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 */
70 /* Flag used to temporarily pause data consumption from testpoints. */
71 int data_consumption_paused
;
74 * Flag to inform the polling thread to quit when all fd hung up. Updated by
75 * the consumer_thread_receive_fds when it notices that all fds has hung up.
76 * Also updated by the signal handler (consumer_should_exit()). Read by the
82 * Global hash table containing respectively metadata and data streams. The
83 * stream element in this ht should only be updated by the metadata poll thread
84 * for the metadata and the data poll thread for the data.
86 static struct lttng_ht
*metadata_ht
;
87 static struct lttng_ht
*data_ht
;
90 * Notify a thread lttng pipe to poll back again. This usually means that some
91 * global state has changed so we just send back the thread in a poll wait
94 static void notify_thread_lttng_pipe(struct lttng_pipe
*pipe
)
96 struct lttng_consumer_stream
*null_stream
= NULL
;
100 (void) lttng_pipe_write(pipe
, &null_stream
, sizeof(null_stream
));
103 static void notify_health_quit_pipe(int *pipe
)
107 ret
= lttng_write(pipe
[1], "4", 1);
109 PERROR("write consumer health quit");
113 static void notify_channel_pipe(struct lttng_consumer_local_data
*ctx
,
114 struct lttng_consumer_channel
*chan
,
116 enum consumer_channel_action action
)
118 struct consumer_channel_msg msg
;
121 memset(&msg
, 0, sizeof(msg
));
126 ret
= lttng_write(ctx
->consumer_channel_pipe
[1], &msg
, sizeof(msg
));
127 if (ret
< sizeof(msg
)) {
128 PERROR("notify_channel_pipe write error");
132 void notify_thread_del_channel(struct lttng_consumer_local_data
*ctx
,
135 notify_channel_pipe(ctx
, NULL
, key
, CONSUMER_CHANNEL_DEL
);
138 static int read_channel_pipe(struct lttng_consumer_local_data
*ctx
,
139 struct lttng_consumer_channel
**chan
,
141 enum consumer_channel_action
*action
)
143 struct consumer_channel_msg msg
;
146 ret
= lttng_read(ctx
->consumer_channel_pipe
[0], &msg
, sizeof(msg
));
147 if (ret
< sizeof(msg
)) {
151 *action
= msg
.action
;
159 * Cleanup the stream list of a channel. Those streams are not yet globally
162 static void clean_channel_stream_list(struct lttng_consumer_channel
*channel
)
164 struct lttng_consumer_stream
*stream
, *stmp
;
168 /* Delete streams that might have been left in the stream list. */
169 cds_list_for_each_entry_safe(stream
, stmp
, &channel
->streams
.head
,
171 cds_list_del(&stream
->send_node
);
173 * Once a stream is added to this list, the buffers were created so we
174 * have a guarantee that this call will succeed. Setting the monitor
175 * mode to 0 so we don't lock nor try to delete the stream from the
179 consumer_stream_destroy(stream
, NULL
);
184 * Find a stream. The consumer_data.lock must be locked during this
187 static struct lttng_consumer_stream
*find_stream(uint64_t key
,
190 struct lttng_ht_iter iter
;
191 struct lttng_ht_node_u64
*node
;
192 struct lttng_consumer_stream
*stream
= NULL
;
196 /* -1ULL keys are lookup failures */
197 if (key
== (uint64_t) -1ULL) {
203 lttng_ht_lookup(ht
, &key
, &iter
);
204 node
= lttng_ht_iter_get_node_u64(&iter
);
206 stream
= caa_container_of(node
, struct lttng_consumer_stream
, node
);
214 static void steal_stream_key(uint64_t key
, struct lttng_ht
*ht
)
216 struct lttng_consumer_stream
*stream
;
219 stream
= find_stream(key
, ht
);
221 stream
->key
= (uint64_t) -1ULL;
223 * We don't want the lookup to match, but we still need
224 * to iterate on this stream when iterating over the hash table. Just
225 * change the node key.
227 stream
->node
.key
= (uint64_t) -1ULL;
233 * Return a channel object for the given key.
235 * RCU read side lock MUST be acquired before calling this function and
236 * protects the channel ptr.
238 struct lttng_consumer_channel
*consumer_find_channel(uint64_t key
)
240 struct lttng_ht_iter iter
;
241 struct lttng_ht_node_u64
*node
;
242 struct lttng_consumer_channel
*channel
= NULL
;
244 /* -1ULL keys are lookup failures */
245 if (key
== (uint64_t) -1ULL) {
249 lttng_ht_lookup(consumer_data
.channel_ht
, &key
, &iter
);
250 node
= lttng_ht_iter_get_node_u64(&iter
);
252 channel
= caa_container_of(node
, struct lttng_consumer_channel
, node
);
259 * There is a possibility that the consumer does not have enough time between
260 * the close of the channel on the session daemon and the cleanup in here thus
261 * once we have a channel add with an existing key, we know for sure that this
262 * channel will eventually get cleaned up by all streams being closed.
264 * This function just nullifies the already existing channel key.
266 static void steal_channel_key(uint64_t key
)
268 struct lttng_consumer_channel
*channel
;
271 channel
= consumer_find_channel(key
);
273 channel
->key
= (uint64_t) -1ULL;
275 * We don't want the lookup to match, but we still need to iterate on
276 * this channel when iterating over the hash table. Just change the
279 channel
->node
.key
= (uint64_t) -1ULL;
284 static void free_channel_rcu(struct rcu_head
*head
)
286 struct lttng_ht_node_u64
*node
=
287 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
288 struct lttng_consumer_channel
*channel
=
289 caa_container_of(node
, struct lttng_consumer_channel
, node
);
291 switch (consumer_data
.type
) {
292 case LTTNG_CONSUMER_KERNEL
:
294 case LTTNG_CONSUMER32_UST
:
295 case LTTNG_CONSUMER64_UST
:
296 lttng_ustconsumer_free_channel(channel
);
299 ERR("Unknown consumer_data type");
306 * RCU protected relayd socket pair free.
308 static void free_relayd_rcu(struct rcu_head
*head
)
310 struct lttng_ht_node_u64
*node
=
311 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
312 struct consumer_relayd_sock_pair
*relayd
=
313 caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
316 * Close all sockets. This is done in the call RCU since we don't want the
317 * socket fds to be reassigned thus potentially creating bad state of the
320 * We do not have to lock the control socket mutex here since at this stage
321 * there is no one referencing to this relayd object.
323 (void) relayd_close(&relayd
->control_sock
);
324 (void) relayd_close(&relayd
->data_sock
);
330 * Destroy and free relayd socket pair object.
332 void consumer_destroy_relayd(struct consumer_relayd_sock_pair
*relayd
)
335 struct lttng_ht_iter iter
;
337 if (relayd
== NULL
) {
341 DBG("Consumer destroy and close relayd socket pair");
343 iter
.iter
.node
= &relayd
->node
.node
;
344 ret
= lttng_ht_del(consumer_data
.relayd_ht
, &iter
);
346 /* We assume the relayd is being or is destroyed */
350 /* RCU free() call */
351 call_rcu(&relayd
->node
.head
, free_relayd_rcu
);
355 * Remove a channel from the global list protected by a mutex. This function is
356 * also responsible for freeing its data structures.
358 void consumer_del_channel(struct lttng_consumer_channel
*channel
)
361 struct lttng_ht_iter iter
;
363 DBG("Consumer delete channel key %" PRIu64
, channel
->key
);
365 pthread_mutex_lock(&consumer_data
.lock
);
366 pthread_mutex_lock(&channel
->lock
);
368 /* Destroy streams that might have been left in the stream list. */
369 clean_channel_stream_list(channel
);
371 if (channel
->live_timer_enabled
== 1) {
372 consumer_timer_live_stop(channel
);
374 if (channel
->monitor_timer_enabled
== 1) {
375 consumer_timer_monitor_stop(channel
);
378 switch (consumer_data
.type
) {
379 case LTTNG_CONSUMER_KERNEL
:
381 case LTTNG_CONSUMER32_UST
:
382 case LTTNG_CONSUMER64_UST
:
383 lttng_ustconsumer_del_channel(channel
);
386 ERR("Unknown consumer_data type");
392 iter
.iter
.node
= &channel
->node
.node
;
393 ret
= lttng_ht_del(consumer_data
.channel_ht
, &iter
);
397 call_rcu(&channel
->node
.head
, free_channel_rcu
);
399 pthread_mutex_unlock(&channel
->lock
);
400 pthread_mutex_unlock(&consumer_data
.lock
);
404 * Iterate over the relayd hash table and destroy each element. Finally,
405 * destroy the whole hash table.
407 static void cleanup_relayd_ht(void)
409 struct lttng_ht_iter iter
;
410 struct consumer_relayd_sock_pair
*relayd
;
414 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
416 consumer_destroy_relayd(relayd
);
421 lttng_ht_destroy(consumer_data
.relayd_ht
);
425 * Update the end point status of all streams having the given network sequence
426 * index (relayd index).
428 * It's atomically set without having the stream mutex locked which is fine
429 * because we handle the write/read race with a pipe wakeup for each thread.
431 static void update_endpoint_status_by_netidx(uint64_t net_seq_idx
,
432 enum consumer_endpoint_status status
)
434 struct lttng_ht_iter iter
;
435 struct lttng_consumer_stream
*stream
;
437 DBG("Consumer set delete flag on stream by idx %" PRIu64
, net_seq_idx
);
441 /* Let's begin with metadata */
442 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
443 if (stream
->net_seq_idx
== net_seq_idx
) {
444 uatomic_set(&stream
->endpoint_status
, status
);
445 DBG("Delete flag set to metadata stream %d", stream
->wait_fd
);
449 /* Follow up by the data streams */
450 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
451 if (stream
->net_seq_idx
== net_seq_idx
) {
452 uatomic_set(&stream
->endpoint_status
, status
);
453 DBG("Delete flag set to data stream %d", stream
->wait_fd
);
460 * Cleanup a relayd object by flagging every associated streams for deletion,
461 * destroying the object meaning removing it from the relayd hash table,
462 * closing the sockets and freeing the memory in a RCU call.
464 * If a local data context is available, notify the threads that the streams'
465 * state have changed.
467 static void cleanup_relayd(struct consumer_relayd_sock_pair
*relayd
,
468 struct lttng_consumer_local_data
*ctx
)
474 DBG("Cleaning up relayd sockets");
476 /* Save the net sequence index before destroying the object */
477 netidx
= relayd
->net_seq_idx
;
480 * Delete the relayd from the relayd hash table, close the sockets and free
481 * the object in a RCU call.
483 consumer_destroy_relayd(relayd
);
485 /* Set inactive endpoint to all streams */
486 update_endpoint_status_by_netidx(netidx
, CONSUMER_ENDPOINT_INACTIVE
);
489 * With a local data context, notify the threads that the streams' state
490 * have changed. The write() action on the pipe acts as an "implicit"
491 * memory barrier ordering the updates of the end point status from the
492 * read of this status which happens AFTER receiving this notify.
495 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
496 notify_thread_lttng_pipe(ctx
->consumer_metadata_pipe
);
501 * Flag a relayd socket pair for destruction. Destroy it if the refcount
504 * RCU read side lock MUST be aquired before calling this function.
506 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair
*relayd
)
510 /* Set destroy flag for this object */
511 uatomic_set(&relayd
->destroy_flag
, 1);
513 /* Destroy the relayd if refcount is 0 */
514 if (uatomic_read(&relayd
->refcount
) == 0) {
515 consumer_destroy_relayd(relayd
);
520 * Completly destroy stream from every visiable data structure and the given
523 * One this call returns, the stream object is not longer usable nor visible.
525 void consumer_del_stream(struct lttng_consumer_stream
*stream
,
528 consumer_stream_destroy(stream
, ht
);
532 * XXX naming of del vs destroy is all mixed up.
534 void consumer_del_stream_for_data(struct lttng_consumer_stream
*stream
)
536 consumer_stream_destroy(stream
, data_ht
);
539 void consumer_del_stream_for_metadata(struct lttng_consumer_stream
*stream
)
541 consumer_stream_destroy(stream
, metadata_ht
);
544 void consumer_stream_update_channel_attributes(
545 struct lttng_consumer_stream
*stream
,
546 struct lttng_consumer_channel
*channel
)
548 stream
->channel_read_only_attributes
.tracefile_size
=
549 channel
->tracefile_size
;
550 memcpy(stream
->channel_read_only_attributes
.path
, channel
->pathname
,
551 sizeof(stream
->channel_read_only_attributes
.path
));
554 struct lttng_consumer_stream
*consumer_allocate_stream(uint64_t channel_key
,
556 enum lttng_consumer_stream_state state
,
557 const char *channel_name
,
564 enum consumer_channel_type type
,
565 unsigned int monitor
)
568 struct lttng_consumer_stream
*stream
;
570 stream
= zmalloc(sizeof(*stream
));
571 if (stream
== NULL
) {
572 PERROR("malloc struct lttng_consumer_stream");
579 stream
->key
= stream_key
;
581 stream
->out_fd_offset
= 0;
582 stream
->output_written
= 0;
583 stream
->state
= state
;
586 stream
->net_seq_idx
= relayd_id
;
587 stream
->session_id
= session_id
;
588 stream
->monitor
= monitor
;
589 stream
->endpoint_status
= CONSUMER_ENDPOINT_ACTIVE
;
590 stream
->index_file
= NULL
;
591 stream
->last_sequence_number
= -1ULL;
592 pthread_mutex_init(&stream
->lock
, NULL
);
593 pthread_mutex_init(&stream
->metadata_timer_lock
, NULL
);
595 /* If channel is the metadata, flag this stream as metadata. */
596 if (type
== CONSUMER_CHANNEL_TYPE_METADATA
) {
597 stream
->metadata_flag
= 1;
598 /* Metadata is flat out. */
599 strncpy(stream
->name
, DEFAULT_METADATA_NAME
, sizeof(stream
->name
));
600 /* Live rendez-vous point. */
601 pthread_cond_init(&stream
->metadata_rdv
, NULL
);
602 pthread_mutex_init(&stream
->metadata_rdv_lock
, NULL
);
604 /* Format stream name to <channel_name>_<cpu_number> */
605 ret
= snprintf(stream
->name
, sizeof(stream
->name
), "%s_%d",
608 PERROR("snprintf stream name");
613 /* Key is always the wait_fd for streams. */
614 lttng_ht_node_init_u64(&stream
->node
, stream
->key
);
616 /* Init node per channel id key */
617 lttng_ht_node_init_u64(&stream
->node_channel_id
, channel_key
);
619 /* Init session id node with the stream session id */
620 lttng_ht_node_init_u64(&stream
->node_session_id
, stream
->session_id
);
622 DBG3("Allocated stream %s (key %" PRIu64
", chan_key %" PRIu64
623 " relayd_id %" PRIu64
", session_id %" PRIu64
,
624 stream
->name
, stream
->key
, channel_key
,
625 stream
->net_seq_idx
, stream
->session_id
);
641 * Add a stream to the global list protected by a mutex.
643 void consumer_add_data_stream(struct lttng_consumer_stream
*stream
)
645 struct lttng_ht
*ht
= data_ht
;
650 DBG3("Adding consumer stream %" PRIu64
, stream
->key
);
652 pthread_mutex_lock(&consumer_data
.lock
);
653 pthread_mutex_lock(&stream
->chan
->lock
);
654 pthread_mutex_lock(&stream
->chan
->timer_lock
);
655 pthread_mutex_lock(&stream
->lock
);
658 /* Steal stream identifier to avoid having streams with the same key */
659 steal_stream_key(stream
->key
, ht
);
661 lttng_ht_add_unique_u64(ht
, &stream
->node
);
663 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
664 &stream
->node_channel_id
);
667 * Add stream to the stream_list_ht of the consumer data. No need to steal
668 * the key since the HT does not use it and we allow to add redundant keys
671 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
674 * When nb_init_stream_left reaches 0, we don't need to trigger any action
675 * in terms of destroying the associated channel, because the action that
676 * causes the count to become 0 also causes a stream to be added. The
677 * channel deletion will thus be triggered by the following removal of this
680 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
681 /* Increment refcount before decrementing nb_init_stream_left */
683 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
686 /* Update consumer data once the node is inserted. */
687 consumer_data
.stream_count
++;
688 consumer_data
.need_update
= 1;
691 pthread_mutex_unlock(&stream
->lock
);
692 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
693 pthread_mutex_unlock(&stream
->chan
->lock
);
694 pthread_mutex_unlock(&consumer_data
.lock
);
697 void consumer_del_data_stream(struct lttng_consumer_stream
*stream
)
699 consumer_del_stream(stream
, data_ht
);
703 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
704 * be acquired before calling this.
706 static int add_relayd(struct consumer_relayd_sock_pair
*relayd
)
709 struct lttng_ht_node_u64
*node
;
710 struct lttng_ht_iter iter
;
714 lttng_ht_lookup(consumer_data
.relayd_ht
,
715 &relayd
->net_seq_idx
, &iter
);
716 node
= lttng_ht_iter_get_node_u64(&iter
);
720 lttng_ht_add_unique_u64(consumer_data
.relayd_ht
, &relayd
->node
);
727 * Allocate and return a consumer relayd socket.
729 static struct consumer_relayd_sock_pair
*consumer_allocate_relayd_sock_pair(
730 uint64_t net_seq_idx
)
732 struct consumer_relayd_sock_pair
*obj
= NULL
;
734 /* net sequence index of -1 is a failure */
735 if (net_seq_idx
== (uint64_t) -1ULL) {
739 obj
= zmalloc(sizeof(struct consumer_relayd_sock_pair
));
741 PERROR("zmalloc relayd sock");
745 obj
->net_seq_idx
= net_seq_idx
;
747 obj
->destroy_flag
= 0;
748 obj
->control_sock
.sock
.fd
= -1;
749 obj
->data_sock
.sock
.fd
= -1;
750 lttng_ht_node_init_u64(&obj
->node
, obj
->net_seq_idx
);
751 pthread_mutex_init(&obj
->ctrl_sock_mutex
, NULL
);
758 * Find a relayd socket pair in the global consumer data.
760 * Return the object if found else NULL.
761 * RCU read-side lock must be held across this call and while using the
764 struct consumer_relayd_sock_pair
*consumer_find_relayd(uint64_t key
)
766 struct lttng_ht_iter iter
;
767 struct lttng_ht_node_u64
*node
;
768 struct consumer_relayd_sock_pair
*relayd
= NULL
;
770 /* Negative keys are lookup failures */
771 if (key
== (uint64_t) -1ULL) {
775 lttng_ht_lookup(consumer_data
.relayd_ht
, &key
,
777 node
= lttng_ht_iter_get_node_u64(&iter
);
779 relayd
= caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
787 * Find a relayd and send the stream
789 * Returns 0 on success, < 0 on error
791 int consumer_send_relayd_stream(struct lttng_consumer_stream
*stream
,
795 struct consumer_relayd_sock_pair
*relayd
;
798 assert(stream
->net_seq_idx
!= -1ULL);
801 /* The stream is not metadata. Get relayd reference if exists. */
803 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
804 if (relayd
!= NULL
) {
805 /* Add stream on the relayd */
806 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
807 ret
= relayd_add_stream(&relayd
->control_sock
, stream
->name
,
808 path
, &stream
->relayd_stream_id
,
809 stream
->chan
->tracefile_size
, stream
->chan
->tracefile_count
);
810 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
815 uatomic_inc(&relayd
->refcount
);
816 stream
->sent_to_relayd
= 1;
818 ERR("Stream %" PRIu64
" relayd ID %" PRIu64
" unknown. Can't send it.",
819 stream
->key
, stream
->net_seq_idx
);
824 DBG("Stream %s with key %" PRIu64
" sent to relayd id %" PRIu64
,
825 stream
->name
, stream
->key
, stream
->net_seq_idx
);
833 * Find a relayd and send the streams sent message
835 * Returns 0 on success, < 0 on error
837 int consumer_send_relayd_streams_sent(uint64_t net_seq_idx
)
840 struct consumer_relayd_sock_pair
*relayd
;
842 assert(net_seq_idx
!= -1ULL);
844 /* The stream is not metadata. Get relayd reference if exists. */
846 relayd
= consumer_find_relayd(net_seq_idx
);
847 if (relayd
!= NULL
) {
848 /* Add stream on the relayd */
849 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
850 ret
= relayd_streams_sent(&relayd
->control_sock
);
851 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
856 ERR("Relayd ID %" PRIu64
" unknown. Can't send streams_sent.",
863 DBG("All streams sent relayd id %" PRIu64
, net_seq_idx
);
871 * Find a relayd and close the stream
873 void close_relayd_stream(struct lttng_consumer_stream
*stream
)
875 struct consumer_relayd_sock_pair
*relayd
;
877 /* The stream is not metadata. Get relayd reference if exists. */
879 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
881 consumer_stream_relayd_close(stream
, relayd
);
887 * Handle stream for relayd transmission if the stream applies for network
888 * streaming where the net sequence index is set.
890 * Return destination file descriptor or negative value on error.
892 static int write_relayd_stream_header(struct lttng_consumer_stream
*stream
,
893 size_t data_size
, unsigned long padding
,
894 struct consumer_relayd_sock_pair
*relayd
)
897 struct lttcomm_relayd_data_hdr data_hdr
;
903 /* Reset data header */
904 memset(&data_hdr
, 0, sizeof(data_hdr
));
906 if (stream
->metadata_flag
) {
907 /* Caller MUST acquire the relayd control socket lock */
908 ret
= relayd_send_metadata(&relayd
->control_sock
, data_size
);
913 /* Metadata are always sent on the control socket. */
914 outfd
= relayd
->control_sock
.sock
.fd
;
916 /* Set header with stream information */
917 data_hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
918 data_hdr
.data_size
= htobe32(data_size
);
919 data_hdr
.padding_size
= htobe32(padding
);
921 * Note that net_seq_num below is assigned with the *current* value of
922 * next_net_seq_num and only after that the next_net_seq_num will be
923 * increment. This is why when issuing a command on the relayd using
924 * this next value, 1 should always be substracted in order to compare
925 * the last seen sequence number on the relayd side to the last sent.
927 data_hdr
.net_seq_num
= htobe64(stream
->next_net_seq_num
);
928 /* Other fields are zeroed previously */
930 ret
= relayd_send_data_hdr(&relayd
->data_sock
, &data_hdr
,
936 ++stream
->next_net_seq_num
;
938 /* Set to go on data socket */
939 outfd
= relayd
->data_sock
.sock
.fd
;
947 * Allocate and return a new lttng_consumer_channel object using the given key
948 * to initialize the hash table node.
950 * On error, return NULL.
952 struct lttng_consumer_channel
*consumer_allocate_channel(uint64_t key
,
954 const char *pathname
,
959 enum lttng_event_output output
,
960 uint64_t tracefile_size
,
961 uint64_t tracefile_count
,
962 uint64_t session_id_per_pid
,
963 unsigned int monitor
,
964 unsigned int live_timer_interval
,
965 const char *root_shm_path
,
966 const char *shm_path
)
968 struct lttng_consumer_channel
*channel
;
970 channel
= zmalloc(sizeof(*channel
));
971 if (channel
== NULL
) {
972 PERROR("malloc struct lttng_consumer_channel");
977 channel
->refcount
= 0;
978 channel
->session_id
= session_id
;
979 channel
->session_id_per_pid
= session_id_per_pid
;
982 channel
->relayd_id
= relayd_id
;
983 channel
->tracefile_size
= tracefile_size
;
984 channel
->tracefile_count
= tracefile_count
;
985 channel
->monitor
= monitor
;
986 channel
->live_timer_interval
= live_timer_interval
;
987 pthread_mutex_init(&channel
->lock
, NULL
);
988 pthread_mutex_init(&channel
->timer_lock
, NULL
);
991 case LTTNG_EVENT_SPLICE
:
992 channel
->output
= CONSUMER_CHANNEL_SPLICE
;
994 case LTTNG_EVENT_MMAP
:
995 channel
->output
= CONSUMER_CHANNEL_MMAP
;
1005 * In monitor mode, the streams associated with the channel will be put in
1006 * a special list ONLY owned by this channel. So, the refcount is set to 1
1007 * here meaning that the channel itself has streams that are referenced.
1009 * On a channel deletion, once the channel is no longer visible, the
1010 * refcount is decremented and checked for a zero value to delete it. With
1011 * streams in no monitor mode, it will now be safe to destroy the channel.
1013 if (!channel
->monitor
) {
1014 channel
->refcount
= 1;
1017 strncpy(channel
->pathname
, pathname
, sizeof(channel
->pathname
));
1018 channel
->pathname
[sizeof(channel
->pathname
) - 1] = '\0';
1020 strncpy(channel
->name
, name
, sizeof(channel
->name
));
1021 channel
->name
[sizeof(channel
->name
) - 1] = '\0';
1023 if (root_shm_path
) {
1024 strncpy(channel
->root_shm_path
, root_shm_path
, sizeof(channel
->root_shm_path
));
1025 channel
->root_shm_path
[sizeof(channel
->root_shm_path
) - 1] = '\0';
1028 strncpy(channel
->shm_path
, shm_path
, sizeof(channel
->shm_path
));
1029 channel
->shm_path
[sizeof(channel
->shm_path
) - 1] = '\0';
1032 lttng_ht_node_init_u64(&channel
->node
, channel
->key
);
1034 channel
->wait_fd
= -1;
1036 CDS_INIT_LIST_HEAD(&channel
->streams
.head
);
1038 DBG("Allocated channel (key %" PRIu64
")", channel
->key
);
1045 * Add a channel to the global list protected by a mutex.
1047 * Always return 0 indicating success.
1049 int consumer_add_channel(struct lttng_consumer_channel
*channel
,
1050 struct lttng_consumer_local_data
*ctx
)
1052 pthread_mutex_lock(&consumer_data
.lock
);
1053 pthread_mutex_lock(&channel
->lock
);
1054 pthread_mutex_lock(&channel
->timer_lock
);
1057 * This gives us a guarantee that the channel we are about to add to the
1058 * channel hash table will be unique. See this function comment on the why
1059 * we need to steel the channel key at this stage.
1061 steal_channel_key(channel
->key
);
1064 lttng_ht_add_unique_u64(consumer_data
.channel_ht
, &channel
->node
);
1067 pthread_mutex_unlock(&channel
->timer_lock
);
1068 pthread_mutex_unlock(&channel
->lock
);
1069 pthread_mutex_unlock(&consumer_data
.lock
);
1071 if (channel
->wait_fd
!= -1 && channel
->type
== CONSUMER_CHANNEL_TYPE_DATA
) {
1072 notify_channel_pipe(ctx
, channel
, -1, CONSUMER_CHANNEL_ADD
);
1079 * Allocate the pollfd structure and the local view of the out fds to avoid
1080 * doing a lookup in the linked list and concurrency issues when writing is
1081 * needed. Called with consumer_data.lock held.
1083 * Returns the number of fds in the structures.
1085 static int update_poll_array(struct lttng_consumer_local_data
*ctx
,
1086 struct pollfd
**pollfd
, struct lttng_consumer_stream
**local_stream
,
1087 struct lttng_ht
*ht
, int *nb_inactive_fd
)
1090 struct lttng_ht_iter iter
;
1091 struct lttng_consumer_stream
*stream
;
1096 assert(local_stream
);
1098 DBG("Updating poll fd array");
1099 *nb_inactive_fd
= 0;
1101 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1103 * Only active streams with an active end point can be added to the
1104 * poll set and local stream storage of the thread.
1106 * There is a potential race here for endpoint_status to be updated
1107 * just after the check. However, this is OK since the stream(s) will
1108 * be deleted once the thread is notified that the end point state has
1109 * changed where this function will be called back again.
1111 * We track the number of inactive FDs because they still need to be
1112 * closed by the polling thread after a wakeup on the data_pipe or
1115 if (stream
->state
!= LTTNG_CONSUMER_ACTIVE_STREAM
||
1116 stream
->endpoint_status
== CONSUMER_ENDPOINT_INACTIVE
) {
1117 (*nb_inactive_fd
)++;
1121 * This clobbers way too much the debug output. Uncomment that if you
1122 * need it for debugging purposes.
1124 * DBG("Active FD %d", stream->wait_fd);
1126 (*pollfd
)[i
].fd
= stream
->wait_fd
;
1127 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1128 local_stream
[i
] = stream
;
1134 * Insert the consumer_data_pipe at the end of the array and don't
1135 * increment i so nb_fd is the number of real FD.
1137 (*pollfd
)[i
].fd
= lttng_pipe_get_readfd(ctx
->consumer_data_pipe
);
1138 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1140 (*pollfd
)[i
+ 1].fd
= lttng_pipe_get_readfd(ctx
->consumer_wakeup_pipe
);
1141 (*pollfd
)[i
+ 1].events
= POLLIN
| POLLPRI
;
1146 * Poll on the should_quit pipe and the command socket return -1 on
1147 * error, 1 if should exit, 0 if data is available on the command socket
1149 int lttng_consumer_poll_socket(struct pollfd
*consumer_sockpoll
)
1154 num_rdy
= poll(consumer_sockpoll
, 2, -1);
1155 if (num_rdy
== -1) {
1157 * Restart interrupted system call.
1159 if (errno
== EINTR
) {
1162 PERROR("Poll error");
1165 if (consumer_sockpoll
[0].revents
& (POLLIN
| POLLPRI
)) {
1166 DBG("consumer_should_quit wake up");
1173 * Set the error socket.
1175 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data
*ctx
,
1178 ctx
->consumer_error_socket
= sock
;
1182 * Set the command socket path.
1184 void lttng_consumer_set_command_sock_path(
1185 struct lttng_consumer_local_data
*ctx
, char *sock
)
1187 ctx
->consumer_command_sock_path
= sock
;
1191 * Send return code to the session daemon.
1192 * If the socket is not defined, we return 0, it is not a fatal error
1194 int lttng_consumer_send_error(struct lttng_consumer_local_data
*ctx
, int cmd
)
1196 if (ctx
->consumer_error_socket
> 0) {
1197 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
1198 sizeof(enum lttcomm_sessiond_command
));
1205 * Close all the tracefiles and stream fds and MUST be called when all
1206 * instances are destroyed i.e. when all threads were joined and are ended.
1208 void lttng_consumer_cleanup(void)
1210 struct lttng_ht_iter iter
;
1211 struct lttng_consumer_channel
*channel
;
1215 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
, channel
,
1217 consumer_del_channel(channel
);
1222 lttng_ht_destroy(consumer_data
.channel_ht
);
1224 cleanup_relayd_ht();
1226 lttng_ht_destroy(consumer_data
.stream_per_chan_id_ht
);
1229 * This HT contains streams that are freed by either the metadata thread or
1230 * the data thread so we do *nothing* on the hash table and simply destroy
1233 lttng_ht_destroy(consumer_data
.stream_list_ht
);
1237 * Called from signal handler.
1239 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
1243 CMM_STORE_SHARED(consumer_quit
, 1);
1244 ret
= lttng_write(ctx
->consumer_should_quit
[1], "4", 1);
1246 PERROR("write consumer quit");
1249 DBG("Consumer flag that it should quit");
1254 * Flush pending writes to trace output disk file.
1257 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1261 int outfd
= stream
->out_fd
;
1264 * This does a blocking write-and-wait on any page that belongs to the
1265 * subbuffer prior to the one we just wrote.
1266 * Don't care about error values, as these are just hints and ways to
1267 * limit the amount of page cache used.
1269 if (orig_offset
< stream
->max_sb_size
) {
1272 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1273 stream
->max_sb_size
,
1274 SYNC_FILE_RANGE_WAIT_BEFORE
1275 | SYNC_FILE_RANGE_WRITE
1276 | SYNC_FILE_RANGE_WAIT_AFTER
);
1278 * Give hints to the kernel about how we access the file:
1279 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1282 * We need to call fadvise again after the file grows because the
1283 * kernel does not seem to apply fadvise to non-existing parts of the
1286 * Call fadvise _after_ having waited for the page writeback to
1287 * complete because the dirty page writeback semantic is not well
1288 * defined. So it can be expected to lead to lower throughput in
1291 ret
= posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1292 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1293 if (ret
&& ret
!= -ENOSYS
) {
1295 PERROR("posix_fadvise on fd %i", outfd
);
1300 * Initialise the necessary environnement :
1301 * - create a new context
1302 * - create the poll_pipe
1303 * - create the should_quit pipe (for signal handler)
1304 * - create the thread pipe (for splice)
1306 * Takes a function pointer as argument, this function is called when data is
1307 * available on a buffer. This function is responsible to do the
1308 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1309 * buffer configuration and then kernctl_put_next_subbuf at the end.
1311 * Returns a pointer to the new context or NULL on error.
1313 struct lttng_consumer_local_data
*lttng_consumer_create(
1314 enum lttng_consumer_type type
,
1315 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1316 struct lttng_consumer_local_data
*ctx
),
1317 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1318 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1319 int (*update_stream
)(uint64_t stream_key
, uint32_t state
))
1322 struct lttng_consumer_local_data
*ctx
;
1324 assert(consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1325 consumer_data
.type
== type
);
1326 consumer_data
.type
= type
;
1328 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
1330 PERROR("allocating context");
1334 ctx
->consumer_error_socket
= -1;
1335 ctx
->consumer_metadata_socket
= -1;
1336 pthread_mutex_init(&ctx
->metadata_socket_lock
, NULL
);
1337 /* assign the callbacks */
1338 ctx
->on_buffer_ready
= buffer_ready
;
1339 ctx
->on_recv_channel
= recv_channel
;
1340 ctx
->on_recv_stream
= recv_stream
;
1341 ctx
->on_update_stream
= update_stream
;
1343 ctx
->consumer_data_pipe
= lttng_pipe_open(0);
1344 if (!ctx
->consumer_data_pipe
) {
1345 goto error_poll_pipe
;
1348 ctx
->consumer_wakeup_pipe
= lttng_pipe_open(0);
1349 if (!ctx
->consumer_wakeup_pipe
) {
1350 goto error_wakeup_pipe
;
1353 ret
= pipe(ctx
->consumer_should_quit
);
1355 PERROR("Error creating recv pipe");
1356 goto error_quit_pipe
;
1359 ret
= pipe(ctx
->consumer_channel_pipe
);
1361 PERROR("Error creating channel pipe");
1362 goto error_channel_pipe
;
1365 ctx
->consumer_metadata_pipe
= lttng_pipe_open(0);
1366 if (!ctx
->consumer_metadata_pipe
) {
1367 goto error_metadata_pipe
;
1370 ctx
->channel_monitor_pipe
= -1;
1374 error_metadata_pipe
:
1375 utils_close_pipe(ctx
->consumer_channel_pipe
);
1377 utils_close_pipe(ctx
->consumer_should_quit
);
1379 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1381 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1389 * Iterate over all streams of the hashtable and free them properly.
1391 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1393 struct lttng_ht_iter iter
;
1394 struct lttng_consumer_stream
*stream
;
1401 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1403 * Ignore return value since we are currently cleaning up so any error
1406 (void) consumer_del_stream(stream
, ht
);
1410 lttng_ht_destroy(ht
);
1414 * Iterate over all streams of the metadata hashtable and free them
1417 static void destroy_metadata_stream_ht(struct lttng_ht
*ht
)
1419 struct lttng_ht_iter iter
;
1420 struct lttng_consumer_stream
*stream
;
1427 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1429 * Ignore return value since we are currently cleaning up so any error
1432 (void) consumer_del_metadata_stream(stream
, ht
);
1436 lttng_ht_destroy(ht
);
1440 * Close all fds associated with the instance and free the context.
1442 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1446 DBG("Consumer destroying it. Closing everything.");
1452 destroy_data_stream_ht(data_ht
);
1453 destroy_metadata_stream_ht(metadata_ht
);
1455 ret
= close(ctx
->consumer_error_socket
);
1459 ret
= close(ctx
->consumer_metadata_socket
);
1463 utils_close_pipe(ctx
->consumer_channel_pipe
);
1464 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1465 lttng_pipe_destroy(ctx
->consumer_metadata_pipe
);
1466 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1467 utils_close_pipe(ctx
->consumer_should_quit
);
1469 unlink(ctx
->consumer_command_sock_path
);
1474 * Write the metadata stream id on the specified file descriptor.
1476 static int write_relayd_metadata_id(int fd
,
1477 struct lttng_consumer_stream
*stream
,
1478 struct consumer_relayd_sock_pair
*relayd
, unsigned long padding
)
1481 struct lttcomm_relayd_metadata_payload hdr
;
1483 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1484 hdr
.padding_size
= htobe32(padding
);
1485 ret
= lttng_write(fd
, (void *) &hdr
, sizeof(hdr
));
1486 if (ret
< sizeof(hdr
)) {
1488 * This error means that the fd's end is closed so ignore the PERROR
1489 * not to clubber the error output since this can happen in a normal
1492 if (errno
!= EPIPE
) {
1493 PERROR("write metadata stream id");
1495 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1497 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1498 * handle writting the missing part so report that as an error and
1499 * don't lie to the caller.
1504 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1505 stream
->relayd_stream_id
, padding
);
1512 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1513 * core function for writing trace buffers to either the local filesystem or
1516 * It must be called with the stream lock held.
1518 * Careful review MUST be put if any changes occur!
1520 * Returns the number of bytes written
1522 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1523 struct lttng_consumer_local_data
*ctx
,
1524 struct lttng_consumer_stream
*stream
, unsigned long len
,
1525 unsigned long padding
,
1526 struct ctf_packet_index
*index
)
1528 unsigned long mmap_offset
;
1531 off_t orig_offset
= stream
->out_fd_offset
;
1532 /* Default is on the disk */
1533 int outfd
= stream
->out_fd
;
1534 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1535 unsigned int relayd_hang_up
= 0;
1537 /* RCU lock for the relayd pointer */
1540 /* Flag that the current stream if set for network streaming. */
1541 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1542 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1543 if (relayd
== NULL
) {
1549 /* get the offset inside the fd to mmap */
1550 switch (consumer_data
.type
) {
1551 case LTTNG_CONSUMER_KERNEL
:
1552 mmap_base
= stream
->mmap_base
;
1553 ret
= kernctl_get_mmap_read_offset(stream
->wait_fd
, &mmap_offset
);
1555 PERROR("tracer ctl get_mmap_read_offset");
1559 case LTTNG_CONSUMER32_UST
:
1560 case LTTNG_CONSUMER64_UST
:
1561 mmap_base
= lttng_ustctl_get_mmap_base(stream
);
1563 ERR("read mmap get mmap base for stream %s", stream
->name
);
1567 ret
= lttng_ustctl_get_mmap_read_offset(stream
, &mmap_offset
);
1569 PERROR("tracer ctl get_mmap_read_offset");
1575 ERR("Unknown consumer_data type");
1579 /* Handle stream on the relayd if the output is on the network */
1581 unsigned long netlen
= len
;
1584 * Lock the control socket for the complete duration of the function
1585 * since from this point on we will use the socket.
1587 if (stream
->metadata_flag
) {
1588 /* Metadata requires the control socket. */
1589 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1590 if (stream
->reset_metadata_flag
) {
1591 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1592 stream
->relayd_stream_id
,
1593 stream
->metadata_version
);
1598 stream
->reset_metadata_flag
= 0;
1600 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1603 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1608 /* Use the returned socket. */
1611 /* Write metadata stream id before payload */
1612 if (stream
->metadata_flag
) {
1613 ret
= write_relayd_metadata_id(outfd
, stream
, relayd
, padding
);
1620 /* No streaming, we have to set the len with the full padding */
1623 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1624 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1626 ERR("Reset metadata file");
1629 stream
->reset_metadata_flag
= 0;
1633 * Check if we need to change the tracefile before writing the packet.
1635 if (stream
->chan
->tracefile_size
> 0 &&
1636 (stream
->tracefile_size_current
+ len
) >
1637 stream
->chan
->tracefile_size
) {
1638 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1639 stream
->name
, stream
->chan
->tracefile_size
,
1640 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1641 stream
->out_fd
, &(stream
->tracefile_count_current
),
1644 ERR("Rotating output file");
1647 outfd
= stream
->out_fd
;
1649 if (stream
->index_file
) {
1650 lttng_index_file_put(stream
->index_file
);
1651 stream
->index_file
= lttng_index_file_create(stream
->chan
->pathname
,
1652 stream
->name
, stream
->uid
, stream
->gid
,
1653 stream
->chan
->tracefile_size
,
1654 stream
->tracefile_count_current
,
1655 CTF_INDEX_MAJOR
, CTF_INDEX_MINOR
);
1656 if (!stream
->index_file
) {
1661 /* Reset current size because we just perform a rotation. */
1662 stream
->tracefile_size_current
= 0;
1663 stream
->out_fd_offset
= 0;
1666 stream
->tracefile_size_current
+= len
;
1668 index
->offset
= htobe64(stream
->out_fd_offset
);
1673 * This call guarantee that len or less is returned. It's impossible to
1674 * receive a ret value that is bigger than len.
1676 ret
= lttng_write(outfd
, mmap_base
+ mmap_offset
, len
);
1677 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, len
);
1678 if (ret
< 0 || ((size_t) ret
!= len
)) {
1680 * Report error to caller if nothing was written else at least send the
1688 /* Socket operation failed. We consider the relayd dead */
1689 if (errno
== EPIPE
|| errno
== EINVAL
|| errno
== EBADF
) {
1691 * This is possible if the fd is closed on the other side
1692 * (outfd) or any write problem. It can be verbose a bit for a
1693 * normal execution if for instance the relayd is stopped
1694 * abruptly. This can happen so set this to a DBG statement.
1696 DBG("Consumer mmap write detected relayd hang up");
1698 /* Unhandled error, print it and stop function right now. */
1699 PERROR("Error in write mmap (ret %zd != len %lu)", ret
, len
);
1703 stream
->output_written
+= ret
;
1705 /* This call is useless on a socket so better save a syscall. */
1707 /* This won't block, but will start writeout asynchronously */
1708 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, len
,
1709 SYNC_FILE_RANGE_WRITE
);
1710 stream
->out_fd_offset
+= len
;
1711 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1716 * This is a special case that the relayd has closed its socket. Let's
1717 * cleanup the relayd object and all associated streams.
1719 if (relayd
&& relayd_hang_up
) {
1720 cleanup_relayd(relayd
, ctx
);
1724 /* Unlock only if ctrl socket used */
1725 if (relayd
&& stream
->metadata_flag
) {
1726 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1734 * Splice the data from the ring buffer to the tracefile.
1736 * It must be called with the stream lock held.
1738 * Returns the number of bytes spliced.
1740 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1741 struct lttng_consumer_local_data
*ctx
,
1742 struct lttng_consumer_stream
*stream
, unsigned long len
,
1743 unsigned long padding
,
1744 struct ctf_packet_index
*index
)
1746 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1748 off_t orig_offset
= stream
->out_fd_offset
;
1749 int fd
= stream
->wait_fd
;
1750 /* Default is on the disk */
1751 int outfd
= stream
->out_fd
;
1752 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1754 unsigned int relayd_hang_up
= 0;
1756 switch (consumer_data
.type
) {
1757 case LTTNG_CONSUMER_KERNEL
:
1759 case LTTNG_CONSUMER32_UST
:
1760 case LTTNG_CONSUMER64_UST
:
1761 /* Not supported for user space tracing */
1764 ERR("Unknown consumer_data type");
1768 /* RCU lock for the relayd pointer */
1771 /* Flag that the current stream if set for network streaming. */
1772 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1773 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1774 if (relayd
== NULL
) {
1779 splice_pipe
= stream
->splice_pipe
;
1781 /* Write metadata stream id before payload */
1783 unsigned long total_len
= len
;
1785 if (stream
->metadata_flag
) {
1787 * Lock the control socket for the complete duration of the function
1788 * since from this point on we will use the socket.
1790 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1792 if (stream
->reset_metadata_flag
) {
1793 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1794 stream
->relayd_stream_id
,
1795 stream
->metadata_version
);
1800 stream
->reset_metadata_flag
= 0;
1802 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
, relayd
,
1810 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1813 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1819 /* Use the returned socket. */
1822 /* No streaming, we have to set the len with the full padding */
1825 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1826 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1828 ERR("Reset metadata file");
1831 stream
->reset_metadata_flag
= 0;
1834 * Check if we need to change the tracefile before writing the packet.
1836 if (stream
->chan
->tracefile_size
> 0 &&
1837 (stream
->tracefile_size_current
+ len
) >
1838 stream
->chan
->tracefile_size
) {
1839 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1840 stream
->name
, stream
->chan
->tracefile_size
,
1841 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1842 stream
->out_fd
, &(stream
->tracefile_count_current
),
1846 ERR("Rotating output file");
1849 outfd
= stream
->out_fd
;
1851 if (stream
->index_file
) {
1852 lttng_index_file_put(stream
->index_file
);
1853 stream
->index_file
= lttng_index_file_create(stream
->chan
->pathname
,
1854 stream
->name
, stream
->uid
, stream
->gid
,
1855 stream
->chan
->tracefile_size
,
1856 stream
->tracefile_count_current
,
1857 CTF_INDEX_MAJOR
, CTF_INDEX_MINOR
);
1858 if (!stream
->index_file
) {
1863 /* Reset current size because we just perform a rotation. */
1864 stream
->tracefile_size_current
= 0;
1865 stream
->out_fd_offset
= 0;
1868 stream
->tracefile_size_current
+= len
;
1869 index
->offset
= htobe64(stream
->out_fd_offset
);
1873 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1874 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1875 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1876 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1877 DBG("splice chan to pipe, ret %zd", ret_splice
);
1878 if (ret_splice
< 0) {
1881 PERROR("Error in relay splice");
1885 /* Handle stream on the relayd if the output is on the network */
1886 if (relayd
&& stream
->metadata_flag
) {
1887 size_t metadata_payload_size
=
1888 sizeof(struct lttcomm_relayd_metadata_payload
);
1890 /* Update counter to fit the spliced data */
1891 ret_splice
+= metadata_payload_size
;
1892 len
+= metadata_payload_size
;
1894 * We do this so the return value can match the len passed as
1895 * argument to this function.
1897 written
-= metadata_payload_size
;
1900 /* Splice data out */
1901 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1902 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1903 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
1905 if (ret_splice
< 0) {
1910 } else if (ret_splice
> len
) {
1912 * We don't expect this code path to be executed but you never know
1913 * so this is an extra protection agains a buggy splice().
1916 written
+= ret_splice
;
1917 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice
,
1921 /* All good, update current len and continue. */
1925 /* This call is useless on a socket so better save a syscall. */
1927 /* This won't block, but will start writeout asynchronously */
1928 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1929 SYNC_FILE_RANGE_WRITE
);
1930 stream
->out_fd_offset
+= ret_splice
;
1932 stream
->output_written
+= ret_splice
;
1933 written
+= ret_splice
;
1936 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1942 * This is a special case that the relayd has closed its socket. Let's
1943 * cleanup the relayd object and all associated streams.
1945 if (relayd
&& relayd_hang_up
) {
1946 cleanup_relayd(relayd
, ctx
);
1947 /* Skip splice error so the consumer does not fail */
1952 /* send the appropriate error description to sessiond */
1955 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1958 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1961 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
1966 if (relayd
&& stream
->metadata_flag
) {
1967 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1975 * Take a snapshot for a specific fd
1977 * Returns 0 on success, < 0 on error
1979 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
1981 switch (consumer_data
.type
) {
1982 case LTTNG_CONSUMER_KERNEL
:
1983 return lttng_kconsumer_take_snapshot(stream
);
1984 case LTTNG_CONSUMER32_UST
:
1985 case LTTNG_CONSUMER64_UST
:
1986 return lttng_ustconsumer_take_snapshot(stream
);
1988 ERR("Unknown consumer_data type");
1995 * Get the produced position
1997 * Returns 0 on success, < 0 on error
1999 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
2002 switch (consumer_data
.type
) {
2003 case LTTNG_CONSUMER_KERNEL
:
2004 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
2005 case LTTNG_CONSUMER32_UST
:
2006 case LTTNG_CONSUMER64_UST
:
2007 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
2009 ERR("Unknown consumer_data type");
2015 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
2016 int sock
, struct pollfd
*consumer_sockpoll
)
2018 switch (consumer_data
.type
) {
2019 case LTTNG_CONSUMER_KERNEL
:
2020 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2021 case LTTNG_CONSUMER32_UST
:
2022 case LTTNG_CONSUMER64_UST
:
2023 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2025 ERR("Unknown consumer_data type");
2031 void lttng_consumer_close_all_metadata(void)
2033 switch (consumer_data
.type
) {
2034 case LTTNG_CONSUMER_KERNEL
:
2036 * The Kernel consumer has a different metadata scheme so we don't
2037 * close anything because the stream will be closed by the session
2041 case LTTNG_CONSUMER32_UST
:
2042 case LTTNG_CONSUMER64_UST
:
2044 * Close all metadata streams. The metadata hash table is passed and
2045 * this call iterates over it by closing all wakeup fd. This is safe
2046 * because at this point we are sure that the metadata producer is
2047 * either dead or blocked.
2049 lttng_ustconsumer_close_all_metadata(metadata_ht
);
2052 ERR("Unknown consumer_data type");
2058 * Clean up a metadata stream and free its memory.
2060 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
2061 struct lttng_ht
*ht
)
2063 struct lttng_consumer_channel
*free_chan
= NULL
;
2067 * This call should NEVER receive regular stream. It must always be
2068 * metadata stream and this is crucial for data structure synchronization.
2070 assert(stream
->metadata_flag
);
2072 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
2074 pthread_mutex_lock(&consumer_data
.lock
);
2075 pthread_mutex_lock(&stream
->chan
->lock
);
2076 pthread_mutex_lock(&stream
->lock
);
2077 if (stream
->chan
->metadata_cache
) {
2078 /* Only applicable to userspace consumers. */
2079 pthread_mutex_lock(&stream
->chan
->metadata_cache
->lock
);
2082 /* Remove any reference to that stream. */
2083 consumer_stream_delete(stream
, ht
);
2085 /* Close down everything including the relayd if one. */
2086 consumer_stream_close(stream
);
2087 /* Destroy tracer buffers of the stream. */
2088 consumer_stream_destroy_buffers(stream
);
2090 /* Atomically decrement channel refcount since other threads can use it. */
2091 if (!uatomic_sub_return(&stream
->chan
->refcount
, 1)
2092 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
2093 /* Go for channel deletion! */
2094 free_chan
= stream
->chan
;
2098 * Nullify the stream reference so it is not used after deletion. The
2099 * channel lock MUST be acquired before being able to check for a NULL
2102 stream
->chan
->metadata_stream
= NULL
;
2104 if (stream
->chan
->metadata_cache
) {
2105 pthread_mutex_unlock(&stream
->chan
->metadata_cache
->lock
);
2107 pthread_mutex_unlock(&stream
->lock
);
2108 pthread_mutex_unlock(&stream
->chan
->lock
);
2109 pthread_mutex_unlock(&consumer_data
.lock
);
2112 consumer_del_channel(free_chan
);
2115 consumer_stream_free(stream
);
2119 * Action done with the metadata stream when adding it to the consumer internal
2120 * data structures to handle it.
2122 void consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
2124 struct lttng_ht
*ht
= metadata_ht
;
2125 struct lttng_ht_iter iter
;
2126 struct lttng_ht_node_u64
*node
;
2131 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2133 pthread_mutex_lock(&consumer_data
.lock
);
2134 pthread_mutex_lock(&stream
->chan
->lock
);
2135 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2136 pthread_mutex_lock(&stream
->lock
);
2139 * From here, refcounts are updated so be _careful_ when returning an error
2146 * Lookup the stream just to make sure it does not exist in our internal
2147 * state. This should NEVER happen.
2149 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2150 node
= lttng_ht_iter_get_node_u64(&iter
);
2154 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2155 * in terms of destroying the associated channel, because the action that
2156 * causes the count to become 0 also causes a stream to be added. The
2157 * channel deletion will thus be triggered by the following removal of this
2160 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2161 /* Increment refcount before decrementing nb_init_stream_left */
2163 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2166 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2168 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
2169 &stream
->node_channel_id
);
2172 * Add stream to the stream_list_ht of the consumer data. No need to steal
2173 * the key since the HT does not use it and we allow to add redundant keys
2176 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2180 pthread_mutex_unlock(&stream
->lock
);
2181 pthread_mutex_unlock(&stream
->chan
->lock
);
2182 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2183 pthread_mutex_unlock(&consumer_data
.lock
);
2187 * Delete data stream that are flagged for deletion (endpoint_status).
2189 static void validate_endpoint_status_data_stream(void)
2191 struct lttng_ht_iter iter
;
2192 struct lttng_consumer_stream
*stream
;
2194 DBG("Consumer delete flagged data stream");
2197 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2198 /* Validate delete flag of the stream */
2199 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2202 /* Delete it right now */
2203 consumer_del_stream(stream
, data_ht
);
2209 * Delete metadata stream that are flagged for deletion (endpoint_status).
2211 static void validate_endpoint_status_metadata_stream(
2212 struct lttng_poll_event
*pollset
)
2214 struct lttng_ht_iter iter
;
2215 struct lttng_consumer_stream
*stream
;
2217 DBG("Consumer delete flagged metadata stream");
2222 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2223 /* Validate delete flag of the stream */
2224 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2228 * Remove from pollset so the metadata thread can continue without
2229 * blocking on a deleted stream.
2231 lttng_poll_del(pollset
, stream
->wait_fd
);
2233 /* Delete it right now */
2234 consumer_del_metadata_stream(stream
, metadata_ht
);
2240 * Thread polls on metadata file descriptor and write them on disk or on the
2243 void *consumer_thread_metadata_poll(void *data
)
2245 int ret
, i
, pollfd
, err
= -1;
2246 uint32_t revents
, nb_fd
;
2247 struct lttng_consumer_stream
*stream
= NULL
;
2248 struct lttng_ht_iter iter
;
2249 struct lttng_ht_node_u64
*node
;
2250 struct lttng_poll_event events
;
2251 struct lttng_consumer_local_data
*ctx
= data
;
2254 rcu_register_thread();
2256 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2258 if (testpoint(consumerd_thread_metadata
)) {
2259 goto error_testpoint
;
2262 health_code_update();
2264 DBG("Thread metadata poll started");
2266 /* Size is set to 1 for the consumer_metadata pipe */
2267 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2269 ERR("Poll set creation failed");
2273 ret
= lttng_poll_add(&events
,
2274 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2280 DBG("Metadata main loop started");
2284 health_code_update();
2285 health_poll_entry();
2286 DBG("Metadata poll wait");
2287 ret
= lttng_poll_wait(&events
, -1);
2288 DBG("Metadata poll return from wait with %d fd(s)",
2289 LTTNG_POLL_GETNB(&events
));
2291 DBG("Metadata event caught in thread");
2293 if (errno
== EINTR
) {
2294 ERR("Poll EINTR caught");
2297 if (LTTNG_POLL_GETNB(&events
) == 0) {
2298 err
= 0; /* All is OK */
2305 /* From here, the event is a metadata wait fd */
2306 for (i
= 0; i
< nb_fd
; i
++) {
2307 health_code_update();
2309 revents
= LTTNG_POLL_GETEV(&events
, i
);
2310 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2313 /* No activity for this FD (poll implementation). */
2317 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2318 if (revents
& LPOLLIN
) {
2321 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2322 &stream
, sizeof(stream
));
2323 if (pipe_len
< sizeof(stream
)) {
2325 PERROR("read metadata stream");
2328 * Remove the pipe from the poll set and continue the loop
2329 * since their might be data to consume.
2331 lttng_poll_del(&events
,
2332 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2333 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2337 /* A NULL stream means that the state has changed. */
2338 if (stream
== NULL
) {
2339 /* Check for deleted streams. */
2340 validate_endpoint_status_metadata_stream(&events
);
2344 DBG("Adding metadata stream %d to poll set",
2347 /* Add metadata stream to the global poll events list */
2348 lttng_poll_add(&events
, stream
->wait_fd
,
2349 LPOLLIN
| LPOLLPRI
| LPOLLHUP
);
2350 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2351 DBG("Metadata thread pipe hung up");
2353 * Remove the pipe from the poll set and continue the loop
2354 * since their might be data to consume.
2356 lttng_poll_del(&events
,
2357 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2358 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2361 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2365 /* Handle other stream */
2371 uint64_t tmp_id
= (uint64_t) pollfd
;
2373 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2375 node
= lttng_ht_iter_get_node_u64(&iter
);
2378 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2381 if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2382 /* Get the data out of the metadata file descriptor */
2383 DBG("Metadata available on fd %d", pollfd
);
2384 assert(stream
->wait_fd
== pollfd
);
2387 health_code_update();
2389 len
= ctx
->on_buffer_ready(stream
, ctx
);
2391 * We don't check the return value here since if we get
2392 * a negative len, it means an error occurred thus we
2393 * simply remove it from the poll set and free the
2398 /* It's ok to have an unavailable sub-buffer */
2399 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2400 /* Clean up stream from consumer and free it. */
2401 lttng_poll_del(&events
, stream
->wait_fd
);
2402 consumer_del_metadata_stream(stream
, metadata_ht
);
2404 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2405 DBG("Metadata fd %d is hup|err.", pollfd
);
2406 if (!stream
->hangup_flush_done
2407 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2408 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2409 DBG("Attempting to flush and consume the UST buffers");
2410 lttng_ustconsumer_on_stream_hangup(stream
);
2412 /* We just flushed the stream now read it. */
2414 health_code_update();
2416 len
= ctx
->on_buffer_ready(stream
, ctx
);
2418 * We don't check the return value here since if we get
2419 * a negative len, it means an error occurred thus we
2420 * simply remove it from the poll set and free the
2426 lttng_poll_del(&events
, stream
->wait_fd
);
2428 * This call update the channel states, closes file descriptors
2429 * and securely free the stream.
2431 consumer_del_metadata_stream(stream
, metadata_ht
);
2433 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2437 /* Release RCU lock for the stream looked up */
2445 DBG("Metadata poll thread exiting");
2447 lttng_poll_clean(&events
);
2452 ERR("Health error occurred in %s", __func__
);
2454 health_unregister(health_consumerd
);
2455 rcu_unregister_thread();
2460 * This thread polls the fds in the set to consume the data and write
2461 * it to tracefile if necessary.
2463 void *consumer_thread_data_poll(void *data
)
2465 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2466 struct pollfd
*pollfd
= NULL
;
2467 /* local view of the streams */
2468 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2469 /* local view of consumer_data.fds_count */
2470 int nb_fd
= 0, nb_pipes_fd
;
2471 /* Number of FDs with CONSUMER_ENDPOINT_INACTIVE but still open. */
2472 int nb_inactive_fd
= 0;
2473 struct lttng_consumer_local_data
*ctx
= data
;
2476 rcu_register_thread();
2478 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2480 if (testpoint(consumerd_thread_data
)) {
2481 goto error_testpoint
;
2484 health_code_update();
2486 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
*));
2487 if (local_stream
== NULL
) {
2488 PERROR("local_stream malloc");
2493 health_code_update();
2499 * the fds set has been updated, we need to update our
2500 * local array as well
2502 pthread_mutex_lock(&consumer_data
.lock
);
2503 if (consumer_data
.need_update
) {
2508 local_stream
= NULL
;
2511 * Allocate for all fds + 2:
2512 * +1 for the consumer_data_pipe
2513 * +1 for wake up pipe
2516 pollfd
= zmalloc((consumer_data
.stream_count
+ nb_pipes_fd
) * sizeof(struct pollfd
));
2517 if (pollfd
== NULL
) {
2518 PERROR("pollfd malloc");
2519 pthread_mutex_unlock(&consumer_data
.lock
);
2523 local_stream
= zmalloc((consumer_data
.stream_count
+ nb_pipes_fd
) *
2524 sizeof(struct lttng_consumer_stream
*));
2525 if (local_stream
== NULL
) {
2526 PERROR("local_stream malloc");
2527 pthread_mutex_unlock(&consumer_data
.lock
);
2530 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2531 data_ht
, &nb_inactive_fd
);
2533 ERR("Error in allocating pollfd or local_outfds");
2534 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2535 pthread_mutex_unlock(&consumer_data
.lock
);
2539 consumer_data
.need_update
= 0;
2541 pthread_mutex_unlock(&consumer_data
.lock
);
2543 /* No FDs and consumer_quit, consumer_cleanup the thread */
2544 if (nb_fd
== 0 && nb_inactive_fd
== 0 &&
2545 CMM_LOAD_SHARED(consumer_quit
) == 1) {
2546 err
= 0; /* All is OK */
2549 /* poll on the array of fds */
2551 DBG("polling on %d fd", nb_fd
+ nb_pipes_fd
);
2552 if (testpoint(consumerd_thread_data_poll
)) {
2555 health_poll_entry();
2556 num_rdy
= poll(pollfd
, nb_fd
+ nb_pipes_fd
, -1);
2558 DBG("poll num_rdy : %d", num_rdy
);
2559 if (num_rdy
== -1) {
2561 * Restart interrupted system call.
2563 if (errno
== EINTR
) {
2566 PERROR("Poll error");
2567 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2569 } else if (num_rdy
== 0) {
2570 DBG("Polling thread timed out");
2574 if (caa_unlikely(data_consumption_paused
)) {
2575 DBG("Data consumption paused, sleeping...");
2581 * If the consumer_data_pipe triggered poll go directly to the
2582 * beginning of the loop to update the array. We want to prioritize
2583 * array update over low-priority reads.
2585 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2586 ssize_t pipe_readlen
;
2588 DBG("consumer_data_pipe wake up");
2589 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2590 &new_stream
, sizeof(new_stream
));
2591 if (pipe_readlen
< sizeof(new_stream
)) {
2592 PERROR("Consumer data pipe");
2593 /* Continue so we can at least handle the current stream(s). */
2598 * If the stream is NULL, just ignore it. It's also possible that
2599 * the sessiond poll thread changed the consumer_quit state and is
2600 * waking us up to test it.
2602 if (new_stream
== NULL
) {
2603 validate_endpoint_status_data_stream();
2607 /* Continue to update the local streams and handle prio ones */
2611 /* Handle wakeup pipe. */
2612 if (pollfd
[nb_fd
+ 1].revents
& (POLLIN
| POLLPRI
)) {
2614 ssize_t pipe_readlen
;
2616 pipe_readlen
= lttng_pipe_read(ctx
->consumer_wakeup_pipe
, &dummy
,
2618 if (pipe_readlen
< 0) {
2619 PERROR("Consumer data wakeup pipe");
2621 /* We've been awakened to handle stream(s). */
2622 ctx
->has_wakeup
= 0;
2625 /* Take care of high priority channels first. */
2626 for (i
= 0; i
< nb_fd
; i
++) {
2627 health_code_update();
2629 if (local_stream
[i
] == NULL
) {
2632 if (pollfd
[i
].revents
& POLLPRI
) {
2633 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2635 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2636 /* it's ok to have an unavailable sub-buffer */
2637 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2638 /* Clean the stream and free it. */
2639 consumer_del_stream(local_stream
[i
], data_ht
);
2640 local_stream
[i
] = NULL
;
2641 } else if (len
> 0) {
2642 local_stream
[i
]->data_read
= 1;
2648 * If we read high prio channel in this loop, try again
2649 * for more high prio data.
2655 /* Take care of low priority channels. */
2656 for (i
= 0; i
< nb_fd
; i
++) {
2657 health_code_update();
2659 if (local_stream
[i
] == NULL
) {
2662 if ((pollfd
[i
].revents
& POLLIN
) ||
2663 local_stream
[i
]->hangup_flush_done
||
2664 local_stream
[i
]->has_data
) {
2665 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2666 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2667 /* it's ok to have an unavailable sub-buffer */
2668 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2669 /* Clean the stream and free it. */
2670 consumer_del_stream(local_stream
[i
], data_ht
);
2671 local_stream
[i
] = NULL
;
2672 } else if (len
> 0) {
2673 local_stream
[i
]->data_read
= 1;
2678 /* Handle hangup and errors */
2679 for (i
= 0; i
< nb_fd
; i
++) {
2680 health_code_update();
2682 if (local_stream
[i
] == NULL
) {
2685 if (!local_stream
[i
]->hangup_flush_done
2686 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2687 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2688 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2689 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2691 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2692 /* Attempt read again, for the data we just flushed. */
2693 local_stream
[i
]->data_read
= 1;
2696 * If the poll flag is HUP/ERR/NVAL and we have
2697 * read no data in this pass, we can remove the
2698 * stream from its hash table.
2700 if ((pollfd
[i
].revents
& POLLHUP
)) {
2701 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2702 if (!local_stream
[i
]->data_read
) {
2703 consumer_del_stream(local_stream
[i
], data_ht
);
2704 local_stream
[i
] = NULL
;
2707 } else if (pollfd
[i
].revents
& POLLERR
) {
2708 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2709 if (!local_stream
[i
]->data_read
) {
2710 consumer_del_stream(local_stream
[i
], data_ht
);
2711 local_stream
[i
] = NULL
;
2714 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2715 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2716 if (!local_stream
[i
]->data_read
) {
2717 consumer_del_stream(local_stream
[i
], data_ht
);
2718 local_stream
[i
] = NULL
;
2722 if (local_stream
[i
] != NULL
) {
2723 local_stream
[i
]->data_read
= 0;
2730 DBG("polling thread exiting");
2735 * Close the write side of the pipe so epoll_wait() in
2736 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2737 * read side of the pipe. If we close them both, epoll_wait strangely does
2738 * not return and could create a endless wait period if the pipe is the
2739 * only tracked fd in the poll set. The thread will take care of closing
2742 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2747 ERR("Health error occurred in %s", __func__
);
2749 health_unregister(health_consumerd
);
2751 rcu_unregister_thread();
2756 * Close wake-up end of each stream belonging to the channel. This will
2757 * allow the poll() on the stream read-side to detect when the
2758 * write-side (application) finally closes them.
2761 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2763 struct lttng_ht
*ht
;
2764 struct lttng_consumer_stream
*stream
;
2765 struct lttng_ht_iter iter
;
2767 ht
= consumer_data
.stream_per_chan_id_ht
;
2770 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2771 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2772 ht
->match_fct
, &channel
->key
,
2773 &iter
.iter
, stream
, node_channel_id
.node
) {
2775 * Protect against teardown with mutex.
2777 pthread_mutex_lock(&stream
->lock
);
2778 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2781 switch (consumer_data
.type
) {
2782 case LTTNG_CONSUMER_KERNEL
:
2784 case LTTNG_CONSUMER32_UST
:
2785 case LTTNG_CONSUMER64_UST
:
2786 if (stream
->metadata_flag
) {
2787 /* Safe and protected by the stream lock. */
2788 lttng_ustconsumer_close_metadata(stream
->chan
);
2791 * Note: a mutex is taken internally within
2792 * liblttng-ust-ctl to protect timer wakeup_fd
2793 * use from concurrent close.
2795 lttng_ustconsumer_close_stream_wakeup(stream
);
2799 ERR("Unknown consumer_data type");
2803 pthread_mutex_unlock(&stream
->lock
);
2808 static void destroy_channel_ht(struct lttng_ht
*ht
)
2810 struct lttng_ht_iter iter
;
2811 struct lttng_consumer_channel
*channel
;
2819 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2820 ret
= lttng_ht_del(ht
, &iter
);
2825 lttng_ht_destroy(ht
);
2829 * This thread polls the channel fds to detect when they are being
2830 * closed. It closes all related streams if the channel is detected as
2831 * closed. It is currently only used as a shim layer for UST because the
2832 * consumerd needs to keep the per-stream wakeup end of pipes open for
2835 void *consumer_thread_channel_poll(void *data
)
2837 int ret
, i
, pollfd
, err
= -1;
2838 uint32_t revents
, nb_fd
;
2839 struct lttng_consumer_channel
*chan
= NULL
;
2840 struct lttng_ht_iter iter
;
2841 struct lttng_ht_node_u64
*node
;
2842 struct lttng_poll_event events
;
2843 struct lttng_consumer_local_data
*ctx
= data
;
2844 struct lttng_ht
*channel_ht
;
2846 rcu_register_thread();
2848 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
2850 if (testpoint(consumerd_thread_channel
)) {
2851 goto error_testpoint
;
2854 health_code_update();
2856 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2858 /* ENOMEM at this point. Better to bail out. */
2862 DBG("Thread channel poll started");
2864 /* Size is set to 1 for the consumer_channel pipe */
2865 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2867 ERR("Poll set creation failed");
2871 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2877 DBG("Channel main loop started");
2881 health_code_update();
2882 DBG("Channel poll wait");
2883 health_poll_entry();
2884 ret
= lttng_poll_wait(&events
, -1);
2885 DBG("Channel poll return from wait with %d fd(s)",
2886 LTTNG_POLL_GETNB(&events
));
2888 DBG("Channel event caught in thread");
2890 if (errno
== EINTR
) {
2891 ERR("Poll EINTR caught");
2894 if (LTTNG_POLL_GETNB(&events
) == 0) {
2895 err
= 0; /* All is OK */
2902 /* From here, the event is a channel wait fd */
2903 for (i
= 0; i
< nb_fd
; i
++) {
2904 health_code_update();
2906 revents
= LTTNG_POLL_GETEV(&events
, i
);
2907 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2910 /* No activity for this FD (poll implementation). */
2914 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2915 if (revents
& LPOLLIN
) {
2916 enum consumer_channel_action action
;
2919 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
2922 ERR("Error reading channel pipe");
2924 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2929 case CONSUMER_CHANNEL_ADD
:
2930 DBG("Adding channel %d to poll set",
2933 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
2936 lttng_ht_add_unique_u64(channel_ht
,
2937 &chan
->wait_fd_node
);
2939 /* Add channel to the global poll events list */
2940 lttng_poll_add(&events
, chan
->wait_fd
,
2941 LPOLLERR
| LPOLLHUP
);
2943 case CONSUMER_CHANNEL_DEL
:
2946 * This command should never be called if the channel
2947 * has streams monitored by either the data or metadata
2948 * thread. The consumer only notify this thread with a
2949 * channel del. command if it receives a destroy
2950 * channel command from the session daemon that send it
2951 * if a command prior to the GET_CHANNEL failed.
2955 chan
= consumer_find_channel(key
);
2958 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
2961 lttng_poll_del(&events
, chan
->wait_fd
);
2962 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
2963 ret
= lttng_ht_del(channel_ht
, &iter
);
2966 switch (consumer_data
.type
) {
2967 case LTTNG_CONSUMER_KERNEL
:
2969 case LTTNG_CONSUMER32_UST
:
2970 case LTTNG_CONSUMER64_UST
:
2971 health_code_update();
2972 /* Destroy streams that might have been left in the stream list. */
2973 clean_channel_stream_list(chan
);
2976 ERR("Unknown consumer_data type");
2981 * Release our own refcount. Force channel deletion even if
2982 * streams were not initialized.
2984 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
2985 consumer_del_channel(chan
);
2990 case CONSUMER_CHANNEL_QUIT
:
2992 * Remove the pipe from the poll set and continue the loop
2993 * since their might be data to consume.
2995 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2998 ERR("Unknown action");
3001 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
3002 DBG("Channel thread pipe hung up");
3004 * Remove the pipe from the poll set and continue the loop
3005 * since their might be data to consume.
3007 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
3010 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3014 /* Handle other stream */
3020 uint64_t tmp_id
= (uint64_t) pollfd
;
3022 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
3024 node
= lttng_ht_iter_get_node_u64(&iter
);
3027 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
3030 /* Check for error event */
3031 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
3032 DBG("Channel fd %d is hup|err.", pollfd
);
3034 lttng_poll_del(&events
, chan
->wait_fd
);
3035 ret
= lttng_ht_del(channel_ht
, &iter
);
3039 * This will close the wait fd for each stream associated to
3040 * this channel AND monitored by the data/metadata thread thus
3041 * will be clean by the right thread.
3043 consumer_close_channel_streams(chan
);
3045 /* Release our own refcount */
3046 if (!uatomic_sub_return(&chan
->refcount
, 1)
3047 && !uatomic_read(&chan
->nb_init_stream_left
)) {
3048 consumer_del_channel(chan
);
3051 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3056 /* Release RCU lock for the channel looked up */
3064 lttng_poll_clean(&events
);
3066 destroy_channel_ht(channel_ht
);
3069 DBG("Channel poll thread exiting");
3072 ERR("Health error occurred in %s", __func__
);
3074 health_unregister(health_consumerd
);
3075 rcu_unregister_thread();
3079 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
3080 struct pollfd
*sockpoll
, int client_socket
)
3087 ret
= lttng_consumer_poll_socket(sockpoll
);
3091 DBG("Metadata connection on client_socket");
3093 /* Blocking call, waiting for transmission */
3094 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
3095 if (ctx
->consumer_metadata_socket
< 0) {
3096 WARN("On accept metadata");
3107 * This thread listens on the consumerd socket and receives the file
3108 * descriptors from the session daemon.
3110 void *consumer_thread_sessiond_poll(void *data
)
3112 int sock
= -1, client_socket
, ret
, err
= -1;
3114 * structure to poll for incoming data on communication socket avoids
3115 * making blocking sockets.
3117 struct pollfd consumer_sockpoll
[2];
3118 struct lttng_consumer_local_data
*ctx
= data
;
3120 rcu_register_thread();
3122 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
3124 if (testpoint(consumerd_thread_sessiond
)) {
3125 goto error_testpoint
;
3128 health_code_update();
3130 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
3131 unlink(ctx
->consumer_command_sock_path
);
3132 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
3133 if (client_socket
< 0) {
3134 ERR("Cannot create command socket");
3138 ret
= lttcomm_listen_unix_sock(client_socket
);
3143 DBG("Sending ready command to lttng-sessiond");
3144 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3145 /* return < 0 on error, but == 0 is not fatal */
3147 ERR("Error sending ready command to lttng-sessiond");
3151 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3152 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3153 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3154 consumer_sockpoll
[1].fd
= client_socket
;
3155 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3157 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3165 DBG("Connection on client_socket");
3167 /* Blocking call, waiting for transmission */
3168 sock
= lttcomm_accept_unix_sock(client_socket
);
3175 * Setup metadata socket which is the second socket connection on the
3176 * command unix socket.
3178 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3187 /* This socket is not useful anymore. */
3188 ret
= close(client_socket
);
3190 PERROR("close client_socket");
3194 /* update the polling structure to poll on the established socket */
3195 consumer_sockpoll
[1].fd
= sock
;
3196 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3199 health_code_update();
3201 health_poll_entry();
3202 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3211 DBG("Incoming command on sock");
3212 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3215 * This could simply be a session daemon quitting. Don't output
3218 DBG("Communication interrupted on command socket");
3222 if (CMM_LOAD_SHARED(consumer_quit
)) {
3223 DBG("consumer_thread_receive_fds received quit from signal");
3224 err
= 0; /* All is OK */
3227 DBG("received command on sock");
3233 DBG("Consumer thread sessiond poll exiting");
3236 * Close metadata streams since the producer is the session daemon which
3239 * NOTE: for now, this only applies to the UST tracer.
3241 lttng_consumer_close_all_metadata();
3244 * when all fds have hung up, the polling thread
3247 CMM_STORE_SHARED(consumer_quit
, 1);
3250 * Notify the data poll thread to poll back again and test the
3251 * consumer_quit state that we just set so to quit gracefully.
3253 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3255 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3257 notify_health_quit_pipe(health_quit_pipe
);
3259 /* Cleaning up possibly open sockets. */
3263 PERROR("close sock sessiond poll");
3266 if (client_socket
>= 0) {
3267 ret
= close(client_socket
);
3269 PERROR("close client_socket sessiond poll");
3276 ERR("Health error occurred in %s", __func__
);
3278 health_unregister(health_consumerd
);
3280 rcu_unregister_thread();
3284 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3285 struct lttng_consumer_local_data
*ctx
)
3289 pthread_mutex_lock(&stream
->lock
);
3290 if (stream
->metadata_flag
) {
3291 pthread_mutex_lock(&stream
->metadata_rdv_lock
);
3294 switch (consumer_data
.type
) {
3295 case LTTNG_CONSUMER_KERNEL
:
3296 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
3298 case LTTNG_CONSUMER32_UST
:
3299 case LTTNG_CONSUMER64_UST
:
3300 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
3303 ERR("Unknown consumer_data type");
3309 if (stream
->metadata_flag
) {
3310 pthread_cond_broadcast(&stream
->metadata_rdv
);
3311 pthread_mutex_unlock(&stream
->metadata_rdv_lock
);
3313 pthread_mutex_unlock(&stream
->lock
);
3317 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3319 switch (consumer_data
.type
) {
3320 case LTTNG_CONSUMER_KERNEL
:
3321 return lttng_kconsumer_on_recv_stream(stream
);
3322 case LTTNG_CONSUMER32_UST
:
3323 case LTTNG_CONSUMER64_UST
:
3324 return lttng_ustconsumer_on_recv_stream(stream
);
3326 ERR("Unknown consumer_data type");
3333 * Allocate and set consumer data hash tables.
3335 int lttng_consumer_init(void)
3337 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3338 if (!consumer_data
.channel_ht
) {
3342 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3343 if (!consumer_data
.relayd_ht
) {
3347 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3348 if (!consumer_data
.stream_list_ht
) {
3352 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3353 if (!consumer_data
.stream_per_chan_id_ht
) {
3357 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3362 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3374 * Process the ADD_RELAYD command receive by a consumer.
3376 * This will create a relayd socket pair and add it to the relayd hash table.
3377 * The caller MUST acquire a RCU read side lock before calling it.
3379 void consumer_add_relayd_socket(uint64_t net_seq_idx
, int sock_type
,
3380 struct lttng_consumer_local_data
*ctx
, int sock
,
3381 struct pollfd
*consumer_sockpoll
,
3382 struct lttcomm_relayd_sock
*relayd_sock
, uint64_t sessiond_id
,
3383 uint64_t relayd_session_id
)
3385 int fd
= -1, ret
= -1, relayd_created
= 0;
3386 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3387 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3390 assert(relayd_sock
);
3392 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", net_seq_idx
);
3394 /* Get relayd reference if exists. */
3395 relayd
= consumer_find_relayd(net_seq_idx
);
3396 if (relayd
== NULL
) {
3397 assert(sock_type
== LTTNG_STREAM_CONTROL
);
3398 /* Not found. Allocate one. */
3399 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3400 if (relayd
== NULL
) {
3401 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3404 relayd
->sessiond_session_id
= sessiond_id
;
3409 * This code path MUST continue to the consumer send status message to
3410 * we can notify the session daemon and continue our work without
3411 * killing everything.
3415 * relayd key should never be found for control socket.
3417 assert(sock_type
!= LTTNG_STREAM_CONTROL
);
3420 /* First send a status message before receiving the fds. */
3421 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3423 /* Somehow, the session daemon is not responding anymore. */
3424 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3425 goto error_nosignal
;
3428 /* Poll on consumer socket. */
3429 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3431 /* Needing to exit in the middle of a command: error. */
3432 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3433 goto error_nosignal
;
3436 /* Get relayd socket from session daemon */
3437 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3438 if (ret
!= sizeof(fd
)) {
3439 fd
= -1; /* Just in case it gets set with an invalid value. */
3442 * Failing to receive FDs might indicate a major problem such as
3443 * reaching a fd limit during the receive where the kernel returns a
3444 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3445 * don't take any chances and stop everything.
3447 * XXX: Feature request #558 will fix that and avoid this possible
3448 * issue when reaching the fd limit.
3450 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3451 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3455 /* Copy socket information and received FD */
3456 switch (sock_type
) {
3457 case LTTNG_STREAM_CONTROL
:
3458 /* Copy received lttcomm socket */
3459 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3460 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3461 /* Handle create_sock error. */
3463 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3467 * Close the socket created internally by
3468 * lttcomm_create_sock, so we can replace it by the one
3469 * received from sessiond.
3471 if (close(relayd
->control_sock
.sock
.fd
)) {
3475 /* Assign new file descriptor */
3476 relayd
->control_sock
.sock
.fd
= fd
;
3477 fd
= -1; /* For error path */
3478 /* Assign version values. */
3479 relayd
->control_sock
.major
= relayd_sock
->major
;
3480 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3482 relayd
->relayd_session_id
= relayd_session_id
;
3485 case LTTNG_STREAM_DATA
:
3486 /* Copy received lttcomm socket */
3487 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3488 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3489 /* Handle create_sock error. */
3491 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3495 * Close the socket created internally by
3496 * lttcomm_create_sock, so we can replace it by the one
3497 * received from sessiond.
3499 if (close(relayd
->data_sock
.sock
.fd
)) {
3503 /* Assign new file descriptor */
3504 relayd
->data_sock
.sock
.fd
= fd
;
3505 fd
= -1; /* for eventual error paths */
3506 /* Assign version values. */
3507 relayd
->data_sock
.major
= relayd_sock
->major
;
3508 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3511 ERR("Unknown relayd socket type (%d)", sock_type
);
3512 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3516 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3517 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3518 relayd
->net_seq_idx
, fd
);
3520 /* We successfully added the socket. Send status back. */
3521 ret
= consumer_send_status_msg(sock
, ret_code
);
3523 /* Somehow, the session daemon is not responding anymore. */
3524 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3525 goto error_nosignal
;
3529 * Add relayd socket pair to consumer data hashtable. If object already
3530 * exists or on error, the function gracefully returns.
3538 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3539 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3543 /* Close received socket if valid. */
3546 PERROR("close received socket");
3550 if (relayd_created
) {
3556 * Try to lock the stream mutex.
3558 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
3560 static int stream_try_lock(struct lttng_consumer_stream
*stream
)
3567 * Try to lock the stream mutex. On failure, we know that the stream is
3568 * being used else where hence there is data still being extracted.
3570 ret
= pthread_mutex_trylock(&stream
->lock
);
3572 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
3584 * Search for a relayd associated to the session id and return the reference.
3586 * A rcu read side lock MUST be acquire before calling this function and locked
3587 * until the relayd object is no longer necessary.
3589 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3591 struct lttng_ht_iter iter
;
3592 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3594 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3595 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3598 * Check by sessiond id which is unique here where the relayd session
3599 * id might not be when having multiple relayd.
3601 if (relayd
->sessiond_session_id
== id
) {
3602 /* Found the relayd. There can be only one per id. */
3614 * Check if for a given session id there is still data needed to be extract
3617 * Return 1 if data is pending or else 0 meaning ready to be read.
3619 int consumer_data_pending(uint64_t id
)
3622 struct lttng_ht_iter iter
;
3623 struct lttng_ht
*ht
;
3624 struct lttng_consumer_stream
*stream
;
3625 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3626 int (*data_pending
)(struct lttng_consumer_stream
*);
3628 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3631 pthread_mutex_lock(&consumer_data
.lock
);
3633 switch (consumer_data
.type
) {
3634 case LTTNG_CONSUMER_KERNEL
:
3635 data_pending
= lttng_kconsumer_data_pending
;
3637 case LTTNG_CONSUMER32_UST
:
3638 case LTTNG_CONSUMER64_UST
:
3639 data_pending
= lttng_ustconsumer_data_pending
;
3642 ERR("Unknown consumer data type");
3646 /* Ease our life a bit */
3647 ht
= consumer_data
.stream_list_ht
;
3649 relayd
= find_relayd_by_session_id(id
);
3651 /* Send init command for data pending. */
3652 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3653 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3654 relayd
->relayd_session_id
);
3655 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3657 /* Communication error thus the relayd so no data pending. */
3658 goto data_not_pending
;
3662 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3663 ht
->hash_fct(&id
, lttng_ht_seed
),
3665 &iter
.iter
, stream
, node_session_id
.node
) {
3666 /* If this call fails, the stream is being used hence data pending. */
3667 ret
= stream_try_lock(stream
);
3673 * A removed node from the hash table indicates that the stream has
3674 * been deleted thus having a guarantee that the buffers are closed
3675 * on the consumer side. However, data can still be transmitted
3676 * over the network so don't skip the relayd check.
3678 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3680 /* Check the stream if there is data in the buffers. */
3681 ret
= data_pending(stream
);
3683 pthread_mutex_unlock(&stream
->lock
);
3690 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3691 if (stream
->metadata_flag
) {
3692 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3693 stream
->relayd_stream_id
);
3695 ret
= relayd_data_pending(&relayd
->control_sock
,
3696 stream
->relayd_stream_id
,
3697 stream
->next_net_seq_num
- 1);
3699 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3701 pthread_mutex_unlock(&stream
->lock
);
3705 pthread_mutex_unlock(&stream
->lock
);
3709 unsigned int is_data_inflight
= 0;
3711 /* Send init command for data pending. */
3712 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3713 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3714 relayd
->relayd_session_id
, &is_data_inflight
);
3715 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3717 goto data_not_pending
;
3719 if (is_data_inflight
) {
3725 * Finding _no_ node in the hash table and no inflight data means that the
3726 * stream(s) have been removed thus data is guaranteed to be available for
3727 * analysis from the trace files.
3731 /* Data is available to be read by a viewer. */
3732 pthread_mutex_unlock(&consumer_data
.lock
);
3737 /* Data is still being extracted from buffers. */
3738 pthread_mutex_unlock(&consumer_data
.lock
);
3744 * Send a ret code status message to the sessiond daemon.
3746 * Return the sendmsg() return value.
3748 int consumer_send_status_msg(int sock
, int ret_code
)
3750 struct lttcomm_consumer_status_msg msg
;
3752 memset(&msg
, 0, sizeof(msg
));
3753 msg
.ret_code
= ret_code
;
3755 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3759 * Send a channel status message to the sessiond daemon.
3761 * Return the sendmsg() return value.
3763 int consumer_send_status_channel(int sock
,
3764 struct lttng_consumer_channel
*channel
)
3766 struct lttcomm_consumer_status_channel msg
;
3770 memset(&msg
, 0, sizeof(msg
));
3772 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3774 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3775 msg
.key
= channel
->key
;
3776 msg
.stream_count
= channel
->streams
.count
;
3779 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3782 unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos
,
3783 unsigned long produced_pos
, uint64_t nb_packets_per_stream
,
3784 uint64_t max_sb_size
)
3786 unsigned long start_pos
;
3788 if (!nb_packets_per_stream
) {
3789 return consumed_pos
; /* Grab everything */
3791 start_pos
= produced_pos
- offset_align_floor(produced_pos
, max_sb_size
);
3792 start_pos
-= max_sb_size
* nb_packets_per_stream
;
3793 if ((long) (start_pos
- consumed_pos
) < 0) {
3794 return consumed_pos
; /* Grab everything */
3800 int rotate_rename_local(const char *old_path
, const char *new_path
,
3801 uid_t uid
, gid_t gid
)
3808 ret
= utils_mkdir_recursive(new_path
, S_IRWXU
| S_IRWXG
, uid
, gid
);
3810 ERR("Create directory on rotate");
3814 ret
= rename(old_path
, new_path
);
3815 if (ret
< 0 && errno
!= ENOENT
) {
3816 PERROR("Rename completed rotation chunk");
3826 int rotate_rename_relay(const char *old_path
, const char *new_path
,
3830 struct consumer_relayd_sock_pair
*relayd
;
3832 relayd
= consumer_find_relayd(relayd_id
);
3834 ERR("Failed to find relayd while running rotate_rename_relay command");
3839 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3840 ret
= relayd_rotate_rename(&relayd
->control_sock
, old_path
, new_path
);
3841 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3846 int lttng_consumer_rotate_rename(const char *old_path
, const char *new_path
,
3847 uid_t uid
, gid_t gid
, uint64_t relayd_id
)
3849 if (relayd_id
!= -1ULL) {
3850 return rotate_rename_relay(old_path
, new_path
, relayd_id
);
3852 return rotate_rename_local(old_path
, new_path
, uid
, gid
);
3857 int mkdir_local(const char *path
, uid_t uid
, gid_t gid
)
3861 ret
= utils_mkdir_recursive(path
, S_IRWXU
| S_IRWXG
, uid
, gid
);
3863 /* utils_mkdir_recursive logs an error. */
3873 int mkdir_relay(const char *path
, uint64_t relayd_id
)
3876 struct consumer_relayd_sock_pair
*relayd
;
3878 relayd
= consumer_find_relayd(relayd_id
);
3880 ERR("Failed to find relayd");
3885 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3886 ret
= relayd_mkdir(&relayd
->control_sock
, path
);
3887 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3894 int lttng_consumer_mkdir(const char *path
, uid_t uid
, gid_t gid
,
3897 if (relayd_id
!= -1ULL) {
3898 return mkdir_relay(path
, relayd_id
);
3900 return mkdir_local(path
, uid
, gid
);