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_copy_ro_channel_values(struct lttng_consumer_stream
*stream
,
545 struct lttng_consumer_channel
*channel
)
547 stream
->channel_ro_tracefile_size
= channel
->tracefile_size
;
548 memcpy(stream
->channel_ro_pathname
, channel
->pathname
, PATH_MAX
);
551 struct lttng_consumer_stream
*consumer_allocate_stream(uint64_t channel_key
,
553 enum lttng_consumer_stream_state state
,
554 const char *channel_name
,
561 enum consumer_channel_type type
,
562 unsigned int monitor
)
565 struct lttng_consumer_stream
*stream
;
567 stream
= zmalloc(sizeof(*stream
));
568 if (stream
== NULL
) {
569 PERROR("malloc struct lttng_consumer_stream");
576 stream
->key
= stream_key
;
578 stream
->out_fd_offset
= 0;
579 stream
->output_written
= 0;
580 stream
->state
= state
;
583 stream
->net_seq_idx
= relayd_id
;
584 stream
->session_id
= session_id
;
585 stream
->monitor
= monitor
;
586 stream
->endpoint_status
= CONSUMER_ENDPOINT_ACTIVE
;
587 stream
->index_file
= NULL
;
588 stream
->last_sequence_number
= -1ULL;
589 pthread_mutex_init(&stream
->lock
, NULL
);
590 pthread_mutex_init(&stream
->metadata_timer_lock
, NULL
);
592 /* If channel is the metadata, flag this stream as metadata. */
593 if (type
== CONSUMER_CHANNEL_TYPE_METADATA
) {
594 stream
->metadata_flag
= 1;
595 /* Metadata is flat out. */
596 strncpy(stream
->name
, DEFAULT_METADATA_NAME
, sizeof(stream
->name
));
597 /* Live rendez-vous point. */
598 pthread_cond_init(&stream
->metadata_rdv
, NULL
);
599 pthread_mutex_init(&stream
->metadata_rdv_lock
, NULL
);
601 /* Format stream name to <channel_name>_<cpu_number> */
602 ret
= snprintf(stream
->name
, sizeof(stream
->name
), "%s_%d",
605 PERROR("snprintf stream name");
610 /* Key is always the wait_fd for streams. */
611 lttng_ht_node_init_u64(&stream
->node
, stream
->key
);
613 /* Init node per channel id key */
614 lttng_ht_node_init_u64(&stream
->node_channel_id
, channel_key
);
616 /* Init session id node with the stream session id */
617 lttng_ht_node_init_u64(&stream
->node_session_id
, stream
->session_id
);
619 DBG3("Allocated stream %s (key %" PRIu64
", chan_key %" PRIu64
620 " relayd_id %" PRIu64
", session_id %" PRIu64
,
621 stream
->name
, stream
->key
, channel_key
,
622 stream
->net_seq_idx
, stream
->session_id
);
638 * Add a stream to the global list protected by a mutex.
640 int consumer_add_data_stream(struct lttng_consumer_stream
*stream
)
642 struct lttng_ht
*ht
= data_ht
;
648 DBG3("Adding consumer stream %" PRIu64
, stream
->key
);
650 pthread_mutex_lock(&consumer_data
.lock
);
651 pthread_mutex_lock(&stream
->chan
->lock
);
652 pthread_mutex_lock(&stream
->chan
->timer_lock
);
653 pthread_mutex_lock(&stream
->lock
);
656 /* Steal stream identifier to avoid having streams with the same key */
657 steal_stream_key(stream
->key
, ht
);
659 lttng_ht_add_unique_u64(ht
, &stream
->node
);
661 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
662 &stream
->node_channel_id
);
665 * Add stream to the stream_list_ht of the consumer data. No need to steal
666 * the key since the HT does not use it and we allow to add redundant keys
669 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
672 * When nb_init_stream_left reaches 0, we don't need to trigger any action
673 * in terms of destroying the associated channel, because the action that
674 * causes the count to become 0 also causes a stream to be added. The
675 * channel deletion will thus be triggered by the following removal of this
678 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
679 /* Increment refcount before decrementing nb_init_stream_left */
681 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
684 /* Update consumer data once the node is inserted. */
685 consumer_data
.stream_count
++;
686 consumer_data
.need_update
= 1;
689 pthread_mutex_unlock(&stream
->lock
);
690 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
691 pthread_mutex_unlock(&stream
->chan
->lock
);
692 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
)
1090 struct lttng_ht_iter iter
;
1091 struct lttng_consumer_stream
*stream
;
1096 assert(local_stream
);
1098 DBG("Updating poll fd array");
1100 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1102 * Only active streams with an active end point can be added to the
1103 * poll set and local stream storage of the thread.
1105 * There is a potential race here for endpoint_status to be updated
1106 * just after the check. However, this is OK since the stream(s) will
1107 * be deleted once the thread is notified that the end point state has
1108 * changed where this function will be called back again.
1110 if (stream
->state
!= LTTNG_CONSUMER_ACTIVE_STREAM
||
1111 stream
->endpoint_status
== CONSUMER_ENDPOINT_INACTIVE
) {
1115 * This clobbers way too much the debug output. Uncomment that if you
1116 * need it for debugging purposes.
1118 * DBG("Active FD %d", stream->wait_fd);
1120 (*pollfd
)[i
].fd
= stream
->wait_fd
;
1121 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1122 local_stream
[i
] = stream
;
1128 * Insert the consumer_data_pipe at the end of the array and don't
1129 * increment i so nb_fd is the number of real FD.
1131 (*pollfd
)[i
].fd
= lttng_pipe_get_readfd(ctx
->consumer_data_pipe
);
1132 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1134 (*pollfd
)[i
+ 1].fd
= lttng_pipe_get_readfd(ctx
->consumer_wakeup_pipe
);
1135 (*pollfd
)[i
+ 1].events
= POLLIN
| POLLPRI
;
1140 * Poll on the should_quit pipe and the command socket return -1 on
1141 * error, 1 if should exit, 0 if data is available on the command socket
1143 int lttng_consumer_poll_socket(struct pollfd
*consumer_sockpoll
)
1148 num_rdy
= poll(consumer_sockpoll
, 2, -1);
1149 if (num_rdy
== -1) {
1151 * Restart interrupted system call.
1153 if (errno
== EINTR
) {
1156 PERROR("Poll error");
1159 if (consumer_sockpoll
[0].revents
& (POLLIN
| POLLPRI
)) {
1160 DBG("consumer_should_quit wake up");
1167 * Set the error socket.
1169 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data
*ctx
,
1172 ctx
->consumer_error_socket
= sock
;
1176 * Set the command socket path.
1178 void lttng_consumer_set_command_sock_path(
1179 struct lttng_consumer_local_data
*ctx
, char *sock
)
1181 ctx
->consumer_command_sock_path
= sock
;
1185 * Send return code to the session daemon.
1186 * If the socket is not defined, we return 0, it is not a fatal error
1188 int lttng_consumer_send_error(struct lttng_consumer_local_data
*ctx
, int cmd
)
1190 if (ctx
->consumer_error_socket
> 0) {
1191 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
1192 sizeof(enum lttcomm_sessiond_command
));
1199 * Close all the tracefiles and stream fds and MUST be called when all
1200 * instances are destroyed i.e. when all threads were joined and are ended.
1202 void lttng_consumer_cleanup(void)
1204 struct lttng_ht_iter iter
;
1205 struct lttng_consumer_channel
*channel
;
1209 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
, channel
,
1211 consumer_del_channel(channel
);
1216 lttng_ht_destroy(consumer_data
.channel_ht
);
1218 cleanup_relayd_ht();
1220 lttng_ht_destroy(consumer_data
.stream_per_chan_id_ht
);
1223 * This HT contains streams that are freed by either the metadata thread or
1224 * the data thread so we do *nothing* on the hash table and simply destroy
1227 lttng_ht_destroy(consumer_data
.stream_list_ht
);
1231 * Called from signal handler.
1233 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
1237 CMM_STORE_SHARED(consumer_quit
, 1);
1238 ret
= lttng_write(ctx
->consumer_should_quit
[1], "4", 1);
1240 PERROR("write consumer quit");
1243 DBG("Consumer flag that it should quit");
1248 * Flush pending writes to trace output disk file.
1251 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1255 int outfd
= stream
->out_fd
;
1258 * This does a blocking write-and-wait on any page that belongs to the
1259 * subbuffer prior to the one we just wrote.
1260 * Don't care about error values, as these are just hints and ways to
1261 * limit the amount of page cache used.
1263 if (orig_offset
< stream
->max_sb_size
) {
1266 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1267 stream
->max_sb_size
,
1268 SYNC_FILE_RANGE_WAIT_BEFORE
1269 | SYNC_FILE_RANGE_WRITE
1270 | SYNC_FILE_RANGE_WAIT_AFTER
);
1272 * Give hints to the kernel about how we access the file:
1273 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1276 * We need to call fadvise again after the file grows because the
1277 * kernel does not seem to apply fadvise to non-existing parts of the
1280 * Call fadvise _after_ having waited for the page writeback to
1281 * complete because the dirty page writeback semantic is not well
1282 * defined. So it can be expected to lead to lower throughput in
1285 ret
= posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1286 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1287 if (ret
&& ret
!= -ENOSYS
) {
1289 PERROR("posix_fadvise on fd %i", outfd
);
1294 * Initialise the necessary environnement :
1295 * - create a new context
1296 * - create the poll_pipe
1297 * - create the should_quit pipe (for signal handler)
1298 * - create the thread pipe (for splice)
1300 * Takes a function pointer as argument, this function is called when data is
1301 * available on a buffer. This function is responsible to do the
1302 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1303 * buffer configuration and then kernctl_put_next_subbuf at the end.
1305 * Returns a pointer to the new context or NULL on error.
1307 struct lttng_consumer_local_data
*lttng_consumer_create(
1308 enum lttng_consumer_type type
,
1309 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1310 struct lttng_consumer_local_data
*ctx
),
1311 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1312 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1313 int (*update_stream
)(uint64_t stream_key
, uint32_t state
))
1316 struct lttng_consumer_local_data
*ctx
;
1318 assert(consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1319 consumer_data
.type
== type
);
1320 consumer_data
.type
= type
;
1322 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
1324 PERROR("allocating context");
1328 ctx
->consumer_error_socket
= -1;
1329 ctx
->consumer_metadata_socket
= -1;
1330 pthread_mutex_init(&ctx
->metadata_socket_lock
, NULL
);
1331 /* assign the callbacks */
1332 ctx
->on_buffer_ready
= buffer_ready
;
1333 ctx
->on_recv_channel
= recv_channel
;
1334 ctx
->on_recv_stream
= recv_stream
;
1335 ctx
->on_update_stream
= update_stream
;
1337 ctx
->consumer_data_pipe
= lttng_pipe_open(0);
1338 if (!ctx
->consumer_data_pipe
) {
1339 goto error_poll_pipe
;
1342 ctx
->consumer_wakeup_pipe
= lttng_pipe_open(0);
1343 if (!ctx
->consumer_wakeup_pipe
) {
1344 goto error_wakeup_pipe
;
1347 ret
= pipe(ctx
->consumer_should_quit
);
1349 PERROR("Error creating recv pipe");
1350 goto error_quit_pipe
;
1353 ret
= pipe(ctx
->consumer_channel_pipe
);
1355 PERROR("Error creating channel pipe");
1356 goto error_channel_pipe
;
1359 ctx
->consumer_metadata_pipe
= lttng_pipe_open(0);
1360 if (!ctx
->consumer_metadata_pipe
) {
1361 goto error_metadata_pipe
;
1364 ctx
->channel_monitor_pipe
= -1;
1368 error_metadata_pipe
:
1369 utils_close_pipe(ctx
->consumer_channel_pipe
);
1371 utils_close_pipe(ctx
->consumer_should_quit
);
1373 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1375 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1383 * Iterate over all streams of the hashtable and free them properly.
1385 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1387 struct lttng_ht_iter iter
;
1388 struct lttng_consumer_stream
*stream
;
1395 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1397 * Ignore return value since we are currently cleaning up so any error
1400 (void) consumer_del_stream(stream
, ht
);
1404 lttng_ht_destroy(ht
);
1408 * Iterate over all streams of the metadata hashtable and free them
1411 static void destroy_metadata_stream_ht(struct lttng_ht
*ht
)
1413 struct lttng_ht_iter iter
;
1414 struct lttng_consumer_stream
*stream
;
1421 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1423 * Ignore return value since we are currently cleaning up so any error
1426 (void) consumer_del_metadata_stream(stream
, ht
);
1430 lttng_ht_destroy(ht
);
1434 * Close all fds associated with the instance and free the context.
1436 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1440 DBG("Consumer destroying it. Closing everything.");
1446 destroy_data_stream_ht(data_ht
);
1447 destroy_metadata_stream_ht(metadata_ht
);
1449 ret
= close(ctx
->consumer_error_socket
);
1453 ret
= close(ctx
->consumer_metadata_socket
);
1457 utils_close_pipe(ctx
->consumer_channel_pipe
);
1458 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1459 lttng_pipe_destroy(ctx
->consumer_metadata_pipe
);
1460 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1461 utils_close_pipe(ctx
->consumer_should_quit
);
1463 unlink(ctx
->consumer_command_sock_path
);
1468 * Write the metadata stream id on the specified file descriptor.
1470 static int write_relayd_metadata_id(int fd
,
1471 struct lttng_consumer_stream
*stream
,
1472 struct consumer_relayd_sock_pair
*relayd
, unsigned long padding
)
1475 struct lttcomm_relayd_metadata_payload hdr
;
1477 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1478 hdr
.padding_size
= htobe32(padding
);
1479 ret
= lttng_write(fd
, (void *) &hdr
, sizeof(hdr
));
1480 if (ret
< sizeof(hdr
)) {
1482 * This error means that the fd's end is closed so ignore the PERROR
1483 * not to clubber the error output since this can happen in a normal
1486 if (errno
!= EPIPE
) {
1487 PERROR("write metadata stream id");
1489 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1491 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1492 * handle writting the missing part so report that as an error and
1493 * don't lie to the caller.
1498 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1499 stream
->relayd_stream_id
, padding
);
1506 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1507 * core function for writing trace buffers to either the local filesystem or
1510 * It must be called with the stream lock held.
1512 * Careful review MUST be put if any changes occur!
1514 * Returns the number of bytes written
1516 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1517 struct lttng_consumer_local_data
*ctx
,
1518 struct lttng_consumer_stream
*stream
, unsigned long len
,
1519 unsigned long padding
,
1520 struct ctf_packet_index
*index
)
1522 unsigned long mmap_offset
;
1525 off_t orig_offset
= stream
->out_fd_offset
;
1526 /* Default is on the disk */
1527 int outfd
= stream
->out_fd
;
1528 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1529 unsigned int relayd_hang_up
= 0;
1531 /* RCU lock for the relayd pointer */
1534 /* Flag that the current stream if set for network streaming. */
1535 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1536 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1537 if (relayd
== NULL
) {
1543 /* get the offset inside the fd to mmap */
1544 switch (consumer_data
.type
) {
1545 case LTTNG_CONSUMER_KERNEL
:
1546 mmap_base
= stream
->mmap_base
;
1547 ret
= kernctl_get_mmap_read_offset(stream
->wait_fd
, &mmap_offset
);
1549 PERROR("tracer ctl get_mmap_read_offset");
1553 case LTTNG_CONSUMER32_UST
:
1554 case LTTNG_CONSUMER64_UST
:
1555 mmap_base
= lttng_ustctl_get_mmap_base(stream
);
1557 ERR("read mmap get mmap base for stream %s", stream
->name
);
1561 ret
= lttng_ustctl_get_mmap_read_offset(stream
, &mmap_offset
);
1563 PERROR("tracer ctl get_mmap_read_offset");
1569 ERR("Unknown consumer_data type");
1573 /* Handle stream on the relayd if the output is on the network */
1575 unsigned long netlen
= len
;
1578 * Lock the control socket for the complete duration of the function
1579 * since from this point on we will use the socket.
1581 if (stream
->metadata_flag
) {
1582 /* Metadata requires the control socket. */
1583 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1584 if (stream
->reset_metadata_flag
) {
1585 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1586 stream
->relayd_stream_id
,
1587 stream
->metadata_version
);
1592 stream
->reset_metadata_flag
= 0;
1594 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1597 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1602 /* Use the returned socket. */
1605 /* Write metadata stream id before payload */
1606 if (stream
->metadata_flag
) {
1607 ret
= write_relayd_metadata_id(outfd
, stream
, relayd
, padding
);
1614 /* No streaming, we have to set the len with the full padding */
1617 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1618 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1620 ERR("Reset metadata file");
1623 stream
->reset_metadata_flag
= 0;
1627 * Check if we need to change the tracefile before writing the packet.
1629 if (stream
->chan
->tracefile_size
> 0 &&
1630 (stream
->tracefile_size_current
+ len
) >
1631 stream
->chan
->tracefile_size
) {
1632 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1633 stream
->name
, stream
->chan
->tracefile_size
,
1634 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1635 stream
->out_fd
, &(stream
->tracefile_count_current
),
1638 ERR("Rotating output file");
1641 outfd
= stream
->out_fd
;
1643 if (stream
->index_file
) {
1644 lttng_index_file_put(stream
->index_file
);
1645 stream
->index_file
= lttng_index_file_create(stream
->chan
->pathname
,
1646 stream
->name
, stream
->uid
, stream
->gid
,
1647 stream
->chan
->tracefile_size
,
1648 stream
->tracefile_count_current
,
1649 CTF_INDEX_MAJOR
, CTF_INDEX_MINOR
);
1650 if (!stream
->index_file
) {
1655 /* Reset current size because we just perform a rotation. */
1656 stream
->tracefile_size_current
= 0;
1657 stream
->out_fd_offset
= 0;
1660 stream
->tracefile_size_current
+= len
;
1662 index
->offset
= htobe64(stream
->out_fd_offset
);
1667 * This call guarantee that len or less is returned. It's impossible to
1668 * receive a ret value that is bigger than len.
1670 ret
= lttng_write(outfd
, mmap_base
+ mmap_offset
, len
);
1671 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, len
);
1672 if (ret
< 0 || ((size_t) ret
!= len
)) {
1674 * Report error to caller if nothing was written else at least send the
1682 /* Socket operation failed. We consider the relayd dead */
1683 if (errno
== EPIPE
|| errno
== EINVAL
|| errno
== EBADF
) {
1685 * This is possible if the fd is closed on the other side
1686 * (outfd) or any write problem. It can be verbose a bit for a
1687 * normal execution if for instance the relayd is stopped
1688 * abruptly. This can happen so set this to a DBG statement.
1690 DBG("Consumer mmap write detected relayd hang up");
1692 /* Unhandled error, print it and stop function right now. */
1693 PERROR("Error in write mmap (ret %zd != len %lu)", ret
, len
);
1697 stream
->output_written
+= ret
;
1699 /* This call is useless on a socket so better save a syscall. */
1701 /* This won't block, but will start writeout asynchronously */
1702 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, len
,
1703 SYNC_FILE_RANGE_WRITE
);
1704 stream
->out_fd_offset
+= len
;
1705 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1710 * This is a special case that the relayd has closed its socket. Let's
1711 * cleanup the relayd object and all associated streams.
1713 if (relayd
&& relayd_hang_up
) {
1714 cleanup_relayd(relayd
, ctx
);
1718 /* Unlock only if ctrl socket used */
1719 if (relayd
&& stream
->metadata_flag
) {
1720 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1728 * Splice the data from the ring buffer to the tracefile.
1730 * It must be called with the stream lock held.
1732 * Returns the number of bytes spliced.
1734 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1735 struct lttng_consumer_local_data
*ctx
,
1736 struct lttng_consumer_stream
*stream
, unsigned long len
,
1737 unsigned long padding
,
1738 struct ctf_packet_index
*index
)
1740 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1742 off_t orig_offset
= stream
->out_fd_offset
;
1743 int fd
= stream
->wait_fd
;
1744 /* Default is on the disk */
1745 int outfd
= stream
->out_fd
;
1746 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1748 unsigned int relayd_hang_up
= 0;
1750 switch (consumer_data
.type
) {
1751 case LTTNG_CONSUMER_KERNEL
:
1753 case LTTNG_CONSUMER32_UST
:
1754 case LTTNG_CONSUMER64_UST
:
1755 /* Not supported for user space tracing */
1758 ERR("Unknown consumer_data type");
1762 /* RCU lock for the relayd pointer */
1765 /* Flag that the current stream if set for network streaming. */
1766 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1767 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1768 if (relayd
== NULL
) {
1773 splice_pipe
= stream
->splice_pipe
;
1775 /* Write metadata stream id before payload */
1777 unsigned long total_len
= len
;
1779 if (stream
->metadata_flag
) {
1781 * Lock the control socket for the complete duration of the function
1782 * since from this point on we will use the socket.
1784 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1786 if (stream
->reset_metadata_flag
) {
1787 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1788 stream
->relayd_stream_id
,
1789 stream
->metadata_version
);
1794 stream
->reset_metadata_flag
= 0;
1796 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
, relayd
,
1804 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1807 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1813 /* Use the returned socket. */
1816 /* No streaming, we have to set the len with the full padding */
1819 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1820 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1822 ERR("Reset metadata file");
1825 stream
->reset_metadata_flag
= 0;
1828 * Check if we need to change the tracefile before writing the packet.
1830 if (stream
->chan
->tracefile_size
> 0 &&
1831 (stream
->tracefile_size_current
+ len
) >
1832 stream
->chan
->tracefile_size
) {
1833 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1834 stream
->name
, stream
->chan
->tracefile_size
,
1835 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1836 stream
->out_fd
, &(stream
->tracefile_count_current
),
1840 ERR("Rotating output file");
1843 outfd
= stream
->out_fd
;
1845 if (stream
->index_file
) {
1846 lttng_index_file_put(stream
->index_file
);
1847 stream
->index_file
= lttng_index_file_create(stream
->chan
->pathname
,
1848 stream
->name
, stream
->uid
, stream
->gid
,
1849 stream
->chan
->tracefile_size
,
1850 stream
->tracefile_count_current
,
1851 CTF_INDEX_MAJOR
, CTF_INDEX_MINOR
);
1852 if (!stream
->index_file
) {
1857 /* Reset current size because we just perform a rotation. */
1858 stream
->tracefile_size_current
= 0;
1859 stream
->out_fd_offset
= 0;
1862 stream
->tracefile_size_current
+= len
;
1863 index
->offset
= htobe64(stream
->out_fd_offset
);
1867 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1868 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1869 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1870 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1871 DBG("splice chan to pipe, ret %zd", ret_splice
);
1872 if (ret_splice
< 0) {
1875 PERROR("Error in relay splice");
1879 /* Handle stream on the relayd if the output is on the network */
1880 if (relayd
&& stream
->metadata_flag
) {
1881 size_t metadata_payload_size
=
1882 sizeof(struct lttcomm_relayd_metadata_payload
);
1884 /* Update counter to fit the spliced data */
1885 ret_splice
+= metadata_payload_size
;
1886 len
+= metadata_payload_size
;
1888 * We do this so the return value can match the len passed as
1889 * argument to this function.
1891 written
-= metadata_payload_size
;
1894 /* Splice data out */
1895 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1896 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1897 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
1899 if (ret_splice
< 0) {
1904 } else if (ret_splice
> len
) {
1906 * We don't expect this code path to be executed but you never know
1907 * so this is an extra protection agains a buggy splice().
1910 written
+= ret_splice
;
1911 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice
,
1915 /* All good, update current len and continue. */
1919 /* This call is useless on a socket so better save a syscall. */
1921 /* This won't block, but will start writeout asynchronously */
1922 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1923 SYNC_FILE_RANGE_WRITE
);
1924 stream
->out_fd_offset
+= ret_splice
;
1926 stream
->output_written
+= ret_splice
;
1927 written
+= ret_splice
;
1930 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1936 * This is a special case that the relayd has closed its socket. Let's
1937 * cleanup the relayd object and all associated streams.
1939 if (relayd
&& relayd_hang_up
) {
1940 cleanup_relayd(relayd
, ctx
);
1941 /* Skip splice error so the consumer does not fail */
1946 /* send the appropriate error description to sessiond */
1949 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1952 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1955 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
1960 if (relayd
&& stream
->metadata_flag
) {
1961 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1969 * Sample the snapshot positions for a specific fd
1971 * Returns 0 on success, < 0 on error
1973 int lttng_consumer_sample_snapshot_positions(struct lttng_consumer_stream
*stream
)
1975 switch (consumer_data
.type
) {
1976 case LTTNG_CONSUMER_KERNEL
:
1977 return lttng_kconsumer_sample_snapshot_positions(stream
);
1978 case LTTNG_CONSUMER32_UST
:
1979 case LTTNG_CONSUMER64_UST
:
1980 return lttng_ustconsumer_sample_snapshot_positions(stream
);
1982 ERR("Unknown consumer_data type");
1988 * Take a snapshot for a specific fd
1990 * Returns 0 on success, < 0 on error
1992 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
1994 switch (consumer_data
.type
) {
1995 case LTTNG_CONSUMER_KERNEL
:
1996 return lttng_kconsumer_take_snapshot(stream
);
1997 case LTTNG_CONSUMER32_UST
:
1998 case LTTNG_CONSUMER64_UST
:
1999 return lttng_ustconsumer_take_snapshot(stream
);
2001 ERR("Unknown consumer_data type");
2008 * Get the produced position
2010 * Returns 0 on success, < 0 on error
2012 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
2015 switch (consumer_data
.type
) {
2016 case LTTNG_CONSUMER_KERNEL
:
2017 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
2018 case LTTNG_CONSUMER32_UST
:
2019 case LTTNG_CONSUMER64_UST
:
2020 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
2022 ERR("Unknown consumer_data type");
2029 * Get the consumed position
2031 * Returns 0 on success, < 0 on error
2033 int lttng_consumer_get_consumed_snapshot(struct lttng_consumer_stream
*stream
,
2036 switch (consumer_data
.type
) {
2037 case LTTNG_CONSUMER_KERNEL
:
2038 return lttng_kconsumer_get_consumed_snapshot(stream
, pos
);
2039 case LTTNG_CONSUMER32_UST
:
2040 case LTTNG_CONSUMER64_UST
:
2041 return lttng_ustconsumer_get_consumed_snapshot(stream
, pos
);
2043 ERR("Unknown consumer_data type");
2049 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
2050 int sock
, struct pollfd
*consumer_sockpoll
)
2052 switch (consumer_data
.type
) {
2053 case LTTNG_CONSUMER_KERNEL
:
2054 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2055 case LTTNG_CONSUMER32_UST
:
2056 case LTTNG_CONSUMER64_UST
:
2057 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2059 ERR("Unknown consumer_data type");
2065 void lttng_consumer_close_all_metadata(void)
2067 switch (consumer_data
.type
) {
2068 case LTTNG_CONSUMER_KERNEL
:
2070 * The Kernel consumer has a different metadata scheme so we don't
2071 * close anything because the stream will be closed by the session
2075 case LTTNG_CONSUMER32_UST
:
2076 case LTTNG_CONSUMER64_UST
:
2078 * Close all metadata streams. The metadata hash table is passed and
2079 * this call iterates over it by closing all wakeup fd. This is safe
2080 * because at this point we are sure that the metadata producer is
2081 * either dead or blocked.
2083 lttng_ustconsumer_close_all_metadata(metadata_ht
);
2086 ERR("Unknown consumer_data type");
2092 * Clean up a metadata stream and free its memory.
2094 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
2095 struct lttng_ht
*ht
)
2097 struct lttng_consumer_channel
*free_chan
= NULL
;
2101 * This call should NEVER receive regular stream. It must always be
2102 * metadata stream and this is crucial for data structure synchronization.
2104 assert(stream
->metadata_flag
);
2106 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
2108 pthread_mutex_lock(&consumer_data
.lock
);
2109 pthread_mutex_lock(&stream
->chan
->lock
);
2110 pthread_mutex_lock(&stream
->lock
);
2111 if (stream
->chan
->metadata_cache
) {
2112 /* Only applicable to userspace consumers. */
2113 pthread_mutex_lock(&stream
->chan
->metadata_cache
->lock
);
2116 /* Remove any reference to that stream. */
2117 consumer_stream_delete(stream
, ht
);
2119 /* Close down everything including the relayd if one. */
2120 consumer_stream_close(stream
);
2121 /* Destroy tracer buffers of the stream. */
2122 consumer_stream_destroy_buffers(stream
);
2124 /* Atomically decrement channel refcount since other threads can use it. */
2125 if (!uatomic_sub_return(&stream
->chan
->refcount
, 1)
2126 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
2127 /* Go for channel deletion! */
2128 free_chan
= stream
->chan
;
2132 * Nullify the stream reference so it is not used after deletion. The
2133 * channel lock MUST be acquired before being able to check for a NULL
2136 stream
->chan
->metadata_stream
= NULL
;
2138 if (stream
->chan
->metadata_cache
) {
2139 pthread_mutex_unlock(&stream
->chan
->metadata_cache
->lock
);
2141 pthread_mutex_unlock(&stream
->lock
);
2142 pthread_mutex_unlock(&stream
->chan
->lock
);
2143 pthread_mutex_unlock(&consumer_data
.lock
);
2146 consumer_del_channel(free_chan
);
2149 consumer_stream_free(stream
);
2153 * Action done with the metadata stream when adding it to the consumer internal
2154 * data structures to handle it.
2156 int consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
2158 struct lttng_ht
*ht
= metadata_ht
;
2160 struct lttng_ht_iter iter
;
2161 struct lttng_ht_node_u64
*node
;
2166 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2168 pthread_mutex_lock(&consumer_data
.lock
);
2169 pthread_mutex_lock(&stream
->chan
->lock
);
2170 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2171 pthread_mutex_lock(&stream
->lock
);
2174 * From here, refcounts are updated so be _careful_ when returning an error
2181 * Lookup the stream just to make sure it does not exist in our internal
2182 * state. This should NEVER happen.
2184 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2185 node
= lttng_ht_iter_get_node_u64(&iter
);
2189 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2190 * in terms of destroying the associated channel, because the action that
2191 * causes the count to become 0 also causes a stream to be added. The
2192 * channel deletion will thus be triggered by the following removal of this
2195 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2196 /* Increment refcount before decrementing nb_init_stream_left */
2198 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2201 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2203 lttng_ht_add_unique_u64(consumer_data
.stream_per_chan_id_ht
,
2204 &stream
->node_channel_id
);
2207 * Add stream to the stream_list_ht of the consumer data. No need to steal
2208 * the key since the HT does not use it and we allow to add redundant keys
2211 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2215 pthread_mutex_unlock(&stream
->lock
);
2216 pthread_mutex_unlock(&stream
->chan
->lock
);
2217 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2218 pthread_mutex_unlock(&consumer_data
.lock
);
2223 * Delete data stream that are flagged for deletion (endpoint_status).
2225 static void validate_endpoint_status_data_stream(void)
2227 struct lttng_ht_iter iter
;
2228 struct lttng_consumer_stream
*stream
;
2230 DBG("Consumer delete flagged data stream");
2233 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2234 /* Validate delete flag of the stream */
2235 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2238 /* Delete it right now */
2239 consumer_del_stream(stream
, data_ht
);
2245 * Delete metadata stream that are flagged for deletion (endpoint_status).
2247 static void validate_endpoint_status_metadata_stream(
2248 struct lttng_poll_event
*pollset
)
2250 struct lttng_ht_iter iter
;
2251 struct lttng_consumer_stream
*stream
;
2253 DBG("Consumer delete flagged metadata stream");
2258 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2259 /* Validate delete flag of the stream */
2260 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2264 * Remove from pollset so the metadata thread can continue without
2265 * blocking on a deleted stream.
2267 lttng_poll_del(pollset
, stream
->wait_fd
);
2269 /* Delete it right now */
2270 consumer_del_metadata_stream(stream
, metadata_ht
);
2276 * Thread polls on metadata file descriptor and write them on disk or on the
2279 void *consumer_thread_metadata_poll(void *data
)
2281 int ret
, i
, pollfd
, err
= -1;
2282 uint32_t revents
, nb_fd
;
2283 struct lttng_consumer_stream
*stream
= NULL
;
2284 struct lttng_ht_iter iter
;
2285 struct lttng_ht_node_u64
*node
;
2286 struct lttng_poll_event events
;
2287 struct lttng_consumer_local_data
*ctx
= data
;
2290 rcu_register_thread();
2292 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2294 if (testpoint(consumerd_thread_metadata
)) {
2295 goto error_testpoint
;
2298 health_code_update();
2300 DBG("Thread metadata poll started");
2302 /* Size is set to 1 for the consumer_metadata pipe */
2303 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2305 ERR("Poll set creation failed");
2309 ret
= lttng_poll_add(&events
,
2310 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2316 DBG("Metadata main loop started");
2320 health_code_update();
2321 health_poll_entry();
2322 DBG("Metadata poll wait");
2323 ret
= lttng_poll_wait(&events
, -1);
2324 DBG("Metadata poll return from wait with %d fd(s)",
2325 LTTNG_POLL_GETNB(&events
));
2327 DBG("Metadata event caught in thread");
2329 if (errno
== EINTR
) {
2330 ERR("Poll EINTR caught");
2333 if (LTTNG_POLL_GETNB(&events
) == 0) {
2334 err
= 0; /* All is OK */
2341 /* From here, the event is a metadata wait fd */
2342 for (i
= 0; i
< nb_fd
; i
++) {
2343 health_code_update();
2345 revents
= LTTNG_POLL_GETEV(&events
, i
);
2346 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2349 /* No activity for this FD (poll implementation). */
2353 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2354 if (revents
& LPOLLIN
) {
2357 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2358 &stream
, sizeof(stream
));
2359 if (pipe_len
< sizeof(stream
)) {
2361 PERROR("read metadata stream");
2364 * Remove the pipe from the poll set and continue the loop
2365 * since their might be data to consume.
2367 lttng_poll_del(&events
,
2368 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2369 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2373 /* A NULL stream means that the state has changed. */
2374 if (stream
== NULL
) {
2375 /* Check for deleted streams. */
2376 validate_endpoint_status_metadata_stream(&events
);
2380 DBG("Adding metadata stream %d to poll set",
2383 /* Add metadata stream to the global poll events list */
2384 lttng_poll_add(&events
, stream
->wait_fd
,
2385 LPOLLIN
| LPOLLPRI
| LPOLLHUP
);
2386 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2387 DBG("Metadata thread pipe hung up");
2389 * Remove the pipe from the poll set and continue the loop
2390 * since their might be data to consume.
2392 lttng_poll_del(&events
,
2393 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2394 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2397 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2401 /* Handle other stream */
2407 uint64_t tmp_id
= (uint64_t) pollfd
;
2409 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2411 node
= lttng_ht_iter_get_node_u64(&iter
);
2414 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2417 if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2418 /* Get the data out of the metadata file descriptor */
2419 DBG("Metadata available on fd %d", pollfd
);
2420 assert(stream
->wait_fd
== pollfd
);
2423 health_code_update();
2425 len
= ctx
->on_buffer_ready(stream
, ctx
);
2427 * We don't check the return value here since if we get
2428 * a negative len, it means an error occurred thus we
2429 * simply remove it from the poll set and free the
2434 /* It's ok to have an unavailable sub-buffer */
2435 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2436 /* Clean up stream from consumer and free it. */
2437 lttng_poll_del(&events
, stream
->wait_fd
);
2438 consumer_del_metadata_stream(stream
, metadata_ht
);
2440 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2441 DBG("Metadata fd %d is hup|err.", pollfd
);
2442 if (!stream
->hangup_flush_done
2443 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2444 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2445 DBG("Attempting to flush and consume the UST buffers");
2446 lttng_ustconsumer_on_stream_hangup(stream
);
2448 /* We just flushed the stream now read it. */
2450 health_code_update();
2452 len
= ctx
->on_buffer_ready(stream
, ctx
);
2454 * We don't check the return value here since if we get
2455 * a negative len, it means an error occurred thus we
2456 * simply remove it from the poll set and free the
2462 lttng_poll_del(&events
, stream
->wait_fd
);
2464 * This call update the channel states, closes file descriptors
2465 * and securely free the stream.
2467 consumer_del_metadata_stream(stream
, metadata_ht
);
2469 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2473 /* Release RCU lock for the stream looked up */
2481 DBG("Metadata poll thread exiting");
2483 lttng_poll_clean(&events
);
2488 ERR("Health error occurred in %s", __func__
);
2490 health_unregister(health_consumerd
);
2491 rcu_unregister_thread();
2496 * This thread polls the fds in the set to consume the data and write
2497 * it to tracefile if necessary.
2499 void *consumer_thread_data_poll(void *data
)
2501 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2502 struct pollfd
*pollfd
= NULL
;
2503 /* local view of the streams */
2504 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2505 /* local view of consumer_data.fds_count */
2507 struct lttng_consumer_local_data
*ctx
= data
;
2510 rcu_register_thread();
2512 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2514 if (testpoint(consumerd_thread_data
)) {
2515 goto error_testpoint
;
2518 health_code_update();
2520 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
*));
2521 if (local_stream
== NULL
) {
2522 PERROR("local_stream malloc");
2527 health_code_update();
2533 * the fds set has been updated, we need to update our
2534 * local array as well
2536 pthread_mutex_lock(&consumer_data
.lock
);
2537 if (consumer_data
.need_update
) {
2542 local_stream
= NULL
;
2545 * Allocate for all fds +1 for the consumer_data_pipe and +1 for
2548 pollfd
= zmalloc((consumer_data
.stream_count
+ 2) * sizeof(struct pollfd
));
2549 if (pollfd
== NULL
) {
2550 PERROR("pollfd malloc");
2551 pthread_mutex_unlock(&consumer_data
.lock
);
2555 local_stream
= zmalloc((consumer_data
.stream_count
+ 2) *
2556 sizeof(struct lttng_consumer_stream
*));
2557 if (local_stream
== NULL
) {
2558 PERROR("local_stream malloc");
2559 pthread_mutex_unlock(&consumer_data
.lock
);
2562 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2565 ERR("Error in allocating pollfd or local_outfds");
2566 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2567 pthread_mutex_unlock(&consumer_data
.lock
);
2571 consumer_data
.need_update
= 0;
2573 pthread_mutex_unlock(&consumer_data
.lock
);
2575 /* No FDs and consumer_quit, consumer_cleanup the thread */
2576 if (nb_fd
== 0 && CMM_LOAD_SHARED(consumer_quit
) == 1) {
2577 err
= 0; /* All is OK */
2580 /* poll on the array of fds */
2582 DBG("polling on %d fd", nb_fd
+ 2);
2583 if (testpoint(consumerd_thread_data_poll
)) {
2586 health_poll_entry();
2587 num_rdy
= poll(pollfd
, nb_fd
+ 2, -1);
2589 DBG("poll num_rdy : %d", num_rdy
);
2590 if (num_rdy
== -1) {
2592 * Restart interrupted system call.
2594 if (errno
== EINTR
) {
2597 PERROR("Poll error");
2598 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2600 } else if (num_rdy
== 0) {
2601 DBG("Polling thread timed out");
2605 if (caa_unlikely(data_consumption_paused
)) {
2606 DBG("Data consumption paused, sleeping...");
2612 * If the consumer_data_pipe triggered poll go directly to the
2613 * beginning of the loop to update the array. We want to prioritize
2614 * array update over low-priority reads.
2616 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2617 ssize_t pipe_readlen
;
2619 DBG("consumer_data_pipe wake up");
2620 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2621 &new_stream
, sizeof(new_stream
));
2622 if (pipe_readlen
< sizeof(new_stream
)) {
2623 PERROR("Consumer data pipe");
2624 /* Continue so we can at least handle the current stream(s). */
2629 * If the stream is NULL, just ignore it. It's also possible that
2630 * the sessiond poll thread changed the consumer_quit state and is
2631 * waking us up to test it.
2633 if (new_stream
== NULL
) {
2634 validate_endpoint_status_data_stream();
2638 /* Continue to update the local streams and handle prio ones */
2642 /* Handle wakeup pipe. */
2643 if (pollfd
[nb_fd
+ 1].revents
& (POLLIN
| POLLPRI
)) {
2645 ssize_t pipe_readlen
;
2647 pipe_readlen
= lttng_pipe_read(ctx
->consumer_wakeup_pipe
, &dummy
,
2649 if (pipe_readlen
< 0) {
2650 PERROR("Consumer data wakeup pipe");
2652 /* We've been awakened to handle stream(s). */
2653 ctx
->has_wakeup
= 0;
2656 /* Take care of high priority channels first. */
2657 for (i
= 0; i
< nb_fd
; i
++) {
2658 health_code_update();
2660 if (local_stream
[i
] == NULL
) {
2663 if (pollfd
[i
].revents
& POLLPRI
) {
2664 DBG("Urgent 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;
2679 * If we read high prio channel in this loop, try again
2680 * for more high prio data.
2686 /* Take care of low priority channels. */
2687 for (i
= 0; i
< nb_fd
; i
++) {
2688 health_code_update();
2690 if (local_stream
[i
] == NULL
) {
2693 if ((pollfd
[i
].revents
& POLLIN
) ||
2694 local_stream
[i
]->hangup_flush_done
||
2695 local_stream
[i
]->has_data
) {
2696 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2697 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2698 /* it's ok to have an unavailable sub-buffer */
2699 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2700 /* Clean the stream and free it. */
2701 consumer_del_stream(local_stream
[i
], data_ht
);
2702 local_stream
[i
] = NULL
;
2703 } else if (len
> 0) {
2704 local_stream
[i
]->data_read
= 1;
2709 /* Handle hangup and errors */
2710 for (i
= 0; i
< nb_fd
; i
++) {
2711 health_code_update();
2713 if (local_stream
[i
] == NULL
) {
2716 if (!local_stream
[i
]->hangup_flush_done
2717 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2718 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2719 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2720 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2722 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2723 /* Attempt read again, for the data we just flushed. */
2724 local_stream
[i
]->data_read
= 1;
2727 * If the poll flag is HUP/ERR/NVAL and we have
2728 * read no data in this pass, we can remove the
2729 * stream from its hash table.
2731 if ((pollfd
[i
].revents
& POLLHUP
)) {
2732 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2733 if (!local_stream
[i
]->data_read
) {
2734 consumer_del_stream(local_stream
[i
], data_ht
);
2735 local_stream
[i
] = NULL
;
2738 } else if (pollfd
[i
].revents
& POLLERR
) {
2739 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2740 if (!local_stream
[i
]->data_read
) {
2741 consumer_del_stream(local_stream
[i
], data_ht
);
2742 local_stream
[i
] = NULL
;
2745 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2746 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2747 if (!local_stream
[i
]->data_read
) {
2748 consumer_del_stream(local_stream
[i
], data_ht
);
2749 local_stream
[i
] = NULL
;
2753 if (local_stream
[i
] != NULL
) {
2754 local_stream
[i
]->data_read
= 0;
2761 DBG("polling thread exiting");
2766 * Close the write side of the pipe so epoll_wait() in
2767 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2768 * read side of the pipe. If we close them both, epoll_wait strangely does
2769 * not return and could create a endless wait period if the pipe is the
2770 * only tracked fd in the poll set. The thread will take care of closing
2773 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2778 ERR("Health error occurred in %s", __func__
);
2780 health_unregister(health_consumerd
);
2782 rcu_unregister_thread();
2787 * Close wake-up end of each stream belonging to the channel. This will
2788 * allow the poll() on the stream read-side to detect when the
2789 * write-side (application) finally closes them.
2792 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2794 struct lttng_ht
*ht
;
2795 struct lttng_consumer_stream
*stream
;
2796 struct lttng_ht_iter iter
;
2798 ht
= consumer_data
.stream_per_chan_id_ht
;
2801 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2802 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2803 ht
->match_fct
, &channel
->key
,
2804 &iter
.iter
, stream
, node_channel_id
.node
) {
2806 * Protect against teardown with mutex.
2808 pthread_mutex_lock(&stream
->lock
);
2809 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2812 switch (consumer_data
.type
) {
2813 case LTTNG_CONSUMER_KERNEL
:
2815 case LTTNG_CONSUMER32_UST
:
2816 case LTTNG_CONSUMER64_UST
:
2817 if (stream
->metadata_flag
) {
2818 /* Safe and protected by the stream lock. */
2819 lttng_ustconsumer_close_metadata(stream
->chan
);
2822 * Note: a mutex is taken internally within
2823 * liblttng-ust-ctl to protect timer wakeup_fd
2824 * use from concurrent close.
2826 lttng_ustconsumer_close_stream_wakeup(stream
);
2830 ERR("Unknown consumer_data type");
2834 pthread_mutex_unlock(&stream
->lock
);
2839 static void destroy_channel_ht(struct lttng_ht
*ht
)
2841 struct lttng_ht_iter iter
;
2842 struct lttng_consumer_channel
*channel
;
2850 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2851 ret
= lttng_ht_del(ht
, &iter
);
2856 lttng_ht_destroy(ht
);
2860 * This thread polls the channel fds to detect when they are being
2861 * closed. It closes all related streams if the channel is detected as
2862 * closed. It is currently only used as a shim layer for UST because the
2863 * consumerd needs to keep the per-stream wakeup end of pipes open for
2866 void *consumer_thread_channel_poll(void *data
)
2868 int ret
, i
, pollfd
, err
= -1;
2869 uint32_t revents
, nb_fd
;
2870 struct lttng_consumer_channel
*chan
= NULL
;
2871 struct lttng_ht_iter iter
;
2872 struct lttng_ht_node_u64
*node
;
2873 struct lttng_poll_event events
;
2874 struct lttng_consumer_local_data
*ctx
= data
;
2875 struct lttng_ht
*channel_ht
;
2877 rcu_register_thread();
2879 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
2881 if (testpoint(consumerd_thread_channel
)) {
2882 goto error_testpoint
;
2885 health_code_update();
2887 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2889 /* ENOMEM at this point. Better to bail out. */
2893 DBG("Thread channel poll started");
2895 /* Size is set to 1 for the consumer_channel pipe */
2896 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2898 ERR("Poll set creation failed");
2902 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2908 DBG("Channel main loop started");
2912 health_code_update();
2913 DBG("Channel poll wait");
2914 health_poll_entry();
2915 ret
= lttng_poll_wait(&events
, -1);
2916 DBG("Channel poll return from wait with %d fd(s)",
2917 LTTNG_POLL_GETNB(&events
));
2919 DBG("Channel event caught in thread");
2921 if (errno
== EINTR
) {
2922 ERR("Poll EINTR caught");
2925 if (LTTNG_POLL_GETNB(&events
) == 0) {
2926 err
= 0; /* All is OK */
2933 /* From here, the event is a channel wait fd */
2934 for (i
= 0; i
< nb_fd
; i
++) {
2935 health_code_update();
2937 revents
= LTTNG_POLL_GETEV(&events
, i
);
2938 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2941 /* No activity for this FD (poll implementation). */
2945 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2946 if (revents
& LPOLLIN
) {
2947 enum consumer_channel_action action
;
2950 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
2953 ERR("Error reading channel pipe");
2955 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2960 case CONSUMER_CHANNEL_ADD
:
2961 DBG("Adding channel %d to poll set",
2964 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
2967 lttng_ht_add_unique_u64(channel_ht
,
2968 &chan
->wait_fd_node
);
2970 /* Add channel to the global poll events list */
2971 lttng_poll_add(&events
, chan
->wait_fd
,
2972 LPOLLERR
| LPOLLHUP
);
2974 case CONSUMER_CHANNEL_DEL
:
2977 * This command should never be called if the channel
2978 * has streams monitored by either the data or metadata
2979 * thread. The consumer only notify this thread with a
2980 * channel del. command if it receives a destroy
2981 * channel command from the session daemon that send it
2982 * if a command prior to the GET_CHANNEL failed.
2986 chan
= consumer_find_channel(key
);
2989 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
2992 lttng_poll_del(&events
, chan
->wait_fd
);
2993 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
2994 ret
= lttng_ht_del(channel_ht
, &iter
);
2997 switch (consumer_data
.type
) {
2998 case LTTNG_CONSUMER_KERNEL
:
3000 case LTTNG_CONSUMER32_UST
:
3001 case LTTNG_CONSUMER64_UST
:
3002 health_code_update();
3003 /* Destroy streams that might have been left in the stream list. */
3004 clean_channel_stream_list(chan
);
3007 ERR("Unknown consumer_data type");
3012 * Release our own refcount. Force channel deletion even if
3013 * streams were not initialized.
3015 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
3016 consumer_del_channel(chan
);
3021 case CONSUMER_CHANNEL_QUIT
:
3023 * Remove the pipe from the poll set and continue the loop
3024 * since their might be data to consume.
3026 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
3029 ERR("Unknown action");
3032 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
3033 DBG("Channel thread pipe hung up");
3035 * Remove the pipe from the poll set and continue the loop
3036 * since their might be data to consume.
3038 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
3041 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3045 /* Handle other stream */
3051 uint64_t tmp_id
= (uint64_t) pollfd
;
3053 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
3055 node
= lttng_ht_iter_get_node_u64(&iter
);
3058 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
3061 /* Check for error event */
3062 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
3063 DBG("Channel fd %d is hup|err.", pollfd
);
3065 lttng_poll_del(&events
, chan
->wait_fd
);
3066 ret
= lttng_ht_del(channel_ht
, &iter
);
3070 * This will close the wait fd for each stream associated to
3071 * this channel AND monitored by the data/metadata thread thus
3072 * will be clean by the right thread.
3074 consumer_close_channel_streams(chan
);
3076 /* Release our own refcount */
3077 if (!uatomic_sub_return(&chan
->refcount
, 1)
3078 && !uatomic_read(&chan
->nb_init_stream_left
)) {
3079 consumer_del_channel(chan
);
3082 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3087 /* Release RCU lock for the channel looked up */
3095 lttng_poll_clean(&events
);
3097 destroy_channel_ht(channel_ht
);
3100 DBG("Channel poll thread exiting");
3103 ERR("Health error occurred in %s", __func__
);
3105 health_unregister(health_consumerd
);
3106 rcu_unregister_thread();
3110 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
3111 struct pollfd
*sockpoll
, int client_socket
)
3118 ret
= lttng_consumer_poll_socket(sockpoll
);
3122 DBG("Metadata connection on client_socket");
3124 /* Blocking call, waiting for transmission */
3125 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
3126 if (ctx
->consumer_metadata_socket
< 0) {
3127 WARN("On accept metadata");
3138 * This thread listens on the consumerd socket and receives the file
3139 * descriptors from the session daemon.
3141 void *consumer_thread_sessiond_poll(void *data
)
3143 int sock
= -1, client_socket
, ret
, err
= -1;
3145 * structure to poll for incoming data on communication socket avoids
3146 * making blocking sockets.
3148 struct pollfd consumer_sockpoll
[2];
3149 struct lttng_consumer_local_data
*ctx
= data
;
3151 rcu_register_thread();
3153 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
3155 if (testpoint(consumerd_thread_sessiond
)) {
3156 goto error_testpoint
;
3159 health_code_update();
3161 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
3162 unlink(ctx
->consumer_command_sock_path
);
3163 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
3164 if (client_socket
< 0) {
3165 ERR("Cannot create command socket");
3169 ret
= lttcomm_listen_unix_sock(client_socket
);
3174 DBG("Sending ready command to lttng-sessiond");
3175 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3176 /* return < 0 on error, but == 0 is not fatal */
3178 ERR("Error sending ready command to lttng-sessiond");
3182 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3183 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3184 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3185 consumer_sockpoll
[1].fd
= client_socket
;
3186 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3188 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3196 DBG("Connection on client_socket");
3198 /* Blocking call, waiting for transmission */
3199 sock
= lttcomm_accept_unix_sock(client_socket
);
3206 * Setup metadata socket which is the second socket connection on the
3207 * command unix socket.
3209 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3218 /* This socket is not useful anymore. */
3219 ret
= close(client_socket
);
3221 PERROR("close client_socket");
3225 /* update the polling structure to poll on the established socket */
3226 consumer_sockpoll
[1].fd
= sock
;
3227 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3230 health_code_update();
3232 health_poll_entry();
3233 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3242 DBG("Incoming command on sock");
3243 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3246 * This could simply be a session daemon quitting. Don't output
3249 DBG("Communication interrupted on command socket");
3253 if (CMM_LOAD_SHARED(consumer_quit
)) {
3254 DBG("consumer_thread_receive_fds received quit from signal");
3255 err
= 0; /* All is OK */
3258 DBG("received command on sock");
3264 DBG("Consumer thread sessiond poll exiting");
3267 * Close metadata streams since the producer is the session daemon which
3270 * NOTE: for now, this only applies to the UST tracer.
3272 lttng_consumer_close_all_metadata();
3275 * when all fds have hung up, the polling thread
3278 CMM_STORE_SHARED(consumer_quit
, 1);
3281 * Notify the data poll thread to poll back again and test the
3282 * consumer_quit state that we just set so to quit gracefully.
3284 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3286 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3288 notify_health_quit_pipe(health_quit_pipe
);
3290 /* Cleaning up possibly open sockets. */
3294 PERROR("close sock sessiond poll");
3297 if (client_socket
>= 0) {
3298 ret
= close(client_socket
);
3300 PERROR("close client_socket sessiond poll");
3307 ERR("Health error occurred in %s", __func__
);
3309 health_unregister(health_consumerd
);
3311 rcu_unregister_thread();
3316 int rotate_notify_sessiond(struct lttng_consumer_local_data
*ctx
,
3322 ret
= write(ctx
->channel_rotate_pipe
, &key
, sizeof(key
));
3323 } while (ret
== -1 && errno
== EINTR
);
3325 PERROR("write to the channel rotate pipe");
3327 DBG("Sent channel rotation notification for channel key %"
3335 * Perform operations that need to be done after a stream has
3336 * rotated and released the stream lock.
3338 * Multiple rotations cannot occur simultaneously, so we know the state of the
3339 * "rotated" stream flag cannot change.
3341 * This MUST be called WITHOUT the stream lock held.
3344 int consumer_post_rotation(struct lttng_consumer_stream
*stream
,
3345 struct lttng_consumer_local_data
*ctx
)
3349 if (!stream
->rotated
) {
3353 pthread_mutex_lock(&stream
->chan
->lock
);
3354 switch (consumer_data
.type
) {
3355 case LTTNG_CONSUMER_KERNEL
:
3357 case LTTNG_CONSUMER32_UST
:
3358 case LTTNG_CONSUMER64_UST
:
3360 * The ust_metadata_pushed counter has been reset to 0, so now
3361 * we can wakeup the metadata thread so it dumps the metadata
3362 * cache to the new file.
3364 if (stream
->metadata_flag
) {
3365 consumer_metadata_wakeup_pipe(stream
->chan
);
3369 ERR("Unknown consumer_data type");
3373 if (--stream
->chan
->nr_stream_rotate_pending
== 0) {
3374 ret
= rotate_notify_sessiond(ctx
, stream
->chan
->key
);
3376 pthread_mutex_unlock(&stream
->chan
->lock
);
3377 stream
->rotated
= 0;
3383 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3384 struct lttng_consumer_local_data
*ctx
)
3389 pthread_mutex_lock(&stream
->lock
);
3390 if (stream
->metadata_flag
) {
3391 pthread_mutex_lock(&stream
->metadata_rdv_lock
);
3394 switch (consumer_data
.type
) {
3395 case LTTNG_CONSUMER_KERNEL
:
3396 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
3398 case LTTNG_CONSUMER32_UST
:
3399 case LTTNG_CONSUMER64_UST
:
3400 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
3403 ERR("Unknown consumer_data type");
3409 if (stream
->metadata_flag
) {
3410 pthread_cond_broadcast(&stream
->metadata_rdv
);
3411 pthread_mutex_unlock(&stream
->metadata_rdv_lock
);
3413 pthread_mutex_unlock(&stream
->lock
);
3415 rotate_ret
= consumer_post_rotation(stream
, ctx
);
3416 if (rotate_ret
< 0) {
3417 ERR("Failed after a rotation");
3424 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3426 switch (consumer_data
.type
) {
3427 case LTTNG_CONSUMER_KERNEL
:
3428 return lttng_kconsumer_on_recv_stream(stream
);
3429 case LTTNG_CONSUMER32_UST
:
3430 case LTTNG_CONSUMER64_UST
:
3431 return lttng_ustconsumer_on_recv_stream(stream
);
3433 ERR("Unknown consumer_data type");
3440 * Allocate and set consumer data hash tables.
3442 int lttng_consumer_init(void)
3444 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3445 if (!consumer_data
.channel_ht
) {
3449 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3450 if (!consumer_data
.relayd_ht
) {
3454 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3455 if (!consumer_data
.stream_list_ht
) {
3459 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3460 if (!consumer_data
.stream_per_chan_id_ht
) {
3464 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3469 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3481 * Process the ADD_RELAYD command receive by a consumer.
3483 * This will create a relayd socket pair and add it to the relayd hash table.
3484 * The caller MUST acquire a RCU read side lock before calling it.
3486 void consumer_add_relayd_socket(uint64_t net_seq_idx
, int sock_type
,
3487 struct lttng_consumer_local_data
*ctx
, int sock
,
3488 struct pollfd
*consumer_sockpoll
,
3489 struct lttcomm_relayd_sock
*relayd_sock
, uint64_t sessiond_id
,
3490 uint64_t relayd_session_id
)
3492 int fd
= -1, ret
= -1, relayd_created
= 0;
3493 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3494 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3497 assert(relayd_sock
);
3499 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", net_seq_idx
);
3501 /* Get relayd reference if exists. */
3502 relayd
= consumer_find_relayd(net_seq_idx
);
3503 if (relayd
== NULL
) {
3504 assert(sock_type
== LTTNG_STREAM_CONTROL
);
3505 /* Not found. Allocate one. */
3506 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3507 if (relayd
== NULL
) {
3508 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3511 relayd
->sessiond_session_id
= sessiond_id
;
3516 * This code path MUST continue to the consumer send status message to
3517 * we can notify the session daemon and continue our work without
3518 * killing everything.
3522 * relayd key should never be found for control socket.
3524 assert(sock_type
!= LTTNG_STREAM_CONTROL
);
3527 /* First send a status message before receiving the fds. */
3528 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3530 /* Somehow, the session daemon is not responding anymore. */
3531 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3532 goto error_nosignal
;
3535 /* Poll on consumer socket. */
3536 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3538 /* Needing to exit in the middle of a command: error. */
3539 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3540 goto error_nosignal
;
3543 /* Get relayd socket from session daemon */
3544 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3545 if (ret
!= sizeof(fd
)) {
3546 fd
= -1; /* Just in case it gets set with an invalid value. */
3549 * Failing to receive FDs might indicate a major problem such as
3550 * reaching a fd limit during the receive where the kernel returns a
3551 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3552 * don't take any chances and stop everything.
3554 * XXX: Feature request #558 will fix that and avoid this possible
3555 * issue when reaching the fd limit.
3557 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3558 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3562 /* Copy socket information and received FD */
3563 switch (sock_type
) {
3564 case LTTNG_STREAM_CONTROL
:
3565 /* Copy received lttcomm socket */
3566 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3567 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3568 /* Handle create_sock error. */
3570 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3574 * Close the socket created internally by
3575 * lttcomm_create_sock, so we can replace it by the one
3576 * received from sessiond.
3578 if (close(relayd
->control_sock
.sock
.fd
)) {
3582 /* Assign new file descriptor */
3583 relayd
->control_sock
.sock
.fd
= fd
;
3584 fd
= -1; /* For error path */
3585 /* Assign version values. */
3586 relayd
->control_sock
.major
= relayd_sock
->major
;
3587 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3589 relayd
->relayd_session_id
= relayd_session_id
;
3592 case LTTNG_STREAM_DATA
:
3593 /* Copy received lttcomm socket */
3594 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3595 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3596 /* Handle create_sock error. */
3598 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3602 * Close the socket created internally by
3603 * lttcomm_create_sock, so we can replace it by the one
3604 * received from sessiond.
3606 if (close(relayd
->data_sock
.sock
.fd
)) {
3610 /* Assign new file descriptor */
3611 relayd
->data_sock
.sock
.fd
= fd
;
3612 fd
= -1; /* for eventual error paths */
3613 /* Assign version values. */
3614 relayd
->data_sock
.major
= relayd_sock
->major
;
3615 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3618 ERR("Unknown relayd socket type (%d)", sock_type
);
3619 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3623 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3624 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3625 relayd
->net_seq_idx
, fd
);
3627 /* We successfully added the socket. Send status back. */
3628 ret
= consumer_send_status_msg(sock
, ret_code
);
3630 /* Somehow, the session daemon is not responding anymore. */
3631 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3632 goto error_nosignal
;
3636 * Add relayd socket pair to consumer data hashtable. If object already
3637 * exists or on error, the function gracefully returns.
3645 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3646 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3650 /* Close received socket if valid. */
3653 PERROR("close received socket");
3657 if (relayd_created
) {
3663 * Try to lock the stream mutex.
3665 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
3667 static int stream_try_lock(struct lttng_consumer_stream
*stream
)
3674 * Try to lock the stream mutex. On failure, we know that the stream is
3675 * being used else where hence there is data still being extracted.
3677 ret
= pthread_mutex_trylock(&stream
->lock
);
3679 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
3691 * Search for a relayd associated to the session id and return the reference.
3693 * A rcu read side lock MUST be acquire before calling this function and locked
3694 * until the relayd object is no longer necessary.
3696 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3698 struct lttng_ht_iter iter
;
3699 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3701 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3702 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3705 * Check by sessiond id which is unique here where the relayd session
3706 * id might not be when having multiple relayd.
3708 if (relayd
->sessiond_session_id
== id
) {
3709 /* Found the relayd. There can be only one per id. */
3721 * Check if for a given session id there is still data needed to be extract
3724 * Return 1 if data is pending or else 0 meaning ready to be read.
3726 int consumer_data_pending(uint64_t id
)
3729 struct lttng_ht_iter iter
;
3730 struct lttng_ht
*ht
;
3731 struct lttng_consumer_stream
*stream
;
3732 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3733 int (*data_pending
)(struct lttng_consumer_stream
*);
3735 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3738 pthread_mutex_lock(&consumer_data
.lock
);
3740 switch (consumer_data
.type
) {
3741 case LTTNG_CONSUMER_KERNEL
:
3742 data_pending
= lttng_kconsumer_data_pending
;
3744 case LTTNG_CONSUMER32_UST
:
3745 case LTTNG_CONSUMER64_UST
:
3746 data_pending
= lttng_ustconsumer_data_pending
;
3749 ERR("Unknown consumer data type");
3753 /* Ease our life a bit */
3754 ht
= consumer_data
.stream_list_ht
;
3756 relayd
= find_relayd_by_session_id(id
);
3758 /* Send init command for data pending. */
3759 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3760 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3761 relayd
->relayd_session_id
);
3762 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3764 /* Communication error thus the relayd so no data pending. */
3765 goto data_not_pending
;
3769 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3770 ht
->hash_fct(&id
, lttng_ht_seed
),
3772 &iter
.iter
, stream
, node_session_id
.node
) {
3773 /* If this call fails, the stream is being used hence data pending. */
3774 ret
= stream_try_lock(stream
);
3780 * A removed node from the hash table indicates that the stream has
3781 * been deleted thus having a guarantee that the buffers are closed
3782 * on the consumer side. However, data can still be transmitted
3783 * over the network so don't skip the relayd check.
3785 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3787 /* Check the stream if there is data in the buffers. */
3788 ret
= data_pending(stream
);
3790 pthread_mutex_unlock(&stream
->lock
);
3797 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3798 if (stream
->metadata_flag
) {
3799 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3800 stream
->relayd_stream_id
);
3802 ret
= relayd_data_pending(&relayd
->control_sock
,
3803 stream
->relayd_stream_id
,
3804 stream
->next_net_seq_num
- 1);
3806 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3808 pthread_mutex_unlock(&stream
->lock
);
3812 pthread_mutex_unlock(&stream
->lock
);
3816 unsigned int is_data_inflight
= 0;
3818 /* Send init command for data pending. */
3819 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3820 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3821 relayd
->relayd_session_id
, &is_data_inflight
);
3822 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3824 goto data_not_pending
;
3826 if (is_data_inflight
) {
3832 * Finding _no_ node in the hash table and no inflight data means that the
3833 * stream(s) have been removed thus data is guaranteed to be available for
3834 * analysis from the trace files.
3838 /* Data is available to be read by a viewer. */
3839 pthread_mutex_unlock(&consumer_data
.lock
);
3844 /* Data is still being extracted from buffers. */
3845 pthread_mutex_unlock(&consumer_data
.lock
);
3851 * Send a ret code status message to the sessiond daemon.
3853 * Return the sendmsg() return value.
3855 int consumer_send_status_msg(int sock
, int ret_code
)
3857 struct lttcomm_consumer_status_msg msg
;
3859 memset(&msg
, 0, sizeof(msg
));
3860 msg
.ret_code
= ret_code
;
3862 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3866 * Send a channel status message to the sessiond daemon.
3868 * Return the sendmsg() return value.
3870 int consumer_send_status_channel(int sock
,
3871 struct lttng_consumer_channel
*channel
)
3873 struct lttcomm_consumer_status_channel msg
;
3877 memset(&msg
, 0, sizeof(msg
));
3879 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3881 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3882 msg
.key
= channel
->key
;
3883 msg
.stream_count
= channel
->streams
.count
;
3886 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3889 unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos
,
3890 unsigned long produced_pos
, uint64_t nb_packets_per_stream
,
3891 uint64_t max_sb_size
)
3893 unsigned long start_pos
;
3895 if (!nb_packets_per_stream
) {
3896 return consumed_pos
; /* Grab everything */
3898 start_pos
= produced_pos
- offset_align_floor(produced_pos
, max_sb_size
);
3899 start_pos
-= max_sb_size
* nb_packets_per_stream
;
3900 if ((long) (start_pos
- consumed_pos
) < 0) {
3901 return consumed_pos
; /* Grab everything */
3907 int consumer_flush_buffer(struct lttng_consumer_stream
*stream
, int producer_active
)
3911 switch (consumer_data
.type
) {
3912 case LTTNG_CONSUMER_KERNEL
:
3913 ret
= kernctl_buffer_flush(stream
->wait_fd
);
3915 ERR("Failed to flush kernel stream");
3919 case LTTNG_CONSUMER32_UST
:
3920 case LTTNG_CONSUMER64_UST
:
3921 lttng_ustctl_flush_buffer(stream
, producer_active
);
3924 ERR("Unknown consumer_data type");
3933 * Sample the rotate position for all the streams of a channel.
3935 * Returns 0 on success, < 0 on error
3937 int lttng_consumer_rotate_channel(uint64_t key
, char *path
,
3938 uint64_t relayd_id
, uint32_t metadata
,
3939 struct lttng_consumer_local_data
*ctx
)
3942 struct lttng_consumer_channel
*channel
;
3943 struct lttng_consumer_stream
*stream
;
3944 struct lttng_ht_iter iter
;
3945 struct lttng_ht
*ht
= consumer_data
.stream_per_chan_id_ht
;
3947 DBG("Kernel consumer sample rotate position for channel %" PRIu64
, key
);
3951 channel
= consumer_find_channel(key
);
3953 ERR("No channel found for key %" PRIu64
, key
);
3957 pthread_mutex_lock(&channel
->lock
);
3958 snprintf(channel
->pathname
, PATH_MAX
, "%s", path
);
3960 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3961 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
3962 ht
->match_fct
, &channel
->key
, &iter
.iter
,
3963 stream
, node_channel_id
.node
) {
3964 health_code_update();
3967 * Lock stream because we are about to change its state.
3969 pthread_mutex_lock(&stream
->lock
);
3970 memcpy(stream
->channel_ro_pathname
, channel
->pathname
, PATH_MAX
);
3971 ret
= lttng_consumer_sample_snapshot_positions(stream
);
3973 ERR("Taking kernel snapshot positions");
3976 uint64_t consumed_pos
;
3978 ret
= lttng_consumer_get_produced_snapshot(stream
,
3979 &stream
->rotate_position
);
3981 ERR("Produced kernel snapshot position");
3984 fprintf(stderr
, "Stream %lu should rotate after %lu to %s\n",
3985 stream
->key
, stream
->rotate_position
,
3987 lttng_consumer_get_consumed_snapshot(stream
,
3989 fprintf(stderr
, "consumed %lu\n", consumed_pos
);
3990 if (consumed_pos
== stream
->rotate_position
) {
3991 stream
->rotate_ready
= 1;
3992 fprintf(stderr
, "Stream %lu ready to rotate to %s\n",
3993 stream
->key
, channel
->pathname
);
3996 channel
->nr_stream_rotate_pending
++;
3998 ret
= consumer_flush_buffer(stream
, 1);
4000 ERR("Failed to flush stream");
4004 pthread_mutex_unlock(&stream
->lock
);
4008 goto end_unlock_channel
;
4011 pthread_mutex_unlock(&stream
->lock
);
4013 pthread_mutex_unlock(&channel
->lock
);
4020 * Performs the stream rotation for the rotate session feature if needed.
4021 * It must be called with the stream and channel locks held.
4023 * Return 0 on success, a negative number of error.
4025 int lttng_consumer_rotate_stream(struct lttng_consumer_local_data
*ctx
,
4026 struct lttng_consumer_stream
*stream
)
4029 unsigned long consumed_pos
;
4031 if (!stream
->rotate_position
&& !stream
->rotate_ready
) {
4037 * If we don't have the rotate_ready flag, check the consumed position
4038 * to determine if we need to rotate.
4040 if (!stream
->rotate_ready
) {
4041 ret
= lttng_consumer_sample_snapshot_positions(stream
);
4043 ERR("Taking kernel snapshot positions");
4047 ret
= lttng_consumer_get_consumed_snapshot(stream
, &consumed_pos
);
4049 ERR("Produced kernel snapshot position");
4053 fprintf(stderr
, "packet %lu, pos %lu\n", stream
->key
, consumed_pos
);
4054 /* Rotate position not reached yet. */
4055 if (consumed_pos
< stream
->rotate_position
) {
4059 fprintf(stderr
, "Rotate position %lu (expected %lu) reached for stream %lu\n",
4060 consumed_pos
, stream
->rotate_position
, stream
->key
);
4062 fprintf(stderr
, "Rotate position reached for stream %lu\n",
4066 ret
= close(stream
->out_fd
);
4068 PERROR("Closing tracefile");
4072 fprintf(stderr
, "Rotating stream %lu to %s/%s\n", stream
->key
,
4073 stream
->channel_ro_pathname
, stream
->name
);
4074 ret
= utils_create_stream_file(stream
->channel_ro_pathname
, stream
->name
,
4075 stream
->channel_ro_tracefile_size
, stream
->tracefile_count_current
,
4076 stream
->uid
, stream
->gid
, NULL
);
4080 stream
->out_fd
= ret
;
4081 stream
->tracefile_size_current
= 0;
4083 if (!stream
->metadata_flag
) {
4084 struct lttng_index_file
*index_file
;
4086 lttng_index_file_put(stream
->index_file
);
4088 index_file
= lttng_index_file_create(stream
->channel_ro_pathname
,
4089 stream
->name
, stream
->uid
, stream
->gid
,
4090 stream
->channel_ro_tracefile_size
,
4091 stream
->tracefile_count_current
,
4092 CTF_INDEX_MAJOR
, CTF_INDEX_MINOR
);
4096 stream
->index_file
= index_file
;
4097 stream
->out_fd_offset
= 0;
4099 switch (consumer_data
.type
) {
4100 case LTTNG_CONSUMER_KERNEL
:
4102 * Reset the position of what has been read from the metadata
4103 * cache to 0 so we can dump it again.
4105 ret
= kernctl_metadata_cache_dump(stream
->wait_fd
);
4107 ERR("Failed to dump the kernel metadata cache after rotation");
4111 case LTTNG_CONSUMER32_UST
:
4112 case LTTNG_CONSUMER64_UST
:
4114 * Reset the position pushed from the metadata cache so it
4115 * will write from the beginning on the next push.
4117 stream
->ust_metadata_pushed
= 0;
4120 ERR("Unknown consumer_data type");
4125 stream
->rotate_position
= 0;
4126 stream
->rotate_ready
= 0;
4127 stream
->rotated
= 1;
4139 * Rotate all the ready streams.
4141 * This is especially important for low throughput streams that have already
4142 * been consumed, we cannot wait for their next packet to perform the
4145 * Returns 0 on success, < 0 on error
4147 int lttng_consumer_rotate_ready_streams(uint64_t key
,
4148 struct lttng_consumer_local_data
*ctx
)
4151 struct lttng_consumer_channel
*channel
;
4152 struct lttng_consumer_stream
*stream
;
4153 struct lttng_ht_iter iter
;
4154 struct lttng_ht
*ht
= consumer_data
.stream_per_chan_id_ht
;
4158 channel
= consumer_find_channel(key
);
4160 ERR("No channel found for key %" PRIu64
, key
);
4164 cds_lfht_for_each_entry_duplicate(ht
->ht
,
4165 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
4166 ht
->match_fct
, &channel
->key
, &iter
.iter
,
4167 stream
, node_channel_id
.node
) {
4168 health_code_update();
4171 * Lock stream because we are about to change its state.
4173 pthread_mutex_lock(&stream
->lock
);
4174 if (stream
->rotate_ready
== 0) {
4175 pthread_mutex_unlock(&stream
->lock
);
4178 ret
= lttng_consumer_rotate_stream(ctx
, stream
);
4180 pthread_mutex_unlock(&stream
->lock
);
4181 ERR("Stream rotation error");
4185 pthread_mutex_unlock(&stream
->lock
);
4186 ret
= consumer_post_rotation(stream
, ctx
);
4188 ERR("Failed after a rotation");