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
)
833 struct lttng_consumer_channel
*channel
;
835 channel
= zmalloc(sizeof(*channel
));
836 if (channel
== NULL
) {
837 PERROR("malloc struct lttng_consumer_channel");
842 channel
->refcount
= 0;
843 channel
->session_id
= session_id
;
846 channel
->relayd_id
= relayd_id
;
847 channel
->output
= output
;
849 strncpy(channel
->pathname
, pathname
, sizeof(channel
->pathname
));
850 channel
->pathname
[sizeof(channel
->pathname
) - 1] = '\0';
852 strncpy(channel
->name
, name
, sizeof(channel
->name
));
853 channel
->name
[sizeof(channel
->name
) - 1] = '\0';
855 lttng_ht_node_init_u64(&channel
->node
, channel
->key
);
857 channel
->wait_fd
= -1;
859 CDS_INIT_LIST_HEAD(&channel
->streams
.head
);
861 DBG("Allocated channel (key %" PRIu64
")", channel
->key
)
868 * Add a channel to the global list protected by a mutex.
870 int consumer_add_channel(struct lttng_consumer_channel
*channel
,
871 struct lttng_consumer_local_data
*ctx
)
874 struct lttng_ht_node_u64
*node
;
875 struct lttng_ht_iter iter
;
877 pthread_mutex_lock(&consumer_data
.lock
);
880 lttng_ht_lookup(consumer_data
.channel_ht
, &channel
->key
, &iter
);
881 node
= lttng_ht_iter_get_node_u64(&iter
);
883 /* Channel already exist. Ignore the insertion */
884 ERR("Consumer add channel key %" PRIu64
" already exists!",
890 lttng_ht_add_unique_u64(consumer_data
.channel_ht
, &channel
->node
);
894 pthread_mutex_unlock(&consumer_data
.lock
);
896 if (!ret
&& channel
->wait_fd
!= -1 &&
897 channel
->metadata_stream
== NULL
) {
898 notify_channel_pipe(ctx
, channel
, CONSUMER_CHANNEL_ADD
);
904 * Allocate the pollfd structure and the local view of the out fds to avoid
905 * doing a lookup in the linked list and concurrency issues when writing is
906 * needed. Called with consumer_data.lock held.
908 * Returns the number of fds in the structures.
910 static int update_poll_array(struct lttng_consumer_local_data
*ctx
,
911 struct pollfd
**pollfd
, struct lttng_consumer_stream
**local_stream
,
915 struct lttng_ht_iter iter
;
916 struct lttng_consumer_stream
*stream
;
921 assert(local_stream
);
923 DBG("Updating poll fd array");
925 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
927 * Only active streams with an active end point can be added to the
928 * poll set and local stream storage of the thread.
930 * There is a potential race here for endpoint_status to be updated
931 * just after the check. However, this is OK since the stream(s) will
932 * be deleted once the thread is notified that the end point state has
933 * changed where this function will be called back again.
935 if (stream
->state
!= LTTNG_CONSUMER_ACTIVE_STREAM
||
936 stream
->endpoint_status
== CONSUMER_ENDPOINT_INACTIVE
) {
940 * This clobbers way too much the debug output. Uncomment that if you
941 * need it for debugging purposes.
943 * DBG("Active FD %d", stream->wait_fd);
945 (*pollfd
)[i
].fd
= stream
->wait_fd
;
946 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
947 local_stream
[i
] = stream
;
953 * Insert the consumer_data_pipe at the end of the array and don't
954 * increment i so nb_fd is the number of real FD.
956 (*pollfd
)[i
].fd
= ctx
->consumer_data_pipe
[0];
957 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
962 * Poll on the should_quit pipe and the command socket return -1 on error and
963 * should exit, 0 if data is available on the command socket
965 int lttng_consumer_poll_socket(struct pollfd
*consumer_sockpoll
)
970 num_rdy
= poll(consumer_sockpoll
, 2, -1);
973 * Restart interrupted system call.
975 if (errno
== EINTR
) {
978 PERROR("Poll error");
981 if (consumer_sockpoll
[0].revents
& (POLLIN
| POLLPRI
)) {
982 DBG("consumer_should_quit wake up");
992 * Set the error socket.
994 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data
*ctx
,
997 ctx
->consumer_error_socket
= sock
;
1001 * Set the command socket path.
1003 void lttng_consumer_set_command_sock_path(
1004 struct lttng_consumer_local_data
*ctx
, char *sock
)
1006 ctx
->consumer_command_sock_path
= sock
;
1010 * Send return code to the session daemon.
1011 * If the socket is not defined, we return 0, it is not a fatal error
1013 int lttng_consumer_send_error(struct lttng_consumer_local_data
*ctx
, int cmd
)
1015 if (ctx
->consumer_error_socket
> 0) {
1016 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
1017 sizeof(enum lttcomm_sessiond_command
));
1024 * Close all the tracefiles and stream fds and MUST be called when all
1025 * instances are destroyed i.e. when all threads were joined and are ended.
1027 void lttng_consumer_cleanup(void)
1029 struct lttng_ht_iter iter
;
1030 struct lttng_consumer_channel
*channel
;
1034 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
, channel
,
1036 consumer_del_channel(channel
);
1041 lttng_ht_destroy(consumer_data
.channel_ht
);
1043 cleanup_relayd_ht();
1045 lttng_ht_destroy(consumer_data
.stream_per_chan_id_ht
);
1048 * This HT contains streams that are freed by either the metadata thread or
1049 * the data thread so we do *nothing* on the hash table and simply destroy
1052 lttng_ht_destroy(consumer_data
.stream_list_ht
);
1056 * Called from signal handler.
1058 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
1063 ret
= write(ctx
->consumer_should_quit
[1], "4", 1);
1064 } while (ret
< 0 && errno
== EINTR
);
1065 if (ret
< 0 || ret
!= 1) {
1066 PERROR("write consumer quit");
1069 DBG("Consumer flag that it should quit");
1072 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1075 int outfd
= stream
->out_fd
;
1078 * This does a blocking write-and-wait on any page that belongs to the
1079 * subbuffer prior to the one we just wrote.
1080 * Don't care about error values, as these are just hints and ways to
1081 * limit the amount of page cache used.
1083 if (orig_offset
< stream
->max_sb_size
) {
1086 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1087 stream
->max_sb_size
,
1088 SYNC_FILE_RANGE_WAIT_BEFORE
1089 | SYNC_FILE_RANGE_WRITE
1090 | SYNC_FILE_RANGE_WAIT_AFTER
);
1092 * Give hints to the kernel about how we access the file:
1093 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1096 * We need to call fadvise again after the file grows because the
1097 * kernel does not seem to apply fadvise to non-existing parts of the
1100 * Call fadvise _after_ having waited for the page writeback to
1101 * complete because the dirty page writeback semantic is not well
1102 * defined. So it can be expected to lead to lower throughput in
1105 posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1106 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1110 * Initialise the necessary environnement :
1111 * - create a new context
1112 * - create the poll_pipe
1113 * - create the should_quit pipe (for signal handler)
1114 * - create the thread pipe (for splice)
1116 * Takes a function pointer as argument, this function is called when data is
1117 * available on a buffer. This function is responsible to do the
1118 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1119 * buffer configuration and then kernctl_put_next_subbuf at the end.
1121 * Returns a pointer to the new context or NULL on error.
1123 struct lttng_consumer_local_data
*lttng_consumer_create(
1124 enum lttng_consumer_type type
,
1125 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1126 struct lttng_consumer_local_data
*ctx
),
1127 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1128 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1129 int (*update_stream
)(int stream_key
, uint32_t state
))
1132 struct lttng_consumer_local_data
*ctx
;
1134 assert(consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1135 consumer_data
.type
== type
);
1136 consumer_data
.type
= type
;
1138 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
1140 PERROR("allocating context");
1144 ctx
->consumer_error_socket
= -1;
1145 ctx
->consumer_metadata_socket
= -1;
1146 /* assign the callbacks */
1147 ctx
->on_buffer_ready
= buffer_ready
;
1148 ctx
->on_recv_channel
= recv_channel
;
1149 ctx
->on_recv_stream
= recv_stream
;
1150 ctx
->on_update_stream
= update_stream
;
1152 ret
= pipe(ctx
->consumer_data_pipe
);
1154 PERROR("Error creating poll pipe");
1155 goto error_poll_pipe
;
1158 /* set read end of the pipe to non-blocking */
1159 ret
= fcntl(ctx
->consumer_data_pipe
[0], F_SETFL
, O_NONBLOCK
);
1161 PERROR("fcntl O_NONBLOCK");
1162 goto error_poll_fcntl
;
1165 /* set write end of the pipe to non-blocking */
1166 ret
= fcntl(ctx
->consumer_data_pipe
[1], F_SETFL
, O_NONBLOCK
);
1168 PERROR("fcntl O_NONBLOCK");
1169 goto error_poll_fcntl
;
1172 ret
= pipe(ctx
->consumer_should_quit
);
1174 PERROR("Error creating recv pipe");
1175 goto error_quit_pipe
;
1178 ret
= pipe(ctx
->consumer_thread_pipe
);
1180 PERROR("Error creating thread pipe");
1181 goto error_thread_pipe
;
1184 ret
= pipe(ctx
->consumer_channel_pipe
);
1186 PERROR("Error creating channel pipe");
1187 goto error_channel_pipe
;
1190 ret
= utils_create_pipe(ctx
->consumer_metadata_pipe
);
1192 goto error_metadata_pipe
;
1195 ret
= utils_create_pipe(ctx
->consumer_splice_metadata_pipe
);
1197 goto error_splice_pipe
;
1203 utils_close_pipe(ctx
->consumer_metadata_pipe
);
1204 error_metadata_pipe
:
1205 utils_close_pipe(ctx
->consumer_channel_pipe
);
1207 utils_close_pipe(ctx
->consumer_thread_pipe
);
1209 utils_close_pipe(ctx
->consumer_should_quit
);
1212 utils_close_pipe(ctx
->consumer_data_pipe
);
1220 * Close all fds associated with the instance and free the context.
1222 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1226 DBG("Consumer destroying it. Closing everything.");
1228 ret
= close(ctx
->consumer_error_socket
);
1232 ret
= close(ctx
->consumer_metadata_socket
);
1236 utils_close_pipe(ctx
->consumer_thread_pipe
);
1237 utils_close_pipe(ctx
->consumer_channel_pipe
);
1238 utils_close_pipe(ctx
->consumer_data_pipe
);
1239 utils_close_pipe(ctx
->consumer_should_quit
);
1240 utils_close_pipe(ctx
->consumer_splice_metadata_pipe
);
1242 unlink(ctx
->consumer_command_sock_path
);
1247 * Write the metadata stream id on the specified file descriptor.
1249 static int write_relayd_metadata_id(int fd
,
1250 struct lttng_consumer_stream
*stream
,
1251 struct consumer_relayd_sock_pair
*relayd
, unsigned long padding
)
1254 struct lttcomm_relayd_metadata_payload hdr
;
1256 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1257 hdr
.padding_size
= htobe32(padding
);
1259 ret
= write(fd
, (void *) &hdr
, sizeof(hdr
));
1260 } while (ret
< 0 && errno
== EINTR
);
1261 if (ret
< 0 || ret
!= sizeof(hdr
)) {
1263 * This error means that the fd's end is closed so ignore the perror
1264 * not to clubber the error output since this can happen in a normal
1267 if (errno
!= EPIPE
) {
1268 PERROR("write metadata stream id");
1270 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1272 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1273 * handle writting the missing part so report that as an error and
1274 * don't lie to the caller.
1279 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1280 stream
->relayd_stream_id
, padding
);
1287 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1288 * core function for writing trace buffers to either the local filesystem or
1291 * It must be called with the stream lock held.
1293 * Careful review MUST be put if any changes occur!
1295 * Returns the number of bytes written
1297 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1298 struct lttng_consumer_local_data
*ctx
,
1299 struct lttng_consumer_stream
*stream
, unsigned long len
,
1300 unsigned long padding
)
1302 unsigned long mmap_offset
;
1304 ssize_t ret
= 0, written
= 0;
1305 off_t orig_offset
= stream
->out_fd_offset
;
1306 /* Default is on the disk */
1307 int outfd
= stream
->out_fd
;
1308 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1309 unsigned int relayd_hang_up
= 0;
1311 /* RCU lock for the relayd pointer */
1314 /* Flag that the current stream if set for network streaming. */
1315 if (stream
->net_seq_idx
!= -1) {
1316 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1317 if (relayd
== NULL
) {
1322 /* get the offset inside the fd to mmap */
1323 switch (consumer_data
.type
) {
1324 case LTTNG_CONSUMER_KERNEL
:
1325 mmap_base
= stream
->mmap_base
;
1326 ret
= kernctl_get_mmap_read_offset(stream
->wait_fd
, &mmap_offset
);
1328 case LTTNG_CONSUMER32_UST
:
1329 case LTTNG_CONSUMER64_UST
:
1330 mmap_base
= lttng_ustctl_get_mmap_base(stream
);
1332 ERR("read mmap get mmap base for stream %s", stream
->name
);
1336 ret
= lttng_ustctl_get_mmap_read_offset(stream
, &mmap_offset
);
1340 ERR("Unknown consumer_data type");
1345 PERROR("tracer ctl get_mmap_read_offset");
1350 /* Handle stream on the relayd if the output is on the network */
1352 unsigned long netlen
= len
;
1355 * Lock the control socket for the complete duration of the function
1356 * since from this point on we will use the socket.
1358 if (stream
->metadata_flag
) {
1359 /* Metadata requires the control socket. */
1360 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1361 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1364 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1366 /* Use the returned socket. */
1369 /* Write metadata stream id before payload */
1370 if (stream
->metadata_flag
) {
1371 ret
= write_relayd_metadata_id(outfd
, stream
, relayd
, padding
);
1374 /* Socket operation failed. We consider the relayd dead */
1375 if (ret
== -EPIPE
|| ret
== -EINVAL
) {
1383 /* Socket operation failed. We consider the relayd dead */
1384 if (ret
== -EPIPE
|| ret
== -EINVAL
) {
1388 /* Else, use the default set before which is the filesystem. */
1391 /* No streaming, we have to set the len with the full padding */
1397 ret
= write(outfd
, mmap_base
+ mmap_offset
, len
);
1398 } while (ret
< 0 && errno
== EINTR
);
1399 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, len
);
1402 * This is possible if the fd is closed on the other side (outfd)
1403 * or any write problem. It can be verbose a bit for a normal
1404 * execution if for instance the relayd is stopped abruptly. This
1405 * can happen so set this to a DBG statement.
1407 DBG("Error in file write mmap");
1411 /* Socket operation failed. We consider the relayd dead */
1412 if (errno
== EPIPE
|| errno
== EINVAL
) {
1417 } else if (ret
> len
) {
1418 PERROR("Error in file write (ret %zd > len %lu)", ret
, len
);
1426 /* This call is useless on a socket so better save a syscall. */
1428 /* This won't block, but will start writeout asynchronously */
1429 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret
,
1430 SYNC_FILE_RANGE_WRITE
);
1431 stream
->out_fd_offset
+= ret
;
1435 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1439 * This is a special case that the relayd has closed its socket. Let's
1440 * cleanup the relayd object and all associated streams.
1442 if (relayd
&& relayd_hang_up
) {
1443 cleanup_relayd(relayd
, ctx
);
1447 /* Unlock only if ctrl socket used */
1448 if (relayd
&& stream
->metadata_flag
) {
1449 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1457 * Splice the data from the ring buffer to the tracefile.
1459 * It must be called with the stream lock held.
1461 * Returns the number of bytes spliced.
1463 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1464 struct lttng_consumer_local_data
*ctx
,
1465 struct lttng_consumer_stream
*stream
, unsigned long len
,
1466 unsigned long padding
)
1468 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1470 off_t orig_offset
= stream
->out_fd_offset
;
1471 int fd
= stream
->wait_fd
;
1472 /* Default is on the disk */
1473 int outfd
= stream
->out_fd
;
1474 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1476 unsigned int relayd_hang_up
= 0;
1478 switch (consumer_data
.type
) {
1479 case LTTNG_CONSUMER_KERNEL
:
1481 case LTTNG_CONSUMER32_UST
:
1482 case LTTNG_CONSUMER64_UST
:
1483 /* Not supported for user space tracing */
1486 ERR("Unknown consumer_data type");
1490 /* RCU lock for the relayd pointer */
1493 /* Flag that the current stream if set for network streaming. */
1494 if (stream
->net_seq_idx
!= -1) {
1495 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1496 if (relayd
== NULL
) {
1502 * Choose right pipe for splice. Metadata and trace data are handled by
1503 * different threads hence the use of two pipes in order not to race or
1504 * corrupt the written data.
1506 if (stream
->metadata_flag
) {
1507 splice_pipe
= ctx
->consumer_splice_metadata_pipe
;
1509 splice_pipe
= ctx
->consumer_thread_pipe
;
1512 /* Write metadata stream id before payload */
1514 int total_len
= len
;
1516 if (stream
->metadata_flag
) {
1518 * Lock the control socket for the complete duration of the function
1519 * since from this point on we will use the socket.
1521 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1523 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
, relayd
,
1527 /* Socket operation failed. We consider the relayd dead */
1528 if (ret
== -EBADF
) {
1529 WARN("Remote relayd disconnected. Stopping");
1536 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1539 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1541 /* Use the returned socket. */
1544 /* Socket operation failed. We consider the relayd dead */
1545 if (ret
== -EBADF
) {
1546 WARN("Remote relayd disconnected. Stopping");
1553 /* No streaming, we have to set the len with the full padding */
1558 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1559 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1560 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1561 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1562 DBG("splice chan to pipe, ret %zd", ret_splice
);
1563 if (ret_splice
< 0) {
1564 PERROR("Error in relay splice");
1566 written
= ret_splice
;
1572 /* Handle stream on the relayd if the output is on the network */
1574 if (stream
->metadata_flag
) {
1575 size_t metadata_payload_size
=
1576 sizeof(struct lttcomm_relayd_metadata_payload
);
1578 /* Update counter to fit the spliced data */
1579 ret_splice
+= metadata_payload_size
;
1580 len
+= metadata_payload_size
;
1582 * We do this so the return value can match the len passed as
1583 * argument to this function.
1585 written
-= metadata_payload_size
;
1589 /* Splice data out */
1590 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1591 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1592 DBG("Consumer splice pipe to file, ret %zd", ret_splice
);
1593 if (ret_splice
< 0) {
1594 PERROR("Error in file splice");
1596 written
= ret_splice
;
1598 /* Socket operation failed. We consider the relayd dead */
1599 if (errno
== EBADF
|| errno
== EPIPE
) {
1600 WARN("Remote relayd disconnected. Stopping");
1606 } else if (ret_splice
> len
) {
1608 PERROR("Wrote more data than requested %zd (len: %lu)",
1610 written
+= ret_splice
;
1616 /* This call is useless on a socket so better save a syscall. */
1618 /* This won't block, but will start writeout asynchronously */
1619 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1620 SYNC_FILE_RANGE_WRITE
);
1621 stream
->out_fd_offset
+= ret_splice
;
1623 written
+= ret_splice
;
1625 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1633 * This is a special case that the relayd has closed its socket. Let's
1634 * cleanup the relayd object and all associated streams.
1636 if (relayd
&& relayd_hang_up
) {
1637 cleanup_relayd(relayd
, ctx
);
1638 /* Skip splice error so the consumer does not fail */
1643 /* send the appropriate error description to sessiond */
1646 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1649 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1652 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
1657 if (relayd
&& stream
->metadata_flag
) {
1658 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1666 * Take a snapshot for a specific fd
1668 * Returns 0 on success, < 0 on error
1670 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
1672 switch (consumer_data
.type
) {
1673 case LTTNG_CONSUMER_KERNEL
:
1674 return lttng_kconsumer_take_snapshot(stream
);
1675 case LTTNG_CONSUMER32_UST
:
1676 case LTTNG_CONSUMER64_UST
:
1677 return lttng_ustconsumer_take_snapshot(stream
);
1679 ERR("Unknown consumer_data type");
1686 * Get the produced position
1688 * Returns 0 on success, < 0 on error
1690 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
1693 switch (consumer_data
.type
) {
1694 case LTTNG_CONSUMER_KERNEL
:
1695 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
1696 case LTTNG_CONSUMER32_UST
:
1697 case LTTNG_CONSUMER64_UST
:
1698 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
1700 ERR("Unknown consumer_data type");
1706 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
1707 int sock
, struct pollfd
*consumer_sockpoll
)
1709 switch (consumer_data
.type
) {
1710 case LTTNG_CONSUMER_KERNEL
:
1711 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1712 case LTTNG_CONSUMER32_UST
:
1713 case LTTNG_CONSUMER64_UST
:
1714 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1716 ERR("Unknown consumer_data type");
1723 * Iterate over all streams of the hashtable and free them properly.
1725 * WARNING: *MUST* be used with data stream only.
1727 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1729 struct lttng_ht_iter iter
;
1730 struct lttng_consumer_stream
*stream
;
1737 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1739 * Ignore return value since we are currently cleaning up so any error
1742 (void) consumer_del_stream(stream
, ht
);
1746 lttng_ht_destroy(ht
);
1750 * Iterate over all streams of the hashtable and free them properly.
1752 * XXX: Should not be only for metadata stream or else use an other name.
1754 static void destroy_stream_ht(struct lttng_ht
*ht
)
1756 struct lttng_ht_iter iter
;
1757 struct lttng_consumer_stream
*stream
;
1764 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1766 * Ignore return value since we are currently cleaning up so any error
1769 (void) consumer_del_metadata_stream(stream
, ht
);
1773 lttng_ht_destroy(ht
);
1776 void lttng_consumer_close_metadata(void)
1778 switch (consumer_data
.type
) {
1779 case LTTNG_CONSUMER_KERNEL
:
1781 * The Kernel consumer has a different metadata scheme so we don't
1782 * close anything because the stream will be closed by the session
1786 case LTTNG_CONSUMER32_UST
:
1787 case LTTNG_CONSUMER64_UST
:
1789 * Close all metadata streams. The metadata hash table is passed and
1790 * this call iterates over it by closing all wakeup fd. This is safe
1791 * because at this point we are sure that the metadata producer is
1792 * either dead or blocked.
1794 lttng_ustconsumer_close_metadata(metadata_ht
);
1797 ERR("Unknown consumer_data type");
1803 * Clean up a metadata stream and free its memory.
1805 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
1806 struct lttng_ht
*ht
)
1809 struct lttng_ht_iter iter
;
1810 struct lttng_consumer_channel
*free_chan
= NULL
;
1811 struct consumer_relayd_sock_pair
*relayd
;
1815 * This call should NEVER receive regular stream. It must always be
1816 * metadata stream and this is crucial for data structure synchronization.
1818 assert(stream
->metadata_flag
);
1820 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
1823 /* Means the stream was allocated but not successfully added */
1824 goto free_stream_rcu
;
1827 pthread_mutex_lock(&consumer_data
.lock
);
1828 pthread_mutex_lock(&stream
->lock
);
1830 switch (consumer_data
.type
) {
1831 case LTTNG_CONSUMER_KERNEL
:
1832 if (stream
->mmap_base
!= NULL
) {
1833 ret
= munmap(stream
->mmap_base
, stream
->mmap_len
);
1835 PERROR("munmap metadata stream");
1839 case LTTNG_CONSUMER32_UST
:
1840 case LTTNG_CONSUMER64_UST
:
1841 lttng_ustconsumer_del_stream(stream
);
1844 ERR("Unknown consumer_data type");
1850 iter
.iter
.node
= &stream
->node
.node
;
1851 ret
= lttng_ht_del(ht
, &iter
);
1854 iter
.iter
.node
= &stream
->node_channel_id
.node
;
1855 ret
= lttng_ht_del(consumer_data
.stream_per_chan_id_ht
, &iter
);
1858 iter
.iter
.node
= &stream
->node_session_id
.node
;
1859 ret
= lttng_ht_del(consumer_data
.stream_list_ht
, &iter
);
1863 if (stream
->out_fd
>= 0) {
1864 ret
= close(stream
->out_fd
);
1870 /* Check and cleanup relayd */
1872 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1873 if (relayd
!= NULL
) {
1874 uatomic_dec(&relayd
->refcount
);
1875 assert(uatomic_read(&relayd
->refcount
) >= 0);
1877 /* Closing streams requires to lock the control socket. */
1878 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1879 ret
= relayd_send_close_stream(&relayd
->control_sock
,
1880 stream
->relayd_stream_id
, stream
->next_net_seq_num
- 1);
1881 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1883 DBG("Unable to close stream on the relayd. Continuing");
1885 * Continue here. There is nothing we can do for the relayd.
1886 * Chances are that the relayd has closed the socket so we just
1887 * continue cleaning up.
1891 /* Both conditions are met, we destroy the relayd. */
1892 if (uatomic_read(&relayd
->refcount
) == 0 &&
1893 uatomic_read(&relayd
->destroy_flag
)) {
1894 destroy_relayd(relayd
);
1899 /* Atomically decrement channel refcount since other threads can use it. */
1900 uatomic_dec(&stream
->chan
->refcount
);
1901 if (!uatomic_read(&stream
->chan
->refcount
)
1902 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
1903 /* Go for channel deletion! */
1904 free_chan
= stream
->chan
;
1908 pthread_mutex_unlock(&stream
->lock
);
1909 pthread_mutex_unlock(&consumer_data
.lock
);
1912 consumer_del_channel(free_chan
);
1916 call_rcu(&stream
->node
.head
, free_stream_rcu
);
1920 * Action done with the metadata stream when adding it to the consumer internal
1921 * data structures to handle it.
1923 static int add_metadata_stream(struct lttng_consumer_stream
*stream
,
1924 struct lttng_ht
*ht
)
1927 struct consumer_relayd_sock_pair
*relayd
;
1928 struct lttng_ht_iter iter
;
1929 struct lttng_ht_node_u64
*node
;
1934 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
1936 pthread_mutex_lock(&consumer_data
.lock
);
1937 pthread_mutex_lock(&stream
->lock
);
1940 * From here, refcounts are updated so be _careful_ when returning an error
1947 * Lookup the stream just to make sure it does not exist in our internal
1948 * state. This should NEVER happen.
1950 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
1951 node
= lttng_ht_iter_get_node_u64(&iter
);
1954 /* Find relayd and, if one is found, increment refcount. */
1955 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1956 if (relayd
!= NULL
) {
1957 uatomic_inc(&relayd
->refcount
);
1960 /* Update channel refcount once added without error(s). */
1961 uatomic_inc(&stream
->chan
->refcount
);
1964 * When nb_init_stream_left reaches 0, we don't need to trigger any action
1965 * in terms of destroying the associated channel, because the action that
1966 * causes the count to become 0 also causes a stream to be added. The
1967 * channel deletion will thus be triggered by the following removal of this
1970 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
1971 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
1974 lttng_ht_add_unique_u64(ht
, &stream
->node
);
1976 lttng_ht_add_unique_u64(consumer_data
.stream_per_chan_id_ht
,
1977 &stream
->node_channel_id
);
1980 * Add stream to the stream_list_ht of the consumer data. No need to steal
1981 * the key since the HT does not use it and we allow to add redundant keys
1984 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
1988 pthread_mutex_unlock(&stream
->lock
);
1989 pthread_mutex_unlock(&consumer_data
.lock
);
1994 * Delete data stream that are flagged for deletion (endpoint_status).
1996 static void validate_endpoint_status_data_stream(void)
1998 struct lttng_ht_iter iter
;
1999 struct lttng_consumer_stream
*stream
;
2001 DBG("Consumer delete flagged data stream");
2004 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2005 /* Validate delete flag of the stream */
2006 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2009 /* Delete it right now */
2010 consumer_del_stream(stream
, data_ht
);
2016 * Delete metadata stream that are flagged for deletion (endpoint_status).
2018 static void validate_endpoint_status_metadata_stream(
2019 struct lttng_poll_event
*pollset
)
2021 struct lttng_ht_iter iter
;
2022 struct lttng_consumer_stream
*stream
;
2024 DBG("Consumer delete flagged metadata stream");
2029 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2030 /* Validate delete flag of the stream */
2031 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2035 * Remove from pollset so the metadata thread can continue without
2036 * blocking on a deleted stream.
2038 lttng_poll_del(pollset
, stream
->wait_fd
);
2040 /* Delete it right now */
2041 consumer_del_metadata_stream(stream
, metadata_ht
);
2047 * Thread polls on metadata file descriptor and write them on disk or on the
2050 void *consumer_thread_metadata_poll(void *data
)
2053 uint32_t revents
, nb_fd
;
2054 struct lttng_consumer_stream
*stream
= NULL
;
2055 struct lttng_ht_iter iter
;
2056 struct lttng_ht_node_u64
*node
;
2057 struct lttng_poll_event events
;
2058 struct lttng_consumer_local_data
*ctx
= data
;
2061 rcu_register_thread();
2063 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2065 /* ENOMEM at this point. Better to bail out. */
2069 DBG("Thread metadata poll started");
2071 /* Size is set to 1 for the consumer_metadata pipe */
2072 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2074 ERR("Poll set creation failed");
2078 ret
= lttng_poll_add(&events
, ctx
->consumer_metadata_pipe
[0], LPOLLIN
);
2084 DBG("Metadata main loop started");
2087 /* Only the metadata pipe is set */
2088 if (LTTNG_POLL_GETNB(&events
) == 0 && consumer_quit
== 1) {
2093 DBG("Metadata poll wait with %d fd(s)", LTTNG_POLL_GETNB(&events
));
2094 ret
= lttng_poll_wait(&events
, -1);
2095 DBG("Metadata event catched in thread");
2097 if (errno
== EINTR
) {
2098 ERR("Poll EINTR catched");
2106 /* From here, the event is a metadata wait fd */
2107 for (i
= 0; i
< nb_fd
; i
++) {
2108 revents
= LTTNG_POLL_GETEV(&events
, i
);
2109 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2111 /* Just don't waste time if no returned events for the fd */
2116 if (pollfd
== ctx
->consumer_metadata_pipe
[0]) {
2117 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2118 DBG("Metadata thread pipe hung up");
2120 * Remove the pipe from the poll set and continue the loop
2121 * since their might be data to consume.
2123 lttng_poll_del(&events
, ctx
->consumer_metadata_pipe
[0]);
2124 ret
= close(ctx
->consumer_metadata_pipe
[0]);
2126 PERROR("close metadata pipe");
2129 } else if (revents
& LPOLLIN
) {
2131 /* Get the stream pointer received */
2132 ret
= read(pollfd
, &stream
, sizeof(stream
));
2133 } while (ret
< 0 && errno
== EINTR
);
2135 ret
< sizeof(struct lttng_consumer_stream
*)) {
2136 PERROR("read metadata stream");
2138 * Let's continue here and hope we can still work
2139 * without stopping the consumer. XXX: Should we?
2144 /* A NULL stream means that the state has changed. */
2145 if (stream
== NULL
) {
2146 /* Check for deleted streams. */
2147 validate_endpoint_status_metadata_stream(&events
);
2151 DBG("Adding metadata stream %d to poll set",
2154 ret
= add_metadata_stream(stream
, metadata_ht
);
2156 ERR("Unable to add metadata stream");
2157 /* Stream was not setup properly. Continuing. */
2158 consumer_del_metadata_stream(stream
, NULL
);
2162 /* Add metadata stream to the global poll events list */
2163 lttng_poll_add(&events
, stream
->wait_fd
,
2164 LPOLLIN
| LPOLLPRI
);
2167 /* Handle other stream */
2173 uint64_t tmp_id
= (uint64_t) pollfd
;
2175 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2177 node
= lttng_ht_iter_get_node_u64(&iter
);
2180 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2183 /* Check for error event */
2184 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2185 DBG("Metadata fd %d is hup|err.", pollfd
);
2186 if (!stream
->hangup_flush_done
2187 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2188 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2189 DBG("Attempting to flush and consume the UST buffers");
2190 lttng_ustconsumer_on_stream_hangup(stream
);
2192 /* We just flushed the stream now read it. */
2194 len
= ctx
->on_buffer_ready(stream
, ctx
);
2196 * We don't check the return value here since if we get
2197 * a negative len, it means an error occured thus we
2198 * simply remove it from the poll set and free the
2204 lttng_poll_del(&events
, stream
->wait_fd
);
2206 * This call update the channel states, closes file descriptors
2207 * and securely free the stream.
2209 consumer_del_metadata_stream(stream
, metadata_ht
);
2210 } else if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2211 /* Get the data out of the metadata file descriptor */
2212 DBG("Metadata available on fd %d", pollfd
);
2213 assert(stream
->wait_fd
== pollfd
);
2215 len
= ctx
->on_buffer_ready(stream
, ctx
);
2216 /* It's ok to have an unavailable sub-buffer */
2217 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2218 /* Clean up stream from consumer and free it. */
2219 lttng_poll_del(&events
, stream
->wait_fd
);
2220 consumer_del_metadata_stream(stream
, metadata_ht
);
2221 } else if (len
> 0) {
2222 stream
->data_read
= 1;
2226 /* Release RCU lock for the stream looked up */
2233 DBG("Metadata poll thread exiting");
2235 lttng_poll_clean(&events
);
2237 destroy_stream_ht(metadata_ht
);
2239 rcu_unregister_thread();
2244 * This thread polls the fds in the set to consume the data and write
2245 * it to tracefile if necessary.
2247 void *consumer_thread_data_poll(void *data
)
2249 int num_rdy
, num_hup
, high_prio
, ret
, i
;
2250 struct pollfd
*pollfd
= NULL
;
2251 /* local view of the streams */
2252 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2253 /* local view of consumer_data.fds_count */
2255 struct lttng_consumer_local_data
*ctx
= data
;
2258 rcu_register_thread();
2260 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2261 if (data_ht
== NULL
) {
2262 /* ENOMEM at this point. Better to bail out. */
2266 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
));
2273 * the fds set has been updated, we need to update our
2274 * local array as well
2276 pthread_mutex_lock(&consumer_data
.lock
);
2277 if (consumer_data
.need_update
) {
2282 local_stream
= NULL
;
2284 /* allocate for all fds + 1 for the consumer_data_pipe */
2285 pollfd
= zmalloc((consumer_data
.stream_count
+ 1) * sizeof(struct pollfd
));
2286 if (pollfd
== NULL
) {
2287 PERROR("pollfd malloc");
2288 pthread_mutex_unlock(&consumer_data
.lock
);
2292 /* allocate for all fds + 1 for the consumer_data_pipe */
2293 local_stream
= zmalloc((consumer_data
.stream_count
+ 1) *
2294 sizeof(struct lttng_consumer_stream
));
2295 if (local_stream
== NULL
) {
2296 PERROR("local_stream malloc");
2297 pthread_mutex_unlock(&consumer_data
.lock
);
2300 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2303 ERR("Error in allocating pollfd or local_outfds");
2304 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2305 pthread_mutex_unlock(&consumer_data
.lock
);
2309 consumer_data
.need_update
= 0;
2311 pthread_mutex_unlock(&consumer_data
.lock
);
2313 /* No FDs and consumer_quit, consumer_cleanup the thread */
2314 if (nb_fd
== 0 && consumer_quit
== 1) {
2317 /* poll on the array of fds */
2319 DBG("polling on %d fd", nb_fd
+ 1);
2320 num_rdy
= poll(pollfd
, nb_fd
+ 1, -1);
2321 DBG("poll num_rdy : %d", num_rdy
);
2322 if (num_rdy
== -1) {
2324 * Restart interrupted system call.
2326 if (errno
== EINTR
) {
2329 PERROR("Poll error");
2330 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2332 } else if (num_rdy
== 0) {
2333 DBG("Polling thread timed out");
2338 * If the consumer_data_pipe triggered poll go directly to the
2339 * beginning of the loop to update the array. We want to prioritize
2340 * array update over low-priority reads.
2342 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2343 ssize_t pipe_readlen
;
2345 DBG("consumer_data_pipe wake up");
2346 /* Consume 1 byte of pipe data */
2348 pipe_readlen
= read(ctx
->consumer_data_pipe
[0], &new_stream
,
2349 sizeof(new_stream
));
2350 } while (pipe_readlen
== -1 && errno
== EINTR
);
2351 if (pipe_readlen
< 0) {
2352 PERROR("read consumer data pipe");
2353 /* Continue so we can at least handle the current stream(s). */
2358 * If the stream is NULL, just ignore it. It's also possible that
2359 * the sessiond poll thread changed the consumer_quit state and is
2360 * waking us up to test it.
2362 if (new_stream
== NULL
) {
2363 validate_endpoint_status_data_stream();
2367 ret
= add_stream(new_stream
, data_ht
);
2369 ERR("Consumer add stream %" PRIu64
" failed. Continuing",
2372 * At this point, if the add_stream fails, it is not in the
2373 * hash table thus passing the NULL value here.
2375 consumer_del_stream(new_stream
, NULL
);
2378 /* Continue to update the local streams and handle prio ones */
2382 /* Take care of high priority channels first. */
2383 for (i
= 0; i
< nb_fd
; i
++) {
2384 if (local_stream
[i
] == NULL
) {
2387 if (pollfd
[i
].revents
& POLLPRI
) {
2388 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2390 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2391 /* it's ok to have an unavailable sub-buffer */
2392 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2393 /* Clean the stream and free it. */
2394 consumer_del_stream(local_stream
[i
], data_ht
);
2395 local_stream
[i
] = NULL
;
2396 } else if (len
> 0) {
2397 local_stream
[i
]->data_read
= 1;
2403 * If we read high prio channel in this loop, try again
2404 * for more high prio data.
2410 /* Take care of low priority channels. */
2411 for (i
= 0; i
< nb_fd
; i
++) {
2412 if (local_stream
[i
] == NULL
) {
2415 if ((pollfd
[i
].revents
& POLLIN
) ||
2416 local_stream
[i
]->hangup_flush_done
) {
2417 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2418 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2419 /* it's ok to have an unavailable sub-buffer */
2420 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2421 /* Clean the stream and free it. */
2422 consumer_del_stream(local_stream
[i
], data_ht
);
2423 local_stream
[i
] = NULL
;
2424 } else if (len
> 0) {
2425 local_stream
[i
]->data_read
= 1;
2430 /* Handle hangup and errors */
2431 for (i
= 0; i
< nb_fd
; i
++) {
2432 if (local_stream
[i
] == NULL
) {
2435 if (!local_stream
[i
]->hangup_flush_done
2436 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2437 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2438 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2439 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2441 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2442 /* Attempt read again, for the data we just flushed. */
2443 local_stream
[i
]->data_read
= 1;
2446 * If the poll flag is HUP/ERR/NVAL and we have
2447 * read no data in this pass, we can remove the
2448 * stream from its hash table.
2450 if ((pollfd
[i
].revents
& POLLHUP
)) {
2451 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2452 if (!local_stream
[i
]->data_read
) {
2453 consumer_del_stream(local_stream
[i
], data_ht
);
2454 local_stream
[i
] = NULL
;
2457 } else if (pollfd
[i
].revents
& POLLERR
) {
2458 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2459 if (!local_stream
[i
]->data_read
) {
2460 consumer_del_stream(local_stream
[i
], data_ht
);
2461 local_stream
[i
] = NULL
;
2464 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2465 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2466 if (!local_stream
[i
]->data_read
) {
2467 consumer_del_stream(local_stream
[i
], data_ht
);
2468 local_stream
[i
] = NULL
;
2472 if (local_stream
[i
] != NULL
) {
2473 local_stream
[i
]->data_read
= 0;
2478 DBG("polling thread exiting");
2483 * Close the write side of the pipe so epoll_wait() in
2484 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2485 * read side of the pipe. If we close them both, epoll_wait strangely does
2486 * not return and could create a endless wait period if the pipe is the
2487 * only tracked fd in the poll set. The thread will take care of closing
2490 ret
= close(ctx
->consumer_metadata_pipe
[1]);
2492 PERROR("close data pipe");
2495 destroy_data_stream_ht(data_ht
);
2497 rcu_unregister_thread();
2502 * Close wake-up end of each stream belonging to the channel. This will
2503 * allow the poll() on the stream read-side to detect when the
2504 * write-side (application) finally closes them.
2507 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2509 struct lttng_ht
*ht
;
2510 struct lttng_consumer_stream
*stream
;
2511 struct lttng_ht_iter iter
;
2513 ht
= consumer_data
.stream_per_chan_id_ht
;
2516 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2517 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2518 ht
->match_fct
, &channel
->key
,
2519 &iter
.iter
, stream
, node_channel_id
.node
) {
2520 switch (consumer_data
.type
) {
2521 case LTTNG_CONSUMER_KERNEL
:
2523 case LTTNG_CONSUMER32_UST
:
2524 case LTTNG_CONSUMER64_UST
:
2526 * Note: a mutex is taken internally within
2527 * liblttng-ust-ctl to protect timer wakeup_fd
2528 * use from concurrent close.
2530 lttng_ustconsumer_close_stream_wakeup(stream
);
2533 ERR("Unknown consumer_data type");
2540 static void destroy_channel_ht(struct lttng_ht
*ht
)
2542 struct lttng_ht_iter iter
;
2543 struct lttng_consumer_channel
*channel
;
2551 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2552 ret
= lttng_ht_del(ht
, &iter
);
2557 lttng_ht_destroy(ht
);
2561 * This thread polls the channel fds to detect when they are being
2562 * closed. It closes all related streams if the channel is detected as
2563 * closed. It is currently only used as a shim layer for UST because the
2564 * consumerd needs to keep the per-stream wakeup end of pipes open for
2567 void *consumer_thread_channel_poll(void *data
)
2570 uint32_t revents
, nb_fd
;
2571 struct lttng_consumer_channel
*chan
= NULL
;
2572 struct lttng_ht_iter iter
;
2573 struct lttng_ht_node_u64
*node
;
2574 struct lttng_poll_event events
;
2575 struct lttng_consumer_local_data
*ctx
= data
;
2576 struct lttng_ht
*channel_ht
;
2578 rcu_register_thread();
2580 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2582 /* ENOMEM at this point. Better to bail out. */
2586 DBG("Thread channel poll started");
2588 /* Size is set to 1 for the consumer_channel pipe */
2589 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2591 ERR("Poll set creation failed");
2595 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2601 DBG("Channel main loop started");
2604 /* Only the channel pipe is set */
2605 if (LTTNG_POLL_GETNB(&events
) == 0 && consumer_quit
== 1) {
2610 DBG("Channel poll wait with %d fd(s)", LTTNG_POLL_GETNB(&events
));
2611 ret
= lttng_poll_wait(&events
, -1);
2612 DBG("Channel event catched in thread");
2614 if (errno
== EINTR
) {
2615 ERR("Poll EINTR catched");
2623 /* From here, the event is a channel wait fd */
2624 for (i
= 0; i
< nb_fd
; i
++) {
2625 revents
= LTTNG_POLL_GETEV(&events
, i
);
2626 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2628 /* Just don't waste time if no returned events for the fd */
2632 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2633 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2634 DBG("Channel thread pipe hung up");
2636 * Remove the pipe from the poll set and continue the loop
2637 * since their might be data to consume.
2639 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2641 } else if (revents
& LPOLLIN
) {
2642 enum consumer_channel_action action
;
2644 ret
= read_channel_pipe(ctx
, &chan
, &action
);
2646 ERR("Error reading channel pipe");
2651 case CONSUMER_CHANNEL_ADD
:
2652 DBG("Adding channel %d to poll set",
2655 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
2657 lttng_ht_add_unique_u64(channel_ht
,
2658 &chan
->wait_fd_node
);
2659 /* Add channel to the global poll events list */
2660 lttng_poll_add(&events
, chan
->wait_fd
,
2661 LPOLLIN
| LPOLLPRI
);
2663 case CONSUMER_CHANNEL_QUIT
:
2665 * Remove the pipe from the poll set and continue the loop
2666 * since their might be data to consume.
2668 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2671 ERR("Unknown action");
2676 /* Handle other stream */
2682 uint64_t tmp_id
= (uint64_t) pollfd
;
2684 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
2686 node
= lttng_ht_iter_get_node_u64(&iter
);
2689 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
2692 /* Check for error event */
2693 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2694 DBG("Channel fd %d is hup|err.", pollfd
);
2696 lttng_poll_del(&events
, chan
->wait_fd
);
2697 ret
= lttng_ht_del(channel_ht
, &iter
);
2699 consumer_close_channel_streams(chan
);
2702 /* Release RCU lock for the channel looked up */
2708 lttng_poll_clean(&events
);
2710 destroy_channel_ht(channel_ht
);
2712 DBG("Channel poll thread exiting");
2713 rcu_unregister_thread();
2717 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
2718 struct pollfd
*sockpoll
, int client_socket
)
2725 if (lttng_consumer_poll_socket(sockpoll
) < 0) {
2729 DBG("Metadata connection on client_socket");
2731 /* Blocking call, waiting for transmission */
2732 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
2733 if (ctx
->consumer_metadata_socket
< 0) {
2734 WARN("On accept metadata");
2745 * This thread listens on the consumerd socket and receives the file
2746 * descriptors from the session daemon.
2748 void *consumer_thread_sessiond_poll(void *data
)
2750 int sock
= -1, client_socket
, ret
;
2752 * structure to poll for incoming data on communication socket avoids
2753 * making blocking sockets.
2755 struct pollfd consumer_sockpoll
[2];
2756 struct lttng_consumer_local_data
*ctx
= data
;
2758 rcu_register_thread();
2760 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
2761 unlink(ctx
->consumer_command_sock_path
);
2762 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
2763 if (client_socket
< 0) {
2764 ERR("Cannot create command socket");
2768 ret
= lttcomm_listen_unix_sock(client_socket
);
2773 DBG("Sending ready command to lttng-sessiond");
2774 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
2775 /* return < 0 on error, but == 0 is not fatal */
2777 ERR("Error sending ready command to lttng-sessiond");
2781 ret
= fcntl(client_socket
, F_SETFL
, O_NONBLOCK
);
2783 PERROR("fcntl O_NONBLOCK");
2787 /* prepare the FDs to poll : to client socket and the should_quit pipe */
2788 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
2789 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
2790 consumer_sockpoll
[1].fd
= client_socket
;
2791 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
2793 if (lttng_consumer_poll_socket(consumer_sockpoll
) < 0) {
2796 DBG("Connection on client_socket");
2798 /* Blocking call, waiting for transmission */
2799 sock
= lttcomm_accept_unix_sock(client_socket
);
2804 ret
= fcntl(sock
, F_SETFL
, O_NONBLOCK
);
2806 PERROR("fcntl O_NONBLOCK");
2811 * Setup metadata socket which is the second socket connection on the
2812 * command unix socket.
2814 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
2819 /* This socket is not useful anymore. */
2820 ret
= close(client_socket
);
2822 PERROR("close client_socket");
2826 /* update the polling structure to poll on the established socket */
2827 consumer_sockpoll
[1].fd
= sock
;
2828 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
2831 if (lttng_consumer_poll_socket(consumer_sockpoll
) < 0) {
2834 DBG("Incoming command on sock");
2835 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2836 if (ret
== -ENOENT
) {
2837 DBG("Received STOP command");
2842 * This could simply be a session daemon quitting. Don't output
2845 DBG("Communication interrupted on command socket");
2848 if (consumer_quit
) {
2849 DBG("consumer_thread_receive_fds received quit from signal");
2852 DBG("received command on sock");
2855 DBG("Consumer thread sessiond poll exiting");
2858 * Close metadata streams since the producer is the session daemon which
2861 * NOTE: for now, this only applies to the UST tracer.
2863 lttng_consumer_close_metadata();
2866 * when all fds have hung up, the polling thread
2872 * Notify the data poll thread to poll back again and test the
2873 * consumer_quit state that we just set so to quit gracefully.
2875 notify_thread_pipe(ctx
->consumer_data_pipe
[1]);
2877 notify_channel_pipe(ctx
, NULL
, CONSUMER_CHANNEL_QUIT
);
2879 /* Cleaning up possibly open sockets. */
2883 PERROR("close sock sessiond poll");
2886 if (client_socket
>= 0) {
2889 PERROR("close client_socket sessiond poll");
2893 rcu_unregister_thread();
2897 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
2898 struct lttng_consumer_local_data
*ctx
)
2902 pthread_mutex_lock(&stream
->lock
);
2904 switch (consumer_data
.type
) {
2905 case LTTNG_CONSUMER_KERNEL
:
2906 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
2908 case LTTNG_CONSUMER32_UST
:
2909 case LTTNG_CONSUMER64_UST
:
2910 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
2913 ERR("Unknown consumer_data type");
2919 pthread_mutex_unlock(&stream
->lock
);
2923 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
2925 switch (consumer_data
.type
) {
2926 case LTTNG_CONSUMER_KERNEL
:
2927 return lttng_kconsumer_on_recv_stream(stream
);
2928 case LTTNG_CONSUMER32_UST
:
2929 case LTTNG_CONSUMER64_UST
:
2930 return lttng_ustconsumer_on_recv_stream(stream
);
2932 ERR("Unknown consumer_data type");
2939 * Allocate and set consumer data hash tables.
2941 void lttng_consumer_init(void)
2943 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2944 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2945 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2946 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2950 * Process the ADD_RELAYD command receive by a consumer.
2952 * This will create a relayd socket pair and add it to the relayd hash table.
2953 * The caller MUST acquire a RCU read side lock before calling it.
2955 int consumer_add_relayd_socket(int net_seq_idx
, int sock_type
,
2956 struct lttng_consumer_local_data
*ctx
, int sock
,
2957 struct pollfd
*consumer_sockpoll
, struct lttcomm_sock
*relayd_sock
,
2958 unsigned int sessiond_id
)
2960 int fd
= -1, ret
= -1, relayd_created
= 0;
2961 enum lttng_error_code ret_code
= LTTNG_OK
;
2962 struct consumer_relayd_sock_pair
*relayd
;
2964 DBG("Consumer adding relayd socket (idx: %d)", net_seq_idx
);
2966 /* First send a status message before receiving the fds. */
2967 ret
= consumer_send_status_msg(sock
, ret_code
);
2969 /* Somehow, the session daemon is not responding anymore. */
2973 /* Get relayd reference if exists. */
2974 relayd
= consumer_find_relayd(net_seq_idx
);
2975 if (relayd
== NULL
) {
2976 /* Not found. Allocate one. */
2977 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
2978 if (relayd
== NULL
) {
2979 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_OUTFD_ERROR
);
2983 relayd
->sessiond_session_id
= (uint64_t) sessiond_id
;
2987 /* Poll on consumer socket. */
2988 if (lttng_consumer_poll_socket(consumer_sockpoll
) < 0) {
2993 /* Get relayd socket from session daemon */
2994 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
2995 if (ret
!= sizeof(fd
)) {
2996 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
2998 fd
= -1; /* Just in case it gets set with an invalid value. */
3002 /* We have the fds without error. Send status back. */
3003 ret
= consumer_send_status_msg(sock
, ret_code
);
3005 /* Somehow, the session daemon is not responding anymore. */
3009 /* Copy socket information and received FD */
3010 switch (sock_type
) {
3011 case LTTNG_STREAM_CONTROL
:
3012 /* Copy received lttcomm socket */
3013 lttcomm_copy_sock(&relayd
->control_sock
, relayd_sock
);
3014 ret
= lttcomm_create_sock(&relayd
->control_sock
);
3015 /* Immediately try to close the created socket if valid. */
3016 if (relayd
->control_sock
.fd
>= 0) {
3017 if (close(relayd
->control_sock
.fd
)) {
3018 PERROR("close relayd control socket");
3021 /* Handle create_sock error. */
3026 /* Assign new file descriptor */
3027 relayd
->control_sock
.fd
= fd
;
3030 * Create a session on the relayd and store the returned id. Lock the
3031 * control socket mutex if the relayd was NOT created before.
3033 if (!relayd_created
) {
3034 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3036 ret
= relayd_create_session(&relayd
->control_sock
,
3037 &relayd
->relayd_session_id
);
3038 if (!relayd_created
) {
3039 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3043 * Close all sockets of a relayd object. It will be freed if it was
3044 * created at the error code path or else it will be garbage
3047 (void) relayd_close(&relayd
->control_sock
);
3048 (void) relayd_close(&relayd
->data_sock
);
3053 case LTTNG_STREAM_DATA
:
3054 /* Copy received lttcomm socket */
3055 lttcomm_copy_sock(&relayd
->data_sock
, relayd_sock
);
3056 ret
= lttcomm_create_sock(&relayd
->data_sock
);
3057 /* Immediately try to close the created socket if valid. */
3058 if (relayd
->data_sock
.fd
>= 0) {
3059 if (close(relayd
->data_sock
.fd
)) {
3060 PERROR("close relayd data socket");
3063 /* Handle create_sock error. */
3068 /* Assign new file descriptor */
3069 relayd
->data_sock
.fd
= fd
;
3072 ERR("Unknown relayd socket type (%d)", sock_type
);
3077 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3078 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3079 relayd
->net_seq_idx
, fd
);
3082 * Add relayd socket pair to consumer data hashtable. If object already
3083 * exists or on error, the function gracefully returns.
3091 /* Close received socket if valid. */
3094 PERROR("close received socket");
3099 if (relayd_created
) {
3107 * Try to lock the stream mutex.
3109 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
3111 static int stream_try_lock(struct lttng_consumer_stream
*stream
)
3118 * Try to lock the stream mutex. On failure, we know that the stream is
3119 * being used else where hence there is data still being extracted.
3121 ret
= pthread_mutex_trylock(&stream
->lock
);
3123 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
3135 * Search for a relayd associated to the session id and return the reference.
3137 * A rcu read side lock MUST be acquire before calling this function and locked
3138 * until the relayd object is no longer necessary.
3140 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3142 struct lttng_ht_iter iter
;
3143 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3145 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3146 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3149 * Check by sessiond id which is unique here where the relayd session
3150 * id might not be when having multiple relayd.
3152 if (relayd
->sessiond_session_id
== id
) {
3153 /* Found the relayd. There can be only one per id. */
3165 * Check if for a given session id there is still data needed to be extract
3168 * Return 1 if data is pending or else 0 meaning ready to be read.
3170 int consumer_data_pending(uint64_t id
)
3173 struct lttng_ht_iter iter
;
3174 struct lttng_ht
*ht
;
3175 struct lttng_consumer_stream
*stream
;
3176 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3177 int (*data_pending
)(struct lttng_consumer_stream
*);
3179 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3182 pthread_mutex_lock(&consumer_data
.lock
);
3184 switch (consumer_data
.type
) {
3185 case LTTNG_CONSUMER_KERNEL
:
3186 data_pending
= lttng_kconsumer_data_pending
;
3188 case LTTNG_CONSUMER32_UST
:
3189 case LTTNG_CONSUMER64_UST
:
3190 data_pending
= lttng_ustconsumer_data_pending
;
3193 ERR("Unknown consumer data type");
3197 /* Ease our life a bit */
3198 ht
= consumer_data
.stream_list_ht
;
3200 relayd
= find_relayd_by_session_id(id
);
3202 /* Send init command for data pending. */
3203 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3204 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3205 relayd
->relayd_session_id
);
3206 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3208 /* Communication error thus the relayd so no data pending. */
3209 goto data_not_pending
;
3213 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3214 ht
->hash_fct(&id
, lttng_ht_seed
),
3216 &iter
.iter
, stream
, node_session_id
.node
) {
3217 /* If this call fails, the stream is being used hence data pending. */
3218 ret
= stream_try_lock(stream
);
3224 * A removed node from the hash table indicates that the stream has
3225 * been deleted thus having a guarantee that the buffers are closed
3226 * on the consumer side. However, data can still be transmitted
3227 * over the network so don't skip the relayd check.
3229 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3231 /* Check the stream if there is data in the buffers. */
3232 ret
= data_pending(stream
);
3234 pthread_mutex_unlock(&stream
->lock
);
3241 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3242 if (stream
->metadata_flag
) {
3243 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3244 stream
->relayd_stream_id
);
3246 ret
= relayd_data_pending(&relayd
->control_sock
,
3247 stream
->relayd_stream_id
,
3248 stream
->next_net_seq_num
- 1);
3250 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3252 pthread_mutex_unlock(&stream
->lock
);
3256 pthread_mutex_unlock(&stream
->lock
);
3260 unsigned int is_data_inflight
= 0;
3262 /* Send init command for data pending. */
3263 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3264 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3265 relayd
->relayd_session_id
, &is_data_inflight
);
3266 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3268 goto data_not_pending
;
3270 if (is_data_inflight
) {
3276 * Finding _no_ node in the hash table and no inflight data means that the
3277 * stream(s) have been removed thus data is guaranteed to be available for
3278 * analysis from the trace files.
3282 /* Data is available to be read by a viewer. */
3283 pthread_mutex_unlock(&consumer_data
.lock
);
3288 /* Data is still being extracted from buffers. */
3289 pthread_mutex_unlock(&consumer_data
.lock
);
3295 * Send a ret code status message to the sessiond daemon.
3297 * Return the sendmsg() return value.
3299 int consumer_send_status_msg(int sock
, int ret_code
)
3301 struct lttcomm_consumer_status_msg msg
;
3303 msg
.ret_code
= ret_code
;
3305 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3309 * Send a channel status message to the sessiond daemon.
3311 * Return the sendmsg() return value.
3313 int consumer_send_status_channel(int sock
,
3314 struct lttng_consumer_channel
*channel
)
3316 struct lttcomm_consumer_status_channel msg
;
3321 msg
.ret_code
= -LTTNG_ERR_UST_CHAN_FAIL
;
3323 msg
.ret_code
= LTTNG_OK
;
3324 msg
.key
= channel
->key
;
3325 msg
.stream_count
= channel
->streams
.count
;
3328 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));