2897 ERR("Error reading channel pipe");
2899 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2904 case CONSUMER_CHANNEL_ADD
:
2905 DBG("Adding channel %d to poll set",
2908 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
2911 lttng_ht_add_unique_u64(channel_ht
,
2912 &chan
->wait_fd_node
);
2914 /* Add channel to the global poll events list */
2915 lttng_poll_add(&events
, chan
->wait_fd
,
2916 LPOLLERR
| LPOLLHUP
);
2918 case CONSUMER_CHANNEL_DEL
:
2921 * This command should never be called if the channel
2922 * has streams monitored by either the data or metadata
2923 * thread. The consumer only notify this thread with a
2924 * channel del. command if it receives a destroy
2925 * channel command from the session daemon that send it
2926 * if a command prior to the GET_CHANNEL failed.
2930 chan
= consumer_find_channel(key
);
2933 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
2936 lttng_poll_del(&events
, chan
->wait_fd
);
2937 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
2938 ret
= lttng_ht_del(channel_ht
, &iter
);
2941 switch (consumer_data
.type
) {
2942 case LTTNG_CONSUMER_KERNEL
:
2944 case LTTNG_CONSUMER32_UST
:
2945 case LTTNG_CONSUMER64_UST
:
2946 health_code_update();
2947 /* Destroy streams that might have been left in the stream list. */
2948 clean_channel_stream_list(chan
);
2951 ERR("Unknown consumer_data type");
2956 * Release our own refcount. Force channel deletion even if
2957 * streams were not initialized.
2959 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
2960 consumer_del_channel(chan
);
2965 case CONSUMER_CHANNEL_QUIT
:
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("Unknown action");
2976 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2977 DBG("Channel thread pipe hung up");
2979 * Remove the pipe from the poll set and continue the loop
2980 * since their might be data to consume.
2982 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2985 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2989 /* Handle other stream */
2995 uint64_t tmp_id
= (uint64_t) pollfd
;
2997 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
2999 node
= lttng_ht_iter_get_node_u64(&iter
);
3002 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
3005 /* Check for error event */
3006 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
3007 DBG("Channel fd %d is hup|err.", pollfd
);
3009 lttng_poll_del(&events
, chan
->wait_fd
);
3010 ret
= lttng_ht_del(channel_ht
, &iter
);
3014 * This will close the wait fd for each stream associated to
3015 * this channel AND monitored by the data/metadata thread thus
3016 * will be clean by the right thread.
3018 consumer_close_channel_streams(chan
);
3020 /* Release our own refcount */
3021 if (!uatomic_sub_return(&chan
->refcount
, 1)
3022 && !uatomic_read(&chan
->nb_init_stream_left
)) {
3023 consumer_del_channel(chan
);
3026 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3031 /* Release RCU lock for the channel looked up */
3039 lttng_poll_clean(&events
);
3041 destroy_channel_ht(channel_ht
);
3044 DBG("Channel poll thread exiting");
3047 ERR("Health error occurred in %s", __func__
);
3049 health_unregister(health_consumerd
);
3050 rcu_unregister_thread();
3054 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
3055 struct pollfd
*sockpoll
, int client_socket
)
3062 ret
= lttng_consumer_poll_socket(sockpoll
);
3066 DBG("Metadata connection on client_socket");
3068 /* Blocking call, waiting for transmission */
3069 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
3070 if (ctx
->consumer_metadata_socket
< 0) {
3071 WARN("On accept metadata");
3082 * This thread listens on the consumerd socket and receives the file
3083 * descriptors from the session daemon.
3085 void *consumer_thread_sessiond_poll(void *data
)
3087 int sock
= -1, client_socket
, ret
, err
= -1;
3089 * structure to poll for incoming data on communication socket avoids
3090 * making blocking sockets.
3092 struct pollfd consumer_sockpoll
[2];
3093 struct lttng_consumer_local_data
*ctx
= data
;
3095 rcu_register_thread();
3097 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
3099 if (testpoint(consumerd_thread_sessiond
)) {
3100 goto error_testpoint
;
3103 health_code_update();
3105 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
3106 unlink(ctx
->consumer_command_sock_path
);
3107 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
3108 if (client_socket
< 0) {
3109 ERR("Cannot create command socket");
3113 ret
= lttcomm_listen_unix_sock(client_socket
);
3118 DBG("Sending ready command to lttng-sessiond");
3119 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3120 /* return < 0 on error, but == 0 is not fatal */
3122 ERR("Error sending ready command to lttng-sessiond");
3126 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3127 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3128 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3129 consumer_sockpoll
[1].fd
= client_socket
;
3130 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3132 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3140 DBG("Connection on client_socket");
3142 /* Blocking call, waiting for transmission */
3143 sock
= lttcomm_accept_unix_sock(client_socket
);
3150 * Setup metadata socket which is the second socket connection on the
3151 * command unix socket.
3153 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3162 /* This socket is not useful anymore. */
3163 ret
= close(client_socket
);
3165 PERROR("close client_socket");
3169 /* update the polling structure to poll on the established socket */
3170 consumer_sockpoll
[1].fd
= sock
;
3171 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3174 health_code_update();
3176 health_poll_entry();
3177 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3186 DBG("Incoming command on sock");
3187 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3190 * This could simply be a session daemon quitting. Don't output
3193 DBG("Communication interrupted on command socket");
3197 if (consumer_quit
) {
3198 DBG("consumer_thread_receive_fds received quit from signal");
3199 err
= 0; /* All is OK */
3202 DBG("received command on sock");
3208 DBG("Consumer thread sessiond poll exiting");
3211 * Close metadata streams since the producer is the session daemon which
3214 * NOTE: for now, this only applies to the UST tracer.
3216 lttng_consumer_close_all_metadata();
3219 * when all fds have hung up, the polling thread
3225 * Notify the data poll thread to poll back again and test the
3226 * consumer_quit state that we just set so to quit gracefully.
3228 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3230 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3232 notify_health_quit_pipe(health_quit_pipe
);
3234 /* Cleaning up possibly open sockets. */
3238 PERROR("close sock sessiond poll");
3241 if (client_socket
>= 0) {
3242 ret
= close(client_socket
);
3244 PERROR("close client_socket sessiond poll");
3251 ERR("Health error occurred in %s", __func__
);
3253 health_unregister(health_consumerd
);
3255 rcu_unregister_thread();
3259 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3260 struct lttng_consumer_local_data
*ctx
)
3264 pthread_mutex_lock(&stream
->lock
);
3265 if (stream
->metadata_flag
) {
3266 pthread_mutex_lock(&stream
->metadata_rdv_lock
);
3269 switch (consumer_data
.type
) {
3270 case LTTNG_CONSUMER_KERNEL
:
3271 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
3273 case LTTNG_CONSUMER32_UST
:
3274 case LTTNG_CONSUMER64_UST
:
3275 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
3278 ERR("Unknown consumer_data type");
3284 if (stream
->metadata_flag
) {
3285 pthread_cond_broadcast(&stream
->metadata_rdv
);
3286 pthread_mutex_unlock(&stream
->metadata_rdv_lock
);
3288 pthread_mutex_unlock(&stream
->lock
);
3292 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3294 switch (consumer_data
.type
) {
3295 case LTTNG_CONSUMER_KERNEL
:
3296 return lttng_kconsumer_on_recv_stream(stream
);
3297 case LTTNG_CONSUMER32_UST
:
3298 case LTTNG_CONSUMER64_UST
:
3299 return lttng_ustconsumer_on_recv_stream(stream
);
3301 ERR("Unknown consumer_data type");
3308 * Allocate and set consumer data hash tables.
3310 int lttng_consumer_init(void)
3312 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3313 if (!consumer_data
.channel_ht
) {
3317 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3318 if (!consumer_data
.relayd_ht
) {
3322 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3323 if (!consumer_data
.stream_list_ht
) {
3327 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3328 if (!consumer_data
.stream_per_chan_id_ht
) {
3332 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3337 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3349 * Process the ADD_RELAYD command receive by a consumer.
3351 * This will create a relayd socket pair and add it to the relayd hash table.
3352 * The caller MUST acquire a RCU read side lock before calling it.
3354 int consumer_add_relayd_socket(uint64_t net_seq_idx
, int sock_type
,
3355 struct lttng_consumer_local_data
*ctx
, int sock
,
3356 struct pollfd
*consumer_sockpoll
,
3357 struct lttcomm_relayd_sock
*relayd_sock
, uint64_t sessiond_id
,
3358 uint64_t relayd_session_id
)
3360 int fd
= -1, ret
= -1, relayd_created
= 0;
3361 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3362 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3365 assert(relayd_sock
);
3367 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", net_seq_idx
);
3369 /* Get relayd reference if exists. */
3370 relayd
= consumer_find_relayd(net_seq_idx
);
3371 if (relayd
== NULL
) {
3372 assert(sock_type
== LTTNG_STREAM_CONTROL
);
3373 /* Not found. Allocate one. */
3374 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3375 if (relayd
== NULL
) {
3377 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3380 relayd
->sessiond_session_id
= sessiond_id
;
3385 * This code path MUST continue to the consumer send status message to
3386 * we can notify the session daemon and continue our work without
3387 * killing everything.
3391 * relayd key should never be found for control socket.
3393 assert(sock_type
!= LTTNG_STREAM_CONTROL
);
3396 /* First send a status message before receiving the fds. */
3397 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3399 /* Somehow, the session daemon is not responding anymore. */
3400 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3401 goto error_nosignal
;
3404 /* Poll on consumer socket. */
3405 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3407 /* Needing to exit in the middle of a command: error. */
3408 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3410 goto error_nosignal
;
3413 /* Get relayd socket from session daemon */
3414 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3415 if (ret
!= sizeof(fd
)) {
3417 fd
= -1; /* Just in case it gets set with an invalid value. */
3420 * Failing to receive FDs might indicate a major problem such as
3421 * reaching a fd limit during the receive where the kernel returns a
3422 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3423 * don't take any chances and stop everything.
3425 * XXX: Feature request #558 will fix that and avoid this possible
3426 * issue when reaching the fd limit.
3428 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3429 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3433 /* Copy socket information and received FD */
3434 switch (sock_type
) {
3435 case LTTNG_STREAM_CONTROL
:
3436 /* Copy received lttcomm socket */
3437 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3438 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3439 /* Handle create_sock error. */
3441 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3445 * Close the socket created internally by
3446 * lttcomm_create_sock, so we can replace it by the one
3447 * received from sessiond.
3449 if (close(relayd
->control_sock
.sock
.fd
)) {
3453 /* Assign new file descriptor */
3454 relayd
->control_sock
.sock
.fd
= fd
;
3455 fd
= -1; /* For error path */
3456 /* Assign version values. */
3457 relayd
->control_sock
.major
= relayd_sock
->major
;
3458 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3460 relayd
->relayd_session_id
= relayd_session_id
;
3463 case LTTNG_STREAM_DATA
:
3464 /* Copy received lttcomm socket */
3465 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3466 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3467 /* Handle create_sock error. */
3469 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3473 * Close the socket created internally by
3474 * lttcomm_create_sock, so we can replace it by the one
3475 * received from sessiond.
3477 if (close(relayd
->data_sock
.sock
.fd
)) {
3481 /* Assign new file descriptor */
3482 relayd
->data_sock
.sock
.fd
= fd
;
3483 fd
= -1; /* for eventual error paths */
3484 /* Assign version values. */
3485 relayd
->data_sock
.major
= relayd_sock
->major
;
3486 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3489 ERR("Unknown relayd socket type (%d)", sock_type
);
3491 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3495 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3496 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3497 relayd
->net_seq_idx
, fd
);
3499 /* We successfully added the socket. Send status back. */
3500 ret
= consumer_send_status_msg(sock
, ret_code
);
3502 /* Somehow, the session daemon is not responding anymore. */
3503 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3504 goto error_nosignal
;
3508 * Add relayd socket pair to consumer data hashtable. If object already
3509 * exists or on error, the function gracefully returns.
3517 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3518 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3522 /* Close received socket if valid. */
3525 PERROR("close received socket");
3529 if (relayd_created
) {
3537 * Try to lock the stream mutex.
3539 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
3541 static int stream_try_lock(struct lttng_consumer_stream
*stream
)
3548 * Try to lock the stream mutex. On failure, we know that the stream is
3549 * being used else where hence there is data still being extracted.
3551 ret
= pthread_mutex_trylock(&stream
->lock
);
3553 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
3565 * Search for a relayd associated to the session id and return the reference.
3567 * A rcu read side lock MUST be acquire before calling this function and locked
3568 * until the relayd object is no longer necessary.
3570 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3572 struct lttng_ht_iter iter
;
3573 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3575 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3576 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3579 * Check by sessiond id which is unique here where the relayd session
3580 * id might not be when having multiple relayd.
3582 if (relayd
->sessiond_session_id
== id
) {
3583 /* Found the relayd. There can be only one per id. */
3595 * Check if for a given session id there is still data needed to be extract
3598 * Return 1 if data is pending or else 0 meaning ready to be read.
3600 int consumer_data_pending(uint64_t id
)
3603 struct lttng_ht_iter iter
;
3604 struct lttng_ht
*ht
;
3605 struct lttng_consumer_stream
*stream
;
3606 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3607 int (*data_pending
)(struct lttng_consumer_stream
*);
3609 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3612 pthread_mutex_lock(&consumer_data
.lock
);
3614 switch (consumer_data
.type
) {
3615 case LTTNG_CONSUMER_KERNEL
:
3616 data_pending
= lttng_kconsumer_data_pending
;
3618 case LTTNG_CONSUMER32_UST
:
3619 case LTTNG_CONSUMER64_UST
:
3620 data_pending
= lttng_ustconsumer_data_pending
;
3623 ERR("Unknown consumer data type");
3627 /* Ease our life a bit */
3628 ht
= consumer_data
.stream_list_ht
;
3630 relayd
= find_relayd_by_session_id(id
);
3632 /* Send init command for data pending. */
3633 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3634 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3635 relayd
->relayd_session_id
);
3636 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3638 /* Communication error thus the relayd so no data pending. */
3639 goto data_not_pending
;
3643 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3644 ht
->hash_fct(&id
, lttng_ht_seed
),
3646 &iter
.iter
, stream
, node_session_id
.node
) {
3647 /* If this call fails, the stream is being used hence data pending. */
3648 ret
= stream_try_lock(stream
);
3654 * A removed node from the hash table indicates that the stream has
3655 * been deleted thus having a guarantee that the buffers are closed
3656 * on the consumer side. However, data can still be transmitted
3657 * over the network so don't skip the relayd check.
3659 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3661 /* Check the stream if there is data in the buffers. */
3662 ret
= data_pending(stream
);
3664 pthread_mutex_unlock(&stream
->lock
);
3671 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3672 if (stream
->metadata_flag
) {
3673 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3674 stream
->relayd_stream_id
);
3676 ret
= relayd_data_pending(&relayd
->control_sock
,
3677 stream
->relayd_stream_id
,
3678 stream
->next_net_seq_num
- 1);
3680 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3682 pthread_mutex_unlock(&stream
->lock
);
3686 pthread_mutex_unlock(&stream
->lock
);
3690 unsigned int is_data_inflight
= 0;
3692 /* Send init command for data pending. */
3693 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3694 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3695 relayd
->relayd_session_id
, &is_data_inflight
);
3696 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3698 goto data_not_pending
;
3700 if (is_data_inflight
) {
3706 * Finding _no_ node in the hash table and no inflight data means that the
3707 * stream(s) have been removed thus data is guaranteed to be available for
3708 * analysis from the trace files.
3712 /* Data is available to be read by a viewer. */
3713 pthread_mutex_unlock(&consumer_data
.lock
);
3718 /* Data is still being extracted from buffers. */
3719 pthread_mutex_unlock(&consumer_data
.lock
);
3725 * Send a ret code status message to the sessiond daemon.
3727 * Return the sendmsg() return value.
3729 int consumer_send_status_msg(int sock
, int ret_code
)
3731 struct lttcomm_consumer_status_msg msg
;
3733 memset(&msg
, 0, sizeof(msg
));
3734 msg
.ret_code
= ret_code
;
3736 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3740 * Send a channel status message to the sessiond daemon.
3742 * Return the sendmsg() return value.
3744 int consumer_send_status_channel(int sock
,
3745 struct lttng_consumer_channel
*channel
)
3747 struct lttcomm_consumer_status_channel msg
;
3751 memset(&msg
, 0, sizeof(msg
));
3753 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3755 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3756 msg
.key
= channel
->key
;
3757 msg
.stream_count
= channel
->streams
.count
;
3760 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3763 unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos
,
3764 unsigned long produced_pos
, uint64_t nb_packets_per_stream
,
3765 uint64_t max_sb_size
)
3767 unsigned long start_pos
;
3769 if (!nb_packets_per_stream
) {
3770 return consumed_pos
; /* Grab everything */
3772 start_pos
= produced_pos
- offset_align_floor(produced_pos
, max_sb_size
);
3773 start_pos
-= max_sb_size
* nb_packets_per_stream
;
3774 if ((long) (start_pos
- consumed_pos
) < 0) {
3775 return consumed_pos
; /* Grab everything */