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
;
105 ret
= write(ctx
->consumer_channel_pipe
[1], &msg
, sizeof(msg
));
106 } while (ret
< 0 && errno
== EINTR
);
109 void notify_thread_del_channel(struct lttng_consumer_local_data
*ctx
,
112 notify_channel_pipe(ctx
, NULL
, key
, CONSUMER_CHANNEL_DEL
);
115 static int read_channel_pipe(struct lttng_consumer_local_data
*ctx
,
116 struct lttng_consumer_channel
**chan
,
118 enum consumer_channel_action
*action
)
120 struct consumer_channel_msg msg
;
124 ret
= read(ctx
->consumer_channel_pipe
[0], &msg
, sizeof(msg
));
125 } while (ret
< 0 && errno
== EINTR
);
127 *action
= msg
.action
;
135 * Find a stream. The consumer_data.lock must be locked during this
138 static struct lttng_consumer_stream
*find_stream(uint64_t key
,
141 struct lttng_ht_iter iter
;
142 struct lttng_ht_node_u64
*node
;
143 struct lttng_consumer_stream
*stream
= NULL
;
147 /* -1ULL keys are lookup failures */
148 if (key
== (uint64_t) -1ULL) {
154 lttng_ht_lookup(ht
, &key
, &iter
);
155 node
= lttng_ht_iter_get_node_u64(&iter
);
157 stream
= caa_container_of(node
, struct lttng_consumer_stream
, node
);
165 static void steal_stream_key(int key
, struct lttng_ht
*ht
)
167 struct lttng_consumer_stream
*stream
;
170 stream
= find_stream(key
, ht
);
174 * We don't want the lookup to match, but we still need
175 * to iterate on this stream when iterating over the hash table. Just
176 * change the node key.
178 stream
->node
.key
= -1ULL;
184 * Return a channel object for the given key.
186 * RCU read side lock MUST be acquired before calling this function and
187 * protects the channel ptr.
189 struct lttng_consumer_channel
*consumer_find_channel(uint64_t key
)
191 struct lttng_ht_iter iter
;
192 struct lttng_ht_node_u64
*node
;
193 struct lttng_consumer_channel
*channel
= NULL
;
195 /* -1ULL keys are lookup failures */
196 if (key
== (uint64_t) -1ULL) {
200 lttng_ht_lookup(consumer_data
.channel_ht
, &key
, &iter
);
201 node
= lttng_ht_iter_get_node_u64(&iter
);
203 channel
= caa_container_of(node
, struct lttng_consumer_channel
, node
);
209 static void free_stream_rcu(struct rcu_head
*head
)
211 struct lttng_ht_node_u64
*node
=
212 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
213 struct lttng_consumer_stream
*stream
=
214 caa_container_of(node
, struct lttng_consumer_stream
, node
);
219 static void free_channel_rcu(struct rcu_head
*head
)
221 struct lttng_ht_node_u64
*node
=
222 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
223 struct lttng_consumer_channel
*channel
=
224 caa_container_of(node
, struct lttng_consumer_channel
, node
);
230 * RCU protected relayd socket pair free.
232 static void free_relayd_rcu(struct rcu_head
*head
)
234 struct lttng_ht_node_u64
*node
=
235 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
236 struct consumer_relayd_sock_pair
*relayd
=
237 caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
240 * Close all sockets. This is done in the call RCU since we don't want the
241 * socket fds to be reassigned thus potentially creating bad state of the
244 * We do not have to lock the control socket mutex here since at this stage
245 * there is no one referencing to this relayd object.
247 (void) relayd_close(&relayd
->control_sock
);
248 (void) relayd_close(&relayd
->data_sock
);
254 * Destroy and free relayd socket pair object.
256 * This function MUST be called with the consumer_data lock acquired.
258 static void destroy_relayd(struct consumer_relayd_sock_pair
*relayd
)
261 struct lttng_ht_iter iter
;
263 if (relayd
== NULL
) {
267 DBG("Consumer destroy and close relayd socket pair");
269 iter
.iter
.node
= &relayd
->node
.node
;
270 ret
= lttng_ht_del(consumer_data
.relayd_ht
, &iter
);
272 /* We assume the relayd is being or is destroyed */
276 /* RCU free() call */
277 call_rcu(&relayd
->node
.head
, free_relayd_rcu
);
281 * Remove a channel from the global list protected by a mutex. This function is
282 * also responsible for freeing its data structures.
284 void consumer_del_channel(struct lttng_consumer_channel
*channel
)
287 struct lttng_ht_iter iter
;
289 DBG("Consumer delete channel key %" PRIu64
, channel
->key
);
291 pthread_mutex_lock(&consumer_data
.lock
);
293 switch (consumer_data
.type
) {
294 case LTTNG_CONSUMER_KERNEL
:
296 case LTTNG_CONSUMER32_UST
:
297 case LTTNG_CONSUMER64_UST
:
298 lttng_ustconsumer_del_channel(channel
);
301 ERR("Unknown consumer_data type");
307 iter
.iter
.node
= &channel
->node
.node
;
308 ret
= lttng_ht_del(consumer_data
.channel_ht
, &iter
);
312 call_rcu(&channel
->node
.head
, free_channel_rcu
);
314 pthread_mutex_unlock(&consumer_data
.lock
);
318 * Iterate over the relayd hash table and destroy each element. Finally,
319 * destroy the whole hash table.
321 static void cleanup_relayd_ht(void)
323 struct lttng_ht_iter iter
;
324 struct consumer_relayd_sock_pair
*relayd
;
328 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
330 destroy_relayd(relayd
);
335 lttng_ht_destroy(consumer_data
.relayd_ht
);
339 * Update the end point status of all streams having the given network sequence
340 * index (relayd index).
342 * It's atomically set without having the stream mutex locked which is fine
343 * because we handle the write/read race with a pipe wakeup for each thread.
345 static void update_endpoint_status_by_netidx(int net_seq_idx
,
346 enum consumer_endpoint_status status
)
348 struct lttng_ht_iter iter
;
349 struct lttng_consumer_stream
*stream
;
351 DBG("Consumer set delete flag on stream by idx %d", net_seq_idx
);
355 /* Let's begin with metadata */
356 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
357 if (stream
->net_seq_idx
== net_seq_idx
) {
358 uatomic_set(&stream
->endpoint_status
, status
);
359 DBG("Delete flag set to metadata stream %d", stream
->wait_fd
);
363 /* Follow up by the data streams */
364 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
365 if (stream
->net_seq_idx
== net_seq_idx
) {
366 uatomic_set(&stream
->endpoint_status
, status
);
367 DBG("Delete flag set to data stream %d", stream
->wait_fd
);
374 * Cleanup a relayd object by flagging every associated streams for deletion,
375 * destroying the object meaning removing it from the relayd hash table,
376 * closing the sockets and freeing the memory in a RCU call.
378 * If a local data context is available, notify the threads that the streams'
379 * state have changed.
381 static void cleanup_relayd(struct consumer_relayd_sock_pair
*relayd
,
382 struct lttng_consumer_local_data
*ctx
)
388 DBG("Cleaning up relayd sockets");
390 /* Save the net sequence index before destroying the object */
391 netidx
= relayd
->net_seq_idx
;
394 * Delete the relayd from the relayd hash table, close the sockets and free
395 * the object in a RCU call.
397 destroy_relayd(relayd
);
399 /* Set inactive endpoint to all streams */
400 update_endpoint_status_by_netidx(netidx
, CONSUMER_ENDPOINT_INACTIVE
);
403 * With a local data context, notify the threads that the streams' state
404 * have changed. The write() action on the pipe acts as an "implicit"
405 * memory barrier ordering the updates of the end point status from the
406 * read of this status which happens AFTER receiving this notify.
409 notify_thread_pipe(ctx
->consumer_data_pipe
[1]);
410 notify_thread_pipe(ctx
->consumer_metadata_pipe
[1]);
415 * Flag a relayd socket pair for destruction. Destroy it if the refcount
418 * RCU read side lock MUST be aquired before calling this function.
420 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair
*relayd
)
424 /* Set destroy flag for this object */
425 uatomic_set(&relayd
->destroy_flag
, 1);
427 /* Destroy the relayd if refcount is 0 */
428 if (uatomic_read(&relayd
->refcount
) == 0) {
429 destroy_relayd(relayd
);
434 * Remove a stream from the global list protected by a mutex. This
435 * function is also responsible for freeing its data structures.
437 void consumer_del_stream(struct lttng_consumer_stream
*stream
,
441 struct lttng_ht_iter iter
;
442 struct lttng_consumer_channel
*free_chan
= NULL
;
443 struct consumer_relayd_sock_pair
*relayd
;
447 DBG("Consumer del stream %d", stream
->wait_fd
);
450 /* Means the stream was allocated but not successfully added */
451 goto free_stream_rcu
;
454 pthread_mutex_lock(&consumer_data
.lock
);
455 pthread_mutex_lock(&stream
->lock
);
457 switch (consumer_data
.type
) {
458 case LTTNG_CONSUMER_KERNEL
:
459 if (stream
->mmap_base
!= NULL
) {
460 ret
= munmap(stream
->mmap_base
, stream
->mmap_len
);
466 case LTTNG_CONSUMER32_UST
:
467 case LTTNG_CONSUMER64_UST
:
468 lttng_ustconsumer_del_stream(stream
);
471 ERR("Unknown consumer_data type");
477 iter
.iter
.node
= &stream
->node
.node
;
478 ret
= lttng_ht_del(ht
, &iter
);
481 iter
.iter
.node
= &stream
->node_channel_id
.node
;
482 ret
= lttng_ht_del(consumer_data
.stream_per_chan_id_ht
, &iter
);
485 iter
.iter
.node
= &stream
->node_session_id
.node
;
486 ret
= lttng_ht_del(consumer_data
.stream_list_ht
, &iter
);
490 assert(consumer_data
.stream_count
> 0);
491 consumer_data
.stream_count
--;
493 if (stream
->out_fd
>= 0) {
494 ret
= close(stream
->out_fd
);
500 /* Check and cleanup relayd */
502 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
503 if (relayd
!= NULL
) {
504 uatomic_dec(&relayd
->refcount
);
505 assert(uatomic_read(&relayd
->refcount
) >= 0);
507 /* Closing streams requires to lock the control socket. */
508 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
509 ret
= relayd_send_close_stream(&relayd
->control_sock
,
510 stream
->relayd_stream_id
,
511 stream
->next_net_seq_num
- 1);
512 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
514 DBG("Unable to close stream on the relayd. Continuing");
516 * Continue here. There is nothing we can do for the relayd.
517 * Chances are that the relayd has closed the socket so we just
518 * continue cleaning up.
522 /* Both conditions are met, we destroy the relayd. */
523 if (uatomic_read(&relayd
->refcount
) == 0 &&
524 uatomic_read(&relayd
->destroy_flag
)) {
525 destroy_relayd(relayd
);
530 if (!uatomic_sub_return(&stream
->chan
->refcount
, 1)
531 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
532 free_chan
= stream
->chan
;
536 consumer_data
.need_update
= 1;
537 pthread_mutex_unlock(&stream
->lock
);
538 pthread_mutex_unlock(&consumer_data
.lock
);
541 consumer_del_channel(free_chan
);
545 call_rcu(&stream
->node
.head
, free_stream_rcu
);
548 struct lttng_consumer_stream
*consumer_allocate_stream(uint64_t channel_key
,
550 enum lttng_consumer_stream_state state
,
551 const char *channel_name
,
558 enum consumer_channel_type type
)
561 struct lttng_consumer_stream
*stream
;
563 stream
= zmalloc(sizeof(*stream
));
564 if (stream
== NULL
) {
565 PERROR("malloc struct lttng_consumer_stream");
572 stream
->key
= stream_key
;
574 stream
->out_fd_offset
= 0;
575 stream
->state
= state
;
578 stream
->net_seq_idx
= relayd_id
;
579 stream
->session_id
= session_id
;
580 pthread_mutex_init(&stream
->lock
, NULL
);
582 /* If channel is the metadata, flag this stream as metadata. */
583 if (type
== CONSUMER_CHANNEL_TYPE_METADATA
) {
584 stream
->metadata_flag
= 1;
585 /* Metadata is flat out. */
586 strncpy(stream
->name
, DEFAULT_METADATA_NAME
, sizeof(stream
->name
));
588 /* Format stream name to <channel_name>_<cpu_number> */
589 ret
= snprintf(stream
->name
, sizeof(stream
->name
), "%s_%d",
592 PERROR("snprintf stream name");
597 /* Key is always the wait_fd for streams. */
598 lttng_ht_node_init_u64(&stream
->node
, stream
->key
);
600 /* Init node per channel id key */
601 lttng_ht_node_init_u64(&stream
->node_channel_id
, channel_key
);
603 /* Init session id node with the stream session id */
604 lttng_ht_node_init_u64(&stream
->node_session_id
, stream
->session_id
);
606 DBG3("Allocated stream %s (key %" PRIu64
", chan_key %" PRIu64
" relayd_id %" PRIu64
", session_id %" PRIu64
,
607 stream
->name
, stream
->key
, channel_key
, stream
->net_seq_idx
, stream
->session_id
);
623 * Add a stream to the global list protected by a mutex.
625 static int add_stream(struct lttng_consumer_stream
*stream
,
629 struct consumer_relayd_sock_pair
*relayd
;
634 DBG3("Adding consumer stream %" PRIu64
, stream
->key
);
636 pthread_mutex_lock(&consumer_data
.lock
);
637 pthread_mutex_lock(&stream
->lock
);
640 /* Steal stream identifier to avoid having streams with the same key */
641 steal_stream_key(stream
->key
, ht
);
643 lttng_ht_add_unique_u64(ht
, &stream
->node
);
645 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
646 &stream
->node_channel_id
);
649 * Add stream to the stream_list_ht of the consumer data. No need to steal
650 * the key since the HT does not use it and we allow to add redundant keys
653 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
655 /* Check and cleanup relayd */
656 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
657 if (relayd
!= NULL
) {
658 uatomic_inc(&relayd
->refcount
);
661 /* Update channel refcount once added without error(s). */
662 uatomic_inc(&stream
->chan
->refcount
);
665 * When nb_init_stream_left reaches 0, we don't need to trigger any action
666 * in terms of destroying the associated channel, because the action that
667 * causes the count to become 0 also causes a stream to be added. The
668 * channel deletion will thus be triggered by the following removal of this
671 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
672 /* Increment refcount before decrementing nb_init_stream_left */
674 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
677 /* Update consumer data once the node is inserted. */
678 consumer_data
.stream_count
++;
679 consumer_data
.need_update
= 1;
682 pthread_mutex_unlock(&stream
->lock
);
683 pthread_mutex_unlock(&consumer_data
.lock
);
689 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
690 * be acquired before calling this.
692 static int add_relayd(struct consumer_relayd_sock_pair
*relayd
)
695 struct lttng_ht_node_u64
*node
;
696 struct lttng_ht_iter iter
;
700 lttng_ht_lookup(consumer_data
.relayd_ht
,
701 &relayd
->net_seq_idx
, &iter
);
702 node
= lttng_ht_iter_get_node_u64(&iter
);
706 lttng_ht_add_unique_u64(consumer_data
.relayd_ht
, &relayd
->node
);
713 * Allocate and return a consumer relayd socket.
715 struct consumer_relayd_sock_pair
*consumer_allocate_relayd_sock_pair(
718 struct consumer_relayd_sock_pair
*obj
= NULL
;
720 /* Negative net sequence index is a failure */
721 if (net_seq_idx
< 0) {
725 obj
= zmalloc(sizeof(struct consumer_relayd_sock_pair
));
727 PERROR("zmalloc relayd sock");
731 obj
->net_seq_idx
= net_seq_idx
;
733 obj
->destroy_flag
= 0;
734 obj
->control_sock
.sock
.fd
= -1;
735 obj
->data_sock
.sock
.fd
= -1;
736 lttng_ht_node_init_u64(&obj
->node
, obj
->net_seq_idx
);
737 pthread_mutex_init(&obj
->ctrl_sock_mutex
, NULL
);
744 * Find a relayd socket pair in the global consumer data.
746 * Return the object if found else NULL.
747 * RCU read-side lock must be held across this call and while using the
750 struct consumer_relayd_sock_pair
*consumer_find_relayd(uint64_t key
)
752 struct lttng_ht_iter iter
;
753 struct lttng_ht_node_u64
*node
;
754 struct consumer_relayd_sock_pair
*relayd
= NULL
;
756 /* Negative keys are lookup failures */
757 if (key
== (uint64_t) -1ULL) {
761 lttng_ht_lookup(consumer_data
.relayd_ht
, &key
,
763 node
= lttng_ht_iter_get_node_u64(&iter
);
765 relayd
= caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
773 * Handle stream for relayd transmission if the stream applies for network
774 * streaming where the net sequence index is set.
776 * Return destination file descriptor or negative value on error.
778 static int write_relayd_stream_header(struct lttng_consumer_stream
*stream
,
779 size_t data_size
, unsigned long padding
,
780 struct consumer_relayd_sock_pair
*relayd
)
783 struct lttcomm_relayd_data_hdr data_hdr
;
789 /* Reset data header */
790 memset(&data_hdr
, 0, sizeof(data_hdr
));
792 if (stream
->metadata_flag
) {
793 /* Caller MUST acquire the relayd control socket lock */
794 ret
= relayd_send_metadata(&relayd
->control_sock
, data_size
);
799 /* Metadata are always sent on the control socket. */
800 outfd
= relayd
->control_sock
.sock
.fd
;
802 /* Set header with stream information */
803 data_hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
804 data_hdr
.data_size
= htobe32(data_size
);
805 data_hdr
.padding_size
= htobe32(padding
);
807 * Note that net_seq_num below is assigned with the *current* value of
808 * next_net_seq_num and only after that the next_net_seq_num will be
809 * increment. This is why when issuing a command on the relayd using
810 * this next value, 1 should always be substracted in order to compare
811 * the last seen sequence number on the relayd side to the last sent.
813 data_hdr
.net_seq_num
= htobe64(stream
->next_net_seq_num
);
814 /* Other fields are zeroed previously */
816 ret
= relayd_send_data_hdr(&relayd
->data_sock
, &data_hdr
,
822 ++stream
->next_net_seq_num
;
824 /* Set to go on data socket */
825 outfd
= relayd
->data_sock
.sock
.fd
;
833 * Allocate and return a new lttng_consumer_channel object using the given key
834 * to initialize the hash table node.
836 * On error, return NULL.
838 struct lttng_consumer_channel
*consumer_allocate_channel(uint64_t key
,
840 const char *pathname
,
845 enum lttng_event_output output
,
846 uint64_t tracefile_size
,
847 uint64_t tracefile_count
)
849 struct lttng_consumer_channel
*channel
;
851 channel
= zmalloc(sizeof(*channel
));
852 if (channel
== NULL
) {
853 PERROR("malloc struct lttng_consumer_channel");
858 channel
->refcount
= 0;
859 channel
->session_id
= session_id
;
862 channel
->relayd_id
= relayd_id
;
863 channel
->output
= output
;
864 channel
->tracefile_size
= tracefile_size
;
865 channel
->tracefile_count
= tracefile_count
;
867 strncpy(channel
->pathname
, pathname
, sizeof(channel
->pathname
));
868 channel
->pathname
[sizeof(channel
->pathname
) - 1] = '\0';
870 strncpy(channel
->name
, name
, sizeof(channel
->name
));
871 channel
->name
[sizeof(channel
->name
) - 1] = '\0';
873 lttng_ht_node_init_u64(&channel
->node
, channel
->key
);
875 channel
->wait_fd
= -1;
877 CDS_INIT_LIST_HEAD(&channel
->streams
.head
);
879 DBG("Allocated channel (key %" PRIu64
")", channel
->key
)
886 * Add a channel to the global list protected by a mutex.
888 int consumer_add_channel(struct lttng_consumer_channel
*channel
,
889 struct lttng_consumer_local_data
*ctx
)
892 struct lttng_ht_node_u64
*node
;
893 struct lttng_ht_iter iter
;
895 pthread_mutex_lock(&consumer_data
.lock
);
898 lttng_ht_lookup(consumer_data
.channel_ht
, &channel
->key
, &iter
);
899 node
= lttng_ht_iter_get_node_u64(&iter
);
901 /* Channel already exist. Ignore the insertion */
902 ERR("Consumer add channel key %" PRIu64
" already exists!",
908 lttng_ht_add_unique_u64(consumer_data
.channel_ht
, &channel
->node
);
912 pthread_mutex_unlock(&consumer_data
.lock
);
914 if (!ret
&& channel
->wait_fd
!= -1 &&
915 channel
->metadata_stream
== NULL
) {
916 notify_channel_pipe(ctx
, channel
, -1, CONSUMER_CHANNEL_ADD
);
922 * Allocate the pollfd structure and the local view of the out fds to avoid
923 * doing a lookup in the linked list and concurrency issues when writing is
924 * needed. Called with consumer_data.lock held.
926 * Returns the number of fds in the structures.
928 static int update_poll_array(struct lttng_consumer_local_data
*ctx
,
929 struct pollfd
**pollfd
, struct lttng_consumer_stream
**local_stream
,
933 struct lttng_ht_iter iter
;
934 struct lttng_consumer_stream
*stream
;
939 assert(local_stream
);
941 DBG("Updating poll fd array");
943 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
945 * Only active streams with an active end point can be added to the
946 * poll set and local stream storage of the thread.
948 * There is a potential race here for endpoint_status to be updated
949 * just after the check. However, this is OK since the stream(s) will
950 * be deleted once the thread is notified that the end point state has
951 * changed where this function will be called back again.
953 if (stream
->state
!= LTTNG_CONSUMER_ACTIVE_STREAM
||
954 stream
->endpoint_status
== CONSUMER_ENDPOINT_INACTIVE
) {
958 * This clobbers way too much the debug output. Uncomment that if you
959 * need it for debugging purposes.
961 * DBG("Active FD %d", stream->wait_fd);
963 (*pollfd
)[i
].fd
= stream
->wait_fd
;
964 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
965 local_stream
[i
] = stream
;
971 * Insert the consumer_data_pipe at the end of the array and don't
972 * increment i so nb_fd is the number of real FD.
974 (*pollfd
)[i
].fd
= ctx
->consumer_data_pipe
[0];
975 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
980 * Poll on the should_quit pipe and the command socket return -1 on error and
981 * should exit, 0 if data is available on the command socket
983 int lttng_consumer_poll_socket(struct pollfd
*consumer_sockpoll
)
988 num_rdy
= poll(consumer_sockpoll
, 2, -1);
991 * Restart interrupted system call.
993 if (errno
== EINTR
) {
996 PERROR("Poll error");
999 if (consumer_sockpoll
[0].revents
& (POLLIN
| POLLPRI
)) {
1000 DBG("consumer_should_quit wake up");
1010 * Set the error socket.
1012 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data
*ctx
,
1015 ctx
->consumer_error_socket
= sock
;
1019 * Set the command socket path.
1021 void lttng_consumer_set_command_sock_path(
1022 struct lttng_consumer_local_data
*ctx
, char *sock
)
1024 ctx
->consumer_command_sock_path
= sock
;
1028 * Send return code to the session daemon.
1029 * If the socket is not defined, we return 0, it is not a fatal error
1031 int lttng_consumer_send_error(struct lttng_consumer_local_data
*ctx
, int cmd
)
1033 if (ctx
->consumer_error_socket
> 0) {
1034 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
1035 sizeof(enum lttcomm_sessiond_command
));
1042 * Close all the tracefiles and stream fds and MUST be called when all
1043 * instances are destroyed i.e. when all threads were joined and are ended.
1045 void lttng_consumer_cleanup(void)
1047 struct lttng_ht_iter iter
;
1048 struct lttng_consumer_channel
*channel
;
1052 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
, channel
,
1054 consumer_del_channel(channel
);
1059 lttng_ht_destroy(consumer_data
.channel_ht
);
1061 cleanup_relayd_ht();
1063 lttng_ht_destroy(consumer_data
.stream_per_chan_id_ht
);
1066 * This HT contains streams that are freed by either the metadata thread or
1067 * the data thread so we do *nothing* on the hash table and simply destroy
1070 lttng_ht_destroy(consumer_data
.stream_list_ht
);
1074 * Called from signal handler.
1076 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
1081 ret
= write(ctx
->consumer_should_quit
[1], "4", 1);
1082 } while (ret
< 0 && errno
== EINTR
);
1083 if (ret
< 0 || ret
!= 1) {
1084 PERROR("write consumer quit");
1087 DBG("Consumer flag that it should quit");
1090 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1093 int outfd
= stream
->out_fd
;
1096 * This does a blocking write-and-wait on any page that belongs to the
1097 * subbuffer prior to the one we just wrote.
1098 * Don't care about error values, as these are just hints and ways to
1099 * limit the amount of page cache used.
1101 if (orig_offset
< stream
->max_sb_size
) {
1104 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1105 stream
->max_sb_size
,
1106 SYNC_FILE_RANGE_WAIT_BEFORE
1107 | SYNC_FILE_RANGE_WRITE
1108 | SYNC_FILE_RANGE_WAIT_AFTER
);
1110 * Give hints to the kernel about how we access the file:
1111 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1114 * We need to call fadvise again after the file grows because the
1115 * kernel does not seem to apply fadvise to non-existing parts of the
1118 * Call fadvise _after_ having waited for the page writeback to
1119 * complete because the dirty page writeback semantic is not well
1120 * defined. So it can be expected to lead to lower throughput in
1123 posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1124 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1128 * Initialise the necessary environnement :
1129 * - create a new context
1130 * - create the poll_pipe
1131 * - create the should_quit pipe (for signal handler)
1132 * - create the thread pipe (for splice)
1134 * Takes a function pointer as argument, this function is called when data is
1135 * available on a buffer. This function is responsible to do the
1136 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1137 * buffer configuration and then kernctl_put_next_subbuf at the end.
1139 * Returns a pointer to the new context or NULL on error.
1141 struct lttng_consumer_local_data
*lttng_consumer_create(
1142 enum lttng_consumer_type type
,
1143 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1144 struct lttng_consumer_local_data
*ctx
),
1145 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1146 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1147 int (*update_stream
)(int stream_key
, uint32_t state
))
1150 struct lttng_consumer_local_data
*ctx
;
1152 assert(consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1153 consumer_data
.type
== type
);
1154 consumer_data
.type
= type
;
1156 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
1158 PERROR("allocating context");
1162 ctx
->consumer_error_socket
= -1;
1163 ctx
->consumer_metadata_socket
= -1;
1164 /* assign the callbacks */
1165 ctx
->on_buffer_ready
= buffer_ready
;
1166 ctx
->on_recv_channel
= recv_channel
;
1167 ctx
->on_recv_stream
= recv_stream
;
1168 ctx
->on_update_stream
= update_stream
;
1170 ret
= pipe(ctx
->consumer_data_pipe
);
1172 PERROR("Error creating poll pipe");
1173 goto error_poll_pipe
;
1176 /* set read end of the pipe to non-blocking */
1177 ret
= fcntl(ctx
->consumer_data_pipe
[0], F_SETFL
, O_NONBLOCK
);
1179 PERROR("fcntl O_NONBLOCK");
1180 goto error_poll_fcntl
;
1183 /* set write end of the pipe to non-blocking */
1184 ret
= fcntl(ctx
->consumer_data_pipe
[1], F_SETFL
, O_NONBLOCK
);
1186 PERROR("fcntl O_NONBLOCK");
1187 goto error_poll_fcntl
;
1190 ret
= pipe(ctx
->consumer_should_quit
);
1192 PERROR("Error creating recv pipe");
1193 goto error_quit_pipe
;
1196 ret
= pipe(ctx
->consumer_thread_pipe
);
1198 PERROR("Error creating thread pipe");
1199 goto error_thread_pipe
;
1202 ret
= pipe(ctx
->consumer_channel_pipe
);
1204 PERROR("Error creating channel pipe");
1205 goto error_channel_pipe
;
1208 ret
= utils_create_pipe(ctx
->consumer_metadata_pipe
);
1210 goto error_metadata_pipe
;
1213 ret
= utils_create_pipe(ctx
->consumer_splice_metadata_pipe
);
1215 goto error_splice_pipe
;
1221 utils_close_pipe(ctx
->consumer_metadata_pipe
);
1222 error_metadata_pipe
:
1223 utils_close_pipe(ctx
->consumer_channel_pipe
);
1225 utils_close_pipe(ctx
->consumer_thread_pipe
);
1227 utils_close_pipe(ctx
->consumer_should_quit
);
1230 utils_close_pipe(ctx
->consumer_data_pipe
);
1238 * Close all fds associated with the instance and free the context.
1240 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1244 DBG("Consumer destroying it. Closing everything.");
1246 ret
= close(ctx
->consumer_error_socket
);
1250 ret
= close(ctx
->consumer_metadata_socket
);
1254 utils_close_pipe(ctx
->consumer_thread_pipe
);
1255 utils_close_pipe(ctx
->consumer_channel_pipe
);
1256 utils_close_pipe(ctx
->consumer_data_pipe
);
1257 utils_close_pipe(ctx
->consumer_should_quit
);
1258 utils_close_pipe(ctx
->consumer_splice_metadata_pipe
);
1260 unlink(ctx
->consumer_command_sock_path
);
1265 * Write the metadata stream id on the specified file descriptor.
1267 static int write_relayd_metadata_id(int fd
,
1268 struct lttng_consumer_stream
*stream
,
1269 struct consumer_relayd_sock_pair
*relayd
, unsigned long padding
)
1272 struct lttcomm_relayd_metadata_payload hdr
;
1274 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1275 hdr
.padding_size
= htobe32(padding
);
1277 ret
= write(fd
, (void *) &hdr
, sizeof(hdr
));
1278 } while (ret
< 0 && errno
== EINTR
);
1279 if (ret
< 0 || ret
!= sizeof(hdr
)) {
1281 * This error means that the fd's end is closed so ignore the perror
1282 * not to clubber the error output since this can happen in a normal
1285 if (errno
!= EPIPE
) {
1286 PERROR("write metadata stream id");
1288 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1290 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1291 * handle writting the missing part so report that as an error and
1292 * don't lie to the caller.
1297 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1298 stream
->relayd_stream_id
, padding
);
1305 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1306 * core function for writing trace buffers to either the local filesystem or
1309 * It must be called with the stream lock held.
1311 * Careful review MUST be put if any changes occur!
1313 * Returns the number of bytes written
1315 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1316 struct lttng_consumer_local_data
*ctx
,
1317 struct lttng_consumer_stream
*stream
, unsigned long len
,
1318 unsigned long padding
)
1320 unsigned long mmap_offset
;
1322 ssize_t ret
= 0, written
= 0;
1323 off_t orig_offset
= stream
->out_fd_offset
;
1324 /* Default is on the disk */
1325 int outfd
= stream
->out_fd
;
1326 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1327 unsigned int relayd_hang_up
= 0;
1329 /* RCU lock for the relayd pointer */
1332 /* Flag that the current stream if set for network streaming. */
1333 if (stream
->net_seq_idx
!= -1) {
1334 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1335 if (relayd
== NULL
) {
1340 /* get the offset inside the fd to mmap */
1341 switch (consumer_data
.type
) {
1342 case LTTNG_CONSUMER_KERNEL
:
1343 mmap_base
= stream
->mmap_base
;
1344 ret
= kernctl_get_mmap_read_offset(stream
->wait_fd
, &mmap_offset
);
1346 case LTTNG_CONSUMER32_UST
:
1347 case LTTNG_CONSUMER64_UST
:
1348 mmap_base
= lttng_ustctl_get_mmap_base(stream
);
1350 ERR("read mmap get mmap base for stream %s", stream
->name
);
1354 ret
= lttng_ustctl_get_mmap_read_offset(stream
, &mmap_offset
);
1358 ERR("Unknown consumer_data type");
1363 PERROR("tracer ctl get_mmap_read_offset");
1368 /* Handle stream on the relayd if the output is on the network */
1370 unsigned long netlen
= len
;
1373 * Lock the control socket for the complete duration of the function
1374 * since from this point on we will use the socket.
1376 if (stream
->metadata_flag
) {
1377 /* Metadata requires the control socket. */
1378 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1379 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1382 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1384 /* Use the returned socket. */
1387 /* Write metadata stream id before payload */
1388 if (stream
->metadata_flag
) {
1389 ret
= write_relayd_metadata_id(outfd
, stream
, relayd
, padding
);
1392 /* Socket operation failed. We consider the relayd dead */
1393 if (ret
== -EPIPE
|| ret
== -EINVAL
) {
1401 /* Socket operation failed. We consider the relayd dead */
1402 if (ret
== -EPIPE
|| ret
== -EINVAL
) {
1406 /* Else, use the default set before which is the filesystem. */
1409 /* No streaming, we have to set the len with the full padding */
1413 * Check if we need to change the tracefile before writing the packet.
1415 if (stream
->chan
->tracefile_size
> 0 &&
1416 (stream
->tracefile_size_current
+ len
) >
1417 stream
->chan
->tracefile_size
) {
1418 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1419 stream
->name
, stream
->chan
->tracefile_size
,
1420 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1421 stream
->out_fd
, &(stream
->tracefile_count_current
));
1423 ERR("Rotating output file");
1426 outfd
= stream
->out_fd
= ret
;
1427 /* Reset current size because we just perform a rotation. */
1428 stream
->tracefile_size_current
= 0;
1430 stream
->tracefile_size_current
+= len
;
1435 ret
= write(outfd
, mmap_base
+ mmap_offset
, len
);
1436 } while (ret
< 0 && errno
== EINTR
);
1437 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, len
);
1440 * This is possible if the fd is closed on the other side (outfd)
1441 * or any write problem. It can be verbose a bit for a normal
1442 * execution if for instance the relayd is stopped abruptly. This
1443 * can happen so set this to a DBG statement.
1445 DBG("Error in file write mmap");
1449 /* Socket operation failed. We consider the relayd dead */
1450 if (errno
== EPIPE
|| errno
== EINVAL
) {
1455 } else if (ret
> len
) {
1456 PERROR("Error in file write (ret %zd > len %lu)", ret
, len
);
1464 /* This call is useless on a socket so better save a syscall. */
1466 /* This won't block, but will start writeout asynchronously */
1467 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret
,
1468 SYNC_FILE_RANGE_WRITE
);
1469 stream
->out_fd_offset
+= ret
;
1473 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1477 * This is a special case that the relayd has closed its socket. Let's
1478 * cleanup the relayd object and all associated streams.
1480 if (relayd
&& relayd_hang_up
) {
1481 cleanup_relayd(relayd
, ctx
);
1485 /* Unlock only if ctrl socket used */
1486 if (relayd
&& stream
->metadata_flag
) {
1487 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1495 * Splice the data from the ring buffer to the tracefile.
1497 * It must be called with the stream lock held.
1499 * Returns the number of bytes spliced.
1501 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1502 struct lttng_consumer_local_data
*ctx
,
1503 struct lttng_consumer_stream
*stream
, unsigned long len
,
1504 unsigned long padding
)
1506 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1508 off_t orig_offset
= stream
->out_fd_offset
;
1509 int fd
= stream
->wait_fd
;
1510 /* Default is on the disk */
1511 int outfd
= stream
->out_fd
;
1512 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1514 unsigned int relayd_hang_up
= 0;
1516 switch (consumer_data
.type
) {
1517 case LTTNG_CONSUMER_KERNEL
:
1519 case LTTNG_CONSUMER32_UST
:
1520 case LTTNG_CONSUMER64_UST
:
1521 /* Not supported for user space tracing */
1524 ERR("Unknown consumer_data type");
1528 /* RCU lock for the relayd pointer */
1531 /* Flag that the current stream if set for network streaming. */
1532 if (stream
->net_seq_idx
!= -1) {
1533 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1534 if (relayd
== NULL
) {
1540 * Choose right pipe for splice. Metadata and trace data are handled by
1541 * different threads hence the use of two pipes in order not to race or
1542 * corrupt the written data.
1544 if (stream
->metadata_flag
) {
1545 splice_pipe
= ctx
->consumer_splice_metadata_pipe
;
1547 splice_pipe
= ctx
->consumer_thread_pipe
;
1550 /* Write metadata stream id before payload */
1552 int total_len
= len
;
1554 if (stream
->metadata_flag
) {
1556 * Lock the control socket for the complete duration of the function
1557 * since from this point on we will use the socket.
1559 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1561 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
, relayd
,
1565 /* Socket operation failed. We consider the relayd dead */
1566 if (ret
== -EBADF
) {
1567 WARN("Remote relayd disconnected. Stopping");
1574 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1577 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1579 /* Use the returned socket. */
1582 /* Socket operation failed. We consider the relayd dead */
1583 if (ret
== -EBADF
) {
1584 WARN("Remote relayd disconnected. Stopping");
1591 /* No streaming, we have to set the len with the full padding */
1595 * Check if we need to change the tracefile before writing the packet.
1597 if (stream
->chan
->tracefile_size
> 0 &&
1598 (stream
->tracefile_size_current
+ len
) >
1599 stream
->chan
->tracefile_size
) {
1600 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1601 stream
->name
, stream
->chan
->tracefile_size
,
1602 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1603 stream
->out_fd
, &(stream
->tracefile_count_current
));
1605 ERR("Rotating output file");
1608 outfd
= stream
->out_fd
= ret
;
1609 /* Reset current size because we just perform a rotation. */
1610 stream
->tracefile_size_current
= 0;
1612 stream
->tracefile_size_current
+= len
;
1616 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1617 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1618 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1619 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1620 DBG("splice chan to pipe, ret %zd", ret_splice
);
1621 if (ret_splice
< 0) {
1622 PERROR("Error in relay splice");
1624 written
= ret_splice
;
1630 /* Handle stream on the relayd if the output is on the network */
1632 if (stream
->metadata_flag
) {
1633 size_t metadata_payload_size
=
1634 sizeof(struct lttcomm_relayd_metadata_payload
);
1636 /* Update counter to fit the spliced data */
1637 ret_splice
+= metadata_payload_size
;
1638 len
+= metadata_payload_size
;
1640 * We do this so the return value can match the len passed as
1641 * argument to this function.
1643 written
-= metadata_payload_size
;
1647 /* Splice data out */
1648 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1649 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1650 DBG("Consumer splice pipe to file, ret %zd", ret_splice
);
1651 if (ret_splice
< 0) {
1652 PERROR("Error in file splice");
1654 written
= ret_splice
;
1656 /* Socket operation failed. We consider the relayd dead */
1657 if (errno
== EBADF
|| errno
== EPIPE
) {
1658 WARN("Remote relayd disconnected. Stopping");
1664 } else if (ret_splice
> len
) {
1666 PERROR("Wrote more data than requested %zd (len: %lu)",
1668 written
+= ret_splice
;
1674 /* This call is useless on a socket so better save a syscall. */
1676 /* This won't block, but will start writeout asynchronously */
1677 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1678 SYNC_FILE_RANGE_WRITE
);
1679 stream
->out_fd_offset
+= ret_splice
;
1681 written
+= ret_splice
;
1683 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1691 * This is a special case that the relayd has closed its socket. Let's
1692 * cleanup the relayd object and all associated streams.
1694 if (relayd
&& relayd_hang_up
) {
1695 cleanup_relayd(relayd
, ctx
);
1696 /* Skip splice error so the consumer does not fail */
1701 /* send the appropriate error description to sessiond */
1704 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1707 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1710 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
1715 if (relayd
&& stream
->metadata_flag
) {
1716 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1724 * Take a snapshot for a specific fd
1726 * Returns 0 on success, < 0 on error
1728 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
1730 switch (consumer_data
.type
) {
1731 case LTTNG_CONSUMER_KERNEL
:
1732 return lttng_kconsumer_take_snapshot(stream
);
1733 case LTTNG_CONSUMER32_UST
:
1734 case LTTNG_CONSUMER64_UST
:
1735 return lttng_ustconsumer_take_snapshot(stream
);
1737 ERR("Unknown consumer_data type");
1744 * Get the produced position
1746 * Returns 0 on success, < 0 on error
1748 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
1751 switch (consumer_data
.type
) {
1752 case LTTNG_CONSUMER_KERNEL
:
1753 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
1754 case LTTNG_CONSUMER32_UST
:
1755 case LTTNG_CONSUMER64_UST
:
1756 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
1758 ERR("Unknown consumer_data type");
1764 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
1765 int sock
, struct pollfd
*consumer_sockpoll
)
1767 switch (consumer_data
.type
) {
1768 case LTTNG_CONSUMER_KERNEL
:
1769 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1770 case LTTNG_CONSUMER32_UST
:
1771 case LTTNG_CONSUMER64_UST
:
1772 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1774 ERR("Unknown consumer_data type");
1781 * Iterate over all streams of the hashtable and free them properly.
1783 * WARNING: *MUST* be used with data stream only.
1785 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1787 struct lttng_ht_iter iter
;
1788 struct lttng_consumer_stream
*stream
;
1795 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1797 * Ignore return value since we are currently cleaning up so any error
1800 (void) consumer_del_stream(stream
, ht
);
1804 lttng_ht_destroy(ht
);
1808 * Iterate over all streams of the hashtable and free them properly.
1810 * XXX: Should not be only for metadata stream or else use an other name.
1812 static void destroy_stream_ht(struct lttng_ht
*ht
)
1814 struct lttng_ht_iter iter
;
1815 struct lttng_consumer_stream
*stream
;
1822 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1824 * Ignore return value since we are currently cleaning up so any error
1827 (void) consumer_del_metadata_stream(stream
, ht
);
1831 lttng_ht_destroy(ht
);
1834 void lttng_consumer_close_metadata(void)
1836 switch (consumer_data
.type
) {
1837 case LTTNG_CONSUMER_KERNEL
:
1839 * The Kernel consumer has a different metadata scheme so we don't
1840 * close anything because the stream will be closed by the session
1844 case LTTNG_CONSUMER32_UST
:
1845 case LTTNG_CONSUMER64_UST
:
1847 * Close all metadata streams. The metadata hash table is passed and
1848 * this call iterates over it by closing all wakeup fd. This is safe
1849 * because at this point we are sure that the metadata producer is
1850 * either dead or blocked.
1852 lttng_ustconsumer_close_metadata(metadata_ht
);
1855 ERR("Unknown consumer_data type");
1861 * Clean up a metadata stream and free its memory.
1863 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
1864 struct lttng_ht
*ht
)
1867 struct lttng_ht_iter iter
;
1868 struct lttng_consumer_channel
*free_chan
= NULL
;
1869 struct consumer_relayd_sock_pair
*relayd
;
1873 * This call should NEVER receive regular stream. It must always be
1874 * metadata stream and this is crucial for data structure synchronization.
1876 assert(stream
->metadata_flag
);
1878 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
1881 /* Means the stream was allocated but not successfully added */
1882 goto free_stream_rcu
;
1885 pthread_mutex_lock(&consumer_data
.lock
);
1886 pthread_mutex_lock(&stream
->lock
);
1888 switch (consumer_data
.type
) {
1889 case LTTNG_CONSUMER_KERNEL
:
1890 if (stream
->mmap_base
!= NULL
) {
1891 ret
= munmap(stream
->mmap_base
, stream
->mmap_len
);
1893 PERROR("munmap metadata stream");
1897 case LTTNG_CONSUMER32_UST
:
1898 case LTTNG_CONSUMER64_UST
:
1899 lttng_ustconsumer_del_stream(stream
);
1902 ERR("Unknown consumer_data type");
1908 iter
.iter
.node
= &stream
->node
.node
;
1909 ret
= lttng_ht_del(ht
, &iter
);
1912 iter
.iter
.node
= &stream
->node_channel_id
.node
;
1913 ret
= lttng_ht_del(consumer_data
.stream_per_chan_id_ht
, &iter
);
1916 iter
.iter
.node
= &stream
->node_session_id
.node
;
1917 ret
= lttng_ht_del(consumer_data
.stream_list_ht
, &iter
);
1921 if (stream
->out_fd
>= 0) {
1922 ret
= close(stream
->out_fd
);
1928 /* Check and cleanup relayd */
1930 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1931 if (relayd
!= NULL
) {
1932 uatomic_dec(&relayd
->refcount
);
1933 assert(uatomic_read(&relayd
->refcount
) >= 0);
1935 /* Closing streams requires to lock the control socket. */
1936 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1937 ret
= relayd_send_close_stream(&relayd
->control_sock
,
1938 stream
->relayd_stream_id
, stream
->next_net_seq_num
- 1);
1939 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1941 DBG("Unable to close stream on the relayd. Continuing");
1943 * Continue here. There is nothing we can do for the relayd.
1944 * Chances are that the relayd has closed the socket so we just
1945 * continue cleaning up.
1949 /* Both conditions are met, we destroy the relayd. */
1950 if (uatomic_read(&relayd
->refcount
) == 0 &&
1951 uatomic_read(&relayd
->destroy_flag
)) {
1952 destroy_relayd(relayd
);
1957 /* Atomically decrement channel refcount since other threads can use it. */
1958 if (!uatomic_sub_return(&stream
->chan
->refcount
, 1)
1959 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
1960 /* Go for channel deletion! */
1961 free_chan
= stream
->chan
;
1965 pthread_mutex_unlock(&stream
->lock
);
1966 pthread_mutex_unlock(&consumer_data
.lock
);
1969 consumer_del_channel(free_chan
);
1973 call_rcu(&stream
->node
.head
, free_stream_rcu
);
1977 * Action done with the metadata stream when adding it to the consumer internal
1978 * data structures to handle it.
1980 static int add_metadata_stream(struct lttng_consumer_stream
*stream
,
1981 struct lttng_ht
*ht
)
1984 struct consumer_relayd_sock_pair
*relayd
;
1985 struct lttng_ht_iter iter
;
1986 struct lttng_ht_node_u64
*node
;
1991 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
1993 pthread_mutex_lock(&consumer_data
.lock
);
1994 pthread_mutex_lock(&stream
->lock
);
1997 * From here, refcounts are updated so be _careful_ when returning an error
2004 * Lookup the stream just to make sure it does not exist in our internal
2005 * state. This should NEVER happen.
2007 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2008 node
= lttng_ht_iter_get_node_u64(&iter
);
2011 /* Find relayd and, if one is found, increment refcount. */
2012 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
2013 if (relayd
!= NULL
) {
2014 uatomic_inc(&relayd
->refcount
);
2017 /* Update channel refcount once added without error(s). */
2018 uatomic_inc(&stream
->chan
->refcount
);
2021 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2022 * in terms of destroying the associated channel, because the action that
2023 * causes the count to become 0 also causes a stream to be added. The
2024 * channel deletion will thus be triggered by the following removal of this
2027 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2028 /* Increment refcount before decrementing nb_init_stream_left */
2030 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2033 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2035 lttng_ht_add_unique_u64(consumer_data
.stream_per_chan_id_ht
,
2036 &stream
->node_channel_id
);
2039 * Add stream to the stream_list_ht of the consumer data. No need to steal
2040 * the key since the HT does not use it and we allow to add redundant keys
2043 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2047 pthread_mutex_unlock(&stream
->lock
);
2048 pthread_mutex_unlock(&consumer_data
.lock
);
2053 * Delete data stream that are flagged for deletion (endpoint_status).
2055 static void validate_endpoint_status_data_stream(void)
2057 struct lttng_ht_iter iter
;
2058 struct lttng_consumer_stream
*stream
;
2060 DBG("Consumer delete flagged data stream");
2063 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2064 /* Validate delete flag of the stream */
2065 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2068 /* Delete it right now */
2069 consumer_del_stream(stream
, data_ht
);
2075 * Delete metadata stream that are flagged for deletion (endpoint_status).
2077 static void validate_endpoint_status_metadata_stream(
2078 struct lttng_poll_event
*pollset
)
2080 struct lttng_ht_iter iter
;
2081 struct lttng_consumer_stream
*stream
;
2083 DBG("Consumer delete flagged metadata stream");
2088 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2089 /* Validate delete flag of the stream */
2090 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2094 * Remove from pollset so the metadata thread can continue without
2095 * blocking on a deleted stream.
2097 lttng_poll_del(pollset
, stream
->wait_fd
);
2099 /* Delete it right now */
2100 consumer_del_metadata_stream(stream
, metadata_ht
);
2106 * Thread polls on metadata file descriptor and write them on disk or on the
2109 void *consumer_thread_metadata_poll(void *data
)
2112 uint32_t revents
, nb_fd
;
2113 struct lttng_consumer_stream
*stream
= NULL
;
2114 struct lttng_ht_iter iter
;
2115 struct lttng_ht_node_u64
*node
;
2116 struct lttng_poll_event events
;
2117 struct lttng_consumer_local_data
*ctx
= data
;
2120 rcu_register_thread();
2122 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2124 /* ENOMEM at this point. Better to bail out. */
2128 DBG("Thread metadata poll started");
2130 /* Size is set to 1 for the consumer_metadata pipe */
2131 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2133 ERR("Poll set creation failed");
2137 ret
= lttng_poll_add(&events
, ctx
->consumer_metadata_pipe
[0], LPOLLIN
);
2143 DBG("Metadata main loop started");
2146 /* Only the metadata pipe is set */
2147 if (LTTNG_POLL_GETNB(&events
) == 0 && consumer_quit
== 1) {
2152 DBG("Metadata poll wait with %d fd(s)", LTTNG_POLL_GETNB(&events
));
2153 ret
= lttng_poll_wait(&events
, -1);
2154 DBG("Metadata event catched in thread");
2156 if (errno
== EINTR
) {
2157 ERR("Poll EINTR catched");
2165 /* From here, the event is a metadata wait fd */
2166 for (i
= 0; i
< nb_fd
; i
++) {
2167 revents
= LTTNG_POLL_GETEV(&events
, i
);
2168 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2170 /* Just don't waste time if no returned events for the fd */
2175 if (pollfd
== ctx
->consumer_metadata_pipe
[0]) {
2176 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2177 DBG("Metadata thread pipe hung up");
2179 * Remove the pipe from the poll set and continue the loop
2180 * since their might be data to consume.
2182 lttng_poll_del(&events
, ctx
->consumer_metadata_pipe
[0]);
2183 ret
= close(ctx
->consumer_metadata_pipe
[0]);
2185 PERROR("close metadata pipe");
2188 } else if (revents
& LPOLLIN
) {
2190 /* Get the stream pointer received */
2191 ret
= read(pollfd
, &stream
, sizeof(stream
));
2192 } while (ret
< 0 && errno
== EINTR
);
2194 ret
< sizeof(struct lttng_consumer_stream
*)) {
2195 PERROR("read metadata stream");
2197 * Let's continue here and hope we can still work
2198 * without stopping the consumer. XXX: Should we?
2203 /* A NULL stream means that the state has changed. */
2204 if (stream
== NULL
) {
2205 /* Check for deleted streams. */
2206 validate_endpoint_status_metadata_stream(&events
);
2210 DBG("Adding metadata stream %d to poll set",
2213 ret
= add_metadata_stream(stream
, metadata_ht
);
2215 ERR("Unable to add metadata stream");
2216 /* Stream was not setup properly. Continuing. */
2217 consumer_del_metadata_stream(stream
, NULL
);
2221 /* Add metadata stream to the global poll events list */
2222 lttng_poll_add(&events
, stream
->wait_fd
,
2223 LPOLLIN
| LPOLLPRI
);
2226 /* Handle other stream */
2232 uint64_t tmp_id
= (uint64_t) pollfd
;
2234 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2236 node
= lttng_ht_iter_get_node_u64(&iter
);
2239 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2242 /* Check for error event */
2243 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2244 DBG("Metadata fd %d is hup|err.", pollfd
);
2245 if (!stream
->hangup_flush_done
2246 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2247 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2248 DBG("Attempting to flush and consume the UST buffers");
2249 lttng_ustconsumer_on_stream_hangup(stream
);
2251 /* We just flushed the stream now read it. */
2253 len
= ctx
->on_buffer_ready(stream
, ctx
);
2255 * We don't check the return value here since if we get
2256 * a negative len, it means an error occured thus we
2257 * simply remove it from the poll set and free the
2263 lttng_poll_del(&events
, stream
->wait_fd
);
2265 * This call update the channel states, closes file descriptors
2266 * and securely free the stream.
2268 consumer_del_metadata_stream(stream
, metadata_ht
);
2269 } else if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2270 /* Get the data out of the metadata file descriptor */
2271 DBG("Metadata available on fd %d", pollfd
);
2272 assert(stream
->wait_fd
== pollfd
);
2274 len
= ctx
->on_buffer_ready(stream
, ctx
);
2275 /* It's ok to have an unavailable sub-buffer */
2276 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2277 /* Clean up stream from consumer and free it. */
2278 lttng_poll_del(&events
, stream
->wait_fd
);
2279 consumer_del_metadata_stream(stream
, metadata_ht
);
2280 } else if (len
> 0) {
2281 stream
->data_read
= 1;
2285 /* Release RCU lock for the stream looked up */
2292 DBG("Metadata poll thread exiting");
2294 lttng_poll_clean(&events
);
2296 destroy_stream_ht(metadata_ht
);
2298 rcu_unregister_thread();
2303 * This thread polls the fds in the set to consume the data and write
2304 * it to tracefile if necessary.
2306 void *consumer_thread_data_poll(void *data
)
2308 int num_rdy
, num_hup
, high_prio
, ret
, i
;
2309 struct pollfd
*pollfd
= NULL
;
2310 /* local view of the streams */
2311 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2312 /* local view of consumer_data.fds_count */
2314 struct lttng_consumer_local_data
*ctx
= data
;
2317 rcu_register_thread();
2319 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2320 if (data_ht
== NULL
) {
2321 /* ENOMEM at this point. Better to bail out. */
2325 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
));
2332 * the fds set has been updated, we need to update our
2333 * local array as well
2335 pthread_mutex_lock(&consumer_data
.lock
);
2336 if (consumer_data
.need_update
) {
2341 local_stream
= NULL
;
2343 /* allocate for all fds + 1 for the consumer_data_pipe */
2344 pollfd
= zmalloc((consumer_data
.stream_count
+ 1) * sizeof(struct pollfd
));
2345 if (pollfd
== NULL
) {
2346 PERROR("pollfd malloc");
2347 pthread_mutex_unlock(&consumer_data
.lock
);
2351 /* allocate for all fds + 1 for the consumer_data_pipe */
2352 local_stream
= zmalloc((consumer_data
.stream_count
+ 1) *
2353 sizeof(struct lttng_consumer_stream
));
2354 if (local_stream
== NULL
) {
2355 PERROR("local_stream malloc");
2356 pthread_mutex_unlock(&consumer_data
.lock
);
2359 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2362 ERR("Error in allocating pollfd or local_outfds");
2363 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2364 pthread_mutex_unlock(&consumer_data
.lock
);
2368 consumer_data
.need_update
= 0;
2370 pthread_mutex_unlock(&consumer_data
.lock
);
2372 /* No FDs and consumer_quit, consumer_cleanup the thread */
2373 if (nb_fd
== 0 && consumer_quit
== 1) {
2376 /* poll on the array of fds */
2378 DBG("polling on %d fd", nb_fd
+ 1);
2379 num_rdy
= poll(pollfd
, nb_fd
+ 1, -1);
2380 DBG("poll num_rdy : %d", num_rdy
);
2381 if (num_rdy
== -1) {
2383 * Restart interrupted system call.
2385 if (errno
== EINTR
) {
2388 PERROR("Poll error");
2389 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2391 } else if (num_rdy
== 0) {
2392 DBG("Polling thread timed out");
2397 * If the consumer_data_pipe triggered poll go directly to the
2398 * beginning of the loop to update the array. We want to prioritize
2399 * array update over low-priority reads.
2401 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2402 ssize_t pipe_readlen
;
2404 DBG("consumer_data_pipe wake up");
2405 /* Consume 1 byte of pipe data */
2407 pipe_readlen
= read(ctx
->consumer_data_pipe
[0], &new_stream
,
2408 sizeof(new_stream
));
2409 } while (pipe_readlen
== -1 && errno
== EINTR
);
2410 if (pipe_readlen
< 0) {
2411 PERROR("read consumer data pipe");
2412 /* Continue so we can at least handle the current stream(s). */
2417 * If the stream is NULL, just ignore it. It's also possible that
2418 * the sessiond poll thread changed the consumer_quit state and is
2419 * waking us up to test it.
2421 if (new_stream
== NULL
) {
2422 validate_endpoint_status_data_stream();
2426 ret
= add_stream(new_stream
, data_ht
);
2428 ERR("Consumer add stream %" PRIu64
" failed. Continuing",
2431 * At this point, if the add_stream fails, it is not in the
2432 * hash table thus passing the NULL value here.
2434 consumer_del_stream(new_stream
, NULL
);
2437 /* Continue to update the local streams and handle prio ones */
2441 /* Take care of high priority channels first. */
2442 for (i
= 0; i
< nb_fd
; i
++) {
2443 if (local_stream
[i
] == NULL
) {
2446 if (pollfd
[i
].revents
& POLLPRI
) {
2447 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2449 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2450 /* it's ok to have an unavailable sub-buffer */
2451 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2452 /* Clean the stream and free it. */
2453 consumer_del_stream(local_stream
[i
], data_ht
);
2454 local_stream
[i
] = NULL
;
2455 } else if (len
> 0) {
2456 local_stream
[i
]->data_read
= 1;
2462 * If we read high prio channel in this loop, try again
2463 * for more high prio data.
2469 /* Take care of low priority channels. */
2470 for (i
= 0; i
< nb_fd
; i
++) {
2471 if (local_stream
[i
] == NULL
) {
2474 if ((pollfd
[i
].revents
& POLLIN
) ||
2475 local_stream
[i
]->hangup_flush_done
) {
2476 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2477 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2478 /* it's ok to have an unavailable sub-buffer */
2479 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2480 /* Clean the stream and free it. */
2481 consumer_del_stream(local_stream
[i
], data_ht
);
2482 local_stream
[i
] = NULL
;
2483 } else if (len
> 0) {
2484 local_stream
[i
]->data_read
= 1;
2489 /* Handle hangup and errors */
2490 for (i
= 0; i
< nb_fd
; i
++) {
2491 if (local_stream
[i
] == NULL
) {
2494 if (!local_stream
[i
]->hangup_flush_done
2495 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2496 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2497 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2498 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2500 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2501 /* Attempt read again, for the data we just flushed. */
2502 local_stream
[i
]->data_read
= 1;
2505 * If the poll flag is HUP/ERR/NVAL and we have
2506 * read no data in this pass, we can remove the
2507 * stream from its hash table.
2509 if ((pollfd
[i
].revents
& POLLHUP
)) {
2510 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2511 if (!local_stream
[i
]->data_read
) {
2512 consumer_del_stream(local_stream
[i
], data_ht
);
2513 local_stream
[i
] = NULL
;
2516 } else if (pollfd
[i
].revents
& POLLERR
) {
2517 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2518 if (!local_stream
[i
]->data_read
) {
2519 consumer_del_stream(local_stream
[i
], data_ht
);
2520 local_stream
[i
] = NULL
;
2523 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2524 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2525 if (!local_stream
[i
]->data_read
) {
2526 consumer_del_stream(local_stream
[i
], data_ht
);
2527 local_stream
[i
] = NULL
;
2531 if (local_stream
[i
] != NULL
) {
2532 local_stream
[i
]->data_read
= 0;
2537 DBG("polling thread exiting");
2542 * Close the write side of the pipe so epoll_wait() in
2543 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2544 * read side of the pipe. If we close them both, epoll_wait strangely does
2545 * not return and could create a endless wait period if the pipe is the
2546 * only tracked fd in the poll set. The thread will take care of closing
2549 ret
= close(ctx
->consumer_metadata_pipe
[1]);
2551 PERROR("close data pipe");
2554 destroy_data_stream_ht(data_ht
);
2556 rcu_unregister_thread();
2561 * Close wake-up end of each stream belonging to the channel. This will
2562 * allow the poll() on the stream read-side to detect when the
2563 * write-side (application) finally closes them.
2566 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2568 struct lttng_ht
*ht
;
2569 struct lttng_consumer_stream
*stream
;
2570 struct lttng_ht_iter iter
;
2572 ht
= consumer_data
.stream_per_chan_id_ht
;
2575 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2576 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2577 ht
->match_fct
, &channel
->key
,
2578 &iter
.iter
, stream
, node_channel_id
.node
) {
2580 * Protect against teardown with mutex.
2582 pthread_mutex_lock(&stream
->lock
);
2583 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2586 switch (consumer_data
.type
) {
2587 case LTTNG_CONSUMER_KERNEL
:
2589 case LTTNG_CONSUMER32_UST
:
2590 case LTTNG_CONSUMER64_UST
:
2592 * Note: a mutex is taken internally within
2593 * liblttng-ust-ctl to protect timer wakeup_fd
2594 * use from concurrent close.
2596 lttng_ustconsumer_close_stream_wakeup(stream
);
2599 ERR("Unknown consumer_data type");
2603 pthread_mutex_unlock(&stream
->lock
);
2608 static void destroy_channel_ht(struct lttng_ht
*ht
)
2610 struct lttng_ht_iter iter
;
2611 struct lttng_consumer_channel
*channel
;
2619 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2620 ret
= lttng_ht_del(ht
, &iter
);
2625 lttng_ht_destroy(ht
);
2629 * This thread polls the channel fds to detect when they are being
2630 * closed. It closes all related streams if the channel is detected as
2631 * closed. It is currently only used as a shim layer for UST because the
2632 * consumerd needs to keep the per-stream wakeup end of pipes open for
2635 void *consumer_thread_channel_poll(void *data
)
2638 uint32_t revents
, nb_fd
;
2639 struct lttng_consumer_channel
*chan
= NULL
;
2640 struct lttng_ht_iter iter
;
2641 struct lttng_ht_node_u64
*node
;
2642 struct lttng_poll_event events
;
2643 struct lttng_consumer_local_data
*ctx
= data
;
2644 struct lttng_ht
*channel_ht
;
2646 rcu_register_thread();
2648 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2650 /* ENOMEM at this point. Better to bail out. */
2654 DBG("Thread channel poll started");
2656 /* Size is set to 1 for the consumer_channel pipe */
2657 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2659 ERR("Poll set creation failed");
2663 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2669 DBG("Channel main loop started");
2672 /* Only the channel pipe is set */
2673 if (LTTNG_POLL_GETNB(&events
) == 0 && consumer_quit
== 1) {
2678 DBG("Channel poll wait with %d fd(s)", LTTNG_POLL_GETNB(&events
));
2679 ret
= lttng_poll_wait(&events
, -1);
2680 DBG("Channel event catched in thread");
2682 if (errno
== EINTR
) {
2683 ERR("Poll EINTR catched");
2691 /* From here, the event is a channel wait fd */
2692 for (i
= 0; i
< nb_fd
; i
++) {
2693 revents
= LTTNG_POLL_GETEV(&events
, i
);
2694 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2696 /* Just don't waste time if no returned events for the fd */
2700 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2701 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2702 DBG("Channel thread pipe hung up");
2704 * Remove the pipe from the poll set and continue the loop
2705 * since their might be data to consume.
2707 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2709 } else if (revents
& LPOLLIN
) {
2710 enum consumer_channel_action action
;
2713 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
2715 ERR("Error reading channel pipe");
2720 case CONSUMER_CHANNEL_ADD
:
2721 DBG("Adding channel %d to poll set",
2724 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
2726 lttng_ht_add_unique_u64(channel_ht
,
2727 &chan
->wait_fd_node
);
2728 /* Add channel to the global poll events list */
2729 lttng_poll_add(&events
, chan
->wait_fd
,
2730 LPOLLIN
| LPOLLPRI
);
2732 case CONSUMER_CHANNEL_DEL
:
2734 chan
= consumer_find_channel(key
);
2736 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
2739 lttng_poll_del(&events
, chan
->wait_fd
);
2740 ret
= lttng_ht_del(channel_ht
, &iter
);
2742 consumer_close_channel_streams(chan
);
2745 * Release our own refcount. Force channel deletion even if
2746 * streams were not initialized.
2748 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
2749 consumer_del_channel(chan
);
2753 case CONSUMER_CHANNEL_QUIT
:
2755 * Remove the pipe from the poll set and continue the loop
2756 * since their might be data to consume.
2758 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2761 ERR("Unknown action");
2766 /* Handle other stream */
2772 uint64_t tmp_id
= (uint64_t) pollfd
;
2774 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
2776 node
= lttng_ht_iter_get_node_u64(&iter
);
2779 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
2782 /* Check for error event */
2783 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2784 DBG("Channel fd %d is hup|err.", pollfd
);
2786 lttng_poll_del(&events
, chan
->wait_fd
);
2787 ret
= lttng_ht_del(channel_ht
, &iter
);
2789 consumer_close_channel_streams(chan
);
2791 /* Release our own refcount */
2792 if (!uatomic_sub_return(&chan
->refcount
, 1)
2793 && !uatomic_read(&chan
->nb_init_stream_left
)) {
2794 consumer_del_channel(chan
);
2798 /* Release RCU lock for the channel looked up */
2804 lttng_poll_clean(&events
);
2806 destroy_channel_ht(channel_ht
);
2808 DBG("Channel poll thread exiting");
2809 rcu_unregister_thread();
2813 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
2814 struct pollfd
*sockpoll
, int client_socket
)
2821 if (lttng_consumer_poll_socket(sockpoll
) < 0) {
2825 DBG("Metadata connection on client_socket");
2827 /* Blocking call, waiting for transmission */
2828 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
2829 if (ctx
->consumer_metadata_socket
< 0) {
2830 WARN("On accept metadata");
2841 * This thread listens on the consumerd socket and receives the file
2842 * descriptors from the session daemon.
2844 void *consumer_thread_sessiond_poll(void *data
)
2846 int sock
= -1, client_socket
, ret
;
2848 * structure to poll for incoming data on communication socket avoids
2849 * making blocking sockets.
2851 struct pollfd consumer_sockpoll
[2];
2852 struct lttng_consumer_local_data
*ctx
= data
;
2854 rcu_register_thread();
2856 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
2857 unlink(ctx
->consumer_command_sock_path
);
2858 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
2859 if (client_socket
< 0) {
2860 ERR("Cannot create command socket");
2864 ret
= lttcomm_listen_unix_sock(client_socket
);
2869 DBG("Sending ready command to lttng-sessiond");
2870 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
2871 /* return < 0 on error, but == 0 is not fatal */
2873 ERR("Error sending ready command to lttng-sessiond");
2877 ret
= fcntl(client_socket
, F_SETFL
, O_NONBLOCK
);
2879 PERROR("fcntl O_NONBLOCK");
2883 /* prepare the FDs to poll : to client socket and the should_quit pipe */
2884 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
2885 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
2886 consumer_sockpoll
[1].fd
= client_socket
;
2887 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
2889 if (lttng_consumer_poll_socket(consumer_sockpoll
) < 0) {
2892 DBG("Connection on client_socket");
2894 /* Blocking call, waiting for transmission */
2895 sock
= lttcomm_accept_unix_sock(client_socket
);
2900 ret
= fcntl(sock
, F_SETFL
, O_NONBLOCK
);
2902 PERROR("fcntl O_NONBLOCK");
2907 * Setup metadata socket which is the second socket connection on the
2908 * command unix socket.
2910 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
2915 /* This socket is not useful anymore. */
2916 ret
= close(client_socket
);
2918 PERROR("close client_socket");
2922 /* update the polling structure to poll on the established socket */
2923 consumer_sockpoll
[1].fd
= sock
;
2924 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
2927 if (lttng_consumer_poll_socket(consumer_sockpoll
) < 0) {
2930 DBG("Incoming command on sock");
2931 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2932 if (ret
== -ENOENT
) {
2933 DBG("Received STOP command");
2938 * This could simply be a session daemon quitting. Don't output
2941 DBG("Communication interrupted on command socket");
2944 if (consumer_quit
) {
2945 DBG("consumer_thread_receive_fds received quit from signal");
2948 DBG("received command on sock");
2951 DBG("Consumer thread sessiond poll exiting");
2954 * Close metadata streams since the producer is the session daemon which
2957 * NOTE: for now, this only applies to the UST tracer.
2959 lttng_consumer_close_metadata();
2962 * when all fds have hung up, the polling thread
2968 * Notify the data poll thread to poll back again and test the
2969 * consumer_quit state that we just set so to quit gracefully.
2971 notify_thread_pipe(ctx
->consumer_data_pipe
[1]);
2973 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
2975 /* Cleaning up possibly open sockets. */
2979 PERROR("close sock sessiond poll");
2982 if (client_socket
>= 0) {
2985 PERROR("close client_socket sessiond poll");
2989 rcu_unregister_thread();
2993 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
2994 struct lttng_consumer_local_data
*ctx
)
2998 pthread_mutex_lock(&stream
->lock
);
3000 switch (consumer_data
.type
) {
3001 case LTTNG_CONSUMER_KERNEL
:
3002 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
3004 case LTTNG_CONSUMER32_UST
:
3005 case LTTNG_CONSUMER64_UST
:
3006 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
3009 ERR("Unknown consumer_data type");
3015 pthread_mutex_unlock(&stream
->lock
);
3019 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3021 switch (consumer_data
.type
) {
3022 case LTTNG_CONSUMER_KERNEL
:
3023 return lttng_kconsumer_on_recv_stream(stream
);
3024 case LTTNG_CONSUMER32_UST
:
3025 case LTTNG_CONSUMER64_UST
:
3026 return lttng_ustconsumer_on_recv_stream(stream
);
3028 ERR("Unknown consumer_data type");
3035 * Allocate and set consumer data hash tables.
3037 void lttng_consumer_init(void)
3039 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3040 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3041 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3042 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3046 * Process the ADD_RELAYD command receive by a consumer.
3048 * This will create a relayd socket pair and add it to the relayd hash table.
3049 * The caller MUST acquire a RCU read side lock before calling it.
3051 int consumer_add_relayd_socket(int net_seq_idx
, int sock_type
,
3052 struct lttng_consumer_local_data
*ctx
, int sock
,
3053 struct pollfd
*consumer_sockpoll
,
3054 struct lttcomm_relayd_sock
*relayd_sock
, unsigned int sessiond_id
)
3056 int fd
= -1, ret
= -1, relayd_created
= 0;
3057 enum lttng_error_code ret_code
= LTTNG_OK
;
3058 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3061 assert(relayd_sock
);
3063 DBG("Consumer adding relayd socket (idx: %d)", net_seq_idx
);
3065 /* First send a status message before receiving the fds. */
3066 ret
= consumer_send_status_msg(sock
, ret_code
);
3068 /* Somehow, the session daemon is not responding anymore. */
3072 /* Get relayd reference if exists. */
3073 relayd
= consumer_find_relayd(net_seq_idx
);
3074 if (relayd
== NULL
) {
3075 /* Not found. Allocate one. */
3076 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3077 if (relayd
== NULL
) {
3078 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_OUTFD_ERROR
);
3082 relayd
->sessiond_session_id
= (uint64_t) sessiond_id
;
3086 /* Poll on consumer socket. */
3087 if (lttng_consumer_poll_socket(consumer_sockpoll
) < 0) {
3092 /* Get relayd socket from session daemon */
3093 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3094 if (ret
!= sizeof(fd
)) {
3095 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3097 fd
= -1; /* Just in case it gets set with an invalid value. */
3101 /* We have the fds without error. Send status back. */
3102 ret
= consumer_send_status_msg(sock
, ret_code
);
3104 /* Somehow, the session daemon is not responding anymore. */
3108 /* Copy socket information and received FD */
3109 switch (sock_type
) {
3110 case LTTNG_STREAM_CONTROL
:
3111 /* Copy received lttcomm socket */
3112 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3113 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3114 /* Immediately try to close the created socket if valid. */
3115 if (relayd
->control_sock
.sock
.fd
>= 0) {
3116 if (close(relayd
->control_sock
.sock
.fd
)) {
3117 PERROR("close relayd control socket");
3120 /* Handle create_sock error. */
3125 /* Assign new file descriptor */
3126 relayd
->control_sock
.sock
.fd
= fd
;
3127 /* Assign version values. */
3128 relayd
->control_sock
.major
= relayd_sock
->major
;
3129 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3132 * Create a session on the relayd and store the returned id. Lock the
3133 * control socket mutex if the relayd was NOT created before.
3135 if (!relayd_created
) {
3136 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3138 ret
= relayd_create_session(&relayd
->control_sock
,
3139 &relayd
->relayd_session_id
);
3140 if (!relayd_created
) {
3141 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3145 * Close all sockets of a relayd object. It will be freed if it was
3146 * created at the error code path or else it will be garbage
3149 (void) relayd_close(&relayd
->control_sock
);
3150 (void) relayd_close(&relayd
->data_sock
);
3155 case LTTNG_STREAM_DATA
:
3156 /* Copy received lttcomm socket */
3157 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3158 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3159 /* Immediately try to close the created socket if valid. */
3160 if (relayd
->data_sock
.sock
.fd
>= 0) {
3161 if (close(relayd
->data_sock
.sock
.fd
)) {
3162 PERROR("close relayd data socket");
3165 /* Handle create_sock error. */
3170 /* Assign new file descriptor */
3171 relayd
->data_sock
.sock
.fd
= fd
;
3172 /* Assign version values. */
3173 relayd
->data_sock
.major
= relayd_sock
->major
;
3174 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3177 ERR("Unknown relayd socket type (%d)", sock_type
);
3182 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3183 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3184 relayd
->net_seq_idx
, fd
);
3187 * Add relayd socket pair to consumer data hashtable. If object already
3188 * exists or on error, the function gracefully returns.
3196 /* Close received socket if valid. */
3199 PERROR("close received socket");
3204 if (relayd_created
) {
3212 * Try to lock the stream mutex.
3214 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
3216 static int stream_try_lock(struct lttng_consumer_stream
*stream
)
3223 * Try to lock the stream mutex. On failure, we know that the stream is
3224 * being used else where hence there is data still being extracted.
3226 ret
= pthread_mutex_trylock(&stream
->lock
);
3228 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
3240 * Search for a relayd associated to the session id and return the reference.
3242 * A rcu read side lock MUST be acquire before calling this function and locked
3243 * until the relayd object is no longer necessary.
3245 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3247 struct lttng_ht_iter iter
;
3248 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3250 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3251 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3254 * Check by sessiond id which is unique here where the relayd session
3255 * id might not be when having multiple relayd.
3257 if (relayd
->sessiond_session_id
== id
) {
3258 /* Found the relayd. There can be only one per id. */
3270 * Check if for a given session id there is still data needed to be extract
3273 * Return 1 if data is pending or else 0 meaning ready to be read.
3275 int consumer_data_pending(uint64_t id
)
3278 struct lttng_ht_iter iter
;
3279 struct lttng_ht
*ht
;
3280 struct lttng_consumer_stream
*stream
;
3281 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3282 int (*data_pending
)(struct lttng_consumer_stream
*);
3284 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3287 pthread_mutex_lock(&consumer_data
.lock
);
3289 switch (consumer_data
.type
) {
3290 case LTTNG_CONSUMER_KERNEL
:
3291 data_pending
= lttng_kconsumer_data_pending
;
3293 case LTTNG_CONSUMER32_UST
:
3294 case LTTNG_CONSUMER64_UST
:
3295 data_pending
= lttng_ustconsumer_data_pending
;
3298 ERR("Unknown consumer data type");
3302 /* Ease our life a bit */
3303 ht
= consumer_data
.stream_list_ht
;
3305 relayd
= find_relayd_by_session_id(id
);
3307 /* Send init command for data pending. */
3308 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3309 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3310 relayd
->relayd_session_id
);
3311 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3313 /* Communication error thus the relayd so no data pending. */
3314 goto data_not_pending
;
3318 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3319 ht
->hash_fct(&id
, lttng_ht_seed
),
3321 &iter
.iter
, stream
, node_session_id
.node
) {
3322 /* If this call fails, the stream is being used hence data pending. */
3323 ret
= stream_try_lock(stream
);
3329 * A removed node from the hash table indicates that the stream has
3330 * been deleted thus having a guarantee that the buffers are closed
3331 * on the consumer side. However, data can still be transmitted
3332 * over the network so don't skip the relayd check.
3334 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3336 /* Check the stream if there is data in the buffers. */
3337 ret
= data_pending(stream
);
3339 pthread_mutex_unlock(&stream
->lock
);
3346 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3347 if (stream
->metadata_flag
) {
3348 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3349 stream
->relayd_stream_id
);
3351 ret
= relayd_data_pending(&relayd
->control_sock
,
3352 stream
->relayd_stream_id
,
3353 stream
->next_net_seq_num
- 1);
3355 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3357 pthread_mutex_unlock(&stream
->lock
);
3361 pthread_mutex_unlock(&stream
->lock
);
3365 unsigned int is_data_inflight
= 0;
3367 /* Send init command for data pending. */
3368 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3369 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3370 relayd
->relayd_session_id
, &is_data_inflight
);
3371 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3373 goto data_not_pending
;
3375 if (is_data_inflight
) {
3381 * Finding _no_ node in the hash table and no inflight data means that the
3382 * stream(s) have been removed thus data is guaranteed to be available for
3383 * analysis from the trace files.
3387 /* Data is available to be read by a viewer. */
3388 pthread_mutex_unlock(&consumer_data
.lock
);
3393 /* Data is still being extracted from buffers. */
3394 pthread_mutex_unlock(&consumer_data
.lock
);
3400 * Send a ret code status message to the sessiond daemon.
3402 * Return the sendmsg() return value.
3404 int consumer_send_status_msg(int sock
, int ret_code
)
3406 struct lttcomm_consumer_status_msg msg
;
3408 msg
.ret_code
= ret_code
;
3410 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3414 * Send a channel status message to the sessiond daemon.
3416 * Return the sendmsg() return value.
3418 int consumer_send_status_channel(int sock
,
3419 struct lttng_consumer_channel
*channel
)
3421 struct lttcomm_consumer_status_channel msg
;
3426 msg
.ret_code
= -LTTNG_ERR_UST_CHAN_FAIL
;
3428 msg
.ret_code
= LTTNG_OK
;
3429 msg
.key
= channel
->key
;
3430 msg
.stream_count
= channel
->streams
.count
;
3433 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));