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 <common/common.h>
34 #include <common/utils.h>
35 #include <common/compat/poll.h>
36 #include <common/kernel-ctl/kernel-ctl.h>
37 #include <common/sessiond-comm/relayd.h>
38 #include <common/sessiond-comm/sessiond-comm.h>
39 #include <common/kernel-consumer/kernel-consumer.h>
40 #include <common/relayd/relayd.h>
41 #include <common/ust-consumer/ust-consumer.h>
45 struct lttng_consumer_global_data consumer_data
= {
48 .type
= LTTNG_CONSUMER_UNKNOWN
,
51 enum consumer_channel_action
{
53 CONSUMER_CHANNEL_QUIT
,
56 struct consumer_channel_msg
{
57 enum consumer_channel_action action
;
58 struct lttng_consumer_channel
*chan
;
62 * Flag to inform the polling thread to quit when all fd hung up. Updated by
63 * the consumer_thread_receive_fds when it notices that all fds has hung up.
64 * Also updated by the signal handler (consumer_should_exit()). Read by the
67 volatile int consumer_quit
;
70 * Global hash table containing respectively metadata and data streams. The
71 * stream element in this ht should only be updated by the metadata poll thread
72 * for the metadata and the data poll thread for the data.
74 static struct lttng_ht
*metadata_ht
;
75 static struct lttng_ht
*data_ht
;
78 * Notify a thread pipe to poll back again. This usually means that some global
79 * state has changed so we just send back the thread in a poll wait call.
81 static void notify_thread_pipe(int wpipe
)
86 struct lttng_consumer_stream
*null_stream
= NULL
;
88 ret
= write(wpipe
, &null_stream
, sizeof(null_stream
));
89 } while (ret
< 0 && errno
== EINTR
);
92 static void notify_channel_pipe(struct lttng_consumer_local_data
*ctx
,
93 struct lttng_consumer_channel
*chan
,
94 enum consumer_channel_action action
)
96 struct consumer_channel_msg msg
;
102 ret
= write(ctx
->consumer_channel_pipe
[1], &msg
, sizeof(msg
));
103 } while (ret
< 0 && errno
== EINTR
);
106 static int read_channel_pipe(struct lttng_consumer_local_data
*ctx
,
107 struct lttng_consumer_channel
**chan
,
108 enum consumer_channel_action
*action
)
110 struct consumer_channel_msg msg
;
114 ret
= read(ctx
->consumer_channel_pipe
[0], &msg
, sizeof(msg
));
115 } while (ret
< 0 && errno
== EINTR
);
117 *action
= msg
.action
;
124 * Find a stream. The consumer_data.lock must be locked during this
127 static struct lttng_consumer_stream
*find_stream(uint64_t key
,
130 struct lttng_ht_iter iter
;
131 struct lttng_ht_node_u64
*node
;
132 struct lttng_consumer_stream
*stream
= NULL
;
136 /* -1ULL keys are lookup failures */
137 if (key
== (uint64_t) -1ULL) {
143 lttng_ht_lookup(ht
, &key
, &iter
);
144 node
= lttng_ht_iter_get_node_u64(&iter
);
146 stream
= caa_container_of(node
, struct lttng_consumer_stream
, node
);
154 static void steal_stream_key(int key
, struct lttng_ht
*ht
)
156 struct lttng_consumer_stream
*stream
;
159 stream
= find_stream(key
, ht
);
163 * We don't want the lookup to match, but we still need
164 * to iterate on this stream when iterating over the hash table. Just
165 * change the node key.
167 stream
->node
.key
= -1ULL;
173 * Return a channel object for the given key.
175 * RCU read side lock MUST be acquired before calling this function and
176 * protects the channel ptr.
178 struct lttng_consumer_channel
*consumer_find_channel(uint64_t key
)
180 struct lttng_ht_iter iter
;
181 struct lttng_ht_node_u64
*node
;
182 struct lttng_consumer_channel
*channel
= NULL
;
184 /* -1ULL keys are lookup failures */
185 if (key
== (uint64_t) -1ULL) {
189 lttng_ht_lookup(consumer_data
.channel_ht
, &key
, &iter
);
190 node
= lttng_ht_iter_get_node_u64(&iter
);
192 channel
= caa_container_of(node
, struct lttng_consumer_channel
, node
);
198 static void free_stream_rcu(struct rcu_head
*head
)
200 struct lttng_ht_node_u64
*node
=
201 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
202 struct lttng_consumer_stream
*stream
=
203 caa_container_of(node
, struct lttng_consumer_stream
, node
);
208 static void free_channel_rcu(struct rcu_head
*head
)
210 struct lttng_ht_node_u64
*node
=
211 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
212 struct lttng_consumer_channel
*channel
=
213 caa_container_of(node
, struct lttng_consumer_channel
, node
);
219 * RCU protected relayd socket pair free.
221 static void free_relayd_rcu(struct rcu_head
*head
)
223 struct lttng_ht_node_u64
*node
=
224 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
225 struct consumer_relayd_sock_pair
*relayd
=
226 caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
229 * Close all sockets. This is done in the call RCU since we don't want the
230 * socket fds to be reassigned thus potentially creating bad state of the
233 * We do not have to lock the control socket mutex here since at this stage
234 * there is no one referencing to this relayd object.
236 (void) relayd_close(&relayd
->control_sock
);
237 (void) relayd_close(&relayd
->data_sock
);
243 * Destroy and free relayd socket pair object.
245 * This function MUST be called with the consumer_data lock acquired.
247 static void destroy_relayd(struct consumer_relayd_sock_pair
*relayd
)
250 struct lttng_ht_iter iter
;
252 if (relayd
== NULL
) {
256 DBG("Consumer destroy and close relayd socket pair");
258 iter
.iter
.node
= &relayd
->node
.node
;
259 ret
= lttng_ht_del(consumer_data
.relayd_ht
, &iter
);
261 /* We assume the relayd is being or is destroyed */
265 /* RCU free() call */
266 call_rcu(&relayd
->node
.head
, free_relayd_rcu
);
270 * Remove a channel from the global list protected by a mutex. This function is
271 * also responsible for freeing its data structures.
273 void consumer_del_channel(struct lttng_consumer_channel
*channel
)
276 struct lttng_ht_iter iter
;
278 DBG("Consumer delete channel key %" PRIu64
, channel
->key
);
280 pthread_mutex_lock(&consumer_data
.lock
);
282 switch (consumer_data
.type
) {
283 case LTTNG_CONSUMER_KERNEL
:
285 case LTTNG_CONSUMER32_UST
:
286 case LTTNG_CONSUMER64_UST
:
287 lttng_ustconsumer_del_channel(channel
);
290 ERR("Unknown consumer_data type");
296 iter
.iter
.node
= &channel
->node
.node
;
297 ret
= lttng_ht_del(consumer_data
.channel_ht
, &iter
);
301 call_rcu(&channel
->node
.head
, free_channel_rcu
);
303 pthread_mutex_unlock(&consumer_data
.lock
);
307 * Iterate over the relayd hash table and destroy each element. Finally,
308 * destroy the whole hash table.
310 static void cleanup_relayd_ht(void)
312 struct lttng_ht_iter iter
;
313 struct consumer_relayd_sock_pair
*relayd
;
317 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
319 destroy_relayd(relayd
);
322 lttng_ht_destroy(consumer_data
.relayd_ht
);
328 * Update the end point status of all streams having the given network sequence
329 * index (relayd index).
331 * It's atomically set without having the stream mutex locked which is fine
332 * because we handle the write/read race with a pipe wakeup for each thread.
334 static void update_endpoint_status_by_netidx(int net_seq_idx
,
335 enum consumer_endpoint_status status
)
337 struct lttng_ht_iter iter
;
338 struct lttng_consumer_stream
*stream
;
340 DBG("Consumer set delete flag on stream by idx %d", net_seq_idx
);
344 /* Let's begin with metadata */
345 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
346 if (stream
->net_seq_idx
== net_seq_idx
) {
347 uatomic_set(&stream
->endpoint_status
, status
);
348 DBG("Delete flag set to metadata stream %d", stream
->wait_fd
);
352 /* Follow up by the data streams */
353 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
354 if (stream
->net_seq_idx
== net_seq_idx
) {
355 uatomic_set(&stream
->endpoint_status
, status
);
356 DBG("Delete flag set to data stream %d", stream
->wait_fd
);
363 * Cleanup a relayd object by flagging every associated streams for deletion,
364 * destroying the object meaning removing it from the relayd hash table,
365 * closing the sockets and freeing the memory in a RCU call.
367 * If a local data context is available, notify the threads that the streams'
368 * state have changed.
370 static void cleanup_relayd(struct consumer_relayd_sock_pair
*relayd
,
371 struct lttng_consumer_local_data
*ctx
)
377 DBG("Cleaning up relayd sockets");
379 /* Save the net sequence index before destroying the object */
380 netidx
= relayd
->net_seq_idx
;
383 * Delete the relayd from the relayd hash table, close the sockets and free
384 * the object in a RCU call.
386 destroy_relayd(relayd
);
388 /* Set inactive endpoint to all streams */
389 update_endpoint_status_by_netidx(netidx
, CONSUMER_ENDPOINT_INACTIVE
);
392 * With a local data context, notify the threads that the streams' state
393 * have changed. The write() action on the pipe acts as an "implicit"
394 * memory barrier ordering the updates of the end point status from the
395 * read of this status which happens AFTER receiving this notify.
398 notify_thread_pipe(ctx
->consumer_data_pipe
[1]);
399 notify_thread_pipe(ctx
->consumer_metadata_pipe
[1]);
404 * Flag a relayd socket pair for destruction. Destroy it if the refcount
407 * RCU read side lock MUST be aquired before calling this function.
409 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair
*relayd
)
413 /* Set destroy flag for this object */
414 uatomic_set(&relayd
->destroy_flag
, 1);
416 /* Destroy the relayd if refcount is 0 */
417 if (uatomic_read(&relayd
->refcount
) == 0) {
418 destroy_relayd(relayd
);
423 * Remove a stream from the global list protected by a mutex. This
424 * function is also responsible for freeing its data structures.
426 void consumer_del_stream(struct lttng_consumer_stream
*stream
,
430 struct lttng_ht_iter iter
;
431 struct lttng_consumer_channel
*free_chan
= NULL
;
432 struct consumer_relayd_sock_pair
*relayd
;
436 DBG("Consumer del stream %d", stream
->wait_fd
);
439 /* Means the stream was allocated but not successfully added */
440 goto free_stream_rcu
;
443 pthread_mutex_lock(&consumer_data
.lock
);
444 pthread_mutex_lock(&stream
->lock
);
446 switch (consumer_data
.type
) {
447 case LTTNG_CONSUMER_KERNEL
:
448 if (stream
->mmap_base
!= NULL
) {
449 ret
= munmap(stream
->mmap_base
, stream
->mmap_len
);
455 case LTTNG_CONSUMER32_UST
:
456 case LTTNG_CONSUMER64_UST
:
457 lttng_ustconsumer_del_stream(stream
);
460 ERR("Unknown consumer_data type");
466 iter
.iter
.node
= &stream
->node
.node
;
467 ret
= lttng_ht_del(ht
, &iter
);
470 iter
.iter
.node
= &stream
->node_channel_id
.node
;
471 ret
= lttng_ht_del(consumer_data
.stream_per_chan_id_ht
, &iter
);
474 iter
.iter
.node
= &stream
->node_session_id
.node
;
475 ret
= lttng_ht_del(consumer_data
.stream_list_ht
, &iter
);
479 assert(consumer_data
.stream_count
> 0);
480 consumer_data
.stream_count
--;
482 if (stream
->out_fd
>= 0) {
483 ret
= close(stream
->out_fd
);
489 /* Check and cleanup relayd */
491 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
492 if (relayd
!= NULL
) {
493 uatomic_dec(&relayd
->refcount
);
494 assert(uatomic_read(&relayd
->refcount
) >= 0);
496 /* Closing streams requires to lock the control socket. */
497 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
498 ret
= relayd_send_close_stream(&relayd
->control_sock
,
499 stream
->relayd_stream_id
,
500 stream
->next_net_seq_num
- 1);
501 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
503 DBG("Unable to close stream on the relayd. Continuing");
505 * Continue here. There is nothing we can do for the relayd.
506 * Chances are that the relayd has closed the socket so we just
507 * continue cleaning up.
511 /* Both conditions are met, we destroy the relayd. */
512 if (uatomic_read(&relayd
->refcount
) == 0 &&
513 uatomic_read(&relayd
->destroy_flag
)) {
514 destroy_relayd(relayd
);
519 uatomic_dec(&stream
->chan
->refcount
);
520 if (!uatomic_read(&stream
->chan
->refcount
)
521 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
522 free_chan
= stream
->chan
;
526 consumer_data
.need_update
= 1;
527 pthread_mutex_unlock(&stream
->lock
);
528 pthread_mutex_unlock(&consumer_data
.lock
);
531 consumer_del_channel(free_chan
);
535 call_rcu(&stream
->node
.head
, free_stream_rcu
);
538 struct lttng_consumer_stream
*consumer_allocate_stream(uint64_t channel_key
,
540 enum lttng_consumer_stream_state state
,
541 const char *channel_name
,
548 enum consumer_channel_type type
)
551 struct lttng_consumer_stream
*stream
;
553 stream
= zmalloc(sizeof(*stream
));
554 if (stream
== NULL
) {
555 PERROR("malloc struct lttng_consumer_stream");
562 stream
->key
= stream_key
;
564 stream
->out_fd_offset
= 0;
565 stream
->state
= state
;
568 stream
->net_seq_idx
= relayd_id
;
569 stream
->session_id
= session_id
;
570 pthread_mutex_init(&stream
->lock
, NULL
);
572 /* If channel is the metadata, flag this stream as metadata. */
573 if (type
== CONSUMER_CHANNEL_TYPE_METADATA
) {
574 stream
->metadata_flag
= 1;
575 /* Metadata is flat out. */
576 strncpy(stream
->name
, DEFAULT_METADATA_NAME
, sizeof(stream
->name
));
578 /* Format stream name to <channel_name>_<cpu_number> */
579 ret
= snprintf(stream
->name
, sizeof(stream
->name
), "%s_%d",
582 PERROR("snprintf stream name");
587 /* Key is always the wait_fd for streams. */
588 lttng_ht_node_init_u64(&stream
->node
, stream
->key
);
590 /* Init node per channel id key */
591 lttng_ht_node_init_u64(&stream
->node_channel_id
, channel_key
);
593 /* Init session id node with the stream session id */
594 lttng_ht_node_init_u64(&stream
->node_session_id
, stream
->session_id
);
596 DBG3("Allocated stream %s (key %" PRIu64
", chan_key %" PRIu64
" relayd_id %" PRIu64
", session_id %" PRIu64
,
597 stream
->name
, stream
->key
, channel_key
, stream
->net_seq_idx
, stream
->session_id
);
613 * Add a stream to the global list protected by a mutex.
615 static int add_stream(struct lttng_consumer_stream
*stream
,
619 struct consumer_relayd_sock_pair
*relayd
;
624 DBG3("Adding consumer stream %" PRIu64
, stream
->key
);
626 pthread_mutex_lock(&consumer_data
.lock
);
627 pthread_mutex_lock(&stream
->lock
);
630 /* Steal stream identifier to avoid having streams with the same key */
631 steal_stream_key(stream
->key
, ht
);
633 lttng_ht_add_unique_u64(ht
, &stream
->node
);
635 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
636 &stream
->node_channel_id
);
639 * Add stream to the stream_list_ht of the consumer data. No need to steal
640 * the key since the HT does not use it and we allow to add redundant keys
643 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
645 /* Check and cleanup relayd */
646 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
647 if (relayd
!= NULL
) {
648 uatomic_inc(&relayd
->refcount
);
651 /* Update channel refcount once added without error(s). */
652 uatomic_inc(&stream
->chan
->refcount
);
655 * When nb_init_stream_left reaches 0, we don't need to trigger any action
656 * in terms of destroying the associated channel, because the action that
657 * causes the count to become 0 also causes a stream to be added. The
658 * channel deletion will thus be triggered by the following removal of this
661 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
662 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
665 /* Update consumer data once the node is inserted. */
666 consumer_data
.stream_count
++;
667 consumer_data
.need_update
= 1;
670 pthread_mutex_unlock(&stream
->lock
);
671 pthread_mutex_unlock(&consumer_data
.lock
);
677 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
678 * be acquired before calling this.
680 static int add_relayd(struct consumer_relayd_sock_pair
*relayd
)
683 struct lttng_ht_node_u64
*node
;
684 struct lttng_ht_iter iter
;
688 lttng_ht_lookup(consumer_data
.relayd_ht
,
689 &relayd
->net_seq_idx
, &iter
);
690 node
= lttng_ht_iter_get_node_u64(&iter
);
694 lttng_ht_add_unique_u64(consumer_data
.relayd_ht
, &relayd
->node
);
701 * Allocate and return a consumer relayd socket.
703 struct consumer_relayd_sock_pair
*consumer_allocate_relayd_sock_pair(
706 struct consumer_relayd_sock_pair
*obj
= NULL
;
708 /* Negative net sequence index is a failure */
709 if (net_seq_idx
< 0) {
713 obj
= zmalloc(sizeof(struct consumer_relayd_sock_pair
));
715 PERROR("zmalloc relayd sock");
719 obj
->net_seq_idx
= net_seq_idx
;
721 obj
->destroy_flag
= 0;
722 lttng_ht_node_init_u64(&obj
->node
, obj
->net_seq_idx
);
723 pthread_mutex_init(&obj
->ctrl_sock_mutex
, NULL
);
730 * Find a relayd socket pair in the global consumer data.
732 * Return the object if found else NULL.
733 * RCU read-side lock must be held across this call and while using the
736 struct consumer_relayd_sock_pair
*consumer_find_relayd(uint64_t key
)
738 struct lttng_ht_iter iter
;
739 struct lttng_ht_node_u64
*node
;
740 struct consumer_relayd_sock_pair
*relayd
= NULL
;
742 /* Negative keys are lookup failures */
743 if (key
== (uint64_t) -1ULL) {
747 lttng_ht_lookup(consumer_data
.relayd_ht
, &key
,
749 node
= lttng_ht_iter_get_node_u64(&iter
);
751 relayd
= caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
759 * Handle stream for relayd transmission if the stream applies for network
760 * streaming where the net sequence index is set.
762 * Return destination file descriptor or negative value on error.
764 static int write_relayd_stream_header(struct lttng_consumer_stream
*stream
,
765 size_t data_size
, unsigned long padding
,
766 struct consumer_relayd_sock_pair
*relayd
)
769 struct lttcomm_relayd_data_hdr data_hdr
;
775 /* Reset data header */
776 memset(&data_hdr
, 0, sizeof(data_hdr
));
778 if (stream
->metadata_flag
) {
779 /* Caller MUST acquire the relayd control socket lock */
780 ret
= relayd_send_metadata(&relayd
->control_sock
, data_size
);
785 /* Metadata are always sent on the control socket. */
786 outfd
= relayd
->control_sock
.fd
;
788 /* Set header with stream information */
789 data_hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
790 data_hdr
.data_size
= htobe32(data_size
);
791 data_hdr
.padding_size
= htobe32(padding
);
793 * Note that net_seq_num below is assigned with the *current* value of
794 * next_net_seq_num and only after that the next_net_seq_num will be
795 * increment. This is why when issuing a command on the relayd using
796 * this next value, 1 should always be substracted in order to compare
797 * the last seen sequence number on the relayd side to the last sent.
799 data_hdr
.net_seq_num
= htobe64(stream
->next_net_seq_num
);
800 /* Other fields are zeroed previously */
802 ret
= relayd_send_data_hdr(&relayd
->data_sock
, &data_hdr
,
808 ++stream
->next_net_seq_num
;
810 /* Set to go on data socket */
811 outfd
= relayd
->data_sock
.fd
;
819 * Allocate and return a new lttng_consumer_channel object using the given key
820 * to initialize the hash table node.
822 * On error, return NULL.
824 struct lttng_consumer_channel
*consumer_allocate_channel(uint64_t key
,
826 const char *pathname
,
831 enum lttng_event_output output
,
832 uint64_t tracefile_size
,
833 uint64_t tracefile_count
)
835 struct lttng_consumer_channel
*channel
;
837 channel
= zmalloc(sizeof(*channel
));
838 if (channel
== NULL
) {
839 PERROR("malloc struct lttng_consumer_channel");
844 channel
->refcount
= 0;
845 channel
->session_id
= session_id
;
848 channel
->relayd_id
= relayd_id
;
849 channel
->output
= output
;
850 channel
->tracefile_size
= tracefile_size
;
851 channel
->tracefile_count
= tracefile_count
;
853 strncpy(channel
->pathname
, pathname
, sizeof(channel
->pathname
));
854 channel
->pathname
[sizeof(channel
->pathname
) - 1] = '\0';
856 strncpy(channel
->name
, name
, sizeof(channel
->name
));
857 channel
->name
[sizeof(channel
->name
) - 1] = '\0';
859 lttng_ht_node_init_u64(&channel
->node
, channel
->key
);
861 channel
->wait_fd
= -1;
863 CDS_INIT_LIST_HEAD(&channel
->streams
.head
);
865 DBG("Allocated channel (key %" PRIu64
")", channel
->key
)
872 * Add a channel to the global list protected by a mutex.
874 int consumer_add_channel(struct lttng_consumer_channel
*channel
,
875 struct lttng_consumer_local_data
*ctx
)
878 struct lttng_ht_node_u64
*node
;
879 struct lttng_ht_iter iter
;
881 pthread_mutex_lock(&consumer_data
.lock
);
884 lttng_ht_lookup(consumer_data
.channel_ht
, &channel
->key
, &iter
);
885 node
= lttng_ht_iter_get_node_u64(&iter
);
887 /* Channel already exist. Ignore the insertion */
888 ERR("Consumer add channel key %" PRIu64
" already exists!",
894 lttng_ht_add_unique_u64(consumer_data
.channel_ht
, &channel
->node
);
898 pthread_mutex_unlock(&consumer_data
.lock
);
900 if (!ret
&& channel
->wait_fd
!= -1 &&
901 channel
->metadata_stream
== NULL
) {
902 notify_channel_pipe(ctx
, channel
, CONSUMER_CHANNEL_ADD
);
908 * Allocate the pollfd structure and the local view of the out fds to avoid
909 * doing a lookup in the linked list and concurrency issues when writing is
910 * needed. Called with consumer_data.lock held.
912 * Returns the number of fds in the structures.
914 static int update_poll_array(struct lttng_consumer_local_data
*ctx
,
915 struct pollfd
**pollfd
, struct lttng_consumer_stream
**local_stream
,
919 struct lttng_ht_iter iter
;
920 struct lttng_consumer_stream
*stream
;
925 assert(local_stream
);
927 DBG("Updating poll fd array");
929 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
931 * Only active streams with an active end point can be added to the
932 * poll set and local stream storage of the thread.
934 * There is a potential race here for endpoint_status to be updated
935 * just after the check. However, this is OK since the stream(s) will
936 * be deleted once the thread is notified that the end point state has
937 * changed where this function will be called back again.
939 if (stream
->state
!= LTTNG_CONSUMER_ACTIVE_STREAM
||
940 stream
->endpoint_status
== CONSUMER_ENDPOINT_INACTIVE
) {
944 * This clobbers way too much the debug output. Uncomment that if you
945 * need it for debugging purposes.
947 * DBG("Active FD %d", stream->wait_fd);
949 (*pollfd
)[i
].fd
= stream
->wait_fd
;
950 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
951 local_stream
[i
] = stream
;
957 * Insert the consumer_data_pipe at the end of the array and don't
958 * increment i so nb_fd is the number of real FD.
960 (*pollfd
)[i
].fd
= ctx
->consumer_data_pipe
[0];
961 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
966 * Poll on the should_quit pipe and the command socket return -1 on error and
967 * should exit, 0 if data is available on the command socket
969 int lttng_consumer_poll_socket(struct pollfd
*consumer_sockpoll
)
974 num_rdy
= poll(consumer_sockpoll
, 2, -1);
977 * Restart interrupted system call.
979 if (errno
== EINTR
) {
982 PERROR("Poll error");
985 if (consumer_sockpoll
[0].revents
& (POLLIN
| POLLPRI
)) {
986 DBG("consumer_should_quit wake up");
996 * Set the error socket.
998 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data
*ctx
,
1001 ctx
->consumer_error_socket
= sock
;
1005 * Set the command socket path.
1007 void lttng_consumer_set_command_sock_path(
1008 struct lttng_consumer_local_data
*ctx
, char *sock
)
1010 ctx
->consumer_command_sock_path
= sock
;
1014 * Send return code to the session daemon.
1015 * If the socket is not defined, we return 0, it is not a fatal error
1017 int lttng_consumer_send_error(struct lttng_consumer_local_data
*ctx
, int cmd
)
1019 if (ctx
->consumer_error_socket
> 0) {
1020 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
1021 sizeof(enum lttcomm_sessiond_command
));
1028 * Close all the tracefiles and stream fds and MUST be called when all
1029 * instances are destroyed i.e. when all threads were joined and are ended.
1031 void lttng_consumer_cleanup(void)
1033 struct lttng_ht_iter iter
;
1034 struct lttng_consumer_channel
*channel
;
1038 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
, channel
,
1040 consumer_del_channel(channel
);
1045 lttng_ht_destroy(consumer_data
.channel_ht
);
1047 cleanup_relayd_ht();
1049 lttng_ht_destroy(consumer_data
.stream_per_chan_id_ht
);
1052 * This HT contains streams that are freed by either the metadata thread or
1053 * the data thread so we do *nothing* on the hash table and simply destroy
1056 lttng_ht_destroy(consumer_data
.stream_list_ht
);
1060 * Called from signal handler.
1062 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
1067 ret
= write(ctx
->consumer_should_quit
[1], "4", 1);
1068 } while (ret
< 0 && errno
== EINTR
);
1069 if (ret
< 0 || ret
!= 1) {
1070 PERROR("write consumer quit");
1073 DBG("Consumer flag that it should quit");
1076 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1079 int outfd
= stream
->out_fd
;
1082 * This does a blocking write-and-wait on any page that belongs to the
1083 * subbuffer prior to the one we just wrote.
1084 * Don't care about error values, as these are just hints and ways to
1085 * limit the amount of page cache used.
1087 if (orig_offset
< stream
->max_sb_size
) {
1090 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1091 stream
->max_sb_size
,
1092 SYNC_FILE_RANGE_WAIT_BEFORE
1093 | SYNC_FILE_RANGE_WRITE
1094 | SYNC_FILE_RANGE_WAIT_AFTER
);
1096 * Give hints to the kernel about how we access the file:
1097 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1100 * We need to call fadvise again after the file grows because the
1101 * kernel does not seem to apply fadvise to non-existing parts of the
1104 * Call fadvise _after_ having waited for the page writeback to
1105 * complete because the dirty page writeback semantic is not well
1106 * defined. So it can be expected to lead to lower throughput in
1109 posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1110 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1114 * Initialise the necessary environnement :
1115 * - create a new context
1116 * - create the poll_pipe
1117 * - create the should_quit pipe (for signal handler)
1118 * - create the thread pipe (for splice)
1120 * Takes a function pointer as argument, this function is called when data is
1121 * available on a buffer. This function is responsible to do the
1122 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1123 * buffer configuration and then kernctl_put_next_subbuf at the end.
1125 * Returns a pointer to the new context or NULL on error.
1127 struct lttng_consumer_local_data
*lttng_consumer_create(
1128 enum lttng_consumer_type type
,
1129 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1130 struct lttng_consumer_local_data
*ctx
),
1131 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1132 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1133 int (*update_stream
)(int stream_key
, uint32_t state
))
1136 struct lttng_consumer_local_data
*ctx
;
1138 assert(consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1139 consumer_data
.type
== type
);
1140 consumer_data
.type
= type
;
1142 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
1144 PERROR("allocating context");
1148 ctx
->consumer_error_socket
= -1;
1149 ctx
->consumer_metadata_socket
= -1;
1150 /* assign the callbacks */
1151 ctx
->on_buffer_ready
= buffer_ready
;
1152 ctx
->on_recv_channel
= recv_channel
;
1153 ctx
->on_recv_stream
= recv_stream
;
1154 ctx
->on_update_stream
= update_stream
;
1156 ret
= pipe(ctx
->consumer_data_pipe
);
1158 PERROR("Error creating poll pipe");
1159 goto error_poll_pipe
;
1162 /* set read end of the pipe to non-blocking */
1163 ret
= fcntl(ctx
->consumer_data_pipe
[0], F_SETFL
, O_NONBLOCK
);
1165 PERROR("fcntl O_NONBLOCK");
1166 goto error_poll_fcntl
;
1169 /* set write end of the pipe to non-blocking */
1170 ret
= fcntl(ctx
->consumer_data_pipe
[1], F_SETFL
, O_NONBLOCK
);
1172 PERROR("fcntl O_NONBLOCK");
1173 goto error_poll_fcntl
;
1176 ret
= pipe(ctx
->consumer_should_quit
);
1178 PERROR("Error creating recv pipe");
1179 goto error_quit_pipe
;
1182 ret
= pipe(ctx
->consumer_thread_pipe
);
1184 PERROR("Error creating thread pipe");
1185 goto error_thread_pipe
;
1188 ret
= pipe(ctx
->consumer_channel_pipe
);
1190 PERROR("Error creating channel pipe");
1191 goto error_channel_pipe
;
1194 ret
= utils_create_pipe(ctx
->consumer_metadata_pipe
);
1196 goto error_metadata_pipe
;
1199 ret
= utils_create_pipe(ctx
->consumer_splice_metadata_pipe
);
1201 goto error_splice_pipe
;
1207 utils_close_pipe(ctx
->consumer_metadata_pipe
);
1208 error_metadata_pipe
:
1209 utils_close_pipe(ctx
->consumer_channel_pipe
);
1211 utils_close_pipe(ctx
->consumer_thread_pipe
);
1213 utils_close_pipe(ctx
->consumer_should_quit
);
1216 utils_close_pipe(ctx
->consumer_data_pipe
);
1224 * Close all fds associated with the instance and free the context.
1226 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1230 DBG("Consumer destroying it. Closing everything.");
1232 ret
= close(ctx
->consumer_error_socket
);
1236 ret
= close(ctx
->consumer_metadata_socket
);
1240 utils_close_pipe(ctx
->consumer_thread_pipe
);
1241 utils_close_pipe(ctx
->consumer_channel_pipe
);
1242 utils_close_pipe(ctx
->consumer_data_pipe
);
1243 utils_close_pipe(ctx
->consumer_should_quit
);
1244 utils_close_pipe(ctx
->consumer_splice_metadata_pipe
);
1246 unlink(ctx
->consumer_command_sock_path
);
1251 * Write the metadata stream id on the specified file descriptor.
1253 static int write_relayd_metadata_id(int fd
,
1254 struct lttng_consumer_stream
*stream
,
1255 struct consumer_relayd_sock_pair
*relayd
, unsigned long padding
)
1258 struct lttcomm_relayd_metadata_payload hdr
;
1260 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1261 hdr
.padding_size
= htobe32(padding
);
1263 ret
= write(fd
, (void *) &hdr
, sizeof(hdr
));
1264 } while (ret
< 0 && errno
== EINTR
);
1265 if (ret
< 0 || ret
!= sizeof(hdr
)) {
1267 * This error means that the fd's end is closed so ignore the perror
1268 * not to clubber the error output since this can happen in a normal
1271 if (errno
!= EPIPE
) {
1272 PERROR("write metadata stream id");
1274 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1276 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1277 * handle writting the missing part so report that as an error and
1278 * don't lie to the caller.
1283 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1284 stream
->relayd_stream_id
, padding
);
1291 * Create the tracefile on disk.
1293 * Return 0 on success or else a negative value.
1295 int lttng_create_output_file(struct lttng_consumer_stream
*stream
)
1298 char full_path
[PATH_MAX
];
1299 char *path_name_id
= NULL
;
1303 assert(stream
->net_seq_idx
== (uint64_t) -1ULL);
1305 ret
= snprintf(full_path
, sizeof(full_path
), "%s/%s",
1306 stream
->chan
->pathname
, stream
->name
);
1308 PERROR("snprintf create output file");
1313 * If we split the trace in multiple files, we have to add the tracefile
1314 * current count at the end of the tracefile name
1316 if (stream
->chan
->tracefile_size
> 0) {
1317 ret
= asprintf(&path_name_id
, "%s_%" PRIu64
, full_path
,
1318 stream
->tracefile_count_current
);
1320 PERROR("Allocating path name ID");
1323 path
= path_name_id
;
1328 ret
= run_as_open(path
, O_WRONLY
| O_CREAT
| O_TRUNC
,
1329 S_IRWXU
| S_IRWXG
| S_IRWXO
, stream
->uid
, stream
->gid
);
1331 PERROR("open stream path %s", path
);
1334 stream
->out_fd
= ret
;
1335 stream
->tracefile_size_current
= 0;
1344 * Change the output tracefile according to the tracefile_size and
1345 * tracefile_count parameters. The stream lock MUST be held before calling this
1346 * function because we are modifying the stream status.
1348 * Return 0 on success or else a negative value.
1350 static int rotate_output_file(struct lttng_consumer_stream
*stream
)
1355 assert(stream
->tracefile_size_current
);
1357 ret
= close(stream
->out_fd
);
1359 PERROR("Closing tracefile");
1363 if (stream
->chan
->tracefile_count
> 0) {
1364 stream
->tracefile_count_current
=
1365 (stream
->tracefile_count_current
+ 1) %
1366 stream
->chan
->tracefile_count
;
1368 stream
->tracefile_count_current
++;
1371 return lttng_create_output_file(stream
);
1378 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1379 * core function for writing trace buffers to either the local filesystem or
1382 * It must be called with the stream lock held.
1384 * Careful review MUST be put if any changes occur!
1386 * Returns the number of bytes written
1388 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1389 struct lttng_consumer_local_data
*ctx
,
1390 struct lttng_consumer_stream
*stream
, unsigned long len
,
1391 unsigned long padding
)
1393 unsigned long mmap_offset
;
1395 ssize_t ret
= 0, written
= 0;
1396 off_t orig_offset
= stream
->out_fd_offset
;
1397 /* Default is on the disk */
1398 int outfd
= stream
->out_fd
;
1399 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1400 unsigned int relayd_hang_up
= 0;
1402 /* RCU lock for the relayd pointer */
1405 /* Flag that the current stream if set for network streaming. */
1406 if (stream
->net_seq_idx
!= -1) {
1407 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1408 if (relayd
== NULL
) {
1413 /* get the offset inside the fd to mmap */
1414 switch (consumer_data
.type
) {
1415 case LTTNG_CONSUMER_KERNEL
:
1416 mmap_base
= stream
->mmap_base
;
1417 ret
= kernctl_get_mmap_read_offset(stream
->wait_fd
, &mmap_offset
);
1419 case LTTNG_CONSUMER32_UST
:
1420 case LTTNG_CONSUMER64_UST
:
1421 mmap_base
= lttng_ustctl_get_mmap_base(stream
);
1423 ERR("read mmap get mmap base for stream %s", stream
->name
);
1427 ret
= lttng_ustctl_get_mmap_read_offset(stream
, &mmap_offset
);
1431 ERR("Unknown consumer_data type");
1436 PERROR("tracer ctl get_mmap_read_offset");
1441 /* Handle stream on the relayd if the output is on the network */
1443 unsigned long netlen
= len
;
1446 * Lock the control socket for the complete duration of the function
1447 * since from this point on we will use the socket.
1449 if (stream
->metadata_flag
) {
1450 /* Metadata requires the control socket. */
1451 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1452 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1455 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1457 /* Use the returned socket. */
1460 /* Write metadata stream id before payload */
1461 if (stream
->metadata_flag
) {
1462 ret
= write_relayd_metadata_id(outfd
, stream
, relayd
, padding
);
1465 /* Socket operation failed. We consider the relayd dead */
1466 if (ret
== -EPIPE
|| ret
== -EINVAL
) {
1474 /* Socket operation failed. We consider the relayd dead */
1475 if (ret
== -EPIPE
|| ret
== -EINVAL
) {
1479 /* Else, use the default set before which is the filesystem. */
1482 /* No streaming, we have to set the len with the full padding */
1486 * Check if we need to change the tracefile before writing the packet.
1488 if (stream
->chan
->tracefile_size
> 0 &&
1489 (stream
->tracefile_size_current
+ len
) >
1490 stream
->chan
->tracefile_size
) {
1491 ret
= rotate_output_file(stream
);
1493 ERR("Rotating output file");
1496 outfd
= stream
->out_fd
;
1498 stream
->tracefile_size_current
+= len
;
1503 ret
= write(outfd
, mmap_base
+ mmap_offset
, len
);
1504 } while (ret
< 0 && errno
== EINTR
);
1505 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, len
);
1508 * This is possible if the fd is closed on the other side (outfd)
1509 * or any write problem. It can be verbose a bit for a normal
1510 * execution if for instance the relayd is stopped abruptly. This
1511 * can happen so set this to a DBG statement.
1513 DBG("Error in file write mmap");
1517 /* Socket operation failed. We consider the relayd dead */
1518 if (errno
== EPIPE
|| errno
== EINVAL
) {
1523 } else if (ret
> len
) {
1524 PERROR("Error in file write (ret %zd > len %lu)", ret
, len
);
1532 /* This call is useless on a socket so better save a syscall. */
1534 /* This won't block, but will start writeout asynchronously */
1535 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret
,
1536 SYNC_FILE_RANGE_WRITE
);
1537 stream
->out_fd_offset
+= ret
;
1541 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1545 * This is a special case that the relayd has closed its socket. Let's
1546 * cleanup the relayd object and all associated streams.
1548 if (relayd
&& relayd_hang_up
) {
1549 cleanup_relayd(relayd
, ctx
);
1553 /* Unlock only if ctrl socket used */
1554 if (relayd
&& stream
->metadata_flag
) {
1555 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1563 * Splice the data from the ring buffer to the tracefile.
1565 * It must be called with the stream lock held.
1567 * Returns the number of bytes spliced.
1569 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1570 struct lttng_consumer_local_data
*ctx
,
1571 struct lttng_consumer_stream
*stream
, unsigned long len
,
1572 unsigned long padding
)
1574 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1576 off_t orig_offset
= stream
->out_fd_offset
;
1577 int fd
= stream
->wait_fd
;
1578 /* Default is on the disk */
1579 int outfd
= stream
->out_fd
;
1580 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1582 unsigned int relayd_hang_up
= 0;
1584 switch (consumer_data
.type
) {
1585 case LTTNG_CONSUMER_KERNEL
:
1587 case LTTNG_CONSUMER32_UST
:
1588 case LTTNG_CONSUMER64_UST
:
1589 /* Not supported for user space tracing */
1592 ERR("Unknown consumer_data type");
1596 /* RCU lock for the relayd pointer */
1599 /* Flag that the current stream if set for network streaming. */
1600 if (stream
->net_seq_idx
!= -1) {
1601 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1602 if (relayd
== NULL
) {
1608 * Choose right pipe for splice. Metadata and trace data are handled by
1609 * different threads hence the use of two pipes in order not to race or
1610 * corrupt the written data.
1612 if (stream
->metadata_flag
) {
1613 splice_pipe
= ctx
->consumer_splice_metadata_pipe
;
1615 splice_pipe
= ctx
->consumer_thread_pipe
;
1618 /* Write metadata stream id before payload */
1620 int total_len
= len
;
1622 if (stream
->metadata_flag
) {
1624 * Lock the control socket for the complete duration of the function
1625 * since from this point on we will use the socket.
1627 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1629 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
, relayd
,
1633 /* Socket operation failed. We consider the relayd dead */
1634 if (ret
== -EBADF
) {
1635 WARN("Remote relayd disconnected. Stopping");
1642 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1645 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1647 /* Use the returned socket. */
1650 /* Socket operation failed. We consider the relayd dead */
1651 if (ret
== -EBADF
) {
1652 WARN("Remote relayd disconnected. Stopping");
1659 /* No streaming, we have to set the len with the full padding */
1663 * Check if we need to change the tracefile before writing the packet.
1665 if (stream
->chan
->tracefile_size
> 0 &&
1666 (stream
->tracefile_size_current
+ len
) >
1667 stream
->chan
->tracefile_size
) {
1668 ret
= rotate_output_file(stream
);
1670 ERR("Rotating output file");
1673 outfd
= stream
->out_fd
;
1675 stream
->tracefile_size_current
+= len
;
1679 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1680 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1681 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1682 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1683 DBG("splice chan to pipe, ret %zd", ret_splice
);
1684 if (ret_splice
< 0) {
1685 PERROR("Error in relay splice");
1687 written
= ret_splice
;
1693 /* Handle stream on the relayd if the output is on the network */
1695 if (stream
->metadata_flag
) {
1696 size_t metadata_payload_size
=
1697 sizeof(struct lttcomm_relayd_metadata_payload
);
1699 /* Update counter to fit the spliced data */
1700 ret_splice
+= metadata_payload_size
;
1701 len
+= metadata_payload_size
;
1703 * We do this so the return value can match the len passed as
1704 * argument to this function.
1706 written
-= metadata_payload_size
;
1710 /* Splice data out */
1711 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1712 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1713 DBG("Consumer splice pipe to file, ret %zd", ret_splice
);
1714 if (ret_splice
< 0) {
1715 PERROR("Error in file splice");
1717 written
= ret_splice
;
1719 /* Socket operation failed. We consider the relayd dead */
1720 if (errno
== EBADF
|| errno
== EPIPE
) {
1721 WARN("Remote relayd disconnected. Stopping");
1727 } else if (ret_splice
> len
) {
1729 PERROR("Wrote more data than requested %zd (len: %lu)",
1731 written
+= ret_splice
;
1737 /* This call is useless on a socket so better save a syscall. */
1739 /* This won't block, but will start writeout asynchronously */
1740 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1741 SYNC_FILE_RANGE_WRITE
);
1742 stream
->out_fd_offset
+= ret_splice
;
1744 written
+= ret_splice
;
1746 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1754 * This is a special case that the relayd has closed its socket. Let's
1755 * cleanup the relayd object and all associated streams.
1757 if (relayd
&& relayd_hang_up
) {
1758 cleanup_relayd(relayd
, ctx
);
1759 /* Skip splice error so the consumer does not fail */
1764 /* send the appropriate error description to sessiond */
1767 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1770 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1773 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
1778 if (relayd
&& stream
->metadata_flag
) {
1779 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1787 * Take a snapshot for a specific fd
1789 * Returns 0 on success, < 0 on error
1791 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
1793 switch (consumer_data
.type
) {
1794 case LTTNG_CONSUMER_KERNEL
:
1795 return lttng_kconsumer_take_snapshot(stream
);
1796 case LTTNG_CONSUMER32_UST
:
1797 case LTTNG_CONSUMER64_UST
:
1798 return lttng_ustconsumer_take_snapshot(stream
);
1800 ERR("Unknown consumer_data type");
1807 * Get the produced position
1809 * Returns 0 on success, < 0 on error
1811 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
1814 switch (consumer_data
.type
) {
1815 case LTTNG_CONSUMER_KERNEL
:
1816 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
1817 case LTTNG_CONSUMER32_UST
:
1818 case LTTNG_CONSUMER64_UST
:
1819 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
1821 ERR("Unknown consumer_data type");
1827 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
1828 int sock
, struct pollfd
*consumer_sockpoll
)
1830 switch (consumer_data
.type
) {
1831 case LTTNG_CONSUMER_KERNEL
:
1832 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1833 case LTTNG_CONSUMER32_UST
:
1834 case LTTNG_CONSUMER64_UST
:
1835 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1837 ERR("Unknown consumer_data type");
1844 * Iterate over all streams of the hashtable and free them properly.
1846 * WARNING: *MUST* be used with data stream only.
1848 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1850 struct lttng_ht_iter iter
;
1851 struct lttng_consumer_stream
*stream
;
1858 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1860 * Ignore return value since we are currently cleaning up so any error
1863 (void) consumer_del_stream(stream
, ht
);
1867 lttng_ht_destroy(ht
);
1871 * Iterate over all streams of the hashtable and free them properly.
1873 * XXX: Should not be only for metadata stream or else use an other name.
1875 static void destroy_stream_ht(struct lttng_ht
*ht
)
1877 struct lttng_ht_iter iter
;
1878 struct lttng_consumer_stream
*stream
;
1885 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1887 * Ignore return value since we are currently cleaning up so any error
1890 (void) consumer_del_metadata_stream(stream
, ht
);
1894 lttng_ht_destroy(ht
);
1897 void lttng_consumer_close_metadata(void)
1899 switch (consumer_data
.type
) {
1900 case LTTNG_CONSUMER_KERNEL
:
1902 * The Kernel consumer has a different metadata scheme so we don't
1903 * close anything because the stream will be closed by the session
1907 case LTTNG_CONSUMER32_UST
:
1908 case LTTNG_CONSUMER64_UST
:
1910 * Close all metadata streams. The metadata hash table is passed and
1911 * this call iterates over it by closing all wakeup fd. This is safe
1912 * because at this point we are sure that the metadata producer is
1913 * either dead or blocked.
1915 lttng_ustconsumer_close_metadata(metadata_ht
);
1918 ERR("Unknown consumer_data type");
1924 * Clean up a metadata stream and free its memory.
1926 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
1927 struct lttng_ht
*ht
)
1930 struct lttng_ht_iter iter
;
1931 struct lttng_consumer_channel
*free_chan
= NULL
;
1932 struct consumer_relayd_sock_pair
*relayd
;
1936 * This call should NEVER receive regular stream. It must always be
1937 * metadata stream and this is crucial for data structure synchronization.
1939 assert(stream
->metadata_flag
);
1941 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
1944 /* Means the stream was allocated but not successfully added */
1945 goto free_stream_rcu
;
1948 pthread_mutex_lock(&consumer_data
.lock
);
1949 pthread_mutex_lock(&stream
->lock
);
1951 switch (consumer_data
.type
) {
1952 case LTTNG_CONSUMER_KERNEL
:
1953 if (stream
->mmap_base
!= NULL
) {
1954 ret
= munmap(stream
->mmap_base
, stream
->mmap_len
);
1956 PERROR("munmap metadata stream");
1960 case LTTNG_CONSUMER32_UST
:
1961 case LTTNG_CONSUMER64_UST
:
1962 lttng_ustconsumer_del_stream(stream
);
1965 ERR("Unknown consumer_data type");
1971 iter
.iter
.node
= &stream
->node
.node
;
1972 ret
= lttng_ht_del(ht
, &iter
);
1975 iter
.iter
.node
= &stream
->node_channel_id
.node
;
1976 ret
= lttng_ht_del(consumer_data
.stream_per_chan_id_ht
, &iter
);
1979 iter
.iter
.node
= &stream
->node_session_id
.node
;
1980 ret
= lttng_ht_del(consumer_data
.stream_list_ht
, &iter
);
1984 if (stream
->out_fd
>= 0) {
1985 ret
= close(stream
->out_fd
);
1991 /* Check and cleanup relayd */
1993 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1994 if (relayd
!= NULL
) {
1995 uatomic_dec(&relayd
->refcount
);
1996 assert(uatomic_read(&relayd
->refcount
) >= 0);
1998 /* Closing streams requires to lock the control socket. */
1999 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
2000 ret
= relayd_send_close_stream(&relayd
->control_sock
,
2001 stream
->relayd_stream_id
, stream
->next_net_seq_num
- 1);
2002 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
2004 DBG("Unable to close stream on the relayd. Continuing");
2006 * Continue here. There is nothing we can do for the relayd.
2007 * Chances are that the relayd has closed the socket so we just
2008 * continue cleaning up.
2012 /* Both conditions are met, we destroy the relayd. */
2013 if (uatomic_read(&relayd
->refcount
) == 0 &&
2014 uatomic_read(&relayd
->destroy_flag
)) {
2015 destroy_relayd(relayd
);
2020 /* Atomically decrement channel refcount since other threads can use it. */
2021 uatomic_dec(&stream
->chan
->refcount
);
2022 if (!uatomic_read(&stream
->chan
->refcount
)
2023 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
2024 /* Go for channel deletion! */
2025 free_chan
= stream
->chan
;
2029 pthread_mutex_unlock(&stream
->lock
);
2030 pthread_mutex_unlock(&consumer_data
.lock
);
2033 consumer_del_channel(free_chan
);
2037 call_rcu(&stream
->node
.head
, free_stream_rcu
);
2041 * Action done with the metadata stream when adding it to the consumer internal
2042 * data structures to handle it.
2044 static int add_metadata_stream(struct lttng_consumer_stream
*stream
,
2045 struct lttng_ht
*ht
)
2048 struct consumer_relayd_sock_pair
*relayd
;
2049 struct lttng_ht_iter iter
;
2050 struct lttng_ht_node_u64
*node
;
2055 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2057 pthread_mutex_lock(&consumer_data
.lock
);
2058 pthread_mutex_lock(&stream
->lock
);
2061 * From here, refcounts are updated so be _careful_ when returning an error
2068 * Lookup the stream just to make sure it does not exist in our internal
2069 * state. This should NEVER happen.
2071 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2072 node
= lttng_ht_iter_get_node_u64(&iter
);
2075 /* Find relayd and, if one is found, increment refcount. */
2076 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
2077 if (relayd
!= NULL
) {
2078 uatomic_inc(&relayd
->refcount
);
2081 /* Update channel refcount once added without error(s). */
2082 uatomic_inc(&stream
->chan
->refcount
);
2085 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2086 * in terms of destroying the associated channel, because the action that
2087 * causes the count to become 0 also causes a stream to be added. The
2088 * channel deletion will thus be triggered by the following removal of this
2091 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2092 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2095 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2097 lttng_ht_add_unique_u64(consumer_data
.stream_per_chan_id_ht
,
2098 &stream
->node_channel_id
);
2101 * Add stream to the stream_list_ht of the consumer data. No need to steal
2102 * the key since the HT does not use it and we allow to add redundant keys
2105 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2109 pthread_mutex_unlock(&stream
->lock
);
2110 pthread_mutex_unlock(&consumer_data
.lock
);
2115 * Delete data stream that are flagged for deletion (endpoint_status).
2117 static void validate_endpoint_status_data_stream(void)
2119 struct lttng_ht_iter iter
;
2120 struct lttng_consumer_stream
*stream
;
2122 DBG("Consumer delete flagged data stream");
2125 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2126 /* Validate delete flag of the stream */
2127 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2130 /* Delete it right now */
2131 consumer_del_stream(stream
, data_ht
);
2137 * Delete metadata stream that are flagged for deletion (endpoint_status).
2139 static void validate_endpoint_status_metadata_stream(
2140 struct lttng_poll_event
*pollset
)
2142 struct lttng_ht_iter iter
;
2143 struct lttng_consumer_stream
*stream
;
2145 DBG("Consumer delete flagged metadata stream");
2150 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2151 /* Validate delete flag of the stream */
2152 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2156 * Remove from pollset so the metadata thread can continue without
2157 * blocking on a deleted stream.
2159 lttng_poll_del(pollset
, stream
->wait_fd
);
2161 /* Delete it right now */
2162 consumer_del_metadata_stream(stream
, metadata_ht
);
2168 * Thread polls on metadata file descriptor and write them on disk or on the
2171 void *consumer_thread_metadata_poll(void *data
)
2174 uint32_t revents
, nb_fd
;
2175 struct lttng_consumer_stream
*stream
= NULL
;
2176 struct lttng_ht_iter iter
;
2177 struct lttng_ht_node_u64
*node
;
2178 struct lttng_poll_event events
;
2179 struct lttng_consumer_local_data
*ctx
= data
;
2182 rcu_register_thread();
2184 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2186 /* ENOMEM at this point. Better to bail out. */
2190 DBG("Thread metadata poll started");
2192 /* Size is set to 1 for the consumer_metadata pipe */
2193 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2195 ERR("Poll set creation failed");
2199 ret
= lttng_poll_add(&events
, ctx
->consumer_metadata_pipe
[0], LPOLLIN
);
2205 DBG("Metadata main loop started");
2208 /* Only the metadata pipe is set */
2209 if (LTTNG_POLL_GETNB(&events
) == 0 && consumer_quit
== 1) {
2214 DBG("Metadata poll wait with %d fd(s)", LTTNG_POLL_GETNB(&events
));
2215 ret
= lttng_poll_wait(&events
, -1);
2216 DBG("Metadata event catched in thread");
2218 if (errno
== EINTR
) {
2219 ERR("Poll EINTR catched");
2227 /* From here, the event is a metadata wait fd */
2228 for (i
= 0; i
< nb_fd
; i
++) {
2229 revents
= LTTNG_POLL_GETEV(&events
, i
);
2230 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2232 /* Just don't waste time if no returned events for the fd */
2237 if (pollfd
== ctx
->consumer_metadata_pipe
[0]) {
2238 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2239 DBG("Metadata thread pipe hung up");
2241 * Remove the pipe from the poll set and continue the loop
2242 * since their might be data to consume.
2244 lttng_poll_del(&events
, ctx
->consumer_metadata_pipe
[0]);
2245 ret
= close(ctx
->consumer_metadata_pipe
[0]);
2247 PERROR("close metadata pipe");
2250 } else if (revents
& LPOLLIN
) {
2252 /* Get the stream pointer received */
2253 ret
= read(pollfd
, &stream
, sizeof(stream
));
2254 } while (ret
< 0 && errno
== EINTR
);
2256 ret
< sizeof(struct lttng_consumer_stream
*)) {
2257 PERROR("read metadata stream");
2259 * Let's continue here and hope we can still work
2260 * without stopping the consumer. XXX: Should we?
2265 /* A NULL stream means that the state has changed. */
2266 if (stream
== NULL
) {
2267 /* Check for deleted streams. */
2268 validate_endpoint_status_metadata_stream(&events
);
2272 DBG("Adding metadata stream %d to poll set",
2275 ret
= add_metadata_stream(stream
, metadata_ht
);
2277 ERR("Unable to add metadata stream");
2278 /* Stream was not setup properly. Continuing. */
2279 consumer_del_metadata_stream(stream
, NULL
);
2283 /* Add metadata stream to the global poll events list */
2284 lttng_poll_add(&events
, stream
->wait_fd
,
2285 LPOLLIN
| LPOLLPRI
);
2288 /* Handle other stream */
2294 uint64_t tmp_id
= (uint64_t) pollfd
;
2296 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2298 node
= lttng_ht_iter_get_node_u64(&iter
);
2301 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2304 /* Check for error event */
2305 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2306 DBG("Metadata fd %d is hup|err.", pollfd
);
2307 if (!stream
->hangup_flush_done
2308 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2309 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2310 DBG("Attempting to flush and consume the UST buffers");
2311 lttng_ustconsumer_on_stream_hangup(stream
);
2313 /* We just flushed the stream now read it. */
2315 len
= ctx
->on_buffer_ready(stream
, ctx
);
2317 * We don't check the return value here since if we get
2318 * a negative len, it means an error occured thus we
2319 * simply remove it from the poll set and free the
2325 lttng_poll_del(&events
, stream
->wait_fd
);
2327 * This call update the channel states, closes file descriptors
2328 * and securely free the stream.
2330 consumer_del_metadata_stream(stream
, metadata_ht
);
2331 } else if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2332 /* Get the data out of the metadata file descriptor */
2333 DBG("Metadata available on fd %d", pollfd
);
2334 assert(stream
->wait_fd
== pollfd
);
2336 len
= ctx
->on_buffer_ready(stream
, ctx
);
2337 /* It's ok to have an unavailable sub-buffer */
2338 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2339 /* Clean up stream from consumer and free it. */
2340 lttng_poll_del(&events
, stream
->wait_fd
);
2341 consumer_del_metadata_stream(stream
, metadata_ht
);
2342 } else if (len
> 0) {
2343 stream
->data_read
= 1;
2347 /* Release RCU lock for the stream looked up */
2354 DBG("Metadata poll thread exiting");
2356 lttng_poll_clean(&events
);
2358 destroy_stream_ht(metadata_ht
);
2360 rcu_unregister_thread();
2365 * This thread polls the fds in the set to consume the data and write
2366 * it to tracefile if necessary.
2368 void *consumer_thread_data_poll(void *data
)
2370 int num_rdy
, num_hup
, high_prio
, ret
, i
;
2371 struct pollfd
*pollfd
= NULL
;
2372 /* local view of the streams */
2373 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2374 /* local view of consumer_data.fds_count */
2376 struct lttng_consumer_local_data
*ctx
= data
;
2379 rcu_register_thread();
2381 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2382 if (data_ht
== NULL
) {
2383 /* ENOMEM at this point. Better to bail out. */
2387 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
));
2394 * the fds set has been updated, we need to update our
2395 * local array as well
2397 pthread_mutex_lock(&consumer_data
.lock
);
2398 if (consumer_data
.need_update
) {
2403 local_stream
= NULL
;
2405 /* allocate for all fds + 1 for the consumer_data_pipe */
2406 pollfd
= zmalloc((consumer_data
.stream_count
+ 1) * sizeof(struct pollfd
));
2407 if (pollfd
== NULL
) {
2408 PERROR("pollfd malloc");
2409 pthread_mutex_unlock(&consumer_data
.lock
);
2413 /* allocate for all fds + 1 for the consumer_data_pipe */
2414 local_stream
= zmalloc((consumer_data
.stream_count
+ 1) *
2415 sizeof(struct lttng_consumer_stream
));
2416 if (local_stream
== NULL
) {
2417 PERROR("local_stream malloc");
2418 pthread_mutex_unlock(&consumer_data
.lock
);
2421 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2424 ERR("Error in allocating pollfd or local_outfds");
2425 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2426 pthread_mutex_unlock(&consumer_data
.lock
);
2430 consumer_data
.need_update
= 0;
2432 pthread_mutex_unlock(&consumer_data
.lock
);
2434 /* No FDs and consumer_quit, consumer_cleanup the thread */
2435 if (nb_fd
== 0 && consumer_quit
== 1) {
2438 /* poll on the array of fds */
2440 DBG("polling on %d fd", nb_fd
+ 1);
2441 num_rdy
= poll(pollfd
, nb_fd
+ 1, -1);
2442 DBG("poll num_rdy : %d", num_rdy
);
2443 if (num_rdy
== -1) {
2445 * Restart interrupted system call.
2447 if (errno
== EINTR
) {
2450 PERROR("Poll error");
2451 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2453 } else if (num_rdy
== 0) {
2454 DBG("Polling thread timed out");
2459 * If the consumer_data_pipe triggered poll go directly to the
2460 * beginning of the loop to update the array. We want to prioritize
2461 * array update over low-priority reads.
2463 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2464 ssize_t pipe_readlen
;
2466 DBG("consumer_data_pipe wake up");
2467 /* Consume 1 byte of pipe data */
2469 pipe_readlen
= read(ctx
->consumer_data_pipe
[0], &new_stream
,
2470 sizeof(new_stream
));
2471 } while (pipe_readlen
== -1 && errno
== EINTR
);
2472 if (pipe_readlen
< 0) {
2473 PERROR("read consumer data pipe");
2474 /* Continue so we can at least handle the current stream(s). */
2479 * If the stream is NULL, just ignore it. It's also possible that
2480 * the sessiond poll thread changed the consumer_quit state and is
2481 * waking us up to test it.
2483 if (new_stream
== NULL
) {
2484 validate_endpoint_status_data_stream();
2488 ret
= add_stream(new_stream
, data_ht
);
2490 ERR("Consumer add stream %" PRIu64
" failed. Continuing",
2493 * At this point, if the add_stream fails, it is not in the
2494 * hash table thus passing the NULL value here.
2496 consumer_del_stream(new_stream
, NULL
);
2499 /* Continue to update the local streams and handle prio ones */
2503 /* Take care of high priority channels first. */
2504 for (i
= 0; i
< nb_fd
; i
++) {
2505 if (local_stream
[i
] == NULL
) {
2508 if (pollfd
[i
].revents
& POLLPRI
) {
2509 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2511 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2512 /* it's ok to have an unavailable sub-buffer */
2513 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2514 /* Clean the stream and free it. */
2515 consumer_del_stream(local_stream
[i
], data_ht
);
2516 local_stream
[i
] = NULL
;
2517 } else if (len
> 0) {
2518 local_stream
[i
]->data_read
= 1;
2524 * If we read high prio channel in this loop, try again
2525 * for more high prio data.
2531 /* Take care of low priority channels. */
2532 for (i
= 0; i
< nb_fd
; i
++) {
2533 if (local_stream
[i
] == NULL
) {
2536 if ((pollfd
[i
].revents
& POLLIN
) ||
2537 local_stream
[i
]->hangup_flush_done
) {
2538 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2539 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2540 /* it's ok to have an unavailable sub-buffer */
2541 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2542 /* Clean the stream and free it. */
2543 consumer_del_stream(local_stream
[i
], data_ht
);
2544 local_stream
[i
] = NULL
;
2545 } else if (len
> 0) {
2546 local_stream
[i
]->data_read
= 1;
2551 /* Handle hangup and errors */
2552 for (i
= 0; i
< nb_fd
; i
++) {
2553 if (local_stream
[i
] == NULL
) {
2556 if (!local_stream
[i
]->hangup_flush_done
2557 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2558 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2559 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2560 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2562 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2563 /* Attempt read again, for the data we just flushed. */
2564 local_stream
[i
]->data_read
= 1;
2567 * If the poll flag is HUP/ERR/NVAL and we have
2568 * read no data in this pass, we can remove the
2569 * stream from its hash table.
2571 if ((pollfd
[i
].revents
& POLLHUP
)) {
2572 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2573 if (!local_stream
[i
]->data_read
) {
2574 consumer_del_stream(local_stream
[i
], data_ht
);
2575 local_stream
[i
] = NULL
;
2578 } else if (pollfd
[i
].revents
& POLLERR
) {
2579 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2580 if (!local_stream
[i
]->data_read
) {
2581 consumer_del_stream(local_stream
[i
], data_ht
);
2582 local_stream
[i
] = NULL
;
2585 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2586 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2587 if (!local_stream
[i
]->data_read
) {
2588 consumer_del_stream(local_stream
[i
], data_ht
);
2589 local_stream
[i
] = NULL
;
2593 if (local_stream
[i
] != NULL
) {
2594 local_stream
[i
]->data_read
= 0;
2599 DBG("polling thread exiting");
2604 * Close the write side of the pipe so epoll_wait() in
2605 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2606 * read side of the pipe. If we close them both, epoll_wait strangely does
2607 * not return and could create a endless wait period if the pipe is the
2608 * only tracked fd in the poll set. The thread will take care of closing
2611 ret
= close(ctx
->consumer_metadata_pipe
[1]);
2613 PERROR("close data pipe");
2616 destroy_data_stream_ht(data_ht
);
2618 rcu_unregister_thread();
2623 * Close wake-up end of each stream belonging to the channel. This will
2624 * allow the poll() on the stream read-side to detect when the
2625 * write-side (application) finally closes them.
2628 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2630 struct lttng_ht
*ht
;
2631 struct lttng_consumer_stream
*stream
;
2632 struct lttng_ht_iter iter
;
2634 ht
= consumer_data
.stream_per_chan_id_ht
;
2637 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2638 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2639 ht
->match_fct
, &channel
->key
,
2640 &iter
.iter
, stream
, node_channel_id
.node
) {
2641 switch (consumer_data
.type
) {
2642 case LTTNG_CONSUMER_KERNEL
:
2644 case LTTNG_CONSUMER32_UST
:
2645 case LTTNG_CONSUMER64_UST
:
2647 * Note: a mutex is taken internally within
2648 * liblttng-ust-ctl to protect timer wakeup_fd
2649 * use from concurrent close.
2651 lttng_ustconsumer_close_stream_wakeup(stream
);
2654 ERR("Unknown consumer_data type");
2661 static void destroy_channel_ht(struct lttng_ht
*ht
)
2663 struct lttng_ht_iter iter
;
2664 struct lttng_consumer_channel
*channel
;
2672 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2673 ret
= lttng_ht_del(ht
, &iter
);
2678 lttng_ht_destroy(ht
);
2682 * This thread polls the channel fds to detect when they are being
2683 * closed. It closes all related streams if the channel is detected as
2684 * closed. It is currently only used as a shim layer for UST because the
2685 * consumerd needs to keep the per-stream wakeup end of pipes open for
2688 void *consumer_thread_channel_poll(void *data
)
2691 uint32_t revents
, nb_fd
;
2692 struct lttng_consumer_channel
*chan
= NULL
;
2693 struct lttng_ht_iter iter
;
2694 struct lttng_ht_node_u64
*node
;
2695 struct lttng_poll_event events
;
2696 struct lttng_consumer_local_data
*ctx
= data
;
2697 struct lttng_ht
*channel_ht
;
2699 rcu_register_thread();
2701 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2703 /* ENOMEM at this point. Better to bail out. */
2707 DBG("Thread channel poll started");
2709 /* Size is set to 1 for the consumer_channel pipe */
2710 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2712 ERR("Poll set creation failed");
2716 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2722 DBG("Channel main loop started");
2725 /* Only the channel pipe is set */
2726 if (LTTNG_POLL_GETNB(&events
) == 0 && consumer_quit
== 1) {
2731 DBG("Channel poll wait with %d fd(s)", LTTNG_POLL_GETNB(&events
));
2732 ret
= lttng_poll_wait(&events
, -1);
2733 DBG("Channel event catched in thread");
2735 if (errno
== EINTR
) {
2736 ERR("Poll EINTR catched");
2744 /* From here, the event is a channel wait fd */
2745 for (i
= 0; i
< nb_fd
; i
++) {
2746 revents
= LTTNG_POLL_GETEV(&events
, i
);
2747 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2749 /* Just don't waste time if no returned events for the fd */
2753 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2754 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2755 DBG("Channel thread pipe hung up");
2757 * Remove the pipe from the poll set and continue the loop
2758 * since their might be data to consume.
2760 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2762 } else if (revents
& LPOLLIN
) {
2763 enum consumer_channel_action action
;
2765 ret
= read_channel_pipe(ctx
, &chan
, &action
);
2767 ERR("Error reading channel pipe");
2772 case CONSUMER_CHANNEL_ADD
:
2773 DBG("Adding channel %d to poll set",
2776 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
2778 lttng_ht_add_unique_u64(channel_ht
,
2779 &chan
->wait_fd_node
);
2780 /* Add channel to the global poll events list */
2781 lttng_poll_add(&events
, chan
->wait_fd
,
2782 LPOLLIN
| LPOLLPRI
);
2784 case CONSUMER_CHANNEL_QUIT
:
2786 * Remove the pipe from the poll set and continue the loop
2787 * since their might be data to consume.
2789 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2792 ERR("Unknown action");
2797 /* Handle other stream */
2803 uint64_t tmp_id
= (uint64_t) pollfd
;
2805 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
2807 node
= lttng_ht_iter_get_node_u64(&iter
);
2810 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
2813 /* Check for error event */
2814 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2815 DBG("Channel fd %d is hup|err.", pollfd
);
2817 lttng_poll_del(&events
, chan
->wait_fd
);
2818 ret
= lttng_ht_del(channel_ht
, &iter
);
2820 consumer_close_channel_streams(chan
);
2823 /* Release RCU lock for the channel looked up */
2829 lttng_poll_clean(&events
);
2831 destroy_channel_ht(channel_ht
);
2833 DBG("Channel poll thread exiting");
2834 rcu_unregister_thread();
2838 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
2839 struct pollfd
*sockpoll
, int client_socket
)
2846 if (lttng_consumer_poll_socket(sockpoll
) < 0) {
2850 DBG("Metadata connection on client_socket");
2852 /* Blocking call, waiting for transmission */
2853 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
2854 if (ctx
->consumer_metadata_socket
< 0) {
2855 WARN("On accept metadata");
2866 * This thread listens on the consumerd socket and receives the file
2867 * descriptors from the session daemon.
2869 void *consumer_thread_sessiond_poll(void *data
)
2871 int sock
= -1, client_socket
, ret
;
2873 * structure to poll for incoming data on communication socket avoids
2874 * making blocking sockets.
2876 struct pollfd consumer_sockpoll
[2];
2877 struct lttng_consumer_local_data
*ctx
= data
;
2879 rcu_register_thread();
2881 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
2882 unlink(ctx
->consumer_command_sock_path
);
2883 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
2884 if (client_socket
< 0) {
2885 ERR("Cannot create command socket");
2889 ret
= lttcomm_listen_unix_sock(client_socket
);
2894 DBG("Sending ready command to lttng-sessiond");
2895 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
2896 /* return < 0 on error, but == 0 is not fatal */
2898 ERR("Error sending ready command to lttng-sessiond");
2902 ret
= fcntl(client_socket
, F_SETFL
, O_NONBLOCK
);
2904 PERROR("fcntl O_NONBLOCK");
2908 /* prepare the FDs to poll : to client socket and the should_quit pipe */
2909 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
2910 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
2911 consumer_sockpoll
[1].fd
= client_socket
;
2912 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
2914 if (lttng_consumer_poll_socket(consumer_sockpoll
) < 0) {
2917 DBG("Connection on client_socket");
2919 /* Blocking call, waiting for transmission */
2920 sock
= lttcomm_accept_unix_sock(client_socket
);
2925 ret
= fcntl(sock
, F_SETFL
, O_NONBLOCK
);
2927 PERROR("fcntl O_NONBLOCK");
2932 * Setup metadata socket which is the second socket connection on the
2933 * command unix socket.
2935 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
2940 /* This socket is not useful anymore. */
2941 ret
= close(client_socket
);
2943 PERROR("close client_socket");
2947 /* update the polling structure to poll on the established socket */
2948 consumer_sockpoll
[1].fd
= sock
;
2949 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
2952 if (lttng_consumer_poll_socket(consumer_sockpoll
) < 0) {
2955 DBG("Incoming command on sock");
2956 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2957 if (ret
== -ENOENT
) {
2958 DBG("Received STOP command");
2963 * This could simply be a session daemon quitting. Don't output
2966 DBG("Communication interrupted on command socket");
2969 if (consumer_quit
) {
2970 DBG("consumer_thread_receive_fds received quit from signal");
2973 DBG("received command on sock");
2976 DBG("Consumer thread sessiond poll exiting");
2979 * Close metadata streams since the producer is the session daemon which
2982 * NOTE: for now, this only applies to the UST tracer.
2984 lttng_consumer_close_metadata();
2987 * when all fds have hung up, the polling thread
2993 * Notify the data poll thread to poll back again and test the
2994 * consumer_quit state that we just set so to quit gracefully.
2996 notify_thread_pipe(ctx
->consumer_data_pipe
[1]);
2998 notify_channel_pipe(ctx
, NULL
, CONSUMER_CHANNEL_QUIT
);
3000 /* Cleaning up possibly open sockets. */
3004 PERROR("close sock sessiond poll");
3007 if (client_socket
>= 0) {
3010 PERROR("close client_socket sessiond poll");
3014 rcu_unregister_thread();
3018 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3019 struct lttng_consumer_local_data
*ctx
)
3023 pthread_mutex_lock(&stream
->lock
);
3025 switch (consumer_data
.type
) {
3026 case LTTNG_CONSUMER_KERNEL
:
3027 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
3029 case LTTNG_CONSUMER32_UST
:
3030 case LTTNG_CONSUMER64_UST
:
3031 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
3034 ERR("Unknown consumer_data type");
3040 pthread_mutex_unlock(&stream
->lock
);
3044 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3046 switch (consumer_data
.type
) {
3047 case LTTNG_CONSUMER_KERNEL
:
3048 return lttng_kconsumer_on_recv_stream(stream
);
3049 case LTTNG_CONSUMER32_UST
:
3050 case LTTNG_CONSUMER64_UST
:
3051 return lttng_ustconsumer_on_recv_stream(stream
);
3053 ERR("Unknown consumer_data type");
3060 * Allocate and set consumer data hash tables.
3062 void lttng_consumer_init(void)
3064 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3065 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3066 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3067 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3071 * Process the ADD_RELAYD command receive by a consumer.
3073 * This will create a relayd socket pair and add it to the relayd hash table.
3074 * The caller MUST acquire a RCU read side lock before calling it.
3076 int consumer_add_relayd_socket(int net_seq_idx
, int sock_type
,
3077 struct lttng_consumer_local_data
*ctx
, int sock
,
3078 struct pollfd
*consumer_sockpoll
, struct lttcomm_sock
*relayd_sock
,
3079 unsigned int sessiond_id
)
3081 int fd
= -1, ret
= -1, relayd_created
= 0;
3082 enum lttng_error_code ret_code
= LTTNG_OK
;
3083 struct consumer_relayd_sock_pair
*relayd
;
3085 DBG("Consumer adding relayd socket (idx: %d)", net_seq_idx
);
3087 /* First send a status message before receiving the fds. */
3088 ret
= consumer_send_status_msg(sock
, ret_code
);
3090 /* Somehow, the session daemon is not responding anymore. */
3094 /* Get relayd reference if exists. */
3095 relayd
= consumer_find_relayd(net_seq_idx
);
3096 if (relayd
== NULL
) {
3097 /* Not found. Allocate one. */
3098 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3099 if (relayd
== NULL
) {
3100 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_OUTFD_ERROR
);
3104 relayd
->sessiond_session_id
= (uint64_t) sessiond_id
;
3108 /* Poll on consumer socket. */
3109 if (lttng_consumer_poll_socket(consumer_sockpoll
) < 0) {
3114 /* Get relayd socket from session daemon */
3115 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3116 if (ret
!= sizeof(fd
)) {
3117 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3119 fd
= -1; /* Just in case it gets set with an invalid value. */
3123 /* We have the fds without error. Send status back. */
3124 ret
= consumer_send_status_msg(sock
, ret_code
);
3126 /* Somehow, the session daemon is not responding anymore. */
3130 /* Copy socket information and received FD */
3131 switch (sock_type
) {
3132 case LTTNG_STREAM_CONTROL
:
3133 /* Copy received lttcomm socket */
3134 lttcomm_copy_sock(&relayd
->control_sock
, relayd_sock
);
3135 ret
= lttcomm_create_sock(&relayd
->control_sock
);
3136 /* Immediately try to close the created socket if valid. */
3137 if (relayd
->control_sock
.fd
>= 0) {
3138 if (close(relayd
->control_sock
.fd
)) {
3139 PERROR("close relayd control socket");
3142 /* Handle create_sock error. */
3147 /* Assign new file descriptor */
3148 relayd
->control_sock
.fd
= fd
;
3151 * Create a session on the relayd and store the returned id. Lock the
3152 * control socket mutex if the relayd was NOT created before.
3154 if (!relayd_created
) {
3155 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3157 ret
= relayd_create_session(&relayd
->control_sock
,
3158 &relayd
->relayd_session_id
);
3159 if (!relayd_created
) {
3160 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3164 * Close all sockets of a relayd object. It will be freed if it was
3165 * created at the error code path or else it will be garbage
3168 (void) relayd_close(&relayd
->control_sock
);
3169 (void) relayd_close(&relayd
->data_sock
);
3174 case LTTNG_STREAM_DATA
:
3175 /* Copy received lttcomm socket */
3176 lttcomm_copy_sock(&relayd
->data_sock
, relayd_sock
);
3177 ret
= lttcomm_create_sock(&relayd
->data_sock
);
3178 /* Immediately try to close the created socket if valid. */
3179 if (relayd
->data_sock
.fd
>= 0) {
3180 if (close(relayd
->data_sock
.fd
)) {
3181 PERROR("close relayd data socket");
3184 /* Handle create_sock error. */
3189 /* Assign new file descriptor */
3190 relayd
->data_sock
.fd
= fd
;
3193 ERR("Unknown relayd socket type (%d)", sock_type
);
3198 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3199 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3200 relayd
->net_seq_idx
, fd
);
3203 * Add relayd socket pair to consumer data hashtable. If object already
3204 * exists or on error, the function gracefully returns.
3212 /* Close received socket if valid. */
3215 PERROR("close received socket");
3220 if (relayd_created
) {
3228 * Try to lock the stream mutex.
3230 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
3232 static int stream_try_lock(struct lttng_consumer_stream
*stream
)
3239 * Try to lock the stream mutex. On failure, we know that the stream is
3240 * being used else where hence there is data still being extracted.
3242 ret
= pthread_mutex_trylock(&stream
->lock
);
3244 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
3256 * Search for a relayd associated to the session id and return the reference.
3258 * A rcu read side lock MUST be acquire before calling this function and locked
3259 * until the relayd object is no longer necessary.
3261 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3263 struct lttng_ht_iter iter
;
3264 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3266 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3267 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3270 * Check by sessiond id which is unique here where the relayd session
3271 * id might not be when having multiple relayd.
3273 if (relayd
->sessiond_session_id
== id
) {
3274 /* Found the relayd. There can be only one per id. */
3286 * Check if for a given session id there is still data needed to be extract
3289 * Return 1 if data is pending or else 0 meaning ready to be read.
3291 int consumer_data_pending(uint64_t id
)
3294 struct lttng_ht_iter iter
;
3295 struct lttng_ht
*ht
;
3296 struct lttng_consumer_stream
*stream
;
3297 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3298 int (*data_pending
)(struct lttng_consumer_stream
*);
3300 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3303 pthread_mutex_lock(&consumer_data
.lock
);
3305 switch (consumer_data
.type
) {
3306 case LTTNG_CONSUMER_KERNEL
:
3307 data_pending
= lttng_kconsumer_data_pending
;
3309 case LTTNG_CONSUMER32_UST
:
3310 case LTTNG_CONSUMER64_UST
:
3311 data_pending
= lttng_ustconsumer_data_pending
;
3314 ERR("Unknown consumer data type");
3318 /* Ease our life a bit */
3319 ht
= consumer_data
.stream_list_ht
;
3321 relayd
= find_relayd_by_session_id(id
);
3323 /* Send init command for data pending. */
3324 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3325 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3326 relayd
->relayd_session_id
);
3327 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3329 /* Communication error thus the relayd so no data pending. */
3330 goto data_not_pending
;
3334 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3335 ht
->hash_fct(&id
, lttng_ht_seed
),
3337 &iter
.iter
, stream
, node_session_id
.node
) {
3338 /* If this call fails, the stream is being used hence data pending. */
3339 ret
= stream_try_lock(stream
);
3345 * A removed node from the hash table indicates that the stream has
3346 * been deleted thus having a guarantee that the buffers are closed
3347 * on the consumer side. However, data can still be transmitted
3348 * over the network so don't skip the relayd check.
3350 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3352 /* Check the stream if there is data in the buffers. */
3353 ret
= data_pending(stream
);
3355 pthread_mutex_unlock(&stream
->lock
);
3362 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3363 if (stream
->metadata_flag
) {
3364 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3365 stream
->relayd_stream_id
);
3367 ret
= relayd_data_pending(&relayd
->control_sock
,
3368 stream
->relayd_stream_id
,
3369 stream
->next_net_seq_num
- 1);
3371 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3373 pthread_mutex_unlock(&stream
->lock
);
3377 pthread_mutex_unlock(&stream
->lock
);
3381 unsigned int is_data_inflight
= 0;
3383 /* Send init command for data pending. */
3384 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3385 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3386 relayd
->relayd_session_id
, &is_data_inflight
);
3387 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3389 goto data_not_pending
;
3391 if (is_data_inflight
) {
3397 * Finding _no_ node in the hash table and no inflight data means that the
3398 * stream(s) have been removed thus data is guaranteed to be available for
3399 * analysis from the trace files.
3403 /* Data is available to be read by a viewer. */
3404 pthread_mutex_unlock(&consumer_data
.lock
);
3409 /* Data is still being extracted from buffers. */
3410 pthread_mutex_unlock(&consumer_data
.lock
);
3416 * Send a ret code status message to the sessiond daemon.
3418 * Return the sendmsg() return value.
3420 int consumer_send_status_msg(int sock
, int ret_code
)
3422 struct lttcomm_consumer_status_msg msg
;
3424 msg
.ret_code
= ret_code
;
3426 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3430 * Send a channel status message to the sessiond daemon.
3432 * Return the sendmsg() return value.
3434 int consumer_send_status_channel(int sock
,
3435 struct lttng_consumer_channel
*channel
)
3437 struct lttcomm_consumer_status_channel msg
;
3442 msg
.ret_code
= -LTTNG_ERR_UST_CHAN_FAIL
;
3444 msg
.ret_code
= LTTNG_OK
;
3445 msg
.key
= channel
->key
;
3446 msg
.stream_count
= channel
->streams
.count
;
3449 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));