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
);
326 pthread_mutex_destroy(&relayd
->ctrl_sock_mutex
);
331 * Destroy and free relayd socket pair object.
333 void consumer_destroy_relayd(struct consumer_relayd_sock_pair
*relayd
)
336 struct lttng_ht_iter iter
;
338 if (relayd
== NULL
) {
342 DBG("Consumer destroy and close relayd socket pair");
344 iter
.iter
.node
= &relayd
->node
.node
;
345 ret
= lttng_ht_del(consumer_data
.relayd_ht
, &iter
);
347 /* We assume the relayd is being or is destroyed */
351 /* RCU free() call */
352 call_rcu(&relayd
->node
.head
, free_relayd_rcu
);
356 * Remove a channel from the global list protected by a mutex. This function is
357 * also responsible for freeing its data structures.
359 void consumer_del_channel(struct lttng_consumer_channel
*channel
)
362 struct lttng_ht_iter iter
;
364 DBG("Consumer delete channel key %" PRIu64
, channel
->key
);
366 pthread_mutex_lock(&consumer_data
.lock
);
367 pthread_mutex_lock(&channel
->lock
);
369 /* Destroy streams that might have been left in the stream list. */
370 clean_channel_stream_list(channel
);
372 if (channel
->live_timer_enabled
== 1) {
373 consumer_timer_live_stop(channel
);
375 if (channel
->monitor_timer_enabled
== 1) {
376 consumer_timer_monitor_stop(channel
);
379 switch (consumer_data
.type
) {
380 case LTTNG_CONSUMER_KERNEL
:
382 case LTTNG_CONSUMER32_UST
:
383 case LTTNG_CONSUMER64_UST
:
384 lttng_ustconsumer_del_channel(channel
);
387 ERR("Unknown consumer_data type");
393 iter
.iter
.node
= &channel
->node
.node
;
394 ret
= lttng_ht_del(consumer_data
.channel_ht
, &iter
);
398 call_rcu(&channel
->node
.head
, free_channel_rcu
);
400 pthread_mutex_unlock(&channel
->lock
);
401 pthread_mutex_unlock(&consumer_data
.lock
);
405 * Iterate over the relayd hash table and destroy each element. Finally,
406 * destroy the whole hash table.
408 static void cleanup_relayd_ht(void)
410 struct lttng_ht_iter iter
;
411 struct consumer_relayd_sock_pair
*relayd
;
415 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
417 consumer_destroy_relayd(relayd
);
422 lttng_ht_destroy(consumer_data
.relayd_ht
);
426 * Update the end point status of all streams having the given network sequence
427 * index (relayd index).
429 * It's atomically set without having the stream mutex locked which is fine
430 * because we handle the write/read race with a pipe wakeup for each thread.
432 static void update_endpoint_status_by_netidx(uint64_t net_seq_idx
,
433 enum consumer_endpoint_status status
)
435 struct lttng_ht_iter iter
;
436 struct lttng_consumer_stream
*stream
;
438 DBG("Consumer set delete flag on stream by idx %" PRIu64
, net_seq_idx
);
442 /* Let's begin with metadata */
443 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
444 if (stream
->net_seq_idx
== net_seq_idx
) {
445 uatomic_set(&stream
->endpoint_status
, status
);
446 DBG("Delete flag set to metadata stream %d", stream
->wait_fd
);
450 /* Follow up by the data streams */
451 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
452 if (stream
->net_seq_idx
== net_seq_idx
) {
453 uatomic_set(&stream
->endpoint_status
, status
);
454 DBG("Delete flag set to data stream %d", stream
->wait_fd
);
461 * Cleanup a relayd object by flagging every associated streams for deletion,
462 * destroying the object meaning removing it from the relayd hash table,
463 * closing the sockets and freeing the memory in a RCU call.
465 * If a local data context is available, notify the threads that the streams'
466 * state have changed.
468 static void cleanup_relayd(struct consumer_relayd_sock_pair
*relayd
,
469 struct lttng_consumer_local_data
*ctx
)
475 DBG("Cleaning up relayd sockets");
477 /* Save the net sequence index before destroying the object */
478 netidx
= relayd
->net_seq_idx
;
481 * Delete the relayd from the relayd hash table, close the sockets and free
482 * the object in a RCU call.
484 consumer_destroy_relayd(relayd
);
486 /* Set inactive endpoint to all streams */
487 update_endpoint_status_by_netidx(netidx
, CONSUMER_ENDPOINT_INACTIVE
);
490 * With a local data context, notify the threads that the streams' state
491 * have changed. The write() action on the pipe acts as an "implicit"
492 * memory barrier ordering the updates of the end point status from the
493 * read of this status which happens AFTER receiving this notify.
496 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
497 notify_thread_lttng_pipe(ctx
->consumer_metadata_pipe
);
502 * Flag a relayd socket pair for destruction. Destroy it if the refcount
505 * RCU read side lock MUST be aquired before calling this function.
507 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair
*relayd
)
511 /* Set destroy flag for this object */
512 uatomic_set(&relayd
->destroy_flag
, 1);
514 /* Destroy the relayd if refcount is 0 */
515 if (uatomic_read(&relayd
->refcount
) == 0) {
516 consumer_destroy_relayd(relayd
);
521 * Completly destroy stream from every visiable data structure and the given
524 * One this call returns, the stream object is not longer usable nor visible.
526 void consumer_del_stream(struct lttng_consumer_stream
*stream
,
529 consumer_stream_destroy(stream
, ht
);
533 * XXX naming of del vs destroy is all mixed up.
535 void consumer_del_stream_for_data(struct lttng_consumer_stream
*stream
)
537 consumer_stream_destroy(stream
, data_ht
);
540 void consumer_del_stream_for_metadata(struct lttng_consumer_stream
*stream
)
542 consumer_stream_destroy(stream
, metadata_ht
);
545 struct lttng_consumer_stream
*consumer_allocate_stream(uint64_t channel_key
,
547 enum lttng_consumer_stream_state state
,
548 const char *channel_name
,
555 enum consumer_channel_type type
,
556 unsigned int monitor
)
559 struct lttng_consumer_stream
*stream
;
561 stream
= zmalloc(sizeof(*stream
));
562 if (stream
== NULL
) {
563 PERROR("malloc struct lttng_consumer_stream");
570 stream
->key
= stream_key
;
572 stream
->out_fd_offset
= 0;
573 stream
->output_written
= 0;
574 stream
->state
= state
;
577 stream
->net_seq_idx
= relayd_id
;
578 stream
->session_id
= session_id
;
579 stream
->monitor
= monitor
;
580 stream
->endpoint_status
= CONSUMER_ENDPOINT_ACTIVE
;
581 stream
->index_file
= NULL
;
582 stream
->last_sequence_number
= -1ULL;
583 pthread_mutex_init(&stream
->lock
, NULL
);
584 pthread_mutex_init(&stream
->metadata_timer_lock
, NULL
);
586 /* If channel is the metadata, flag this stream as metadata. */
587 if (type
== CONSUMER_CHANNEL_TYPE_METADATA
) {
588 stream
->metadata_flag
= 1;
589 /* Metadata is flat out. */
590 strncpy(stream
->name
, DEFAULT_METADATA_NAME
, sizeof(stream
->name
));
591 /* Live rendez-vous point. */
592 pthread_cond_init(&stream
->metadata_rdv
, NULL
);
593 pthread_mutex_init(&stream
->metadata_rdv_lock
, NULL
);
595 /* Format stream name to <channel_name>_<cpu_number> */
596 ret
= snprintf(stream
->name
, sizeof(stream
->name
), "%s_%d",
599 PERROR("snprintf stream name");
604 /* Key is always the wait_fd for streams. */
605 lttng_ht_node_init_u64(&stream
->node
, stream
->key
);
607 /* Init node per channel id key */
608 lttng_ht_node_init_u64(&stream
->node_channel_id
, channel_key
);
610 /* Init session id node with the stream session id */
611 lttng_ht_node_init_u64(&stream
->node_session_id
, stream
->session_id
);
613 DBG3("Allocated stream %s (key %" PRIu64
", chan_key %" PRIu64
614 " relayd_id %" PRIu64
", session_id %" PRIu64
,
615 stream
->name
, stream
->key
, channel_key
,
616 stream
->net_seq_idx
, stream
->session_id
);
632 * Add a stream to the global list protected by a mutex.
634 int consumer_add_data_stream(struct lttng_consumer_stream
*stream
)
636 struct lttng_ht
*ht
= data_ht
;
642 DBG3("Adding consumer stream %" PRIu64
, stream
->key
);
644 pthread_mutex_lock(&consumer_data
.lock
);
645 pthread_mutex_lock(&stream
->chan
->lock
);
646 pthread_mutex_lock(&stream
->chan
->timer_lock
);
647 pthread_mutex_lock(&stream
->lock
);
650 /* Steal stream identifier to avoid having streams with the same key */
651 steal_stream_key(stream
->key
, ht
);
653 lttng_ht_add_unique_u64(ht
, &stream
->node
);
655 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
656 &stream
->node_channel_id
);
659 * Add stream to the stream_list_ht of the consumer data. No need to steal
660 * the key since the HT does not use it and we allow to add redundant keys
663 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
666 * When nb_init_stream_left reaches 0, we don't need to trigger any action
667 * in terms of destroying the associated channel, because the action that
668 * causes the count to become 0 also causes a stream to be added. The
669 * channel deletion will thus be triggered by the following removal of this
672 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
673 /* Increment refcount before decrementing nb_init_stream_left */
675 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
678 /* Update consumer data once the node is inserted. */
679 consumer_data
.stream_count
++;
680 consumer_data
.need_update
= 1;
683 pthread_mutex_unlock(&stream
->lock
);
684 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
685 pthread_mutex_unlock(&stream
->chan
->lock
);
686 pthread_mutex_unlock(&consumer_data
.lock
);
691 void consumer_del_data_stream(struct lttng_consumer_stream
*stream
)
693 consumer_del_stream(stream
, data_ht
);
697 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
698 * be acquired before calling this.
700 static int add_relayd(struct consumer_relayd_sock_pair
*relayd
)
703 struct lttng_ht_node_u64
*node
;
704 struct lttng_ht_iter iter
;
708 lttng_ht_lookup(consumer_data
.relayd_ht
,
709 &relayd
->net_seq_idx
, &iter
);
710 node
= lttng_ht_iter_get_node_u64(&iter
);
714 lttng_ht_add_unique_u64(consumer_data
.relayd_ht
, &relayd
->node
);
721 * Allocate and return a consumer relayd socket.
723 static struct consumer_relayd_sock_pair
*consumer_allocate_relayd_sock_pair(
724 uint64_t net_seq_idx
)
726 struct consumer_relayd_sock_pair
*obj
= NULL
;
728 /* net sequence index of -1 is a failure */
729 if (net_seq_idx
== (uint64_t) -1ULL) {
733 obj
= zmalloc(sizeof(struct consumer_relayd_sock_pair
));
735 PERROR("zmalloc relayd sock");
739 obj
->net_seq_idx
= net_seq_idx
;
741 obj
->destroy_flag
= 0;
742 obj
->control_sock
.sock
.fd
= -1;
743 obj
->data_sock
.sock
.fd
= -1;
744 lttng_ht_node_init_u64(&obj
->node
, obj
->net_seq_idx
);
745 pthread_mutex_init(&obj
->ctrl_sock_mutex
, NULL
);
752 * Find a relayd socket pair in the global consumer data.
754 * Return the object if found else NULL.
755 * RCU read-side lock must be held across this call and while using the
758 struct consumer_relayd_sock_pair
*consumer_find_relayd(uint64_t key
)
760 struct lttng_ht_iter iter
;
761 struct lttng_ht_node_u64
*node
;
762 struct consumer_relayd_sock_pair
*relayd
= NULL
;
764 /* Negative keys are lookup failures */
765 if (key
== (uint64_t) -1ULL) {
769 lttng_ht_lookup(consumer_data
.relayd_ht
, &key
,
771 node
= lttng_ht_iter_get_node_u64(&iter
);
773 relayd
= caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
781 * Find a relayd and send the stream
783 * Returns 0 on success, < 0 on error
785 int consumer_send_relayd_stream(struct lttng_consumer_stream
*stream
,
789 struct consumer_relayd_sock_pair
*relayd
;
792 assert(stream
->net_seq_idx
!= -1ULL);
795 /* The stream is not metadata. Get relayd reference if exists. */
797 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
798 if (relayd
!= NULL
) {
799 /* Add stream on the relayd */
800 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
801 ret
= relayd_add_stream(&relayd
->control_sock
, stream
->name
,
802 path
, &stream
->relayd_stream_id
,
803 stream
->chan
->tracefile_size
, stream
->chan
->tracefile_count
);
804 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
809 uatomic_inc(&relayd
->refcount
);
810 stream
->sent_to_relayd
= 1;
812 ERR("Stream %" PRIu64
" relayd ID %" PRIu64
" unknown. Can't send it.",
813 stream
->key
, stream
->net_seq_idx
);
818 DBG("Stream %s with key %" PRIu64
" sent to relayd id %" PRIu64
,
819 stream
->name
, stream
->key
, stream
->net_seq_idx
);
827 * Find a relayd and send the streams sent message
829 * Returns 0 on success, < 0 on error
831 int consumer_send_relayd_streams_sent(uint64_t net_seq_idx
)
834 struct consumer_relayd_sock_pair
*relayd
;
836 assert(net_seq_idx
!= -1ULL);
838 /* The stream is not metadata. Get relayd reference if exists. */
840 relayd
= consumer_find_relayd(net_seq_idx
);
841 if (relayd
!= NULL
) {
842 /* Add stream on the relayd */
843 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
844 ret
= relayd_streams_sent(&relayd
->control_sock
);
845 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
850 ERR("Relayd ID %" PRIu64
" unknown. Can't send streams_sent.",
857 DBG("All streams sent relayd id %" PRIu64
, net_seq_idx
);
865 * Find a relayd and close the stream
867 void close_relayd_stream(struct lttng_consumer_stream
*stream
)
869 struct consumer_relayd_sock_pair
*relayd
;
871 /* The stream is not metadata. Get relayd reference if exists. */
873 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
875 consumer_stream_relayd_close(stream
, relayd
);
881 * Handle stream for relayd transmission if the stream applies for network
882 * streaming where the net sequence index is set.
884 * Return destination file descriptor or negative value on error.
886 static int write_relayd_stream_header(struct lttng_consumer_stream
*stream
,
887 size_t data_size
, unsigned long padding
,
888 struct consumer_relayd_sock_pair
*relayd
)
891 struct lttcomm_relayd_data_hdr data_hdr
;
897 /* Reset data header */
898 memset(&data_hdr
, 0, sizeof(data_hdr
));
900 if (stream
->metadata_flag
) {
901 /* Caller MUST acquire the relayd control socket lock */
902 ret
= relayd_send_metadata(&relayd
->control_sock
, data_size
);
907 /* Metadata are always sent on the control socket. */
908 outfd
= relayd
->control_sock
.sock
.fd
;
910 /* Set header with stream information */
911 data_hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
912 data_hdr
.data_size
= htobe32(data_size
);
913 data_hdr
.padding_size
= htobe32(padding
);
915 * Note that net_seq_num below is assigned with the *current* value of
916 * next_net_seq_num and only after that the next_net_seq_num will be
917 * increment. This is why when issuing a command on the relayd using
918 * this next value, 1 should always be substracted in order to compare
919 * the last seen sequence number on the relayd side to the last sent.
921 data_hdr
.net_seq_num
= htobe64(stream
->next_net_seq_num
);
922 /* Other fields are zeroed previously */
924 ret
= relayd_send_data_hdr(&relayd
->data_sock
, &data_hdr
,
930 ++stream
->next_net_seq_num
;
932 /* Set to go on data socket */
933 outfd
= relayd
->data_sock
.sock
.fd
;
941 * Allocate and return a new lttng_consumer_channel object using the given key
942 * to initialize the hash table node.
944 * On error, return NULL.
946 struct lttng_consumer_channel
*consumer_allocate_channel(uint64_t key
,
948 const char *pathname
,
953 enum lttng_event_output output
,
954 uint64_t tracefile_size
,
955 uint64_t tracefile_count
,
956 uint64_t session_id_per_pid
,
957 unsigned int monitor
,
958 unsigned int live_timer_interval
,
959 const char *root_shm_path
,
960 const char *shm_path
)
962 struct lttng_consumer_channel
*channel
;
964 channel
= zmalloc(sizeof(*channel
));
965 if (channel
== NULL
) {
966 PERROR("malloc struct lttng_consumer_channel");
971 channel
->refcount
= 0;
972 channel
->session_id
= session_id
;
973 channel
->session_id_per_pid
= session_id_per_pid
;
976 channel
->relayd_id
= relayd_id
;
977 channel
->tracefile_size
= tracefile_size
;
978 channel
->tracefile_count
= tracefile_count
;
979 channel
->monitor
= monitor
;
980 channel
->live_timer_interval
= live_timer_interval
;
981 pthread_mutex_init(&channel
->lock
, NULL
);
982 pthread_mutex_init(&channel
->timer_lock
, NULL
);
985 case LTTNG_EVENT_SPLICE
:
986 channel
->output
= CONSUMER_CHANNEL_SPLICE
;
988 case LTTNG_EVENT_MMAP
:
989 channel
->output
= CONSUMER_CHANNEL_MMAP
;
999 * In monitor mode, the streams associated with the channel will be put in
1000 * a special list ONLY owned by this channel. So, the refcount is set to 1
1001 * here meaning that the channel itself has streams that are referenced.
1003 * On a channel deletion, once the channel is no longer visible, the
1004 * refcount is decremented and checked for a zero value to delete it. With
1005 * streams in no monitor mode, it will now be safe to destroy the channel.
1007 if (!channel
->monitor
) {
1008 channel
->refcount
= 1;
1011 strncpy(channel
->pathname
, pathname
, sizeof(channel
->pathname
));
1012 channel
->pathname
[sizeof(channel
->pathname
) - 1] = '\0';
1014 strncpy(channel
->name
, name
, sizeof(channel
->name
));
1015 channel
->name
[sizeof(channel
->name
) - 1] = '\0';
1017 if (root_shm_path
) {
1018 strncpy(channel
->root_shm_path
, root_shm_path
, sizeof(channel
->root_shm_path
));
1019 channel
->root_shm_path
[sizeof(channel
->root_shm_path
) - 1] = '\0';
1022 strncpy(channel
->shm_path
, shm_path
, sizeof(channel
->shm_path
));
1023 channel
->shm_path
[sizeof(channel
->shm_path
) - 1] = '\0';
1026 lttng_ht_node_init_u64(&channel
->node
, channel
->key
);
1028 channel
->wait_fd
= -1;
1030 CDS_INIT_LIST_HEAD(&channel
->streams
.head
);
1032 DBG("Allocated channel (key %" PRIu64
")", channel
->key
);
1039 * Add a channel to the global list protected by a mutex.
1041 * Always return 0 indicating success.
1043 int consumer_add_channel(struct lttng_consumer_channel
*channel
,
1044 struct lttng_consumer_local_data
*ctx
)
1046 pthread_mutex_lock(&consumer_data
.lock
);
1047 pthread_mutex_lock(&channel
->lock
);
1048 pthread_mutex_lock(&channel
->timer_lock
);
1051 * This gives us a guarantee that the channel we are about to add to the
1052 * channel hash table will be unique. See this function comment on the why
1053 * we need to steel the channel key at this stage.
1055 steal_channel_key(channel
->key
);
1058 lttng_ht_add_unique_u64(consumer_data
.channel_ht
, &channel
->node
);
1061 pthread_mutex_unlock(&channel
->timer_lock
);
1062 pthread_mutex_unlock(&channel
->lock
);
1063 pthread_mutex_unlock(&consumer_data
.lock
);
1065 if (channel
->wait_fd
!= -1 && channel
->type
== CONSUMER_CHANNEL_TYPE_DATA
) {
1066 notify_channel_pipe(ctx
, channel
, -1, CONSUMER_CHANNEL_ADD
);
1073 * Allocate the pollfd structure and the local view of the out fds to avoid
1074 * doing a lookup in the linked list and concurrency issues when writing is
1075 * needed. Called with consumer_data.lock held.
1077 * Returns the number of fds in the structures.
1079 static int update_poll_array(struct lttng_consumer_local_data
*ctx
,
1080 struct pollfd
**pollfd
, struct lttng_consumer_stream
**local_stream
,
1081 struct lttng_ht
*ht
, int *nb_inactive_fd
)
1084 struct lttng_ht_iter iter
;
1085 struct lttng_consumer_stream
*stream
;
1090 assert(local_stream
);
1092 DBG("Updating poll fd array");
1093 *nb_inactive_fd
= 0;
1095 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1097 * Only active streams with an active end point can be added to the
1098 * poll set and local stream storage of the thread.
1100 * There is a potential race here for endpoint_status to be updated
1101 * just after the check. However, this is OK since the stream(s) will
1102 * be deleted once the thread is notified that the end point state has
1103 * changed where this function will be called back again.
1105 * We track the number of inactive FDs because they still need to be
1106 * closed by the polling thread after a wakeup on the data_pipe or
1109 if (stream
->state
!= LTTNG_CONSUMER_ACTIVE_STREAM
||
1110 stream
->endpoint_status
== CONSUMER_ENDPOINT_INACTIVE
) {
1111 (*nb_inactive_fd
)++;
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 * Take a snapshot for a specific fd
1971 * Returns 0 on success, < 0 on error
1973 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
1975 switch (consumer_data
.type
) {
1976 case LTTNG_CONSUMER_KERNEL
:
1977 return lttng_kconsumer_take_snapshot(stream
);
1978 case LTTNG_CONSUMER32_UST
:
1979 case LTTNG_CONSUMER64_UST
:
1980 return lttng_ustconsumer_take_snapshot(stream
);
1982 ERR("Unknown consumer_data type");
1989 * Get the produced position
1991 * Returns 0 on success, < 0 on error
1993 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
1996 switch (consumer_data
.type
) {
1997 case LTTNG_CONSUMER_KERNEL
:
1998 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
1999 case LTTNG_CONSUMER32_UST
:
2000 case LTTNG_CONSUMER64_UST
:
2001 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
2003 ERR("Unknown consumer_data type");
2009 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
2010 int sock
, struct pollfd
*consumer_sockpoll
)
2012 switch (consumer_data
.type
) {
2013 case LTTNG_CONSUMER_KERNEL
:
2014 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2015 case LTTNG_CONSUMER32_UST
:
2016 case LTTNG_CONSUMER64_UST
:
2017 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2019 ERR("Unknown consumer_data type");
2025 void lttng_consumer_close_all_metadata(void)
2027 switch (consumer_data
.type
) {
2028 case LTTNG_CONSUMER_KERNEL
:
2030 * The Kernel consumer has a different metadata scheme so we don't
2031 * close anything because the stream will be closed by the session
2035 case LTTNG_CONSUMER32_UST
:
2036 case LTTNG_CONSUMER64_UST
:
2038 * Close all metadata streams. The metadata hash table is passed and
2039 * this call iterates over it by closing all wakeup fd. This is safe
2040 * because at this point we are sure that the metadata producer is
2041 * either dead or blocked.
2043 lttng_ustconsumer_close_all_metadata(metadata_ht
);
2046 ERR("Unknown consumer_data type");
2052 * Clean up a metadata stream and free its memory.
2054 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
2055 struct lttng_ht
*ht
)
2057 struct lttng_consumer_channel
*free_chan
= NULL
;
2061 * This call should NEVER receive regular stream. It must always be
2062 * metadata stream and this is crucial for data structure synchronization.
2064 assert(stream
->metadata_flag
);
2066 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
2068 pthread_mutex_lock(&consumer_data
.lock
);
2069 pthread_mutex_lock(&stream
->chan
->lock
);
2070 pthread_mutex_lock(&stream
->lock
);
2071 if (stream
->chan
->metadata_cache
) {
2072 /* Only applicable to userspace consumers. */
2073 pthread_mutex_lock(&stream
->chan
->metadata_cache
->lock
);
2076 /* Remove any reference to that stream. */
2077 consumer_stream_delete(stream
, ht
);
2079 /* Close down everything including the relayd if one. */
2080 consumer_stream_close(stream
);
2081 /* Destroy tracer buffers of the stream. */
2082 consumer_stream_destroy_buffers(stream
);
2084 /* Atomically decrement channel refcount since other threads can use it. */
2085 if (!uatomic_sub_return(&stream
->chan
->refcount
, 1)
2086 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
2087 /* Go for channel deletion! */
2088 free_chan
= stream
->chan
;
2092 * Nullify the stream reference so it is not used after deletion. The
2093 * channel lock MUST be acquired before being able to check for a NULL
2096 stream
->chan
->metadata_stream
= NULL
;
2098 if (stream
->chan
->metadata_cache
) {
2099 pthread_mutex_unlock(&stream
->chan
->metadata_cache
->lock
);
2101 pthread_mutex_unlock(&stream
->lock
);
2102 pthread_mutex_unlock(&stream
->chan
->lock
);
2103 pthread_mutex_unlock(&consumer_data
.lock
);
2106 consumer_del_channel(free_chan
);
2109 consumer_stream_free(stream
);
2113 * Action done with the metadata stream when adding it to the consumer internal
2114 * data structures to handle it.
2116 int consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
2118 struct lttng_ht
*ht
= metadata_ht
;
2120 struct lttng_ht_iter iter
;
2121 struct lttng_ht_node_u64
*node
;
2126 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2128 pthread_mutex_lock(&consumer_data
.lock
);
2129 pthread_mutex_lock(&stream
->chan
->lock
);
2130 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2131 pthread_mutex_lock(&stream
->lock
);
2134 * From here, refcounts are updated so be _careful_ when returning an error
2141 * Lookup the stream just to make sure it does not exist in our internal
2142 * state. This should NEVER happen.
2144 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2145 node
= lttng_ht_iter_get_node_u64(&iter
);
2149 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2150 * in terms of destroying the associated channel, because the action that
2151 * causes the count to become 0 also causes a stream to be added. The
2152 * channel deletion will thus be triggered by the following removal of this
2155 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2156 /* Increment refcount before decrementing nb_init_stream_left */
2158 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2161 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2163 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
2164 &stream
->node_channel_id
);
2167 * Add stream to the stream_list_ht of the consumer data. No need to steal
2168 * the key since the HT does not use it and we allow to add redundant keys
2171 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2175 pthread_mutex_unlock(&stream
->lock
);
2176 pthread_mutex_unlock(&stream
->chan
->lock
);
2177 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2178 pthread_mutex_unlock(&consumer_data
.lock
);
2183 * Delete data stream that are flagged for deletion (endpoint_status).
2185 static void validate_endpoint_status_data_stream(void)
2187 struct lttng_ht_iter iter
;
2188 struct lttng_consumer_stream
*stream
;
2190 DBG("Consumer delete flagged data stream");
2193 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2194 /* Validate delete flag of the stream */
2195 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2198 /* Delete it right now */
2199 consumer_del_stream(stream
, data_ht
);
2205 * Delete metadata stream that are flagged for deletion (endpoint_status).
2207 static void validate_endpoint_status_metadata_stream(
2208 struct lttng_poll_event
*pollset
)
2210 struct lttng_ht_iter iter
;
2211 struct lttng_consumer_stream
*stream
;
2213 DBG("Consumer delete flagged metadata stream");
2218 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2219 /* Validate delete flag of the stream */
2220 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2224 * Remove from pollset so the metadata thread can continue without
2225 * blocking on a deleted stream.
2227 lttng_poll_del(pollset
, stream
->wait_fd
);
2229 /* Delete it right now */
2230 consumer_del_metadata_stream(stream
, metadata_ht
);
2236 * Thread polls on metadata file descriptor and write them on disk or on the
2239 void *consumer_thread_metadata_poll(void *data
)
2241 int ret
, i
, pollfd
, err
= -1;
2242 uint32_t revents
, nb_fd
;
2243 struct lttng_consumer_stream
*stream
= NULL
;
2244 struct lttng_ht_iter iter
;
2245 struct lttng_ht_node_u64
*node
;
2246 struct lttng_poll_event events
;
2247 struct lttng_consumer_local_data
*ctx
= data
;
2250 rcu_register_thread();
2252 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2254 if (testpoint(consumerd_thread_metadata
)) {
2255 goto error_testpoint
;
2258 health_code_update();
2260 DBG("Thread metadata poll started");
2262 /* Size is set to 1 for the consumer_metadata pipe */
2263 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2265 ERR("Poll set creation failed");
2269 ret
= lttng_poll_add(&events
,
2270 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2276 DBG("Metadata main loop started");
2280 health_code_update();
2281 health_poll_entry();
2282 DBG("Metadata poll wait");
2283 ret
= lttng_poll_wait(&events
, -1);
2284 DBG("Metadata poll return from wait with %d fd(s)",
2285 LTTNG_POLL_GETNB(&events
));
2287 DBG("Metadata event caught in thread");
2289 if (errno
== EINTR
) {
2290 ERR("Poll EINTR caught");
2293 if (LTTNG_POLL_GETNB(&events
) == 0) {
2294 err
= 0; /* All is OK */
2301 /* From here, the event is a metadata wait fd */
2302 for (i
= 0; i
< nb_fd
; i
++) {
2303 health_code_update();
2305 revents
= LTTNG_POLL_GETEV(&events
, i
);
2306 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2309 /* No activity for this FD (poll implementation). */
2313 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2314 if (revents
& LPOLLIN
) {
2317 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2318 &stream
, sizeof(stream
));
2319 if (pipe_len
< sizeof(stream
)) {
2321 PERROR("read metadata stream");
2324 * Remove the pipe from the poll set and continue the loop
2325 * since their might be data to consume.
2327 lttng_poll_del(&events
,
2328 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2329 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2333 /* A NULL stream means that the state has changed. */
2334 if (stream
== NULL
) {
2335 /* Check for deleted streams. */
2336 validate_endpoint_status_metadata_stream(&events
);
2340 DBG("Adding metadata stream %d to poll set",
2343 /* Add metadata stream to the global poll events list */
2344 lttng_poll_add(&events
, stream
->wait_fd
,
2345 LPOLLIN
| LPOLLPRI
| LPOLLHUP
);
2346 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2347 DBG("Metadata thread pipe hung up");
2349 * Remove the pipe from the poll set and continue the loop
2350 * since their might be data to consume.
2352 lttng_poll_del(&events
,
2353 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2354 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2357 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2361 /* Handle other stream */
2367 uint64_t tmp_id
= (uint64_t) pollfd
;
2369 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2371 node
= lttng_ht_iter_get_node_u64(&iter
);
2374 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2377 if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2378 /* Get the data out of the metadata file descriptor */
2379 DBG("Metadata available on fd %d", pollfd
);
2380 assert(stream
->wait_fd
== pollfd
);
2383 health_code_update();
2385 len
= ctx
->on_buffer_ready(stream
, ctx
);
2387 * We don't check the return value here since if we get
2388 * a negative len, it means an error occurred thus we
2389 * simply remove it from the poll set and free the
2394 /* It's ok to have an unavailable sub-buffer */
2395 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2396 /* Clean up stream from consumer and free it. */
2397 lttng_poll_del(&events
, stream
->wait_fd
);
2398 consumer_del_metadata_stream(stream
, metadata_ht
);
2400 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2401 DBG("Metadata fd %d is hup|err.", pollfd
);
2402 if (!stream
->hangup_flush_done
2403 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2404 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2405 DBG("Attempting to flush and consume the UST buffers");
2406 lttng_ustconsumer_on_stream_hangup(stream
);
2408 /* We just flushed the stream now read it. */
2410 health_code_update();
2412 len
= ctx
->on_buffer_ready(stream
, ctx
);
2414 * We don't check the return value here since if we get
2415 * a negative len, it means an error occurred thus we
2416 * simply remove it from the poll set and free the
2422 lttng_poll_del(&events
, stream
->wait_fd
);
2424 * This call update the channel states, closes file descriptors
2425 * and securely free the stream.
2427 consumer_del_metadata_stream(stream
, metadata_ht
);
2429 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2433 /* Release RCU lock for the stream looked up */
2441 DBG("Metadata poll thread exiting");
2443 lttng_poll_clean(&events
);
2448 ERR("Health error occurred in %s", __func__
);
2450 health_unregister(health_consumerd
);
2451 rcu_unregister_thread();
2456 * This thread polls the fds in the set to consume the data and write
2457 * it to tracefile if necessary.
2459 void *consumer_thread_data_poll(void *data
)
2461 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2462 struct pollfd
*pollfd
= NULL
;
2463 /* local view of the streams */
2464 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2465 /* local view of consumer_data.fds_count */
2467 /* Number of FDs with CONSUMER_ENDPOINT_INACTIVE but still open. */
2468 int nb_inactive_fd
= 0;
2469 struct lttng_consumer_local_data
*ctx
= data
;
2472 rcu_register_thread();
2474 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2476 if (testpoint(consumerd_thread_data
)) {
2477 goto error_testpoint
;
2480 health_code_update();
2482 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
*));
2483 if (local_stream
== NULL
) {
2484 PERROR("local_stream malloc");
2489 health_code_update();
2495 * the fds set has been updated, we need to update our
2496 * local array as well
2498 pthread_mutex_lock(&consumer_data
.lock
);
2499 if (consumer_data
.need_update
) {
2504 local_stream
= NULL
;
2507 * Allocate for all fds +1 for the consumer_data_pipe and +1 for
2510 pollfd
= zmalloc((consumer_data
.stream_count
+ 2) * sizeof(struct pollfd
));
2511 if (pollfd
== NULL
) {
2512 PERROR("pollfd malloc");
2513 pthread_mutex_unlock(&consumer_data
.lock
);
2517 local_stream
= zmalloc((consumer_data
.stream_count
+ 2) *
2518 sizeof(struct lttng_consumer_stream
*));
2519 if (local_stream
== NULL
) {
2520 PERROR("local_stream malloc");
2521 pthread_mutex_unlock(&consumer_data
.lock
);
2524 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2525 data_ht
, &nb_inactive_fd
);
2527 ERR("Error in allocating pollfd or local_outfds");
2528 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2529 pthread_mutex_unlock(&consumer_data
.lock
);
2533 consumer_data
.need_update
= 0;
2535 pthread_mutex_unlock(&consumer_data
.lock
);
2537 /* No FDs and consumer_quit, consumer_cleanup the thread */
2538 if (nb_fd
== 0 && nb_inactive_fd
== 0 &&
2539 CMM_LOAD_SHARED(consumer_quit
) == 1) {
2540 err
= 0; /* All is OK */
2543 /* poll on the array of fds */
2545 DBG("polling on %d fd", nb_fd
+ 2);
2546 if (testpoint(consumerd_thread_data_poll
)) {
2549 health_poll_entry();
2550 num_rdy
= poll(pollfd
, nb_fd
+ 2, -1);
2552 DBG("poll num_rdy : %d", num_rdy
);
2553 if (num_rdy
== -1) {
2555 * Restart interrupted system call.
2557 if (errno
== EINTR
) {
2560 PERROR("Poll error");
2561 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2563 } else if (num_rdy
== 0) {
2564 DBG("Polling thread timed out");
2568 if (caa_unlikely(data_consumption_paused
)) {
2569 DBG("Data consumption paused, sleeping...");
2575 * If the consumer_data_pipe triggered poll go directly to the
2576 * beginning of the loop to update the array. We want to prioritize
2577 * array update over low-priority reads.
2579 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2580 ssize_t pipe_readlen
;
2582 DBG("consumer_data_pipe wake up");
2583 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2584 &new_stream
, sizeof(new_stream
));
2585 if (pipe_readlen
< sizeof(new_stream
)) {
2586 PERROR("Consumer data pipe");
2587 /* Continue so we can at least handle the current stream(s). */
2592 * If the stream is NULL, just ignore it. It's also possible that
2593 * the sessiond poll thread changed the consumer_quit state and is
2594 * waking us up to test it.
2596 if (new_stream
== NULL
) {
2597 validate_endpoint_status_data_stream();
2601 /* Continue to update the local streams and handle prio ones */
2605 /* Handle wakeup pipe. */
2606 if (pollfd
[nb_fd
+ 1].revents
& (POLLIN
| POLLPRI
)) {
2608 ssize_t pipe_readlen
;
2610 pipe_readlen
= lttng_pipe_read(ctx
->consumer_wakeup_pipe
, &dummy
,
2612 if (pipe_readlen
< 0) {
2613 PERROR("Consumer data wakeup pipe");
2615 /* We've been awakened to handle stream(s). */
2616 ctx
->has_wakeup
= 0;
2619 /* Take care of high priority channels first. */
2620 for (i
= 0; i
< nb_fd
; i
++) {
2621 health_code_update();
2623 if (local_stream
[i
] == NULL
) {
2626 if (pollfd
[i
].revents
& POLLPRI
) {
2627 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2629 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2630 /* it's ok to have an unavailable sub-buffer */
2631 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2632 /* Clean the stream and free it. */
2633 consumer_del_stream(local_stream
[i
], data_ht
);
2634 local_stream
[i
] = NULL
;
2635 } else if (len
> 0) {
2636 local_stream
[i
]->data_read
= 1;
2642 * If we read high prio channel in this loop, try again
2643 * for more high prio data.
2649 /* Take care of low priority channels. */
2650 for (i
= 0; i
< nb_fd
; i
++) {
2651 health_code_update();
2653 if (local_stream
[i
] == NULL
) {
2656 if ((pollfd
[i
].revents
& POLLIN
) ||
2657 local_stream
[i
]->hangup_flush_done
||
2658 local_stream
[i
]->has_data
) {
2659 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2660 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2661 /* it's ok to have an unavailable sub-buffer */
2662 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2663 /* Clean the stream and free it. */
2664 consumer_del_stream(local_stream
[i
], data_ht
);
2665 local_stream
[i
] = NULL
;
2666 } else if (len
> 0) {
2667 local_stream
[i
]->data_read
= 1;
2672 /* Handle hangup and errors */
2673 for (i
= 0; i
< nb_fd
; i
++) {
2674 health_code_update();
2676 if (local_stream
[i
] == NULL
) {
2679 if (!local_stream
[i
]->hangup_flush_done
2680 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2681 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2682 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2683 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2685 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2686 /* Attempt read again, for the data we just flushed. */
2687 local_stream
[i
]->data_read
= 1;
2690 * If the poll flag is HUP/ERR/NVAL and we have
2691 * read no data in this pass, we can remove the
2692 * stream from its hash table.
2694 if ((pollfd
[i
].revents
& POLLHUP
)) {
2695 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2696 if (!local_stream
[i
]->data_read
) {
2697 consumer_del_stream(local_stream
[i
], data_ht
);
2698 local_stream
[i
] = NULL
;
2701 } else if (pollfd
[i
].revents
& POLLERR
) {
2702 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2703 if (!local_stream
[i
]->data_read
) {
2704 consumer_del_stream(local_stream
[i
], data_ht
);
2705 local_stream
[i
] = NULL
;
2708 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2709 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2710 if (!local_stream
[i
]->data_read
) {
2711 consumer_del_stream(local_stream
[i
], data_ht
);
2712 local_stream
[i
] = NULL
;
2716 if (local_stream
[i
] != NULL
) {
2717 local_stream
[i
]->data_read
= 0;
2724 DBG("polling thread exiting");
2729 * Close the write side of the pipe so epoll_wait() in
2730 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2731 * read side of the pipe. If we close them both, epoll_wait strangely does
2732 * not return and could create a endless wait period if the pipe is the
2733 * only tracked fd in the poll set. The thread will take care of closing
2736 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2741 ERR("Health error occurred in %s", __func__
);
2743 health_unregister(health_consumerd
);
2745 rcu_unregister_thread();
2750 * Close wake-up end of each stream belonging to the channel. This will
2751 * allow the poll() on the stream read-side to detect when the
2752 * write-side (application) finally closes them.
2755 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2757 struct lttng_ht
*ht
;
2758 struct lttng_consumer_stream
*stream
;
2759 struct lttng_ht_iter iter
;
2761 ht
= consumer_data
.stream_per_chan_id_ht
;
2764 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2765 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2766 ht
->match_fct
, &channel
->key
,
2767 &iter
.iter
, stream
, node_channel_id
.node
) {
2769 * Protect against teardown with mutex.
2771 pthread_mutex_lock(&stream
->lock
);
2772 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2775 switch (consumer_data
.type
) {
2776 case LTTNG_CONSUMER_KERNEL
:
2778 case LTTNG_CONSUMER32_UST
:
2779 case LTTNG_CONSUMER64_UST
:
2780 if (stream
->metadata_flag
) {
2781 /* Safe and protected by the stream lock. */
2782 lttng_ustconsumer_close_metadata(stream
->chan
);
2785 * Note: a mutex is taken internally within
2786 * liblttng-ust-ctl to protect timer wakeup_fd
2787 * use from concurrent close.
2789 lttng_ustconsumer_close_stream_wakeup(stream
);
2793 ERR("Unknown consumer_data type");
2797 pthread_mutex_unlock(&stream
->lock
);
2802 static void destroy_channel_ht(struct lttng_ht
*ht
)
2804 struct lttng_ht_iter iter
;
2805 struct lttng_consumer_channel
*channel
;
2813 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2814 ret
= lttng_ht_del(ht
, &iter
);
2819 lttng_ht_destroy(ht
);
2823 * This thread polls the channel fds to detect when they are being
2824 * closed. It closes all related streams if the channel is detected as
2825 * closed. It is currently only used as a shim layer for UST because the
2826 * consumerd needs to keep the per-stream wakeup end of pipes open for
2829 void *consumer_thread_channel_poll(void *data
)
2831 int ret
, i
, pollfd
, err
= -1;
2832 uint32_t revents
, nb_fd
;
2833 struct lttng_consumer_channel
*chan
= NULL
;
2834 struct lttng_ht_iter iter
;
2835 struct lttng_ht_node_u64
*node
;
2836 struct lttng_poll_event events
;
2837 struct lttng_consumer_local_data
*ctx
= data
;
2838 struct lttng_ht
*channel_ht
;
2840 rcu_register_thread();
2842 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
2844 if (testpoint(consumerd_thread_channel
)) {
2845 goto error_testpoint
;
2848 health_code_update();
2850 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2852 /* ENOMEM at this point. Better to bail out. */
2856 DBG("Thread channel poll started");
2858 /* Size is set to 1 for the consumer_channel pipe */
2859 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2861 ERR("Poll set creation failed");
2865 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2871 DBG("Channel main loop started");
2875 health_code_update();
2876 DBG("Channel poll wait");
2877 health_poll_entry();
2878 ret
= lttng_poll_wait(&events
, -1);
2879 DBG("Channel poll return from wait with %d fd(s)",
2880 LTTNG_POLL_GETNB(&events
));
2882 DBG("Channel event caught in thread");
2884 if (errno
== EINTR
) {
2885 ERR("Poll EINTR caught");
2888 if (LTTNG_POLL_GETNB(&events
) == 0) {
2889 err
= 0; /* All is OK */
2896 /* From here, the event is a channel wait fd */
2897 for (i
= 0; i
< nb_fd
; i
++) {
2898 health_code_update();
2900 revents
= LTTNG_POLL_GETEV(&events
, i
);
2901 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2904 /* No activity for this FD (poll implementation). */
2908 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2909 if (revents
& LPOLLIN
) {
2910 enum consumer_channel_action action
;
2913 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
2916 ERR("Error reading channel pipe");
2918 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2923 case CONSUMER_CHANNEL_ADD
:
2924 DBG("Adding channel %d to poll set",
2927 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
2930 lttng_ht_add_unique_u64(channel_ht
,
2931 &chan
->wait_fd_node
);
2933 /* Add channel to the global poll events list */
2934 lttng_poll_add(&events
, chan
->wait_fd
,
2935 LPOLLERR
| LPOLLHUP
);
2937 case CONSUMER_CHANNEL_DEL
:
2940 * This command should never be called if the channel
2941 * has streams monitored by either the data or metadata
2942 * thread. The consumer only notify this thread with a
2943 * channel del. command if it receives a destroy
2944 * channel command from the session daemon that send it
2945 * if a command prior to the GET_CHANNEL failed.
2949 chan
= consumer_find_channel(key
);
2952 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
2955 lttng_poll_del(&events
, chan
->wait_fd
);
2956 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
2957 ret
= lttng_ht_del(channel_ht
, &iter
);
2960 switch (consumer_data
.type
) {
2961 case LTTNG_CONSUMER_KERNEL
:
2963 case LTTNG_CONSUMER32_UST
:
2964 case LTTNG_CONSUMER64_UST
:
2965 health_code_update();
2966 /* Destroy streams that might have been left in the stream list. */
2967 clean_channel_stream_list(chan
);
2970 ERR("Unknown consumer_data type");
2975 * Release our own refcount. Force channel deletion even if
2976 * streams were not initialized.
2978 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
2979 consumer_del_channel(chan
);
2984 case CONSUMER_CHANNEL_QUIT
:
2986 * Remove the pipe from the poll set and continue the loop
2987 * since their might be data to consume.
2989 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2992 ERR("Unknown action");
2995 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2996 DBG("Channel thread pipe hung up");
2998 * Remove the pipe from the poll set and continue the loop
2999 * since their might be data to consume.
3001 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
3004 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3008 /* Handle other stream */
3014 uint64_t tmp_id
= (uint64_t) pollfd
;
3016 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
3018 node
= lttng_ht_iter_get_node_u64(&iter
);
3021 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
3024 /* Check for error event */
3025 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
3026 DBG("Channel fd %d is hup|err.", pollfd
);
3028 lttng_poll_del(&events
, chan
->wait_fd
);
3029 ret
= lttng_ht_del(channel_ht
, &iter
);
3033 * This will close the wait fd for each stream associated to
3034 * this channel AND monitored by the data/metadata thread thus
3035 * will be clean by the right thread.
3037 consumer_close_channel_streams(chan
);
3039 /* Release our own refcount */
3040 if (!uatomic_sub_return(&chan
->refcount
, 1)
3041 && !uatomic_read(&chan
->nb_init_stream_left
)) {
3042 consumer_del_channel(chan
);
3045 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3050 /* Release RCU lock for the channel looked up */
3058 lttng_poll_clean(&events
);
3060 destroy_channel_ht(channel_ht
);
3063 DBG("Channel poll thread exiting");
3066 ERR("Health error occurred in %s", __func__
);
3068 health_unregister(health_consumerd
);
3069 rcu_unregister_thread();
3073 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
3074 struct pollfd
*sockpoll
, int client_socket
)
3081 ret
= lttng_consumer_poll_socket(sockpoll
);
3085 DBG("Metadata connection on client_socket");
3087 /* Blocking call, waiting for transmission */
3088 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
3089 if (ctx
->consumer_metadata_socket
< 0) {
3090 WARN("On accept metadata");
3101 * This thread listens on the consumerd socket and receives the file
3102 * descriptors from the session daemon.
3104 void *consumer_thread_sessiond_poll(void *data
)
3106 int sock
= -1, client_socket
, ret
, err
= -1;
3108 * structure to poll for incoming data on communication socket avoids
3109 * making blocking sockets.
3111 struct pollfd consumer_sockpoll
[2];
3112 struct lttng_consumer_local_data
*ctx
= data
;
3114 rcu_register_thread();
3116 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
3118 if (testpoint(consumerd_thread_sessiond
)) {
3119 goto error_testpoint
;
3122 health_code_update();
3124 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
3125 unlink(ctx
->consumer_command_sock_path
);
3126 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
3127 if (client_socket
< 0) {
3128 ERR("Cannot create command socket");
3132 ret
= lttcomm_listen_unix_sock(client_socket
);
3137 DBG("Sending ready command to lttng-sessiond");
3138 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3139 /* return < 0 on error, but == 0 is not fatal */
3141 ERR("Error sending ready command to lttng-sessiond");
3145 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3146 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3147 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3148 consumer_sockpoll
[1].fd
= client_socket
;
3149 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3151 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3159 DBG("Connection on client_socket");
3161 /* Blocking call, waiting for transmission */
3162 sock
= lttcomm_accept_unix_sock(client_socket
);
3169 * Setup metadata socket which is the second socket connection on the
3170 * command unix socket.
3172 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3181 /* This socket is not useful anymore. */
3182 ret
= close(client_socket
);
3184 PERROR("close client_socket");
3188 /* update the polling structure to poll on the established socket */
3189 consumer_sockpoll
[1].fd
= sock
;
3190 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3193 health_code_update();
3195 health_poll_entry();
3196 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3205 DBG("Incoming command on sock");
3206 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3209 * This could simply be a session daemon quitting. Don't output
3212 DBG("Communication interrupted on command socket");
3216 if (CMM_LOAD_SHARED(consumer_quit
)) {
3217 DBG("consumer_thread_receive_fds received quit from signal");
3218 err
= 0; /* All is OK */
3221 DBG("received command on sock");
3227 DBG("Consumer thread sessiond poll exiting");
3230 * Close metadata streams since the producer is the session daemon which
3233 * NOTE: for now, this only applies to the UST tracer.
3235 lttng_consumer_close_all_metadata();
3238 * when all fds have hung up, the polling thread
3241 CMM_STORE_SHARED(consumer_quit
, 1);
3244 * Notify the data poll thread to poll back again and test the
3245 * consumer_quit state that we just set so to quit gracefully.
3247 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3249 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3251 notify_health_quit_pipe(health_quit_pipe
);
3253 /* Cleaning up possibly open sockets. */
3257 PERROR("close sock sessiond poll");
3260 if (client_socket
>= 0) {
3261 ret
= close(client_socket
);
3263 PERROR("close client_socket sessiond poll");
3270 ERR("Health error occurred in %s", __func__
);
3272 health_unregister(health_consumerd
);
3274 rcu_unregister_thread();
3278 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3279 struct lttng_consumer_local_data
*ctx
)
3283 pthread_mutex_lock(&stream
->lock
);
3284 if (stream
->metadata_flag
) {
3285 pthread_mutex_lock(&stream
->metadata_rdv_lock
);
3288 switch (consumer_data
.type
) {
3289 case LTTNG_CONSUMER_KERNEL
:
3290 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
3292 case LTTNG_CONSUMER32_UST
:
3293 case LTTNG_CONSUMER64_UST
:
3294 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
3297 ERR("Unknown consumer_data type");
3303 if (stream
->metadata_flag
) {
3304 pthread_cond_broadcast(&stream
->metadata_rdv
);
3305 pthread_mutex_unlock(&stream
->metadata_rdv_lock
);
3307 pthread_mutex_unlock(&stream
->lock
);
3311 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3313 switch (consumer_data
.type
) {
3314 case LTTNG_CONSUMER_KERNEL
:
3315 return lttng_kconsumer_on_recv_stream(stream
);
3316 case LTTNG_CONSUMER32_UST
:
3317 case LTTNG_CONSUMER64_UST
:
3318 return lttng_ustconsumer_on_recv_stream(stream
);
3320 ERR("Unknown consumer_data type");
3327 * Allocate and set consumer data hash tables.
3329 int lttng_consumer_init(void)
3331 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3332 if (!consumer_data
.channel_ht
) {
3336 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3337 if (!consumer_data
.relayd_ht
) {
3341 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3342 if (!consumer_data
.stream_list_ht
) {
3346 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3347 if (!consumer_data
.stream_per_chan_id_ht
) {
3351 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3356 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3368 * Process the ADD_RELAYD command receive by a consumer.
3370 * This will create a relayd socket pair and add it to the relayd hash table.
3371 * The caller MUST acquire a RCU read side lock before calling it.
3373 void consumer_add_relayd_socket(uint64_t net_seq_idx
, int sock_type
,
3374 struct lttng_consumer_local_data
*ctx
, int sock
,
3375 struct pollfd
*consumer_sockpoll
,
3376 struct lttcomm_relayd_sock
*relayd_sock
, uint64_t sessiond_id
,
3377 uint64_t relayd_session_id
)
3379 int fd
= -1, ret
= -1, relayd_created
= 0;
3380 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3381 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3384 assert(relayd_sock
);
3386 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", net_seq_idx
);
3388 /* Get relayd reference if exists. */
3389 relayd
= consumer_find_relayd(net_seq_idx
);
3390 if (relayd
== NULL
) {
3391 assert(sock_type
== LTTNG_STREAM_CONTROL
);
3392 /* Not found. Allocate one. */
3393 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3394 if (relayd
== NULL
) {
3395 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3398 relayd
->sessiond_session_id
= sessiond_id
;
3403 * This code path MUST continue to the consumer send status message to
3404 * we can notify the session daemon and continue our work without
3405 * killing everything.
3409 * relayd key should never be found for control socket.
3411 assert(sock_type
!= LTTNG_STREAM_CONTROL
);
3414 /* First send a status message before receiving the fds. */
3415 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3417 /* Somehow, the session daemon is not responding anymore. */
3418 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3419 goto error_nosignal
;
3422 /* Poll on consumer socket. */
3423 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3425 /* Needing to exit in the middle of a command: error. */
3426 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3427 goto error_nosignal
;
3430 /* Get relayd socket from session daemon */
3431 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3432 if (ret
!= sizeof(fd
)) {
3433 fd
= -1; /* Just in case it gets set with an invalid value. */
3436 * Failing to receive FDs might indicate a major problem such as
3437 * reaching a fd limit during the receive where the kernel returns a
3438 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3439 * don't take any chances and stop everything.
3441 * XXX: Feature request #558 will fix that and avoid this possible
3442 * issue when reaching the fd limit.
3444 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3445 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3449 /* Copy socket information and received FD */
3450 switch (sock_type
) {
3451 case LTTNG_STREAM_CONTROL
:
3452 /* Copy received lttcomm socket */
3453 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3454 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3455 /* Handle create_sock error. */
3457 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3461 * Close the socket created internally by
3462 * lttcomm_create_sock, so we can replace it by the one
3463 * received from sessiond.
3465 if (close(relayd
->control_sock
.sock
.fd
)) {
3469 /* Assign new file descriptor */
3470 relayd
->control_sock
.sock
.fd
= fd
;
3471 fd
= -1; /* For error path */
3472 /* Assign version values. */
3473 relayd
->control_sock
.major
= relayd_sock
->major
;
3474 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3476 relayd
->relayd_session_id
= relayd_session_id
;
3479 case LTTNG_STREAM_DATA
:
3480 /* Copy received lttcomm socket */
3481 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3482 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3483 /* Handle create_sock error. */
3485 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3489 * Close the socket created internally by
3490 * lttcomm_create_sock, so we can replace it by the one
3491 * received from sessiond.
3493 if (close(relayd
->data_sock
.sock
.fd
)) {
3497 /* Assign new file descriptor */
3498 relayd
->data_sock
.sock
.fd
= fd
;
3499 fd
= -1; /* for eventual error paths */
3500 /* Assign version values. */
3501 relayd
->data_sock
.major
= relayd_sock
->major
;
3502 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3505 ERR("Unknown relayd socket type (%d)", sock_type
);
3506 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3510 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3511 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3512 relayd
->net_seq_idx
, fd
);
3514 /* We successfully added the socket. Send status back. */
3515 ret
= consumer_send_status_msg(sock
, ret_code
);
3517 /* Somehow, the session daemon is not responding anymore. */
3518 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3519 goto error_nosignal
;
3523 * Add relayd socket pair to consumer data hashtable. If object already
3524 * exists or on error, the function gracefully returns.
3532 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3533 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3537 /* Close received socket if valid. */
3540 PERROR("close received socket");
3544 if (relayd_created
) {
3550 * Try to lock the stream mutex.
3552 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
3554 static int stream_try_lock(struct lttng_consumer_stream
*stream
)
3561 * Try to lock the stream mutex. On failure, we know that the stream is
3562 * being used else where hence there is data still being extracted.
3564 ret
= pthread_mutex_trylock(&stream
->lock
);
3566 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
3578 * Search for a relayd associated to the session id and return the reference.
3580 * A rcu read side lock MUST be acquire before calling this function and locked
3581 * until the relayd object is no longer necessary.
3583 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3585 struct lttng_ht_iter iter
;
3586 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3588 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3589 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3592 * Check by sessiond id which is unique here where the relayd session
3593 * id might not be when having multiple relayd.
3595 if (relayd
->sessiond_session_id
== id
) {
3596 /* Found the relayd. There can be only one per id. */
3608 * Check if for a given session id there is still data needed to be extract
3611 * Return 1 if data is pending or else 0 meaning ready to be read.
3613 int consumer_data_pending(uint64_t id
)
3616 struct lttng_ht_iter iter
;
3617 struct lttng_ht
*ht
;
3618 struct lttng_consumer_stream
*stream
;
3619 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3620 int (*data_pending
)(struct lttng_consumer_stream
*);
3622 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3625 pthread_mutex_lock(&consumer_data
.lock
);
3627 switch (consumer_data
.type
) {
3628 case LTTNG_CONSUMER_KERNEL
:
3629 data_pending
= lttng_kconsumer_data_pending
;
3631 case LTTNG_CONSUMER32_UST
:
3632 case LTTNG_CONSUMER64_UST
:
3633 data_pending
= lttng_ustconsumer_data_pending
;
3636 ERR("Unknown consumer data type");
3640 /* Ease our life a bit */
3641 ht
= consumer_data
.stream_list_ht
;
3643 relayd
= find_relayd_by_session_id(id
);
3645 /* Send init command for data pending. */
3646 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3647 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3648 relayd
->relayd_session_id
);
3649 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3651 /* Communication error thus the relayd so no data pending. */
3652 goto data_not_pending
;
3656 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3657 ht
->hash_fct(&id
, lttng_ht_seed
),
3659 &iter
.iter
, stream
, node_session_id
.node
) {
3660 /* If this call fails, the stream is being used hence data pending. */
3661 ret
= stream_try_lock(stream
);
3667 * A removed node from the hash table indicates that the stream has
3668 * been deleted thus having a guarantee that the buffers are closed
3669 * on the consumer side. However, data can still be transmitted
3670 * over the network so don't skip the relayd check.
3672 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3674 /* Check the stream if there is data in the buffers. */
3675 ret
= data_pending(stream
);
3677 pthread_mutex_unlock(&stream
->lock
);
3684 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3685 if (stream
->metadata_flag
) {
3686 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3687 stream
->relayd_stream_id
);
3689 ret
= relayd_data_pending(&relayd
->control_sock
,
3690 stream
->relayd_stream_id
,
3691 stream
->next_net_seq_num
- 1);
3693 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3695 pthread_mutex_unlock(&stream
->lock
);
3699 pthread_mutex_unlock(&stream
->lock
);
3703 unsigned int is_data_inflight
= 0;
3705 /* Send init command for data pending. */
3706 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3707 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3708 relayd
->relayd_session_id
, &is_data_inflight
);
3709 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3711 goto data_not_pending
;
3713 if (is_data_inflight
) {
3719 * Finding _no_ node in the hash table and no inflight data means that the
3720 * stream(s) have been removed thus data is guaranteed to be available for
3721 * analysis from the trace files.
3725 /* Data is available to be read by a viewer. */
3726 pthread_mutex_unlock(&consumer_data
.lock
);
3731 /* Data is still being extracted from buffers. */
3732 pthread_mutex_unlock(&consumer_data
.lock
);
3738 * Send a ret code status message to the sessiond daemon.
3740 * Return the sendmsg() return value.
3742 int consumer_send_status_msg(int sock
, int ret_code
)
3744 struct lttcomm_consumer_status_msg msg
;
3746 memset(&msg
, 0, sizeof(msg
));
3747 msg
.ret_code
= ret_code
;
3749 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3753 * Send a channel status message to the sessiond daemon.
3755 * Return the sendmsg() return value.
3757 int consumer_send_status_channel(int sock
,
3758 struct lttng_consumer_channel
*channel
)
3760 struct lttcomm_consumer_status_channel msg
;
3764 memset(&msg
, 0, sizeof(msg
));
3766 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3768 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3769 msg
.key
= channel
->key
;
3770 msg
.stream_count
= channel
->streams
.count
;
3773 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3776 unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos
,
3777 unsigned long produced_pos
, uint64_t nb_packets_per_stream
,
3778 uint64_t max_sb_size
)
3780 unsigned long start_pos
;
3782 if (!nb_packets_per_stream
) {
3783 return consumed_pos
; /* Grab everything */
3785 start_pos
= produced_pos
- offset_align_floor(produced_pos
, max_sb_size
);
3786 start_pos
-= max_sb_size
* nb_packets_per_stream
;
3787 if ((long) (start_pos
- consumed_pos
) < 0) {
3788 return consumed_pos
; /* Grab everything */