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
{
54 CONSUMER_CHANNEL_QUIT
,
57 struct consumer_channel_msg
{
58 enum consumer_channel_action action
;
59 struct lttng_consumer_channel
*chan
; /* add */
60 uint64_t key
; /* del */
64 * Flag to inform the polling thread to quit when all fd hung up. Updated by
65 * the consumer_thread_receive_fds when it notices that all fds has hung up.
66 * Also updated by the signal handler (consumer_should_exit()). Read by the
69 volatile int consumer_quit
;
72 * Global hash table containing respectively metadata and data streams. The
73 * stream element in this ht should only be updated by the metadata poll thread
74 * for the metadata and the data poll thread for the data.
76 static struct lttng_ht
*metadata_ht
;
77 static struct lttng_ht
*data_ht
;
80 * Notify a thread pipe to poll back again. This usually means that some global
81 * state has changed so we just send back the thread in a poll wait call.
83 static void notify_thread_pipe(int wpipe
)
88 struct lttng_consumer_stream
*null_stream
= NULL
;
90 ret
= write(wpipe
, &null_stream
, sizeof(null_stream
));
91 } while (ret
< 0 && errno
== EINTR
);
94 static void notify_channel_pipe(struct lttng_consumer_local_data
*ctx
,
95 struct lttng_consumer_channel
*chan
,
97 enum consumer_channel_action action
)
99 struct consumer_channel_msg msg
;
102 memset(&msg
, 0, sizeof(msg
));
107 ret
= write(ctx
->consumer_channel_pipe
[1], &msg
, sizeof(msg
));
108 } while (ret
< 0 && errno
== EINTR
);
111 void notify_thread_del_channel(struct lttng_consumer_local_data
*ctx
,
114 notify_channel_pipe(ctx
, NULL
, key
, CONSUMER_CHANNEL_DEL
);
117 static int read_channel_pipe(struct lttng_consumer_local_data
*ctx
,
118 struct lttng_consumer_channel
**chan
,
120 enum consumer_channel_action
*action
)
122 struct consumer_channel_msg msg
;
126 ret
= read(ctx
->consumer_channel_pipe
[0], &msg
, sizeof(msg
));
127 } while (ret
< 0 && errno
== EINTR
);
129 *action
= msg
.action
;
137 * Find a stream. The consumer_data.lock must be locked during this
140 static struct lttng_consumer_stream
*find_stream(uint64_t key
,
143 struct lttng_ht_iter iter
;
144 struct lttng_ht_node_u64
*node
;
145 struct lttng_consumer_stream
*stream
= NULL
;
149 /* -1ULL keys are lookup failures */
150 if (key
== (uint64_t) -1ULL) {
156 lttng_ht_lookup(ht
, &key
, &iter
);
157 node
= lttng_ht_iter_get_node_u64(&iter
);
159 stream
= caa_container_of(node
, struct lttng_consumer_stream
, node
);
167 static void steal_stream_key(int key
, struct lttng_ht
*ht
)
169 struct lttng_consumer_stream
*stream
;
172 stream
= find_stream(key
, ht
);
176 * We don't want the lookup to match, but we still need
177 * to iterate on this stream when iterating over the hash table. Just
178 * change the node key.
180 stream
->node
.key
= -1ULL;
186 * Return a channel object for the given key.
188 * RCU read side lock MUST be acquired before calling this function and
189 * protects the channel ptr.
191 struct lttng_consumer_channel
*consumer_find_channel(uint64_t key
)
193 struct lttng_ht_iter iter
;
194 struct lttng_ht_node_u64
*node
;
195 struct lttng_consumer_channel
*channel
= NULL
;
197 /* -1ULL keys are lookup failures */
198 if (key
== (uint64_t) -1ULL) {
202 lttng_ht_lookup(consumer_data
.channel_ht
, &key
, &iter
);
203 node
= lttng_ht_iter_get_node_u64(&iter
);
205 channel
= caa_container_of(node
, struct lttng_consumer_channel
, node
);
211 static void free_stream_rcu(struct rcu_head
*head
)
213 struct lttng_ht_node_u64
*node
=
214 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
215 struct lttng_consumer_stream
*stream
=
216 caa_container_of(node
, struct lttng_consumer_stream
, node
);
221 static void free_channel_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 lttng_consumer_channel
*channel
=
226 caa_container_of(node
, struct lttng_consumer_channel
, node
);
232 * RCU protected relayd socket pair free.
234 static void free_relayd_rcu(struct rcu_head
*head
)
236 struct lttng_ht_node_u64
*node
=
237 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
238 struct consumer_relayd_sock_pair
*relayd
=
239 caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
242 * Close all sockets. This is done in the call RCU since we don't want the
243 * socket fds to be reassigned thus potentially creating bad state of the
246 * We do not have to lock the control socket mutex here since at this stage
247 * there is no one referencing to this relayd object.
249 (void) relayd_close(&relayd
->control_sock
);
250 (void) relayd_close(&relayd
->data_sock
);
256 * Destroy and free relayd socket pair object.
258 * This function MUST be called with the consumer_data lock acquired.
260 static void destroy_relayd(struct consumer_relayd_sock_pair
*relayd
)
263 struct lttng_ht_iter iter
;
265 if (relayd
== NULL
) {
269 DBG("Consumer destroy and close relayd socket pair");
271 iter
.iter
.node
= &relayd
->node
.node
;
272 ret
= lttng_ht_del(consumer_data
.relayd_ht
, &iter
);
274 /* We assume the relayd is being or is destroyed */
278 /* RCU free() call */
279 call_rcu(&relayd
->node
.head
, free_relayd_rcu
);
283 * Remove a channel from the global list protected by a mutex. This function is
284 * also responsible for freeing its data structures.
286 void consumer_del_channel(struct lttng_consumer_channel
*channel
)
289 struct lttng_ht_iter iter
;
291 DBG("Consumer delete channel key %" PRIu64
, channel
->key
);
293 pthread_mutex_lock(&consumer_data
.lock
);
295 switch (consumer_data
.type
) {
296 case LTTNG_CONSUMER_KERNEL
:
298 case LTTNG_CONSUMER32_UST
:
299 case LTTNG_CONSUMER64_UST
:
300 lttng_ustconsumer_del_channel(channel
);
303 ERR("Unknown consumer_data type");
309 iter
.iter
.node
= &channel
->node
.node
;
310 ret
= lttng_ht_del(consumer_data
.channel_ht
, &iter
);
314 call_rcu(&channel
->node
.head
, free_channel_rcu
);
316 pthread_mutex_unlock(&consumer_data
.lock
);
320 * Iterate over the relayd hash table and destroy each element. Finally,
321 * destroy the whole hash table.
323 static void cleanup_relayd_ht(void)
325 struct lttng_ht_iter iter
;
326 struct consumer_relayd_sock_pair
*relayd
;
330 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
332 destroy_relayd(relayd
);
337 lttng_ht_destroy(consumer_data
.relayd_ht
);
341 * Update the end point status of all streams having the given network sequence
342 * index (relayd index).
344 * It's atomically set without having the stream mutex locked which is fine
345 * because we handle the write/read race with a pipe wakeup for each thread.
347 static void update_endpoint_status_by_netidx(int net_seq_idx
,
348 enum consumer_endpoint_status status
)
350 struct lttng_ht_iter iter
;
351 struct lttng_consumer_stream
*stream
;
353 DBG("Consumer set delete flag on stream by idx %d", net_seq_idx
);
357 /* Let's begin with metadata */
358 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
359 if (stream
->net_seq_idx
== net_seq_idx
) {
360 uatomic_set(&stream
->endpoint_status
, status
);
361 DBG("Delete flag set to metadata stream %d", stream
->wait_fd
);
365 /* Follow up by the data streams */
366 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
367 if (stream
->net_seq_idx
== net_seq_idx
) {
368 uatomic_set(&stream
->endpoint_status
, status
);
369 DBG("Delete flag set to data stream %d", stream
->wait_fd
);
376 * Cleanup a relayd object by flagging every associated streams for deletion,
377 * destroying the object meaning removing it from the relayd hash table,
378 * closing the sockets and freeing the memory in a RCU call.
380 * If a local data context is available, notify the threads that the streams'
381 * state have changed.
383 static void cleanup_relayd(struct consumer_relayd_sock_pair
*relayd
,
384 struct lttng_consumer_local_data
*ctx
)
390 DBG("Cleaning up relayd sockets");
392 /* Save the net sequence index before destroying the object */
393 netidx
= relayd
->net_seq_idx
;
396 * Delete the relayd from the relayd hash table, close the sockets and free
397 * the object in a RCU call.
399 destroy_relayd(relayd
);
401 /* Set inactive endpoint to all streams */
402 update_endpoint_status_by_netidx(netidx
, CONSUMER_ENDPOINT_INACTIVE
);
405 * With a local data context, notify the threads that the streams' state
406 * have changed. The write() action on the pipe acts as an "implicit"
407 * memory barrier ordering the updates of the end point status from the
408 * read of this status which happens AFTER receiving this notify.
411 notify_thread_pipe(ctx
->consumer_data_pipe
[1]);
412 notify_thread_pipe(ctx
->consumer_metadata_pipe
[1]);
417 * Flag a relayd socket pair for destruction. Destroy it if the refcount
420 * RCU read side lock MUST be aquired before calling this function.
422 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair
*relayd
)
426 /* Set destroy flag for this object */
427 uatomic_set(&relayd
->destroy_flag
, 1);
429 /* Destroy the relayd if refcount is 0 */
430 if (uatomic_read(&relayd
->refcount
) == 0) {
431 destroy_relayd(relayd
);
436 * Remove a stream from the global list protected by a mutex. This
437 * function is also responsible for freeing its data structures.
439 void consumer_del_stream(struct lttng_consumer_stream
*stream
,
443 struct lttng_ht_iter iter
;
444 struct lttng_consumer_channel
*free_chan
= NULL
;
445 struct consumer_relayd_sock_pair
*relayd
;
449 DBG("Consumer del stream %d", stream
->wait_fd
);
452 /* Means the stream was allocated but not successfully added */
453 goto free_stream_rcu
;
456 pthread_mutex_lock(&consumer_data
.lock
);
457 pthread_mutex_lock(&stream
->lock
);
459 switch (consumer_data
.type
) {
460 case LTTNG_CONSUMER_KERNEL
:
461 if (stream
->mmap_base
!= NULL
) {
462 ret
= munmap(stream
->mmap_base
, stream
->mmap_len
);
468 case LTTNG_CONSUMER32_UST
:
469 case LTTNG_CONSUMER64_UST
:
470 lttng_ustconsumer_del_stream(stream
);
473 ERR("Unknown consumer_data type");
479 iter
.iter
.node
= &stream
->node
.node
;
480 ret
= lttng_ht_del(ht
, &iter
);
483 iter
.iter
.node
= &stream
->node_channel_id
.node
;
484 ret
= lttng_ht_del(consumer_data
.stream_per_chan_id_ht
, &iter
);
487 iter
.iter
.node
= &stream
->node_session_id
.node
;
488 ret
= lttng_ht_del(consumer_data
.stream_list_ht
, &iter
);
492 assert(consumer_data
.stream_count
> 0);
493 consumer_data
.stream_count
--;
495 if (stream
->out_fd
>= 0) {
496 ret
= close(stream
->out_fd
);
502 /* Check and cleanup relayd */
504 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
505 if (relayd
!= NULL
) {
506 uatomic_dec(&relayd
->refcount
);
507 assert(uatomic_read(&relayd
->refcount
) >= 0);
509 /* Closing streams requires to lock the control socket. */
510 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
511 ret
= relayd_send_close_stream(&relayd
->control_sock
,
512 stream
->relayd_stream_id
,
513 stream
->next_net_seq_num
- 1);
514 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
516 DBG("Unable to close stream on the relayd. Continuing");
518 * Continue here. There is nothing we can do for the relayd.
519 * Chances are that the relayd has closed the socket so we just
520 * continue cleaning up.
524 /* Both conditions are met, we destroy the relayd. */
525 if (uatomic_read(&relayd
->refcount
) == 0 &&
526 uatomic_read(&relayd
->destroy_flag
)) {
527 destroy_relayd(relayd
);
532 if (!uatomic_sub_return(&stream
->chan
->refcount
, 1)
533 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
534 free_chan
= stream
->chan
;
538 consumer_data
.need_update
= 1;
539 pthread_mutex_unlock(&stream
->lock
);
540 pthread_mutex_unlock(&consumer_data
.lock
);
543 consumer_del_channel(free_chan
);
547 call_rcu(&stream
->node
.head
, free_stream_rcu
);
550 struct lttng_consumer_stream
*consumer_allocate_stream(uint64_t channel_key
,
552 enum lttng_consumer_stream_state state
,
553 const char *channel_name
,
560 enum consumer_channel_type type
)
563 struct lttng_consumer_stream
*stream
;
565 stream
= zmalloc(sizeof(*stream
));
566 if (stream
== NULL
) {
567 PERROR("malloc struct lttng_consumer_stream");
574 stream
->key
= stream_key
;
576 stream
->out_fd_offset
= 0;
577 stream
->state
= state
;
580 stream
->net_seq_idx
= relayd_id
;
581 stream
->session_id
= session_id
;
582 pthread_mutex_init(&stream
->lock
, NULL
);
584 /* If channel is the metadata, flag this stream as metadata. */
585 if (type
== CONSUMER_CHANNEL_TYPE_METADATA
) {
586 stream
->metadata_flag
= 1;
587 /* Metadata is flat out. */
588 strncpy(stream
->name
, DEFAULT_METADATA_NAME
, sizeof(stream
->name
));
590 /* Format stream name to <channel_name>_<cpu_number> */
591 ret
= snprintf(stream
->name
, sizeof(stream
->name
), "%s_%d",
594 PERROR("snprintf stream name");
599 /* Key is always the wait_fd for streams. */
600 lttng_ht_node_init_u64(&stream
->node
, stream
->key
);
602 /* Init node per channel id key */
603 lttng_ht_node_init_u64(&stream
->node_channel_id
, channel_key
);
605 /* Init session id node with the stream session id */
606 lttng_ht_node_init_u64(&stream
->node_session_id
, stream
->session_id
);
608 DBG3("Allocated stream %s (key %" PRIu64
", chan_key %" PRIu64
" relayd_id %" PRIu64
", session_id %" PRIu64
,
609 stream
->name
, stream
->key
, channel_key
, stream
->net_seq_idx
, stream
->session_id
);
625 * Add a stream to the global list protected by a mutex.
627 static int add_stream(struct lttng_consumer_stream
*stream
,
631 struct consumer_relayd_sock_pair
*relayd
;
636 DBG3("Adding consumer stream %" PRIu64
, stream
->key
);
638 pthread_mutex_lock(&consumer_data
.lock
);
639 pthread_mutex_lock(&stream
->lock
);
642 /* Steal stream identifier to avoid having streams with the same key */
643 steal_stream_key(stream
->key
, ht
);
645 lttng_ht_add_unique_u64(ht
, &stream
->node
);
647 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
648 &stream
->node_channel_id
);
651 * Add stream to the stream_list_ht of the consumer data. No need to steal
652 * the key since the HT does not use it and we allow to add redundant keys
655 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
657 /* Check and cleanup relayd */
658 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
659 if (relayd
!= NULL
) {
660 uatomic_inc(&relayd
->refcount
);
664 * When nb_init_stream_left reaches 0, we don't need to trigger any action
665 * in terms of destroying the associated channel, because the action that
666 * causes the count to become 0 also causes a stream to be added. The
667 * channel deletion will thus be triggered by the following removal of this
670 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
671 /* Increment refcount before decrementing nb_init_stream_left */
673 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
676 /* Update consumer data once the node is inserted. */
677 consumer_data
.stream_count
++;
678 consumer_data
.need_update
= 1;
681 pthread_mutex_unlock(&stream
->lock
);
682 pthread_mutex_unlock(&consumer_data
.lock
);
688 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
689 * be acquired before calling this.
691 static int add_relayd(struct consumer_relayd_sock_pair
*relayd
)
694 struct lttng_ht_node_u64
*node
;
695 struct lttng_ht_iter iter
;
699 lttng_ht_lookup(consumer_data
.relayd_ht
,
700 &relayd
->net_seq_idx
, &iter
);
701 node
= lttng_ht_iter_get_node_u64(&iter
);
705 lttng_ht_add_unique_u64(consumer_data
.relayd_ht
, &relayd
->node
);
712 * Allocate and return a consumer relayd socket.
714 struct consumer_relayd_sock_pair
*consumer_allocate_relayd_sock_pair(
717 struct consumer_relayd_sock_pair
*obj
= NULL
;
719 /* Negative net sequence index is a failure */
720 if (net_seq_idx
< 0) {
724 obj
= zmalloc(sizeof(struct consumer_relayd_sock_pair
));
726 PERROR("zmalloc relayd sock");
730 obj
->net_seq_idx
= net_seq_idx
;
732 obj
->destroy_flag
= 0;
733 obj
->control_sock
.sock
.fd
= -1;
734 obj
->data_sock
.sock
.fd
= -1;
735 lttng_ht_node_init_u64(&obj
->node
, obj
->net_seq_idx
);
736 pthread_mutex_init(&obj
->ctrl_sock_mutex
, NULL
);
743 * Find a relayd socket pair in the global consumer data.
745 * Return the object if found else NULL.
746 * RCU read-side lock must be held across this call and while using the
749 struct consumer_relayd_sock_pair
*consumer_find_relayd(uint64_t key
)
751 struct lttng_ht_iter iter
;
752 struct lttng_ht_node_u64
*node
;
753 struct consumer_relayd_sock_pair
*relayd
= NULL
;
755 /* Negative keys are lookup failures */
756 if (key
== (uint64_t) -1ULL) {
760 lttng_ht_lookup(consumer_data
.relayd_ht
, &key
,
762 node
= lttng_ht_iter_get_node_u64(&iter
);
764 relayd
= caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
772 * Handle stream for relayd transmission if the stream applies for network
773 * streaming where the net sequence index is set.
775 * Return destination file descriptor or negative value on error.
777 static int write_relayd_stream_header(struct lttng_consumer_stream
*stream
,
778 size_t data_size
, unsigned long padding
,
779 struct consumer_relayd_sock_pair
*relayd
)
782 struct lttcomm_relayd_data_hdr data_hdr
;
788 /* Reset data header */
789 memset(&data_hdr
, 0, sizeof(data_hdr
));
791 if (stream
->metadata_flag
) {
792 /* Caller MUST acquire the relayd control socket lock */
793 ret
= relayd_send_metadata(&relayd
->control_sock
, data_size
);
798 /* Metadata are always sent on the control socket. */
799 outfd
= relayd
->control_sock
.sock
.fd
;
801 /* Set header with stream information */
802 data_hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
803 data_hdr
.data_size
= htobe32(data_size
);
804 data_hdr
.padding_size
= htobe32(padding
);
806 * Note that net_seq_num below is assigned with the *current* value of
807 * next_net_seq_num and only after that the next_net_seq_num will be
808 * increment. This is why when issuing a command on the relayd using
809 * this next value, 1 should always be substracted in order to compare
810 * the last seen sequence number on the relayd side to the last sent.
812 data_hdr
.net_seq_num
= htobe64(stream
->next_net_seq_num
);
813 /* Other fields are zeroed previously */
815 ret
= relayd_send_data_hdr(&relayd
->data_sock
, &data_hdr
,
821 ++stream
->next_net_seq_num
;
823 /* Set to go on data socket */
824 outfd
= relayd
->data_sock
.sock
.fd
;
832 * Allocate and return a new lttng_consumer_channel object using the given key
833 * to initialize the hash table node.
835 * On error, return NULL.
837 struct lttng_consumer_channel
*consumer_allocate_channel(uint64_t key
,
839 const char *pathname
,
844 enum lttng_event_output output
,
845 uint64_t tracefile_size
,
846 uint64_t tracefile_count
)
848 struct lttng_consumer_channel
*channel
;
850 channel
= zmalloc(sizeof(*channel
));
851 if (channel
== NULL
) {
852 PERROR("malloc struct lttng_consumer_channel");
857 channel
->refcount
= 0;
858 channel
->session_id
= session_id
;
861 channel
->relayd_id
= relayd_id
;
862 channel
->output
= output
;
863 channel
->tracefile_size
= tracefile_size
;
864 channel
->tracefile_count
= tracefile_count
;
866 strncpy(channel
->pathname
, pathname
, sizeof(channel
->pathname
));
867 channel
->pathname
[sizeof(channel
->pathname
) - 1] = '\0';
869 strncpy(channel
->name
, name
, sizeof(channel
->name
));
870 channel
->name
[sizeof(channel
->name
) - 1] = '\0';
872 lttng_ht_node_init_u64(&channel
->node
, channel
->key
);
874 channel
->wait_fd
= -1;
876 CDS_INIT_LIST_HEAD(&channel
->streams
.head
);
878 DBG("Allocated channel (key %" PRIu64
")", channel
->key
)
885 * Add a channel to the global list protected by a mutex.
887 int consumer_add_channel(struct lttng_consumer_channel
*channel
,
888 struct lttng_consumer_local_data
*ctx
)
891 struct lttng_ht_node_u64
*node
;
892 struct lttng_ht_iter iter
;
894 pthread_mutex_lock(&consumer_data
.lock
);
897 lttng_ht_lookup(consumer_data
.channel_ht
, &channel
->key
, &iter
);
898 node
= lttng_ht_iter_get_node_u64(&iter
);
900 /* Channel already exist. Ignore the insertion */
901 ERR("Consumer add channel key %" PRIu64
" already exists!",
907 lttng_ht_add_unique_u64(consumer_data
.channel_ht
, &channel
->node
);
911 pthread_mutex_unlock(&consumer_data
.lock
);
913 if (!ret
&& channel
->wait_fd
!= -1 &&
914 channel
->metadata_stream
== NULL
) {
915 notify_channel_pipe(ctx
, channel
, -1, CONSUMER_CHANNEL_ADD
);
921 * Allocate the pollfd structure and the local view of the out fds to avoid
922 * doing a lookup in the linked list and concurrency issues when writing is
923 * needed. Called with consumer_data.lock held.
925 * Returns the number of fds in the structures.
927 static int update_poll_array(struct lttng_consumer_local_data
*ctx
,
928 struct pollfd
**pollfd
, struct lttng_consumer_stream
**local_stream
,
932 struct lttng_ht_iter iter
;
933 struct lttng_consumer_stream
*stream
;
938 assert(local_stream
);
940 DBG("Updating poll fd array");
942 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
944 * Only active streams with an active end point can be added to the
945 * poll set and local stream storage of the thread.
947 * There is a potential race here for endpoint_status to be updated
948 * just after the check. However, this is OK since the stream(s) will
949 * be deleted once the thread is notified that the end point state has
950 * changed where this function will be called back again.
952 if (stream
->state
!= LTTNG_CONSUMER_ACTIVE_STREAM
||
953 stream
->endpoint_status
== CONSUMER_ENDPOINT_INACTIVE
) {
957 * This clobbers way too much the debug output. Uncomment that if you
958 * need it for debugging purposes.
960 * DBG("Active FD %d", stream->wait_fd);
962 (*pollfd
)[i
].fd
= stream
->wait_fd
;
963 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
964 local_stream
[i
] = stream
;
970 * Insert the consumer_data_pipe at the end of the array and don't
971 * increment i so nb_fd is the number of real FD.
973 (*pollfd
)[i
].fd
= ctx
->consumer_data_pipe
[0];
974 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
979 * Poll on the should_quit pipe and the command socket return -1 on error and
980 * should exit, 0 if data is available on the command socket
982 int lttng_consumer_poll_socket(struct pollfd
*consumer_sockpoll
)
987 num_rdy
= poll(consumer_sockpoll
, 2, -1);
990 * Restart interrupted system call.
992 if (errno
== EINTR
) {
995 PERROR("Poll error");
998 if (consumer_sockpoll
[0].revents
& (POLLIN
| POLLPRI
)) {
999 DBG("consumer_should_quit wake up");
1009 * Set the error socket.
1011 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data
*ctx
,
1014 ctx
->consumer_error_socket
= sock
;
1018 * Set the command socket path.
1020 void lttng_consumer_set_command_sock_path(
1021 struct lttng_consumer_local_data
*ctx
, char *sock
)
1023 ctx
->consumer_command_sock_path
= sock
;
1027 * Send return code to the session daemon.
1028 * If the socket is not defined, we return 0, it is not a fatal error
1030 int lttng_consumer_send_error(struct lttng_consumer_local_data
*ctx
, int cmd
)
1032 if (ctx
->consumer_error_socket
> 0) {
1033 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
1034 sizeof(enum lttcomm_sessiond_command
));
1041 * Close all the tracefiles and stream fds and MUST be called when all
1042 * instances are destroyed i.e. when all threads were joined and are ended.
1044 void lttng_consumer_cleanup(void)
1046 struct lttng_ht_iter iter
;
1047 struct lttng_consumer_channel
*channel
;
1051 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
, channel
,
1053 consumer_del_channel(channel
);
1058 lttng_ht_destroy(consumer_data
.channel_ht
);
1060 cleanup_relayd_ht();
1062 lttng_ht_destroy(consumer_data
.stream_per_chan_id_ht
);
1065 * This HT contains streams that are freed by either the metadata thread or
1066 * the data thread so we do *nothing* on the hash table and simply destroy
1069 lttng_ht_destroy(consumer_data
.stream_list_ht
);
1073 * Called from signal handler.
1075 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
1080 ret
= write(ctx
->consumer_should_quit
[1], "4", 1);
1081 } while (ret
< 0 && errno
== EINTR
);
1082 if (ret
< 0 || ret
!= 1) {
1083 PERROR("write consumer quit");
1086 DBG("Consumer flag that it should quit");
1089 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1092 int outfd
= stream
->out_fd
;
1095 * This does a blocking write-and-wait on any page that belongs to the
1096 * subbuffer prior to the one we just wrote.
1097 * Don't care about error values, as these are just hints and ways to
1098 * limit the amount of page cache used.
1100 if (orig_offset
< stream
->max_sb_size
) {
1103 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1104 stream
->max_sb_size
,
1105 SYNC_FILE_RANGE_WAIT_BEFORE
1106 | SYNC_FILE_RANGE_WRITE
1107 | SYNC_FILE_RANGE_WAIT_AFTER
);
1109 * Give hints to the kernel about how we access the file:
1110 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1113 * We need to call fadvise again after the file grows because the
1114 * kernel does not seem to apply fadvise to non-existing parts of the
1117 * Call fadvise _after_ having waited for the page writeback to
1118 * complete because the dirty page writeback semantic is not well
1119 * defined. So it can be expected to lead to lower throughput in
1122 posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1123 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1127 * Initialise the necessary environnement :
1128 * - create a new context
1129 * - create the poll_pipe
1130 * - create the should_quit pipe (for signal handler)
1131 * - create the thread pipe (for splice)
1133 * Takes a function pointer as argument, this function is called when data is
1134 * available on a buffer. This function is responsible to do the
1135 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1136 * buffer configuration and then kernctl_put_next_subbuf at the end.
1138 * Returns a pointer to the new context or NULL on error.
1140 struct lttng_consumer_local_data
*lttng_consumer_create(
1141 enum lttng_consumer_type type
,
1142 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1143 struct lttng_consumer_local_data
*ctx
),
1144 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1145 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1146 int (*update_stream
)(int stream_key
, uint32_t state
))
1149 struct lttng_consumer_local_data
*ctx
;
1151 assert(consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1152 consumer_data
.type
== type
);
1153 consumer_data
.type
= type
;
1155 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
1157 PERROR("allocating context");
1161 ctx
->consumer_error_socket
= -1;
1162 ctx
->consumer_metadata_socket
= -1;
1163 /* assign the callbacks */
1164 ctx
->on_buffer_ready
= buffer_ready
;
1165 ctx
->on_recv_channel
= recv_channel
;
1166 ctx
->on_recv_stream
= recv_stream
;
1167 ctx
->on_update_stream
= update_stream
;
1169 ret
= pipe(ctx
->consumer_data_pipe
);
1171 PERROR("Error creating poll pipe");
1172 goto error_poll_pipe
;
1175 /* set read end of the pipe to non-blocking */
1176 ret
= fcntl(ctx
->consumer_data_pipe
[0], F_SETFL
, O_NONBLOCK
);
1178 PERROR("fcntl O_NONBLOCK");
1179 goto error_poll_fcntl
;
1182 /* set write end of the pipe to non-blocking */
1183 ret
= fcntl(ctx
->consumer_data_pipe
[1], F_SETFL
, O_NONBLOCK
);
1185 PERROR("fcntl O_NONBLOCK");
1186 goto error_poll_fcntl
;
1189 ret
= pipe(ctx
->consumer_should_quit
);
1191 PERROR("Error creating recv pipe");
1192 goto error_quit_pipe
;
1195 ret
= pipe(ctx
->consumer_thread_pipe
);
1197 PERROR("Error creating thread pipe");
1198 goto error_thread_pipe
;
1201 ret
= pipe(ctx
->consumer_channel_pipe
);
1203 PERROR("Error creating channel pipe");
1204 goto error_channel_pipe
;
1207 ret
= utils_create_pipe(ctx
->consumer_metadata_pipe
);
1209 goto error_metadata_pipe
;
1212 ret
= utils_create_pipe(ctx
->consumer_splice_metadata_pipe
);
1214 goto error_splice_pipe
;
1220 utils_close_pipe(ctx
->consumer_metadata_pipe
);
1221 error_metadata_pipe
:
1222 utils_close_pipe(ctx
->consumer_channel_pipe
);
1224 utils_close_pipe(ctx
->consumer_thread_pipe
);
1226 utils_close_pipe(ctx
->consumer_should_quit
);
1229 utils_close_pipe(ctx
->consumer_data_pipe
);
1237 * Close all fds associated with the instance and free the context.
1239 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1243 DBG("Consumer destroying it. Closing everything.");
1245 ret
= close(ctx
->consumer_error_socket
);
1249 ret
= close(ctx
->consumer_metadata_socket
);
1253 utils_close_pipe(ctx
->consumer_thread_pipe
);
1254 utils_close_pipe(ctx
->consumer_channel_pipe
);
1255 utils_close_pipe(ctx
->consumer_data_pipe
);
1256 utils_close_pipe(ctx
->consumer_should_quit
);
1257 utils_close_pipe(ctx
->consumer_splice_metadata_pipe
);
1259 unlink(ctx
->consumer_command_sock_path
);
1264 * Write the metadata stream id on the specified file descriptor.
1266 static int write_relayd_metadata_id(int fd
,
1267 struct lttng_consumer_stream
*stream
,
1268 struct consumer_relayd_sock_pair
*relayd
, unsigned long padding
)
1271 struct lttcomm_relayd_metadata_payload hdr
;
1273 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1274 hdr
.padding_size
= htobe32(padding
);
1276 ret
= write(fd
, (void *) &hdr
, sizeof(hdr
));
1277 } while (ret
< 0 && errno
== EINTR
);
1278 if (ret
< 0 || ret
!= sizeof(hdr
)) {
1280 * This error means that the fd's end is closed so ignore the perror
1281 * not to clubber the error output since this can happen in a normal
1284 if (errno
!= EPIPE
) {
1285 PERROR("write metadata stream id");
1287 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1289 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1290 * handle writting the missing part so report that as an error and
1291 * don't lie to the caller.
1296 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1297 stream
->relayd_stream_id
, padding
);
1304 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1305 * core function for writing trace buffers to either the local filesystem or
1308 * It must be called with the stream lock held.
1310 * Careful review MUST be put if any changes occur!
1312 * Returns the number of bytes written
1314 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1315 struct lttng_consumer_local_data
*ctx
,
1316 struct lttng_consumer_stream
*stream
, unsigned long len
,
1317 unsigned long padding
)
1319 unsigned long mmap_offset
;
1321 ssize_t ret
= 0, written
= 0;
1322 off_t orig_offset
= stream
->out_fd_offset
;
1323 /* Default is on the disk */
1324 int outfd
= stream
->out_fd
;
1325 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1326 unsigned int relayd_hang_up
= 0;
1328 /* RCU lock for the relayd pointer */
1331 /* Flag that the current stream if set for network streaming. */
1332 if (stream
->net_seq_idx
!= -1) {
1333 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1334 if (relayd
== NULL
) {
1339 /* get the offset inside the fd to mmap */
1340 switch (consumer_data
.type
) {
1341 case LTTNG_CONSUMER_KERNEL
:
1342 mmap_base
= stream
->mmap_base
;
1343 ret
= kernctl_get_mmap_read_offset(stream
->wait_fd
, &mmap_offset
);
1345 case LTTNG_CONSUMER32_UST
:
1346 case LTTNG_CONSUMER64_UST
:
1347 mmap_base
= lttng_ustctl_get_mmap_base(stream
);
1349 ERR("read mmap get mmap base for stream %s", stream
->name
);
1353 ret
= lttng_ustctl_get_mmap_read_offset(stream
, &mmap_offset
);
1357 ERR("Unknown consumer_data type");
1362 PERROR("tracer ctl get_mmap_read_offset");
1367 /* Handle stream on the relayd if the output is on the network */
1369 unsigned long netlen
= len
;
1372 * Lock the control socket for the complete duration of the function
1373 * since from this point on we will use the socket.
1375 if (stream
->metadata_flag
) {
1376 /* Metadata requires the control socket. */
1377 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1378 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1381 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1383 /* Use the returned socket. */
1386 /* Write metadata stream id before payload */
1387 if (stream
->metadata_flag
) {
1388 ret
= write_relayd_metadata_id(outfd
, stream
, relayd
, padding
);
1391 /* Socket operation failed. We consider the relayd dead */
1392 if (ret
== -EPIPE
|| ret
== -EINVAL
) {
1400 /* Socket operation failed. We consider the relayd dead */
1401 if (ret
== -EPIPE
|| ret
== -EINVAL
) {
1405 /* Else, use the default set before which is the filesystem. */
1408 /* No streaming, we have to set the len with the full padding */
1412 * Check if we need to change the tracefile before writing the packet.
1414 if (stream
->chan
->tracefile_size
> 0 &&
1415 (stream
->tracefile_size_current
+ len
) >
1416 stream
->chan
->tracefile_size
) {
1417 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1418 stream
->name
, stream
->chan
->tracefile_size
,
1419 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1420 stream
->out_fd
, &(stream
->tracefile_count_current
));
1422 ERR("Rotating output file");
1425 outfd
= stream
->out_fd
= ret
;
1426 /* Reset current size because we just perform a rotation. */
1427 stream
->tracefile_size_current
= 0;
1429 stream
->tracefile_size_current
+= len
;
1434 ret
= write(outfd
, mmap_base
+ mmap_offset
, len
);
1435 } while (ret
< 0 && errno
== EINTR
);
1436 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, len
);
1439 * This is possible if the fd is closed on the other side (outfd)
1440 * or any write problem. It can be verbose a bit for a normal
1441 * execution if for instance the relayd is stopped abruptly. This
1442 * can happen so set this to a DBG statement.
1444 DBG("Error in file write mmap");
1448 /* Socket operation failed. We consider the relayd dead */
1449 if (errno
== EPIPE
|| errno
== EINVAL
) {
1454 } else if (ret
> len
) {
1455 PERROR("Error in file write (ret %zd > len %lu)", ret
, len
);
1463 /* This call is useless on a socket so better save a syscall. */
1465 /* This won't block, but will start writeout asynchronously */
1466 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret
,
1467 SYNC_FILE_RANGE_WRITE
);
1468 stream
->out_fd_offset
+= ret
;
1472 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1476 * This is a special case that the relayd has closed its socket. Let's
1477 * cleanup the relayd object and all associated streams.
1479 if (relayd
&& relayd_hang_up
) {
1480 cleanup_relayd(relayd
, ctx
);
1484 /* Unlock only if ctrl socket used */
1485 if (relayd
&& stream
->metadata_flag
) {
1486 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1494 * Splice the data from the ring buffer to the tracefile.
1496 * It must be called with the stream lock held.
1498 * Returns the number of bytes spliced.
1500 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1501 struct lttng_consumer_local_data
*ctx
,
1502 struct lttng_consumer_stream
*stream
, unsigned long len
,
1503 unsigned long padding
)
1505 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1507 off_t orig_offset
= stream
->out_fd_offset
;
1508 int fd
= stream
->wait_fd
;
1509 /* Default is on the disk */
1510 int outfd
= stream
->out_fd
;
1511 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1513 unsigned int relayd_hang_up
= 0;
1515 switch (consumer_data
.type
) {
1516 case LTTNG_CONSUMER_KERNEL
:
1518 case LTTNG_CONSUMER32_UST
:
1519 case LTTNG_CONSUMER64_UST
:
1520 /* Not supported for user space tracing */
1523 ERR("Unknown consumer_data type");
1527 /* RCU lock for the relayd pointer */
1530 /* Flag that the current stream if set for network streaming. */
1531 if (stream
->net_seq_idx
!= -1) {
1532 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1533 if (relayd
== NULL
) {
1539 * Choose right pipe for splice. Metadata and trace data are handled by
1540 * different threads hence the use of two pipes in order not to race or
1541 * corrupt the written data.
1543 if (stream
->metadata_flag
) {
1544 splice_pipe
= ctx
->consumer_splice_metadata_pipe
;
1546 splice_pipe
= ctx
->consumer_thread_pipe
;
1549 /* Write metadata stream id before payload */
1551 int total_len
= len
;
1553 if (stream
->metadata_flag
) {
1555 * Lock the control socket for the complete duration of the function
1556 * since from this point on we will use the socket.
1558 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1560 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
, relayd
,
1564 /* Socket operation failed. We consider the relayd dead */
1565 if (ret
== -EBADF
) {
1566 WARN("Remote relayd disconnected. Stopping");
1573 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1576 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1578 /* Use the returned socket. */
1581 /* Socket operation failed. We consider the relayd dead */
1582 if (ret
== -EBADF
) {
1583 WARN("Remote relayd disconnected. Stopping");
1590 /* No streaming, we have to set the len with the full padding */
1594 * Check if we need to change the tracefile before writing the packet.
1596 if (stream
->chan
->tracefile_size
> 0 &&
1597 (stream
->tracefile_size_current
+ len
) >
1598 stream
->chan
->tracefile_size
) {
1599 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1600 stream
->name
, stream
->chan
->tracefile_size
,
1601 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1602 stream
->out_fd
, &(stream
->tracefile_count_current
));
1604 ERR("Rotating output file");
1607 outfd
= stream
->out_fd
= ret
;
1608 /* Reset current size because we just perform a rotation. */
1609 stream
->tracefile_size_current
= 0;
1611 stream
->tracefile_size_current
+= len
;
1615 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1616 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1617 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1618 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1619 DBG("splice chan to pipe, ret %zd", ret_splice
);
1620 if (ret_splice
< 0) {
1621 PERROR("Error in relay splice");
1623 written
= ret_splice
;
1629 /* Handle stream on the relayd if the output is on the network */
1631 if (stream
->metadata_flag
) {
1632 size_t metadata_payload_size
=
1633 sizeof(struct lttcomm_relayd_metadata_payload
);
1635 /* Update counter to fit the spliced data */
1636 ret_splice
+= metadata_payload_size
;
1637 len
+= metadata_payload_size
;
1639 * We do this so the return value can match the len passed as
1640 * argument to this function.
1642 written
-= metadata_payload_size
;
1646 /* Splice data out */
1647 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1648 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1649 DBG("Consumer splice pipe to file, ret %zd", ret_splice
);
1650 if (ret_splice
< 0) {
1651 PERROR("Error in file splice");
1653 written
= ret_splice
;
1655 /* Socket operation failed. We consider the relayd dead */
1656 if (errno
== EBADF
|| errno
== EPIPE
) {
1657 WARN("Remote relayd disconnected. Stopping");
1663 } else if (ret_splice
> len
) {
1665 PERROR("Wrote more data than requested %zd (len: %lu)",
1667 written
+= ret_splice
;
1673 /* This call is useless on a socket so better save a syscall. */
1675 /* This won't block, but will start writeout asynchronously */
1676 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1677 SYNC_FILE_RANGE_WRITE
);
1678 stream
->out_fd_offset
+= ret_splice
;
1680 written
+= ret_splice
;
1682 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1690 * This is a special case that the relayd has closed its socket. Let's
1691 * cleanup the relayd object and all associated streams.
1693 if (relayd
&& relayd_hang_up
) {
1694 cleanup_relayd(relayd
, ctx
);
1695 /* Skip splice error so the consumer does not fail */
1700 /* send the appropriate error description to sessiond */
1703 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1706 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1709 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
1714 if (relayd
&& stream
->metadata_flag
) {
1715 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1723 * Take a snapshot for a specific fd
1725 * Returns 0 on success, < 0 on error
1727 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
1729 switch (consumer_data
.type
) {
1730 case LTTNG_CONSUMER_KERNEL
:
1731 return lttng_kconsumer_take_snapshot(stream
);
1732 case LTTNG_CONSUMER32_UST
:
1733 case LTTNG_CONSUMER64_UST
:
1734 return lttng_ustconsumer_take_snapshot(stream
);
1736 ERR("Unknown consumer_data type");
1743 * Get the produced position
1745 * Returns 0 on success, < 0 on error
1747 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
1750 switch (consumer_data
.type
) {
1751 case LTTNG_CONSUMER_KERNEL
:
1752 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
1753 case LTTNG_CONSUMER32_UST
:
1754 case LTTNG_CONSUMER64_UST
:
1755 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
1757 ERR("Unknown consumer_data type");
1763 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
1764 int sock
, struct pollfd
*consumer_sockpoll
)
1766 switch (consumer_data
.type
) {
1767 case LTTNG_CONSUMER_KERNEL
:
1768 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1769 case LTTNG_CONSUMER32_UST
:
1770 case LTTNG_CONSUMER64_UST
:
1771 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1773 ERR("Unknown consumer_data type");
1780 * Iterate over all streams of the hashtable and free them properly.
1782 * WARNING: *MUST* be used with data stream only.
1784 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1786 struct lttng_ht_iter iter
;
1787 struct lttng_consumer_stream
*stream
;
1794 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1796 * Ignore return value since we are currently cleaning up so any error
1799 (void) consumer_del_stream(stream
, ht
);
1803 lttng_ht_destroy(ht
);
1807 * Iterate over all streams of the hashtable and free them properly.
1809 * XXX: Should not be only for metadata stream or else use an other name.
1811 static void destroy_stream_ht(struct lttng_ht
*ht
)
1813 struct lttng_ht_iter iter
;
1814 struct lttng_consumer_stream
*stream
;
1821 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1823 * Ignore return value since we are currently cleaning up so any error
1826 (void) consumer_del_metadata_stream(stream
, ht
);
1830 lttng_ht_destroy(ht
);
1833 void lttng_consumer_close_metadata(void)
1835 switch (consumer_data
.type
) {
1836 case LTTNG_CONSUMER_KERNEL
:
1838 * The Kernel consumer has a different metadata scheme so we don't
1839 * close anything because the stream will be closed by the session
1843 case LTTNG_CONSUMER32_UST
:
1844 case LTTNG_CONSUMER64_UST
:
1846 * Close all metadata streams. The metadata hash table is passed and
1847 * this call iterates over it by closing all wakeup fd. This is safe
1848 * because at this point we are sure that the metadata producer is
1849 * either dead or blocked.
1851 lttng_ustconsumer_close_metadata(metadata_ht
);
1854 ERR("Unknown consumer_data type");
1860 * Clean up a metadata stream and free its memory.
1862 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
1863 struct lttng_ht
*ht
)
1866 struct lttng_ht_iter iter
;
1867 struct lttng_consumer_channel
*free_chan
= NULL
;
1868 struct consumer_relayd_sock_pair
*relayd
;
1872 * This call should NEVER receive regular stream. It must always be
1873 * metadata stream and this is crucial for data structure synchronization.
1875 assert(stream
->metadata_flag
);
1877 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
1880 /* Means the stream was allocated but not successfully added */
1881 goto free_stream_rcu
;
1884 pthread_mutex_lock(&consumer_data
.lock
);
1885 pthread_mutex_lock(&stream
->lock
);
1887 switch (consumer_data
.type
) {
1888 case LTTNG_CONSUMER_KERNEL
:
1889 if (stream
->mmap_base
!= NULL
) {
1890 ret
= munmap(stream
->mmap_base
, stream
->mmap_len
);
1892 PERROR("munmap metadata stream");
1896 case LTTNG_CONSUMER32_UST
:
1897 case LTTNG_CONSUMER64_UST
:
1898 lttng_ustconsumer_del_stream(stream
);
1901 ERR("Unknown consumer_data type");
1907 iter
.iter
.node
= &stream
->node
.node
;
1908 ret
= lttng_ht_del(ht
, &iter
);
1911 iter
.iter
.node
= &stream
->node_channel_id
.node
;
1912 ret
= lttng_ht_del(consumer_data
.stream_per_chan_id_ht
, &iter
);
1915 iter
.iter
.node
= &stream
->node_session_id
.node
;
1916 ret
= lttng_ht_del(consumer_data
.stream_list_ht
, &iter
);
1920 if (stream
->out_fd
>= 0) {
1921 ret
= close(stream
->out_fd
);
1927 /* Check and cleanup relayd */
1929 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1930 if (relayd
!= NULL
) {
1931 uatomic_dec(&relayd
->refcount
);
1932 assert(uatomic_read(&relayd
->refcount
) >= 0);
1934 /* Closing streams requires to lock the control socket. */
1935 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1936 ret
= relayd_send_close_stream(&relayd
->control_sock
,
1937 stream
->relayd_stream_id
, stream
->next_net_seq_num
- 1);
1938 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1940 DBG("Unable to close stream on the relayd. Continuing");
1942 * Continue here. There is nothing we can do for the relayd.
1943 * Chances are that the relayd has closed the socket so we just
1944 * continue cleaning up.
1948 /* Both conditions are met, we destroy the relayd. */
1949 if (uatomic_read(&relayd
->refcount
) == 0 &&
1950 uatomic_read(&relayd
->destroy_flag
)) {
1951 destroy_relayd(relayd
);
1956 /* Atomically decrement channel refcount since other threads can use it. */
1957 if (!uatomic_sub_return(&stream
->chan
->refcount
, 1)
1958 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
1959 /* Go for channel deletion! */
1960 free_chan
= stream
->chan
;
1964 pthread_mutex_unlock(&stream
->lock
);
1965 pthread_mutex_unlock(&consumer_data
.lock
);
1968 consumer_del_channel(free_chan
);
1972 call_rcu(&stream
->node
.head
, free_stream_rcu
);
1976 * Action done with the metadata stream when adding it to the consumer internal
1977 * data structures to handle it.
1979 static int add_metadata_stream(struct lttng_consumer_stream
*stream
,
1980 struct lttng_ht
*ht
)
1983 struct consumer_relayd_sock_pair
*relayd
;
1984 struct lttng_ht_iter iter
;
1985 struct lttng_ht_node_u64
*node
;
1990 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
1992 pthread_mutex_lock(&consumer_data
.lock
);
1993 pthread_mutex_lock(&stream
->lock
);
1996 * From here, refcounts are updated so be _careful_ when returning an error
2003 * Lookup the stream just to make sure it does not exist in our internal
2004 * state. This should NEVER happen.
2006 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2007 node
= lttng_ht_iter_get_node_u64(&iter
);
2010 /* Find relayd and, if one is found, increment refcount. */
2011 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
2012 if (relayd
!= NULL
) {
2013 uatomic_inc(&relayd
->refcount
);
2016 /* Update channel refcount once added without error(s). */
2017 uatomic_inc(&stream
->chan
->refcount
);
2020 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2021 * in terms of destroying the associated channel, because the action that
2022 * causes the count to become 0 also causes a stream to be added. The
2023 * channel deletion will thus be triggered by the following removal of this
2026 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2027 /* Increment refcount before decrementing nb_init_stream_left */
2029 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2032 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2034 lttng_ht_add_unique_u64(consumer_data
.stream_per_chan_id_ht
,
2035 &stream
->node_channel_id
);
2038 * Add stream to the stream_list_ht of the consumer data. No need to steal
2039 * the key since the HT does not use it and we allow to add redundant keys
2042 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2046 pthread_mutex_unlock(&stream
->lock
);
2047 pthread_mutex_unlock(&consumer_data
.lock
);
2052 * Delete data stream that are flagged for deletion (endpoint_status).
2054 static void validate_endpoint_status_data_stream(void)
2056 struct lttng_ht_iter iter
;
2057 struct lttng_consumer_stream
*stream
;
2059 DBG("Consumer delete flagged data stream");
2062 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2063 /* Validate delete flag of the stream */
2064 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2067 /* Delete it right now */
2068 consumer_del_stream(stream
, data_ht
);
2074 * Delete metadata stream that are flagged for deletion (endpoint_status).
2076 static void validate_endpoint_status_metadata_stream(
2077 struct lttng_poll_event
*pollset
)
2079 struct lttng_ht_iter iter
;
2080 struct lttng_consumer_stream
*stream
;
2082 DBG("Consumer delete flagged metadata stream");
2087 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2088 /* Validate delete flag of the stream */
2089 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2093 * Remove from pollset so the metadata thread can continue without
2094 * blocking on a deleted stream.
2096 lttng_poll_del(pollset
, stream
->wait_fd
);
2098 /* Delete it right now */
2099 consumer_del_metadata_stream(stream
, metadata_ht
);
2105 * Thread polls on metadata file descriptor and write them on disk or on the
2108 void *consumer_thread_metadata_poll(void *data
)
2111 uint32_t revents
, nb_fd
;
2112 struct lttng_consumer_stream
*stream
= NULL
;
2113 struct lttng_ht_iter iter
;
2114 struct lttng_ht_node_u64
*node
;
2115 struct lttng_poll_event events
;
2116 struct lttng_consumer_local_data
*ctx
= data
;
2119 rcu_register_thread();
2121 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2123 /* ENOMEM at this point. Better to bail out. */
2127 DBG("Thread metadata poll started");
2129 /* Size is set to 1 for the consumer_metadata pipe */
2130 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2132 ERR("Poll set creation failed");
2136 ret
= lttng_poll_add(&events
, ctx
->consumer_metadata_pipe
[0], LPOLLIN
);
2142 DBG("Metadata main loop started");
2145 /* Only the metadata pipe is set */
2146 if (LTTNG_POLL_GETNB(&events
) == 0 && consumer_quit
== 1) {
2151 DBG("Metadata poll wait with %d fd(s)", LTTNG_POLL_GETNB(&events
));
2152 ret
= lttng_poll_wait(&events
, -1);
2153 DBG("Metadata event catched in thread");
2155 if (errno
== EINTR
) {
2156 ERR("Poll EINTR catched");
2164 /* From here, the event is a metadata wait fd */
2165 for (i
= 0; i
< nb_fd
; i
++) {
2166 revents
= LTTNG_POLL_GETEV(&events
, i
);
2167 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2169 /* Just don't waste time if no returned events for the fd */
2174 if (pollfd
== ctx
->consumer_metadata_pipe
[0]) {
2175 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2176 DBG("Metadata thread pipe hung up");
2178 * Remove the pipe from the poll set and continue the loop
2179 * since their might be data to consume.
2181 lttng_poll_del(&events
, ctx
->consumer_metadata_pipe
[0]);
2182 ret
= close(ctx
->consumer_metadata_pipe
[0]);
2184 PERROR("close metadata pipe");
2187 } else if (revents
& LPOLLIN
) {
2189 /* Get the stream pointer received */
2190 ret
= read(pollfd
, &stream
, sizeof(stream
));
2191 } while (ret
< 0 && errno
== EINTR
);
2193 ret
< sizeof(struct lttng_consumer_stream
*)) {
2194 PERROR("read metadata stream");
2196 * Let's continue here and hope we can still work
2197 * without stopping the consumer. XXX: Should we?
2202 /* A NULL stream means that the state has changed. */
2203 if (stream
== NULL
) {
2204 /* Check for deleted streams. */
2205 validate_endpoint_status_metadata_stream(&events
);
2209 DBG("Adding metadata stream %d to poll set",
2212 ret
= add_metadata_stream(stream
, metadata_ht
);
2214 ERR("Unable to add metadata stream");
2215 /* Stream was not setup properly. Continuing. */
2216 consumer_del_metadata_stream(stream
, NULL
);
2220 /* Add metadata stream to the global poll events list */
2221 lttng_poll_add(&events
, stream
->wait_fd
,
2222 LPOLLIN
| LPOLLPRI
);
2225 /* Handle other stream */
2231 uint64_t tmp_id
= (uint64_t) pollfd
;
2233 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2235 node
= lttng_ht_iter_get_node_u64(&iter
);
2238 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2241 /* Check for error event */
2242 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2243 DBG("Metadata fd %d is hup|err.", pollfd
);
2244 if (!stream
->hangup_flush_done
2245 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2246 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2247 DBG("Attempting to flush and consume the UST buffers");
2248 lttng_ustconsumer_on_stream_hangup(stream
);
2250 /* We just flushed the stream now read it. */
2252 len
= ctx
->on_buffer_ready(stream
, ctx
);
2254 * We don't check the return value here since if we get
2255 * a negative len, it means an error occured thus we
2256 * simply remove it from the poll set and free the
2262 lttng_poll_del(&events
, stream
->wait_fd
);
2264 * This call update the channel states, closes file descriptors
2265 * and securely free the stream.
2267 consumer_del_metadata_stream(stream
, metadata_ht
);
2268 } else if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2269 /* Get the data out of the metadata file descriptor */
2270 DBG("Metadata available on fd %d", pollfd
);
2271 assert(stream
->wait_fd
== pollfd
);
2273 len
= ctx
->on_buffer_ready(stream
, ctx
);
2274 /* It's ok to have an unavailable sub-buffer */
2275 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2276 /* Clean up stream from consumer and free it. */
2277 lttng_poll_del(&events
, stream
->wait_fd
);
2278 consumer_del_metadata_stream(stream
, metadata_ht
);
2279 } else if (len
> 0) {
2280 stream
->data_read
= 1;
2284 /* Release RCU lock for the stream looked up */
2291 DBG("Metadata poll thread exiting");
2293 lttng_poll_clean(&events
);
2295 destroy_stream_ht(metadata_ht
);
2297 rcu_unregister_thread();
2302 * This thread polls the fds in the set to consume the data and write
2303 * it to tracefile if necessary.
2305 void *consumer_thread_data_poll(void *data
)
2307 int num_rdy
, num_hup
, high_prio
, ret
, i
;
2308 struct pollfd
*pollfd
= NULL
;
2309 /* local view of the streams */
2310 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2311 /* local view of consumer_data.fds_count */
2313 struct lttng_consumer_local_data
*ctx
= data
;
2316 rcu_register_thread();
2318 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2319 if (data_ht
== NULL
) {
2320 /* ENOMEM at this point. Better to bail out. */
2324 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
));
2331 * the fds set has been updated, we need to update our
2332 * local array as well
2334 pthread_mutex_lock(&consumer_data
.lock
);
2335 if (consumer_data
.need_update
) {
2340 local_stream
= NULL
;
2342 /* allocate for all fds + 1 for the consumer_data_pipe */
2343 pollfd
= zmalloc((consumer_data
.stream_count
+ 1) * sizeof(struct pollfd
));
2344 if (pollfd
== NULL
) {
2345 PERROR("pollfd malloc");
2346 pthread_mutex_unlock(&consumer_data
.lock
);
2350 /* allocate for all fds + 1 for the consumer_data_pipe */
2351 local_stream
= zmalloc((consumer_data
.stream_count
+ 1) *
2352 sizeof(struct lttng_consumer_stream
));
2353 if (local_stream
== NULL
) {
2354 PERROR("local_stream malloc");
2355 pthread_mutex_unlock(&consumer_data
.lock
);
2358 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2361 ERR("Error in allocating pollfd or local_outfds");
2362 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2363 pthread_mutex_unlock(&consumer_data
.lock
);
2367 consumer_data
.need_update
= 0;
2369 pthread_mutex_unlock(&consumer_data
.lock
);
2371 /* No FDs and consumer_quit, consumer_cleanup the thread */
2372 if (nb_fd
== 0 && consumer_quit
== 1) {
2375 /* poll on the array of fds */
2377 DBG("polling on %d fd", nb_fd
+ 1);
2378 num_rdy
= poll(pollfd
, nb_fd
+ 1, -1);
2379 DBG("poll num_rdy : %d", num_rdy
);
2380 if (num_rdy
== -1) {
2382 * Restart interrupted system call.
2384 if (errno
== EINTR
) {
2387 PERROR("Poll error");
2388 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2390 } else if (num_rdy
== 0) {
2391 DBG("Polling thread timed out");
2396 * If the consumer_data_pipe triggered poll go directly to the
2397 * beginning of the loop to update the array. We want to prioritize
2398 * array update over low-priority reads.
2400 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2401 ssize_t pipe_readlen
;
2403 DBG("consumer_data_pipe wake up");
2404 /* Consume 1 byte of pipe data */
2406 pipe_readlen
= read(ctx
->consumer_data_pipe
[0], &new_stream
,
2407 sizeof(new_stream
));
2408 } while (pipe_readlen
== -1 && errno
== EINTR
);
2409 if (pipe_readlen
< 0) {
2410 PERROR("read consumer data pipe");
2411 /* Continue so we can at least handle the current stream(s). */
2416 * If the stream is NULL, just ignore it. It's also possible that
2417 * the sessiond poll thread changed the consumer_quit state and is
2418 * waking us up to test it.
2420 if (new_stream
== NULL
) {
2421 validate_endpoint_status_data_stream();
2425 ret
= add_stream(new_stream
, data_ht
);
2427 ERR("Consumer add stream %" PRIu64
" failed. Continuing",
2430 * At this point, if the add_stream fails, it is not in the
2431 * hash table thus passing the NULL value here.
2433 consumer_del_stream(new_stream
, NULL
);
2436 /* Continue to update the local streams and handle prio ones */
2440 /* Take care of high priority channels first. */
2441 for (i
= 0; i
< nb_fd
; i
++) {
2442 if (local_stream
[i
] == NULL
) {
2445 if (pollfd
[i
].revents
& POLLPRI
) {
2446 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2448 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2449 /* it's ok to have an unavailable sub-buffer */
2450 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2451 /* Clean the stream and free it. */
2452 consumer_del_stream(local_stream
[i
], data_ht
);
2453 local_stream
[i
] = NULL
;
2454 } else if (len
> 0) {
2455 local_stream
[i
]->data_read
= 1;
2461 * If we read high prio channel in this loop, try again
2462 * for more high prio data.
2468 /* Take care of low priority channels. */
2469 for (i
= 0; i
< nb_fd
; i
++) {
2470 if (local_stream
[i
] == NULL
) {
2473 if ((pollfd
[i
].revents
& POLLIN
) ||
2474 local_stream
[i
]->hangup_flush_done
) {
2475 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2476 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2477 /* it's ok to have an unavailable sub-buffer */
2478 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2479 /* Clean the stream and free it. */
2480 consumer_del_stream(local_stream
[i
], data_ht
);
2481 local_stream
[i
] = NULL
;
2482 } else if (len
> 0) {
2483 local_stream
[i
]->data_read
= 1;
2488 /* Handle hangup and errors */
2489 for (i
= 0; i
< nb_fd
; i
++) {
2490 if (local_stream
[i
] == NULL
) {
2493 if (!local_stream
[i
]->hangup_flush_done
2494 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2495 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2496 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2497 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2499 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2500 /* Attempt read again, for the data we just flushed. */
2501 local_stream
[i
]->data_read
= 1;
2504 * If the poll flag is HUP/ERR/NVAL and we have
2505 * read no data in this pass, we can remove the
2506 * stream from its hash table.
2508 if ((pollfd
[i
].revents
& POLLHUP
)) {
2509 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2510 if (!local_stream
[i
]->data_read
) {
2511 consumer_del_stream(local_stream
[i
], data_ht
);
2512 local_stream
[i
] = NULL
;
2515 } else if (pollfd
[i
].revents
& POLLERR
) {
2516 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2517 if (!local_stream
[i
]->data_read
) {
2518 consumer_del_stream(local_stream
[i
], data_ht
);
2519 local_stream
[i
] = NULL
;
2522 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2523 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2524 if (!local_stream
[i
]->data_read
) {
2525 consumer_del_stream(local_stream
[i
], data_ht
);
2526 local_stream
[i
] = NULL
;
2530 if (local_stream
[i
] != NULL
) {
2531 local_stream
[i
]->data_read
= 0;
2536 DBG("polling thread exiting");
2541 * Close the write side of the pipe so epoll_wait() in
2542 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2543 * read side of the pipe. If we close them both, epoll_wait strangely does
2544 * not return and could create a endless wait period if the pipe is the
2545 * only tracked fd in the poll set. The thread will take care of closing
2548 ret
= close(ctx
->consumer_metadata_pipe
[1]);
2550 PERROR("close data pipe");
2553 destroy_data_stream_ht(data_ht
);
2555 rcu_unregister_thread();
2560 * Close wake-up end of each stream belonging to the channel. This will
2561 * allow the poll() on the stream read-side to detect when the
2562 * write-side (application) finally closes them.
2565 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2567 struct lttng_ht
*ht
;
2568 struct lttng_consumer_stream
*stream
;
2569 struct lttng_ht_iter iter
;
2571 ht
= consumer_data
.stream_per_chan_id_ht
;
2574 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2575 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2576 ht
->match_fct
, &channel
->key
,
2577 &iter
.iter
, stream
, node_channel_id
.node
) {
2579 * Protect against teardown with mutex.
2581 pthread_mutex_lock(&stream
->lock
);
2582 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2585 switch (consumer_data
.type
) {
2586 case LTTNG_CONSUMER_KERNEL
:
2588 case LTTNG_CONSUMER32_UST
:
2589 case LTTNG_CONSUMER64_UST
:
2591 * Note: a mutex is taken internally within
2592 * liblttng-ust-ctl to protect timer wakeup_fd
2593 * use from concurrent close.
2595 lttng_ustconsumer_close_stream_wakeup(stream
);
2598 ERR("Unknown consumer_data type");
2602 pthread_mutex_unlock(&stream
->lock
);
2607 static void destroy_channel_ht(struct lttng_ht
*ht
)
2609 struct lttng_ht_iter iter
;
2610 struct lttng_consumer_channel
*channel
;
2618 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2619 ret
= lttng_ht_del(ht
, &iter
);
2624 lttng_ht_destroy(ht
);
2628 * This thread polls the channel fds to detect when they are being
2629 * closed. It closes all related streams if the channel is detected as
2630 * closed. It is currently only used as a shim layer for UST because the
2631 * consumerd needs to keep the per-stream wakeup end of pipes open for
2634 void *consumer_thread_channel_poll(void *data
)
2637 uint32_t revents
, nb_fd
;
2638 struct lttng_consumer_channel
*chan
= NULL
;
2639 struct lttng_ht_iter iter
;
2640 struct lttng_ht_node_u64
*node
;
2641 struct lttng_poll_event events
;
2642 struct lttng_consumer_local_data
*ctx
= data
;
2643 struct lttng_ht
*channel_ht
;
2645 rcu_register_thread();
2647 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2649 /* ENOMEM at this point. Better to bail out. */
2653 DBG("Thread channel poll started");
2655 /* Size is set to 1 for the consumer_channel pipe */
2656 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2658 ERR("Poll set creation failed");
2662 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2668 DBG("Channel main loop started");
2671 /* Only the channel pipe is set */
2672 if (LTTNG_POLL_GETNB(&events
) == 0 && consumer_quit
== 1) {
2677 DBG("Channel poll wait with %d fd(s)", LTTNG_POLL_GETNB(&events
));
2678 ret
= lttng_poll_wait(&events
, -1);
2679 DBG("Channel event catched in thread");
2681 if (errno
== EINTR
) {
2682 ERR("Poll EINTR catched");
2690 /* From here, the event is a channel wait fd */
2691 for (i
= 0; i
< nb_fd
; i
++) {
2692 revents
= LTTNG_POLL_GETEV(&events
, i
);
2693 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2695 /* Just don't waste time if no returned events for the fd */
2699 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2700 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2701 DBG("Channel thread pipe hung up");
2703 * Remove the pipe from the poll set and continue the loop
2704 * since their might be data to consume.
2706 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2708 } else if (revents
& LPOLLIN
) {
2709 enum consumer_channel_action action
;
2712 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
2714 ERR("Error reading channel pipe");
2719 case CONSUMER_CHANNEL_ADD
:
2720 DBG("Adding channel %d to poll set",
2723 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
2725 lttng_ht_add_unique_u64(channel_ht
,
2726 &chan
->wait_fd_node
);
2727 /* Add channel to the global poll events list */
2728 lttng_poll_add(&events
, chan
->wait_fd
,
2729 LPOLLIN
| LPOLLPRI
);
2731 case CONSUMER_CHANNEL_DEL
:
2733 chan
= consumer_find_channel(key
);
2735 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
2738 lttng_poll_del(&events
, chan
->wait_fd
);
2739 ret
= lttng_ht_del(channel_ht
, &iter
);
2741 consumer_close_channel_streams(chan
);
2744 * Release our own refcount. Force channel deletion even if
2745 * streams were not initialized.
2747 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
2748 consumer_del_channel(chan
);
2752 case CONSUMER_CHANNEL_QUIT
:
2754 * Remove the pipe from the poll set and continue the loop
2755 * since their might be data to consume.
2757 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2760 ERR("Unknown action");
2765 /* Handle other stream */
2771 uint64_t tmp_id
= (uint64_t) pollfd
;
2773 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
2775 node
= lttng_ht_iter_get_node_u64(&iter
);
2778 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
2781 /* Check for error event */
2782 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2783 DBG("Channel fd %d is hup|err.", pollfd
);
2785 lttng_poll_del(&events
, chan
->wait_fd
);
2786 ret
= lttng_ht_del(channel_ht
, &iter
);
2788 consumer_close_channel_streams(chan
);
2790 /* Release our own refcount */
2791 if (!uatomic_sub_return(&chan
->refcount
, 1)
2792 && !uatomic_read(&chan
->nb_init_stream_left
)) {
2793 consumer_del_channel(chan
);
2797 /* Release RCU lock for the channel looked up */
2803 lttng_poll_clean(&events
);
2805 destroy_channel_ht(channel_ht
);
2807 DBG("Channel poll thread exiting");
2808 rcu_unregister_thread();
2812 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
2813 struct pollfd
*sockpoll
, int client_socket
)
2820 if (lttng_consumer_poll_socket(sockpoll
) < 0) {
2824 DBG("Metadata connection on client_socket");
2826 /* Blocking call, waiting for transmission */
2827 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
2828 if (ctx
->consumer_metadata_socket
< 0) {
2829 WARN("On accept metadata");
2840 * This thread listens on the consumerd socket and receives the file
2841 * descriptors from the session daemon.
2843 void *consumer_thread_sessiond_poll(void *data
)
2845 int sock
= -1, client_socket
, ret
;
2847 * structure to poll for incoming data on communication socket avoids
2848 * making blocking sockets.
2850 struct pollfd consumer_sockpoll
[2];
2851 struct lttng_consumer_local_data
*ctx
= data
;
2853 rcu_register_thread();
2855 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
2856 unlink(ctx
->consumer_command_sock_path
);
2857 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
2858 if (client_socket
< 0) {
2859 ERR("Cannot create command socket");
2863 ret
= lttcomm_listen_unix_sock(client_socket
);
2868 DBG("Sending ready command to lttng-sessiond");
2869 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
2870 /* return < 0 on error, but == 0 is not fatal */
2872 ERR("Error sending ready command to lttng-sessiond");
2876 ret
= fcntl(client_socket
, F_SETFL
, O_NONBLOCK
);
2878 PERROR("fcntl O_NONBLOCK");
2882 /* prepare the FDs to poll : to client socket and the should_quit pipe */
2883 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
2884 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
2885 consumer_sockpoll
[1].fd
= client_socket
;
2886 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
2888 if (lttng_consumer_poll_socket(consumer_sockpoll
) < 0) {
2891 DBG("Connection on client_socket");
2893 /* Blocking call, waiting for transmission */
2894 sock
= lttcomm_accept_unix_sock(client_socket
);
2899 ret
= fcntl(sock
, F_SETFL
, O_NONBLOCK
);
2901 PERROR("fcntl O_NONBLOCK");
2906 * Setup metadata socket which is the second socket connection on the
2907 * command unix socket.
2909 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
2914 /* This socket is not useful anymore. */
2915 ret
= close(client_socket
);
2917 PERROR("close client_socket");
2921 /* update the polling structure to poll on the established socket */
2922 consumer_sockpoll
[1].fd
= sock
;
2923 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
2926 if (lttng_consumer_poll_socket(consumer_sockpoll
) < 0) {
2929 DBG("Incoming command on sock");
2930 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2931 if (ret
== -ENOENT
) {
2932 DBG("Received STOP command");
2937 * This could simply be a session daemon quitting. Don't output
2940 DBG("Communication interrupted on command socket");
2943 if (consumer_quit
) {
2944 DBG("consumer_thread_receive_fds received quit from signal");
2947 DBG("received command on sock");
2950 DBG("Consumer thread sessiond poll exiting");
2953 * Close metadata streams since the producer is the session daemon which
2956 * NOTE: for now, this only applies to the UST tracer.
2958 lttng_consumer_close_metadata();
2961 * when all fds have hung up, the polling thread
2967 * Notify the data poll thread to poll back again and test the
2968 * consumer_quit state that we just set so to quit gracefully.
2970 notify_thread_pipe(ctx
->consumer_data_pipe
[1]);
2972 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
2974 /* Cleaning up possibly open sockets. */
2978 PERROR("close sock sessiond poll");
2981 if (client_socket
>= 0) {
2984 PERROR("close client_socket sessiond poll");
2988 rcu_unregister_thread();
2992 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
2993 struct lttng_consumer_local_data
*ctx
)
2997 pthread_mutex_lock(&stream
->lock
);
2999 switch (consumer_data
.type
) {
3000 case LTTNG_CONSUMER_KERNEL
:
3001 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
3003 case LTTNG_CONSUMER32_UST
:
3004 case LTTNG_CONSUMER64_UST
:
3005 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
3008 ERR("Unknown consumer_data type");
3014 pthread_mutex_unlock(&stream
->lock
);
3018 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3020 switch (consumer_data
.type
) {
3021 case LTTNG_CONSUMER_KERNEL
:
3022 return lttng_kconsumer_on_recv_stream(stream
);
3023 case LTTNG_CONSUMER32_UST
:
3024 case LTTNG_CONSUMER64_UST
:
3025 return lttng_ustconsumer_on_recv_stream(stream
);
3027 ERR("Unknown consumer_data type");
3034 * Allocate and set consumer data hash tables.
3036 void lttng_consumer_init(void)
3038 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3039 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3040 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3041 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3045 * Process the ADD_RELAYD command receive by a consumer.
3047 * This will create a relayd socket pair and add it to the relayd hash table.
3048 * The caller MUST acquire a RCU read side lock before calling it.
3050 int consumer_add_relayd_socket(int net_seq_idx
, int sock_type
,
3051 struct lttng_consumer_local_data
*ctx
, int sock
,
3052 struct pollfd
*consumer_sockpoll
,
3053 struct lttcomm_relayd_sock
*relayd_sock
, unsigned int sessiond_id
)
3055 int fd
= -1, ret
= -1, relayd_created
= 0;
3056 enum lttng_error_code ret_code
= LTTNG_OK
;
3057 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3060 assert(relayd_sock
);
3062 DBG("Consumer adding relayd socket (idx: %d)", net_seq_idx
);
3064 /* First send a status message before receiving the fds. */
3065 ret
= consumer_send_status_msg(sock
, ret_code
);
3067 /* Somehow, the session daemon is not responding anymore. */
3071 /* Get relayd reference if exists. */
3072 relayd
= consumer_find_relayd(net_seq_idx
);
3073 if (relayd
== NULL
) {
3074 /* Not found. Allocate one. */
3075 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3076 if (relayd
== NULL
) {
3077 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_OUTFD_ERROR
);
3081 relayd
->sessiond_session_id
= (uint64_t) sessiond_id
;
3085 /* Poll on consumer socket. */
3086 if (lttng_consumer_poll_socket(consumer_sockpoll
) < 0) {
3091 /* Get relayd socket from session daemon */
3092 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3093 if (ret
!= sizeof(fd
)) {
3094 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3096 fd
= -1; /* Just in case it gets set with an invalid value. */
3100 /* We have the fds without error. Send status back. */
3101 ret
= consumer_send_status_msg(sock
, ret_code
);
3103 /* Somehow, the session daemon is not responding anymore. */
3107 /* Copy socket information and received FD */
3108 switch (sock_type
) {
3109 case LTTNG_STREAM_CONTROL
:
3110 /* Copy received lttcomm socket */
3111 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3112 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3113 /* Immediately try to close the created socket if valid. */
3114 if (relayd
->control_sock
.sock
.fd
>= 0) {
3115 if (close(relayd
->control_sock
.sock
.fd
)) {
3116 PERROR("close relayd control socket");
3119 /* Handle create_sock error. */
3124 /* Assign new file descriptor */
3125 relayd
->control_sock
.sock
.fd
= fd
;
3126 /* Assign version values. */
3127 relayd
->control_sock
.major
= relayd_sock
->major
;
3128 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3131 * Create a session on the relayd and store the returned id. Lock the
3132 * control socket mutex if the relayd was NOT created before.
3134 if (!relayd_created
) {
3135 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3137 ret
= relayd_create_session(&relayd
->control_sock
,
3138 &relayd
->relayd_session_id
);
3139 if (!relayd_created
) {
3140 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3144 * Close all sockets of a relayd object. It will be freed if it was
3145 * created at the error code path or else it will be garbage
3148 (void) relayd_close(&relayd
->control_sock
);
3149 (void) relayd_close(&relayd
->data_sock
);
3154 case LTTNG_STREAM_DATA
:
3155 /* Copy received lttcomm socket */
3156 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3157 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3158 /* Immediately try to close the created socket if valid. */
3159 if (relayd
->data_sock
.sock
.fd
>= 0) {
3160 if (close(relayd
->data_sock
.sock
.fd
)) {
3161 PERROR("close relayd data socket");
3164 /* Handle create_sock error. */
3169 /* Assign new file descriptor */
3170 relayd
->data_sock
.sock
.fd
= fd
;
3171 /* Assign version values. */
3172 relayd
->data_sock
.major
= relayd_sock
->major
;
3173 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3176 ERR("Unknown relayd socket type (%d)", sock_type
);
3181 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3182 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3183 relayd
->net_seq_idx
, fd
);
3186 * Add relayd socket pair to consumer data hashtable. If object already
3187 * exists or on error, the function gracefully returns.
3195 /* Close received socket if valid. */
3198 PERROR("close received socket");
3203 if (relayd_created
) {
3211 * Try to lock the stream mutex.
3213 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
3215 static int stream_try_lock(struct lttng_consumer_stream
*stream
)
3222 * Try to lock the stream mutex. On failure, we know that the stream is
3223 * being used else where hence there is data still being extracted.
3225 ret
= pthread_mutex_trylock(&stream
->lock
);
3227 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
3239 * Search for a relayd associated to the session id and return the reference.
3241 * A rcu read side lock MUST be acquire before calling this function and locked
3242 * until the relayd object is no longer necessary.
3244 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3246 struct lttng_ht_iter iter
;
3247 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3249 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3250 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3253 * Check by sessiond id which is unique here where the relayd session
3254 * id might not be when having multiple relayd.
3256 if (relayd
->sessiond_session_id
== id
) {
3257 /* Found the relayd. There can be only one per id. */
3269 * Check if for a given session id there is still data needed to be extract
3272 * Return 1 if data is pending or else 0 meaning ready to be read.
3274 int consumer_data_pending(uint64_t id
)
3277 struct lttng_ht_iter iter
;
3278 struct lttng_ht
*ht
;
3279 struct lttng_consumer_stream
*stream
;
3280 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3281 int (*data_pending
)(struct lttng_consumer_stream
*);
3283 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3286 pthread_mutex_lock(&consumer_data
.lock
);
3288 switch (consumer_data
.type
) {
3289 case LTTNG_CONSUMER_KERNEL
:
3290 data_pending
= lttng_kconsumer_data_pending
;
3292 case LTTNG_CONSUMER32_UST
:
3293 case LTTNG_CONSUMER64_UST
:
3294 data_pending
= lttng_ustconsumer_data_pending
;
3297 ERR("Unknown consumer data type");
3301 /* Ease our life a bit */
3302 ht
= consumer_data
.stream_list_ht
;
3304 relayd
= find_relayd_by_session_id(id
);
3306 /* Send init command for data pending. */
3307 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3308 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3309 relayd
->relayd_session_id
);
3310 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3312 /* Communication error thus the relayd so no data pending. */
3313 goto data_not_pending
;
3317 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3318 ht
->hash_fct(&id
, lttng_ht_seed
),
3320 &iter
.iter
, stream
, node_session_id
.node
) {
3321 /* If this call fails, the stream is being used hence data pending. */
3322 ret
= stream_try_lock(stream
);
3328 * A removed node from the hash table indicates that the stream has
3329 * been deleted thus having a guarantee that the buffers are closed
3330 * on the consumer side. However, data can still be transmitted
3331 * over the network so don't skip the relayd check.
3333 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3335 /* Check the stream if there is data in the buffers. */
3336 ret
= data_pending(stream
);
3338 pthread_mutex_unlock(&stream
->lock
);
3345 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3346 if (stream
->metadata_flag
) {
3347 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3348 stream
->relayd_stream_id
);
3350 ret
= relayd_data_pending(&relayd
->control_sock
,
3351 stream
->relayd_stream_id
,
3352 stream
->next_net_seq_num
- 1);
3354 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3356 pthread_mutex_unlock(&stream
->lock
);
3360 pthread_mutex_unlock(&stream
->lock
);
3364 unsigned int is_data_inflight
= 0;
3366 /* Send init command for data pending. */
3367 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3368 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3369 relayd
->relayd_session_id
, &is_data_inflight
);
3370 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3372 goto data_not_pending
;
3374 if (is_data_inflight
) {
3380 * Finding _no_ node in the hash table and no inflight data means that the
3381 * stream(s) have been removed thus data is guaranteed to be available for
3382 * analysis from the trace files.
3386 /* Data is available to be read by a viewer. */
3387 pthread_mutex_unlock(&consumer_data
.lock
);
3392 /* Data is still being extracted from buffers. */
3393 pthread_mutex_unlock(&consumer_data
.lock
);
3399 * Send a ret code status message to the sessiond daemon.
3401 * Return the sendmsg() return value.
3403 int consumer_send_status_msg(int sock
, int ret_code
)
3405 struct lttcomm_consumer_status_msg msg
;
3407 msg
.ret_code
= ret_code
;
3409 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3413 * Send a channel status message to the sessiond daemon.
3415 * Return the sendmsg() return value.
3417 int consumer_send_status_channel(int sock
,
3418 struct lttng_consumer_channel
*channel
)
3420 struct lttcomm_consumer_status_channel msg
;
3425 msg
.ret_code
= -LTTNG_ERR_UST_CHAN_FAIL
;
3427 msg
.ret_code
= LTTNG_OK
;
3428 msg
.key
= channel
->key
;
3429 msg
.stream_count
= channel
->streams
.count
;
3432 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));