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 lttng pipe to poll back again. This usually means that some
81 * global state has changed so we just send back the thread in a poll wait
84 static void notify_thread_lttng_pipe(struct lttng_pipe
*pipe
)
86 struct lttng_consumer_stream
*null_stream
= NULL
;
90 (void) lttng_pipe_write(pipe
, &null_stream
, sizeof(null_stream
));
93 static void notify_channel_pipe(struct lttng_consumer_local_data
*ctx
,
94 struct lttng_consumer_channel
*chan
,
96 enum consumer_channel_action action
)
98 struct consumer_channel_msg msg
;
101 memset(&msg
, 0, sizeof(msg
));
106 ret
= write(ctx
->consumer_channel_pipe
[1], &msg
, sizeof(msg
));
107 } while (ret
< 0 && errno
== EINTR
);
110 void notify_thread_del_channel(struct lttng_consumer_local_data
*ctx
,
113 notify_channel_pipe(ctx
, NULL
, key
, CONSUMER_CHANNEL_DEL
);
116 static int read_channel_pipe(struct lttng_consumer_local_data
*ctx
,
117 struct lttng_consumer_channel
**chan
,
119 enum consumer_channel_action
*action
)
121 struct consumer_channel_msg msg
;
125 ret
= read(ctx
->consumer_channel_pipe
[0], &msg
, sizeof(msg
));
126 } while (ret
< 0 && errno
== EINTR
);
128 *action
= msg
.action
;
136 * Find a stream. The consumer_data.lock must be locked during this
139 static struct lttng_consumer_stream
*find_stream(uint64_t key
,
142 struct lttng_ht_iter iter
;
143 struct lttng_ht_node_u64
*node
;
144 struct lttng_consumer_stream
*stream
= NULL
;
148 /* -1ULL keys are lookup failures */
149 if (key
== (uint64_t) -1ULL) {
155 lttng_ht_lookup(ht
, &key
, &iter
);
156 node
= lttng_ht_iter_get_node_u64(&iter
);
158 stream
= caa_container_of(node
, struct lttng_consumer_stream
, node
);
166 static void steal_stream_key(int key
, struct lttng_ht
*ht
)
168 struct lttng_consumer_stream
*stream
;
171 stream
= find_stream(key
, ht
);
175 * We don't want the lookup to match, but we still need
176 * to iterate on this stream when iterating over the hash table. Just
177 * change the node key.
179 stream
->node
.key
= -1ULL;
185 * Return a channel object for the given key.
187 * RCU read side lock MUST be acquired before calling this function and
188 * protects the channel ptr.
190 struct lttng_consumer_channel
*consumer_find_channel(uint64_t key
)
192 struct lttng_ht_iter iter
;
193 struct lttng_ht_node_u64
*node
;
194 struct lttng_consumer_channel
*channel
= NULL
;
196 /* -1ULL keys are lookup failures */
197 if (key
== (uint64_t) -1ULL) {
201 lttng_ht_lookup(consumer_data
.channel_ht
, &key
, &iter
);
202 node
= lttng_ht_iter_get_node_u64(&iter
);
204 channel
= caa_container_of(node
, struct lttng_consumer_channel
, node
);
210 static void free_stream_rcu(struct rcu_head
*head
)
212 struct lttng_ht_node_u64
*node
=
213 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
214 struct lttng_consumer_stream
*stream
=
215 caa_container_of(node
, struct lttng_consumer_stream
, node
);
220 static void free_channel_rcu(struct rcu_head
*head
)
222 struct lttng_ht_node_u64
*node
=
223 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
224 struct lttng_consumer_channel
*channel
=
225 caa_container_of(node
, struct lttng_consumer_channel
, node
);
231 * RCU protected relayd socket pair free.
233 static void free_relayd_rcu(struct rcu_head
*head
)
235 struct lttng_ht_node_u64
*node
=
236 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
237 struct consumer_relayd_sock_pair
*relayd
=
238 caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
241 * Close all sockets. This is done in the call RCU since we don't want the
242 * socket fds to be reassigned thus potentially creating bad state of the
245 * We do not have to lock the control socket mutex here since at this stage
246 * there is no one referencing to this relayd object.
248 (void) relayd_close(&relayd
->control_sock
);
249 (void) relayd_close(&relayd
->data_sock
);
255 * Destroy and free relayd socket pair object.
257 * This function MUST be called with the consumer_data lock acquired.
259 static void destroy_relayd(struct consumer_relayd_sock_pair
*relayd
)
262 struct lttng_ht_iter iter
;
264 if (relayd
== NULL
) {
268 DBG("Consumer destroy and close relayd socket pair");
270 iter
.iter
.node
= &relayd
->node
.node
;
271 ret
= lttng_ht_del(consumer_data
.relayd_ht
, &iter
);
273 /* We assume the relayd is being or is destroyed */
277 /* RCU free() call */
278 call_rcu(&relayd
->node
.head
, free_relayd_rcu
);
282 * Remove a channel from the global list protected by a mutex. This function is
283 * also responsible for freeing its data structures.
285 void consumer_del_channel(struct lttng_consumer_channel
*channel
)
288 struct lttng_ht_iter iter
;
290 DBG("Consumer delete channel key %" PRIu64
, channel
->key
);
292 pthread_mutex_lock(&consumer_data
.lock
);
294 switch (consumer_data
.type
) {
295 case LTTNG_CONSUMER_KERNEL
:
297 case LTTNG_CONSUMER32_UST
:
298 case LTTNG_CONSUMER64_UST
:
299 lttng_ustconsumer_del_channel(channel
);
302 ERR("Unknown consumer_data type");
308 iter
.iter
.node
= &channel
->node
.node
;
309 ret
= lttng_ht_del(consumer_data
.channel_ht
, &iter
);
313 call_rcu(&channel
->node
.head
, free_channel_rcu
);
315 pthread_mutex_unlock(&consumer_data
.lock
);
319 * Iterate over the relayd hash table and destroy each element. Finally,
320 * destroy the whole hash table.
322 static void cleanup_relayd_ht(void)
324 struct lttng_ht_iter iter
;
325 struct consumer_relayd_sock_pair
*relayd
;
329 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
331 destroy_relayd(relayd
);
336 lttng_ht_destroy(consumer_data
.relayd_ht
);
340 * Update the end point status of all streams having the given network sequence
341 * index (relayd index).
343 * It's atomically set without having the stream mutex locked which is fine
344 * because we handle the write/read race with a pipe wakeup for each thread.
346 static void update_endpoint_status_by_netidx(int net_seq_idx
,
347 enum consumer_endpoint_status status
)
349 struct lttng_ht_iter iter
;
350 struct lttng_consumer_stream
*stream
;
352 DBG("Consumer set delete flag on stream by idx %d", net_seq_idx
);
356 /* Let's begin with metadata */
357 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
358 if (stream
->net_seq_idx
== net_seq_idx
) {
359 uatomic_set(&stream
->endpoint_status
, status
);
360 DBG("Delete flag set to metadata stream %d", stream
->wait_fd
);
364 /* Follow up by the data streams */
365 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
366 if (stream
->net_seq_idx
== net_seq_idx
) {
367 uatomic_set(&stream
->endpoint_status
, status
);
368 DBG("Delete flag set to data stream %d", stream
->wait_fd
);
375 * Cleanup a relayd object by flagging every associated streams for deletion,
376 * destroying the object meaning removing it from the relayd hash table,
377 * closing the sockets and freeing the memory in a RCU call.
379 * If a local data context is available, notify the threads that the streams'
380 * state have changed.
382 static void cleanup_relayd(struct consumer_relayd_sock_pair
*relayd
,
383 struct lttng_consumer_local_data
*ctx
)
389 DBG("Cleaning up relayd sockets");
391 /* Save the net sequence index before destroying the object */
392 netidx
= relayd
->net_seq_idx
;
395 * Delete the relayd from the relayd hash table, close the sockets and free
396 * the object in a RCU call.
398 destroy_relayd(relayd
);
400 /* Set inactive endpoint to all streams */
401 update_endpoint_status_by_netidx(netidx
, CONSUMER_ENDPOINT_INACTIVE
);
404 * With a local data context, notify the threads that the streams' state
405 * have changed. The write() action on the pipe acts as an "implicit"
406 * memory barrier ordering the updates of the end point status from the
407 * read of this status which happens AFTER receiving this notify.
410 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
411 notify_thread_lttng_pipe(ctx
->consumer_metadata_pipe
);
416 * Flag a relayd socket pair for destruction. Destroy it if the refcount
419 * RCU read side lock MUST be aquired before calling this function.
421 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair
*relayd
)
425 /* Set destroy flag for this object */
426 uatomic_set(&relayd
->destroy_flag
, 1);
428 /* Destroy the relayd if refcount is 0 */
429 if (uatomic_read(&relayd
->refcount
) == 0) {
430 destroy_relayd(relayd
);
435 * Remove a stream from the global list protected by a mutex. This
436 * function is also responsible for freeing its data structures.
438 void consumer_del_stream(struct lttng_consumer_stream
*stream
,
442 struct lttng_ht_iter iter
;
443 struct lttng_consumer_channel
*free_chan
= NULL
;
444 struct consumer_relayd_sock_pair
*relayd
;
448 DBG("Consumer del stream %d", stream
->wait_fd
);
451 /* Means the stream was allocated but not successfully added */
452 goto free_stream_rcu
;
455 pthread_mutex_lock(&consumer_data
.lock
);
456 pthread_mutex_lock(&stream
->lock
);
458 switch (consumer_data
.type
) {
459 case LTTNG_CONSUMER_KERNEL
:
460 if (stream
->mmap_base
!= NULL
) {
461 ret
= munmap(stream
->mmap_base
, stream
->mmap_len
);
467 case LTTNG_CONSUMER32_UST
:
468 case LTTNG_CONSUMER64_UST
:
469 lttng_ustconsumer_del_stream(stream
);
472 ERR("Unknown consumer_data type");
478 iter
.iter
.node
= &stream
->node
.node
;
479 ret
= lttng_ht_del(ht
, &iter
);
482 iter
.iter
.node
= &stream
->node_channel_id
.node
;
483 ret
= lttng_ht_del(consumer_data
.stream_per_chan_id_ht
, &iter
);
486 iter
.iter
.node
= &stream
->node_session_id
.node
;
487 ret
= lttng_ht_del(consumer_data
.stream_list_ht
, &iter
);
491 assert(consumer_data
.stream_count
> 0);
492 consumer_data
.stream_count
--;
494 if (stream
->out_fd
>= 0) {
495 ret
= close(stream
->out_fd
);
501 /* Check and cleanup relayd */
503 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
504 if (relayd
!= NULL
) {
505 uatomic_dec(&relayd
->refcount
);
506 assert(uatomic_read(&relayd
->refcount
) >= 0);
508 /* Closing streams requires to lock the control socket. */
509 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
510 ret
= relayd_send_close_stream(&relayd
->control_sock
,
511 stream
->relayd_stream_id
,
512 stream
->next_net_seq_num
- 1);
513 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
515 DBG("Unable to close stream on the relayd. Continuing");
517 * Continue here. There is nothing we can do for the relayd.
518 * Chances are that the relayd has closed the socket so we just
519 * continue cleaning up.
523 /* Both conditions are met, we destroy the relayd. */
524 if (uatomic_read(&relayd
->refcount
) == 0 &&
525 uatomic_read(&relayd
->destroy_flag
)) {
526 destroy_relayd(relayd
);
531 if (!uatomic_sub_return(&stream
->chan
->refcount
, 1)
532 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
533 free_chan
= stream
->chan
;
537 consumer_data
.need_update
= 1;
538 pthread_mutex_unlock(&stream
->lock
);
539 pthread_mutex_unlock(&consumer_data
.lock
);
542 consumer_del_channel(free_chan
);
546 call_rcu(&stream
->node
.head
, free_stream_rcu
);
549 struct lttng_consumer_stream
*consumer_allocate_stream(uint64_t channel_key
,
551 enum lttng_consumer_stream_state state
,
552 const char *channel_name
,
559 enum consumer_channel_type type
)
562 struct lttng_consumer_stream
*stream
;
564 stream
= zmalloc(sizeof(*stream
));
565 if (stream
== NULL
) {
566 PERROR("malloc struct lttng_consumer_stream");
573 stream
->key
= stream_key
;
575 stream
->out_fd_offset
= 0;
576 stream
->state
= state
;
579 stream
->net_seq_idx
= relayd_id
;
580 stream
->session_id
= session_id
;
581 pthread_mutex_init(&stream
->lock
, NULL
);
583 /* If channel is the metadata, flag this stream as metadata. */
584 if (type
== CONSUMER_CHANNEL_TYPE_METADATA
) {
585 stream
->metadata_flag
= 1;
586 /* Metadata is flat out. */
587 strncpy(stream
->name
, DEFAULT_METADATA_NAME
, sizeof(stream
->name
));
589 /* Format stream name to <channel_name>_<cpu_number> */
590 ret
= snprintf(stream
->name
, sizeof(stream
->name
), "%s_%d",
593 PERROR("snprintf stream name");
598 /* Key is always the wait_fd for streams. */
599 lttng_ht_node_init_u64(&stream
->node
, stream
->key
);
601 /* Init node per channel id key */
602 lttng_ht_node_init_u64(&stream
->node_channel_id
, channel_key
);
604 /* Init session id node with the stream session id */
605 lttng_ht_node_init_u64(&stream
->node_session_id
, stream
->session_id
);
607 DBG3("Allocated stream %s (key %" PRIu64
", chan_key %" PRIu64
" relayd_id %" PRIu64
", session_id %" PRIu64
,
608 stream
->name
, stream
->key
, channel_key
, stream
->net_seq_idx
, stream
->session_id
);
624 * Add a stream to the global list protected by a mutex.
626 static int add_stream(struct lttng_consumer_stream
*stream
,
630 struct consumer_relayd_sock_pair
*relayd
;
635 DBG3("Adding consumer stream %" PRIu64
, stream
->key
);
637 pthread_mutex_lock(&consumer_data
.lock
);
638 pthread_mutex_lock(&stream
->lock
);
641 /* Steal stream identifier to avoid having streams with the same key */
642 steal_stream_key(stream
->key
, ht
);
644 lttng_ht_add_unique_u64(ht
, &stream
->node
);
646 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
647 &stream
->node_channel_id
);
650 * Add stream to the stream_list_ht of the consumer data. No need to steal
651 * the key since the HT does not use it and we allow to add redundant keys
654 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
656 /* Check and cleanup relayd */
657 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
658 if (relayd
!= NULL
) {
659 uatomic_inc(&relayd
->refcount
);
663 * When nb_init_stream_left reaches 0, we don't need to trigger any action
664 * in terms of destroying the associated channel, because the action that
665 * causes the count to become 0 also causes a stream to be added. The
666 * channel deletion will thus be triggered by the following removal of this
669 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
670 /* Increment refcount before decrementing nb_init_stream_left */
672 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
675 /* Update consumer data once the node is inserted. */
676 consumer_data
.stream_count
++;
677 consumer_data
.need_update
= 1;
680 pthread_mutex_unlock(&stream
->lock
);
681 pthread_mutex_unlock(&consumer_data
.lock
);
687 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
688 * be acquired before calling this.
690 static int add_relayd(struct consumer_relayd_sock_pair
*relayd
)
693 struct lttng_ht_node_u64
*node
;
694 struct lttng_ht_iter iter
;
698 lttng_ht_lookup(consumer_data
.relayd_ht
,
699 &relayd
->net_seq_idx
, &iter
);
700 node
= lttng_ht_iter_get_node_u64(&iter
);
704 lttng_ht_add_unique_u64(consumer_data
.relayd_ht
, &relayd
->node
);
711 * Allocate and return a consumer relayd socket.
713 struct consumer_relayd_sock_pair
*consumer_allocate_relayd_sock_pair(
716 struct consumer_relayd_sock_pair
*obj
= NULL
;
718 /* Negative net sequence index is a failure */
719 if (net_seq_idx
< 0) {
723 obj
= zmalloc(sizeof(struct consumer_relayd_sock_pair
));
725 PERROR("zmalloc relayd sock");
729 obj
->net_seq_idx
= net_seq_idx
;
731 obj
->destroy_flag
= 0;
732 obj
->control_sock
.sock
.fd
= -1;
733 obj
->data_sock
.sock
.fd
= -1;
734 lttng_ht_node_init_u64(&obj
->node
, obj
->net_seq_idx
);
735 pthread_mutex_init(&obj
->ctrl_sock_mutex
, NULL
);
742 * Find a relayd socket pair in the global consumer data.
744 * Return the object if found else NULL.
745 * RCU read-side lock must be held across this call and while using the
748 struct consumer_relayd_sock_pair
*consumer_find_relayd(uint64_t key
)
750 struct lttng_ht_iter iter
;
751 struct lttng_ht_node_u64
*node
;
752 struct consumer_relayd_sock_pair
*relayd
= NULL
;
754 /* Negative keys are lookup failures */
755 if (key
== (uint64_t) -1ULL) {
759 lttng_ht_lookup(consumer_data
.relayd_ht
, &key
,
761 node
= lttng_ht_iter_get_node_u64(&iter
);
763 relayd
= caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
771 * Handle stream for relayd transmission if the stream applies for network
772 * streaming where the net sequence index is set.
774 * Return destination file descriptor or negative value on error.
776 static int write_relayd_stream_header(struct lttng_consumer_stream
*stream
,
777 size_t data_size
, unsigned long padding
,
778 struct consumer_relayd_sock_pair
*relayd
)
781 struct lttcomm_relayd_data_hdr data_hdr
;
787 /* Reset data header */
788 memset(&data_hdr
, 0, sizeof(data_hdr
));
790 if (stream
->metadata_flag
) {
791 /* Caller MUST acquire the relayd control socket lock */
792 ret
= relayd_send_metadata(&relayd
->control_sock
, data_size
);
797 /* Metadata are always sent on the control socket. */
798 outfd
= relayd
->control_sock
.sock
.fd
;
800 /* Set header with stream information */
801 data_hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
802 data_hdr
.data_size
= htobe32(data_size
);
803 data_hdr
.padding_size
= htobe32(padding
);
805 * Note that net_seq_num below is assigned with the *current* value of
806 * next_net_seq_num and only after that the next_net_seq_num will be
807 * increment. This is why when issuing a command on the relayd using
808 * this next value, 1 should always be substracted in order to compare
809 * the last seen sequence number on the relayd side to the last sent.
811 data_hdr
.net_seq_num
= htobe64(stream
->next_net_seq_num
);
812 /* Other fields are zeroed previously */
814 ret
= relayd_send_data_hdr(&relayd
->data_sock
, &data_hdr
,
820 ++stream
->next_net_seq_num
;
822 /* Set to go on data socket */
823 outfd
= relayd
->data_sock
.sock
.fd
;
831 * Allocate and return a new lttng_consumer_channel object using the given key
832 * to initialize the hash table node.
834 * On error, return NULL.
836 struct lttng_consumer_channel
*consumer_allocate_channel(uint64_t key
,
838 const char *pathname
,
843 enum lttng_event_output output
,
844 uint64_t tracefile_size
,
845 uint64_t tracefile_count
)
847 struct lttng_consumer_channel
*channel
;
849 channel
= zmalloc(sizeof(*channel
));
850 if (channel
== NULL
) {
851 PERROR("malloc struct lttng_consumer_channel");
856 channel
->refcount
= 0;
857 channel
->session_id
= session_id
;
860 channel
->relayd_id
= relayd_id
;
861 channel
->output
= output
;
862 channel
->tracefile_size
= tracefile_size
;
863 channel
->tracefile_count
= tracefile_count
;
865 strncpy(channel
->pathname
, pathname
, sizeof(channel
->pathname
));
866 channel
->pathname
[sizeof(channel
->pathname
) - 1] = '\0';
868 strncpy(channel
->name
, name
, sizeof(channel
->name
));
869 channel
->name
[sizeof(channel
->name
) - 1] = '\0';
871 lttng_ht_node_init_u64(&channel
->node
, channel
->key
);
873 channel
->wait_fd
= -1;
875 CDS_INIT_LIST_HEAD(&channel
->streams
.head
);
877 DBG("Allocated channel (key %" PRIu64
")", channel
->key
)
884 * Add a channel to the global list protected by a mutex.
886 int consumer_add_channel(struct lttng_consumer_channel
*channel
,
887 struct lttng_consumer_local_data
*ctx
)
890 struct lttng_ht_node_u64
*node
;
891 struct lttng_ht_iter iter
;
893 pthread_mutex_lock(&consumer_data
.lock
);
896 lttng_ht_lookup(consumer_data
.channel_ht
, &channel
->key
, &iter
);
897 node
= lttng_ht_iter_get_node_u64(&iter
);
899 /* Channel already exist. Ignore the insertion */
900 ERR("Consumer add channel key %" PRIu64
" already exists!",
906 lttng_ht_add_unique_u64(consumer_data
.channel_ht
, &channel
->node
);
910 pthread_mutex_unlock(&consumer_data
.lock
);
912 if (!ret
&& channel
->wait_fd
!= -1 &&
913 channel
->metadata_stream
== NULL
) {
914 notify_channel_pipe(ctx
, channel
, -1, CONSUMER_CHANNEL_ADD
);
920 * Allocate the pollfd structure and the local view of the out fds to avoid
921 * doing a lookup in the linked list and concurrency issues when writing is
922 * needed. Called with consumer_data.lock held.
924 * Returns the number of fds in the structures.
926 static int update_poll_array(struct lttng_consumer_local_data
*ctx
,
927 struct pollfd
**pollfd
, struct lttng_consumer_stream
**local_stream
,
931 struct lttng_ht_iter iter
;
932 struct lttng_consumer_stream
*stream
;
937 assert(local_stream
);
939 DBG("Updating poll fd array");
941 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
943 * Only active streams with an active end point can be added to the
944 * poll set and local stream storage of the thread.
946 * There is a potential race here for endpoint_status to be updated
947 * just after the check. However, this is OK since the stream(s) will
948 * be deleted once the thread is notified that the end point state has
949 * changed where this function will be called back again.
951 if (stream
->state
!= LTTNG_CONSUMER_ACTIVE_STREAM
||
952 stream
->endpoint_status
== CONSUMER_ENDPOINT_INACTIVE
) {
956 * This clobbers way too much the debug output. Uncomment that if you
957 * need it for debugging purposes.
959 * DBG("Active FD %d", stream->wait_fd);
961 (*pollfd
)[i
].fd
= stream
->wait_fd
;
962 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
963 local_stream
[i
] = stream
;
969 * Insert the consumer_data_pipe at the end of the array and don't
970 * increment i so nb_fd is the number of real FD.
972 (*pollfd
)[i
].fd
= lttng_pipe_get_readfd(ctx
->consumer_data_pipe
);
973 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
978 * Poll on the should_quit pipe and the command socket return -1 on error and
979 * should exit, 0 if data is available on the command socket
981 int lttng_consumer_poll_socket(struct pollfd
*consumer_sockpoll
)
986 num_rdy
= poll(consumer_sockpoll
, 2, -1);
989 * Restart interrupted system call.
991 if (errno
== EINTR
) {
994 PERROR("Poll error");
997 if (consumer_sockpoll
[0].revents
& (POLLIN
| POLLPRI
)) {
998 DBG("consumer_should_quit wake up");
1008 * Set the error socket.
1010 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data
*ctx
,
1013 ctx
->consumer_error_socket
= sock
;
1017 * Set the command socket path.
1019 void lttng_consumer_set_command_sock_path(
1020 struct lttng_consumer_local_data
*ctx
, char *sock
)
1022 ctx
->consumer_command_sock_path
= sock
;
1026 * Send return code to the session daemon.
1027 * If the socket is not defined, we return 0, it is not a fatal error
1029 int lttng_consumer_send_error(struct lttng_consumer_local_data
*ctx
, int cmd
)
1031 if (ctx
->consumer_error_socket
> 0) {
1032 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
1033 sizeof(enum lttcomm_sessiond_command
));
1040 * Close all the tracefiles and stream fds and MUST be called when all
1041 * instances are destroyed i.e. when all threads were joined and are ended.
1043 void lttng_consumer_cleanup(void)
1045 struct lttng_ht_iter iter
;
1046 struct lttng_consumer_channel
*channel
;
1050 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
, channel
,
1052 consumer_del_channel(channel
);
1057 lttng_ht_destroy(consumer_data
.channel_ht
);
1059 cleanup_relayd_ht();
1061 lttng_ht_destroy(consumer_data
.stream_per_chan_id_ht
);
1064 * This HT contains streams that are freed by either the metadata thread or
1065 * the data thread so we do *nothing* on the hash table and simply destroy
1068 lttng_ht_destroy(consumer_data
.stream_list_ht
);
1072 * Called from signal handler.
1074 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
1079 ret
= write(ctx
->consumer_should_quit
[1], "4", 1);
1080 } while (ret
< 0 && errno
== EINTR
);
1081 if (ret
< 0 || ret
!= 1) {
1082 PERROR("write consumer quit");
1085 DBG("Consumer flag that it should quit");
1088 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1091 int outfd
= stream
->out_fd
;
1094 * This does a blocking write-and-wait on any page that belongs to the
1095 * subbuffer prior to the one we just wrote.
1096 * Don't care about error values, as these are just hints and ways to
1097 * limit the amount of page cache used.
1099 if (orig_offset
< stream
->max_sb_size
) {
1102 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1103 stream
->max_sb_size
,
1104 SYNC_FILE_RANGE_WAIT_BEFORE
1105 | SYNC_FILE_RANGE_WRITE
1106 | SYNC_FILE_RANGE_WAIT_AFTER
);
1108 * Give hints to the kernel about how we access the file:
1109 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1112 * We need to call fadvise again after the file grows because the
1113 * kernel does not seem to apply fadvise to non-existing parts of the
1116 * Call fadvise _after_ having waited for the page writeback to
1117 * complete because the dirty page writeback semantic is not well
1118 * defined. So it can be expected to lead to lower throughput in
1121 posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1122 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1126 * Initialise the necessary environnement :
1127 * - create a new context
1128 * - create the poll_pipe
1129 * - create the should_quit pipe (for signal handler)
1130 * - create the thread pipe (for splice)
1132 * Takes a function pointer as argument, this function is called when data is
1133 * available on a buffer. This function is responsible to do the
1134 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1135 * buffer configuration and then kernctl_put_next_subbuf at the end.
1137 * Returns a pointer to the new context or NULL on error.
1139 struct lttng_consumer_local_data
*lttng_consumer_create(
1140 enum lttng_consumer_type type
,
1141 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1142 struct lttng_consumer_local_data
*ctx
),
1143 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1144 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1145 int (*update_stream
)(int stream_key
, uint32_t state
))
1148 struct lttng_consumer_local_data
*ctx
;
1150 assert(consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1151 consumer_data
.type
== type
);
1152 consumer_data
.type
= type
;
1154 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
1156 PERROR("allocating context");
1160 ctx
->consumer_error_socket
= -1;
1161 ctx
->consumer_metadata_socket
= -1;
1162 /* assign the callbacks */
1163 ctx
->on_buffer_ready
= buffer_ready
;
1164 ctx
->on_recv_channel
= recv_channel
;
1165 ctx
->on_recv_stream
= recv_stream
;
1166 ctx
->on_update_stream
= update_stream
;
1168 ctx
->consumer_data_pipe
= lttng_pipe_open(0);
1169 if (!ctx
->consumer_data_pipe
) {
1170 goto error_poll_pipe
;
1173 ret
= pipe(ctx
->consumer_should_quit
);
1175 PERROR("Error creating recv pipe");
1176 goto error_quit_pipe
;
1179 ret
= pipe(ctx
->consumer_thread_pipe
);
1181 PERROR("Error creating thread pipe");
1182 goto error_thread_pipe
;
1185 ret
= pipe(ctx
->consumer_channel_pipe
);
1187 PERROR("Error creating channel pipe");
1188 goto error_channel_pipe
;
1191 ctx
->consumer_metadata_pipe
= lttng_pipe_open(0);
1192 if (!ctx
->consumer_metadata_pipe
) {
1193 goto error_metadata_pipe
;
1196 ret
= utils_create_pipe(ctx
->consumer_splice_metadata_pipe
);
1198 goto error_splice_pipe
;
1204 lttng_pipe_destroy(ctx
->consumer_metadata_pipe
);
1205 error_metadata_pipe
:
1206 utils_close_pipe(ctx
->consumer_channel_pipe
);
1208 utils_close_pipe(ctx
->consumer_thread_pipe
);
1210 utils_close_pipe(ctx
->consumer_should_quit
);
1212 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1220 * Close all fds associated with the instance and free the context.
1222 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1226 DBG("Consumer destroying it. Closing everything.");
1228 ret
= close(ctx
->consumer_error_socket
);
1232 ret
= close(ctx
->consumer_metadata_socket
);
1236 utils_close_pipe(ctx
->consumer_thread_pipe
);
1237 utils_close_pipe(ctx
->consumer_channel_pipe
);
1238 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1239 lttng_pipe_destroy(ctx
->consumer_metadata_pipe
);
1240 utils_close_pipe(ctx
->consumer_should_quit
);
1241 utils_close_pipe(ctx
->consumer_splice_metadata_pipe
);
1243 unlink(ctx
->consumer_command_sock_path
);
1248 * Write the metadata stream id on the specified file descriptor.
1250 static int write_relayd_metadata_id(int fd
,
1251 struct lttng_consumer_stream
*stream
,
1252 struct consumer_relayd_sock_pair
*relayd
, unsigned long padding
)
1255 struct lttcomm_relayd_metadata_payload hdr
;
1257 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1258 hdr
.padding_size
= htobe32(padding
);
1260 ret
= write(fd
, (void *) &hdr
, sizeof(hdr
));
1261 } while (ret
< 0 && errno
== EINTR
);
1262 if (ret
< 0 || ret
!= sizeof(hdr
)) {
1264 * This error means that the fd's end is closed so ignore the perror
1265 * not to clubber the error output since this can happen in a normal
1268 if (errno
!= EPIPE
) {
1269 PERROR("write metadata stream id");
1271 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1273 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1274 * handle writting the missing part so report that as an error and
1275 * don't lie to the caller.
1280 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1281 stream
->relayd_stream_id
, padding
);
1288 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1289 * core function for writing trace buffers to either the local filesystem or
1292 * It must be called with the stream lock held.
1294 * Careful review MUST be put if any changes occur!
1296 * Returns the number of bytes written
1298 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1299 struct lttng_consumer_local_data
*ctx
,
1300 struct lttng_consumer_stream
*stream
, unsigned long len
,
1301 unsigned long padding
)
1303 unsigned long mmap_offset
;
1305 ssize_t ret
= 0, written
= 0;
1306 off_t orig_offset
= stream
->out_fd_offset
;
1307 /* Default is on the disk */
1308 int outfd
= stream
->out_fd
;
1309 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1310 unsigned int relayd_hang_up
= 0;
1312 /* RCU lock for the relayd pointer */
1315 /* Flag that the current stream if set for network streaming. */
1316 if (stream
->net_seq_idx
!= -1) {
1317 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1318 if (relayd
== NULL
) {
1323 /* get the offset inside the fd to mmap */
1324 switch (consumer_data
.type
) {
1325 case LTTNG_CONSUMER_KERNEL
:
1326 mmap_base
= stream
->mmap_base
;
1327 ret
= kernctl_get_mmap_read_offset(stream
->wait_fd
, &mmap_offset
);
1329 case LTTNG_CONSUMER32_UST
:
1330 case LTTNG_CONSUMER64_UST
:
1331 mmap_base
= lttng_ustctl_get_mmap_base(stream
);
1333 ERR("read mmap get mmap base for stream %s", stream
->name
);
1337 ret
= lttng_ustctl_get_mmap_read_offset(stream
, &mmap_offset
);
1341 ERR("Unknown consumer_data type");
1346 PERROR("tracer ctl get_mmap_read_offset");
1351 /* Handle stream on the relayd if the output is on the network */
1353 unsigned long netlen
= len
;
1356 * Lock the control socket for the complete duration of the function
1357 * since from this point on we will use the socket.
1359 if (stream
->metadata_flag
) {
1360 /* Metadata requires the control socket. */
1361 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1362 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1365 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1367 /* Use the returned socket. */
1370 /* Write metadata stream id before payload */
1371 if (stream
->metadata_flag
) {
1372 ret
= write_relayd_metadata_id(outfd
, stream
, relayd
, padding
);
1375 /* Socket operation failed. We consider the relayd dead */
1376 if (ret
== -EPIPE
|| ret
== -EINVAL
) {
1384 /* Socket operation failed. We consider the relayd dead */
1385 if (ret
== -EPIPE
|| ret
== -EINVAL
) {
1389 /* Else, use the default set before which is the filesystem. */
1392 /* No streaming, we have to set the len with the full padding */
1396 * Check if we need to change the tracefile before writing the packet.
1398 if (stream
->chan
->tracefile_size
> 0 &&
1399 (stream
->tracefile_size_current
+ len
) >
1400 stream
->chan
->tracefile_size
) {
1401 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1402 stream
->name
, stream
->chan
->tracefile_size
,
1403 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1404 stream
->out_fd
, &(stream
->tracefile_count_current
));
1406 ERR("Rotating output file");
1409 outfd
= stream
->out_fd
= ret
;
1410 /* Reset current size because we just perform a rotation. */
1411 stream
->tracefile_size_current
= 0;
1413 stream
->tracefile_size_current
+= len
;
1418 ret
= write(outfd
, mmap_base
+ mmap_offset
, len
);
1419 } while (ret
< 0 && errno
== EINTR
);
1420 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, len
);
1423 * This is possible if the fd is closed on the other side (outfd)
1424 * or any write problem. It can be verbose a bit for a normal
1425 * execution if for instance the relayd is stopped abruptly. This
1426 * can happen so set this to a DBG statement.
1428 DBG("Error in file write mmap");
1432 /* Socket operation failed. We consider the relayd dead */
1433 if (errno
== EPIPE
|| errno
== EINVAL
) {
1438 } else if (ret
> len
) {
1439 PERROR("Error in file write (ret %zd > len %lu)", ret
, len
);
1447 /* This call is useless on a socket so better save a syscall. */
1449 /* This won't block, but will start writeout asynchronously */
1450 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret
,
1451 SYNC_FILE_RANGE_WRITE
);
1452 stream
->out_fd_offset
+= ret
;
1456 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1460 * This is a special case that the relayd has closed its socket. Let's
1461 * cleanup the relayd object and all associated streams.
1463 if (relayd
&& relayd_hang_up
) {
1464 cleanup_relayd(relayd
, ctx
);
1468 /* Unlock only if ctrl socket used */
1469 if (relayd
&& stream
->metadata_flag
) {
1470 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1478 * Splice the data from the ring buffer to the tracefile.
1480 * It must be called with the stream lock held.
1482 * Returns the number of bytes spliced.
1484 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1485 struct lttng_consumer_local_data
*ctx
,
1486 struct lttng_consumer_stream
*stream
, unsigned long len
,
1487 unsigned long padding
)
1489 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1491 off_t orig_offset
= stream
->out_fd_offset
;
1492 int fd
= stream
->wait_fd
;
1493 /* Default is on the disk */
1494 int outfd
= stream
->out_fd
;
1495 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1497 unsigned int relayd_hang_up
= 0;
1499 switch (consumer_data
.type
) {
1500 case LTTNG_CONSUMER_KERNEL
:
1502 case LTTNG_CONSUMER32_UST
:
1503 case LTTNG_CONSUMER64_UST
:
1504 /* Not supported for user space tracing */
1507 ERR("Unknown consumer_data type");
1511 /* RCU lock for the relayd pointer */
1514 /* Flag that the current stream if set for network streaming. */
1515 if (stream
->net_seq_idx
!= -1) {
1516 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1517 if (relayd
== NULL
) {
1523 * Choose right pipe for splice. Metadata and trace data are handled by
1524 * different threads hence the use of two pipes in order not to race or
1525 * corrupt the written data.
1527 if (stream
->metadata_flag
) {
1528 splice_pipe
= ctx
->consumer_splice_metadata_pipe
;
1530 splice_pipe
= ctx
->consumer_thread_pipe
;
1533 /* Write metadata stream id before payload */
1535 int total_len
= len
;
1537 if (stream
->metadata_flag
) {
1539 * Lock the control socket for the complete duration of the function
1540 * since from this point on we will use the socket.
1542 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1544 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
, relayd
,
1548 /* Socket operation failed. We consider the relayd dead */
1549 if (ret
== -EBADF
) {
1550 WARN("Remote relayd disconnected. Stopping");
1557 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1560 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1562 /* Use the returned socket. */
1565 /* Socket operation failed. We consider the relayd dead */
1566 if (ret
== -EBADF
) {
1567 WARN("Remote relayd disconnected. Stopping");
1574 /* No streaming, we have to set the len with the full padding */
1578 * Check if we need to change the tracefile before writing the packet.
1580 if (stream
->chan
->tracefile_size
> 0 &&
1581 (stream
->tracefile_size_current
+ len
) >
1582 stream
->chan
->tracefile_size
) {
1583 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1584 stream
->name
, stream
->chan
->tracefile_size
,
1585 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1586 stream
->out_fd
, &(stream
->tracefile_count_current
));
1588 ERR("Rotating output file");
1591 outfd
= stream
->out_fd
= ret
;
1592 /* Reset current size because we just perform a rotation. */
1593 stream
->tracefile_size_current
= 0;
1595 stream
->tracefile_size_current
+= len
;
1599 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1600 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1601 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1602 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1603 DBG("splice chan to pipe, ret %zd", ret_splice
);
1604 if (ret_splice
< 0) {
1605 PERROR("Error in relay splice");
1607 written
= ret_splice
;
1613 /* Handle stream on the relayd if the output is on the network */
1615 if (stream
->metadata_flag
) {
1616 size_t metadata_payload_size
=
1617 sizeof(struct lttcomm_relayd_metadata_payload
);
1619 /* Update counter to fit the spliced data */
1620 ret_splice
+= metadata_payload_size
;
1621 len
+= metadata_payload_size
;
1623 * We do this so the return value can match the len passed as
1624 * argument to this function.
1626 written
-= metadata_payload_size
;
1630 /* Splice data out */
1631 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1632 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1633 DBG("Consumer splice pipe to file, ret %zd", ret_splice
);
1634 if (ret_splice
< 0) {
1635 PERROR("Error in file splice");
1637 written
= ret_splice
;
1639 /* Socket operation failed. We consider the relayd dead */
1640 if (errno
== EBADF
|| errno
== EPIPE
) {
1641 WARN("Remote relayd disconnected. Stopping");
1647 } else if (ret_splice
> len
) {
1649 PERROR("Wrote more data than requested %zd (len: %lu)",
1651 written
+= ret_splice
;
1657 /* This call is useless on a socket so better save a syscall. */
1659 /* This won't block, but will start writeout asynchronously */
1660 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1661 SYNC_FILE_RANGE_WRITE
);
1662 stream
->out_fd_offset
+= ret_splice
;
1664 written
+= ret_splice
;
1666 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1674 * This is a special case that the relayd has closed its socket. Let's
1675 * cleanup the relayd object and all associated streams.
1677 if (relayd
&& relayd_hang_up
) {
1678 cleanup_relayd(relayd
, ctx
);
1679 /* Skip splice error so the consumer does not fail */
1684 /* send the appropriate error description to sessiond */
1687 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1690 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1693 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
1698 if (relayd
&& stream
->metadata_flag
) {
1699 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1707 * Take a snapshot for a specific fd
1709 * Returns 0 on success, < 0 on error
1711 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
1713 switch (consumer_data
.type
) {
1714 case LTTNG_CONSUMER_KERNEL
:
1715 return lttng_kconsumer_take_snapshot(stream
);
1716 case LTTNG_CONSUMER32_UST
:
1717 case LTTNG_CONSUMER64_UST
:
1718 return lttng_ustconsumer_take_snapshot(stream
);
1720 ERR("Unknown consumer_data type");
1727 * Get the produced position
1729 * Returns 0 on success, < 0 on error
1731 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
1734 switch (consumer_data
.type
) {
1735 case LTTNG_CONSUMER_KERNEL
:
1736 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
1737 case LTTNG_CONSUMER32_UST
:
1738 case LTTNG_CONSUMER64_UST
:
1739 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
1741 ERR("Unknown consumer_data type");
1747 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
1748 int sock
, struct pollfd
*consumer_sockpoll
)
1750 switch (consumer_data
.type
) {
1751 case LTTNG_CONSUMER_KERNEL
:
1752 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1753 case LTTNG_CONSUMER32_UST
:
1754 case LTTNG_CONSUMER64_UST
:
1755 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1757 ERR("Unknown consumer_data type");
1764 * Iterate over all streams of the hashtable and free them properly.
1766 * WARNING: *MUST* be used with data stream only.
1768 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1770 struct lttng_ht_iter iter
;
1771 struct lttng_consumer_stream
*stream
;
1778 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1780 * Ignore return value since we are currently cleaning up so any error
1783 (void) consumer_del_stream(stream
, ht
);
1787 lttng_ht_destroy(ht
);
1791 * Iterate over all streams of the hashtable and free them properly.
1793 * XXX: Should not be only for metadata stream or else use an other name.
1795 static void destroy_stream_ht(struct lttng_ht
*ht
)
1797 struct lttng_ht_iter iter
;
1798 struct lttng_consumer_stream
*stream
;
1805 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1807 * Ignore return value since we are currently cleaning up so any error
1810 (void) consumer_del_metadata_stream(stream
, ht
);
1814 lttng_ht_destroy(ht
);
1817 void lttng_consumer_close_metadata(void)
1819 switch (consumer_data
.type
) {
1820 case LTTNG_CONSUMER_KERNEL
:
1822 * The Kernel consumer has a different metadata scheme so we don't
1823 * close anything because the stream will be closed by the session
1827 case LTTNG_CONSUMER32_UST
:
1828 case LTTNG_CONSUMER64_UST
:
1830 * Close all metadata streams. The metadata hash table is passed and
1831 * this call iterates over it by closing all wakeup fd. This is safe
1832 * because at this point we are sure that the metadata producer is
1833 * either dead or blocked.
1835 lttng_ustconsumer_close_metadata(metadata_ht
);
1838 ERR("Unknown consumer_data type");
1844 * Clean up a metadata stream and free its memory.
1846 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
1847 struct lttng_ht
*ht
)
1850 struct lttng_ht_iter iter
;
1851 struct lttng_consumer_channel
*free_chan
= NULL
;
1852 struct consumer_relayd_sock_pair
*relayd
;
1856 * This call should NEVER receive regular stream. It must always be
1857 * metadata stream and this is crucial for data structure synchronization.
1859 assert(stream
->metadata_flag
);
1861 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
1864 /* Means the stream was allocated but not successfully added */
1865 goto free_stream_rcu
;
1868 pthread_mutex_lock(&consumer_data
.lock
);
1869 pthread_mutex_lock(&stream
->lock
);
1871 switch (consumer_data
.type
) {
1872 case LTTNG_CONSUMER_KERNEL
:
1873 if (stream
->mmap_base
!= NULL
) {
1874 ret
= munmap(stream
->mmap_base
, stream
->mmap_len
);
1876 PERROR("munmap metadata stream");
1880 case LTTNG_CONSUMER32_UST
:
1881 case LTTNG_CONSUMER64_UST
:
1882 lttng_ustconsumer_del_stream(stream
);
1885 ERR("Unknown consumer_data type");
1891 iter
.iter
.node
= &stream
->node
.node
;
1892 ret
= lttng_ht_del(ht
, &iter
);
1895 iter
.iter
.node
= &stream
->node_channel_id
.node
;
1896 ret
= lttng_ht_del(consumer_data
.stream_per_chan_id_ht
, &iter
);
1899 iter
.iter
.node
= &stream
->node_session_id
.node
;
1900 ret
= lttng_ht_del(consumer_data
.stream_list_ht
, &iter
);
1904 if (stream
->out_fd
>= 0) {
1905 ret
= close(stream
->out_fd
);
1911 /* Check and cleanup relayd */
1913 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1914 if (relayd
!= NULL
) {
1915 uatomic_dec(&relayd
->refcount
);
1916 assert(uatomic_read(&relayd
->refcount
) >= 0);
1918 /* Closing streams requires to lock the control socket. */
1919 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1920 ret
= relayd_send_close_stream(&relayd
->control_sock
,
1921 stream
->relayd_stream_id
, stream
->next_net_seq_num
- 1);
1922 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1924 DBG("Unable to close stream on the relayd. Continuing");
1926 * Continue here. There is nothing we can do for the relayd.
1927 * Chances are that the relayd has closed the socket so we just
1928 * continue cleaning up.
1932 /* Both conditions are met, we destroy the relayd. */
1933 if (uatomic_read(&relayd
->refcount
) == 0 &&
1934 uatomic_read(&relayd
->destroy_flag
)) {
1935 destroy_relayd(relayd
);
1940 /* Atomically decrement channel refcount since other threads can use it. */
1941 if (!uatomic_sub_return(&stream
->chan
->refcount
, 1)
1942 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
1943 /* Go for channel deletion! */
1944 free_chan
= stream
->chan
;
1948 pthread_mutex_unlock(&stream
->lock
);
1949 pthread_mutex_unlock(&consumer_data
.lock
);
1952 consumer_del_channel(free_chan
);
1956 call_rcu(&stream
->node
.head
, free_stream_rcu
);
1960 * Action done with the metadata stream when adding it to the consumer internal
1961 * data structures to handle it.
1963 static int add_metadata_stream(struct lttng_consumer_stream
*stream
,
1964 struct lttng_ht
*ht
)
1967 struct consumer_relayd_sock_pair
*relayd
;
1968 struct lttng_ht_iter iter
;
1969 struct lttng_ht_node_u64
*node
;
1974 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
1976 pthread_mutex_lock(&consumer_data
.lock
);
1977 pthread_mutex_lock(&stream
->lock
);
1980 * From here, refcounts are updated so be _careful_ when returning an error
1987 * Lookup the stream just to make sure it does not exist in our internal
1988 * state. This should NEVER happen.
1990 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
1991 node
= lttng_ht_iter_get_node_u64(&iter
);
1994 /* Find relayd and, if one is found, increment refcount. */
1995 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1996 if (relayd
!= NULL
) {
1997 uatomic_inc(&relayd
->refcount
);
2000 /* Update channel refcount once added without error(s). */
2001 uatomic_inc(&stream
->chan
->refcount
);
2004 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2005 * in terms of destroying the associated channel, because the action that
2006 * causes the count to become 0 also causes a stream to be added. The
2007 * channel deletion will thus be triggered by the following removal of this
2010 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2011 /* Increment refcount before decrementing nb_init_stream_left */
2013 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2016 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2018 lttng_ht_add_unique_u64(consumer_data
.stream_per_chan_id_ht
,
2019 &stream
->node_channel_id
);
2022 * Add stream to the stream_list_ht of the consumer data. No need to steal
2023 * the key since the HT does not use it and we allow to add redundant keys
2026 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2030 pthread_mutex_unlock(&stream
->lock
);
2031 pthread_mutex_unlock(&consumer_data
.lock
);
2036 * Delete data stream that are flagged for deletion (endpoint_status).
2038 static void validate_endpoint_status_data_stream(void)
2040 struct lttng_ht_iter iter
;
2041 struct lttng_consumer_stream
*stream
;
2043 DBG("Consumer delete flagged data stream");
2046 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2047 /* Validate delete flag of the stream */
2048 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2051 /* Delete it right now */
2052 consumer_del_stream(stream
, data_ht
);
2058 * Delete metadata stream that are flagged for deletion (endpoint_status).
2060 static void validate_endpoint_status_metadata_stream(
2061 struct lttng_poll_event
*pollset
)
2063 struct lttng_ht_iter iter
;
2064 struct lttng_consumer_stream
*stream
;
2066 DBG("Consumer delete flagged metadata stream");
2071 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2072 /* Validate delete flag of the stream */
2073 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2077 * Remove from pollset so the metadata thread can continue without
2078 * blocking on a deleted stream.
2080 lttng_poll_del(pollset
, stream
->wait_fd
);
2082 /* Delete it right now */
2083 consumer_del_metadata_stream(stream
, metadata_ht
);
2089 * Thread polls on metadata file descriptor and write them on disk or on the
2092 void *consumer_thread_metadata_poll(void *data
)
2095 uint32_t revents
, nb_fd
;
2096 struct lttng_consumer_stream
*stream
= NULL
;
2097 struct lttng_ht_iter iter
;
2098 struct lttng_ht_node_u64
*node
;
2099 struct lttng_poll_event events
;
2100 struct lttng_consumer_local_data
*ctx
= data
;
2103 rcu_register_thread();
2105 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2107 /* ENOMEM at this point. Better to bail out. */
2111 DBG("Thread metadata poll started");
2113 /* Size is set to 1 for the consumer_metadata pipe */
2114 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2116 ERR("Poll set creation failed");
2120 ret
= lttng_poll_add(&events
,
2121 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2127 DBG("Metadata main loop started");
2130 /* Only the metadata pipe is set */
2131 if (LTTNG_POLL_GETNB(&events
) == 0 && consumer_quit
== 1) {
2136 DBG("Metadata poll wait with %d fd(s)", LTTNG_POLL_GETNB(&events
));
2137 ret
= lttng_poll_wait(&events
, -1);
2138 DBG("Metadata event catched in thread");
2140 if (errno
== EINTR
) {
2141 ERR("Poll EINTR catched");
2149 /* From here, the event is a metadata wait fd */
2150 for (i
= 0; i
< nb_fd
; i
++) {
2151 revents
= LTTNG_POLL_GETEV(&events
, i
);
2152 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2154 /* Just don't waste time if no returned events for the fd */
2159 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2160 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2161 DBG("Metadata thread pipe hung up");
2163 * Remove the pipe from the poll set and continue the loop
2164 * since their might be data to consume.
2166 lttng_poll_del(&events
,
2167 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2168 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2170 } else if (revents
& LPOLLIN
) {
2173 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2174 &stream
, sizeof(stream
));
2176 ERR("read metadata stream, ret: %ld", pipe_len
);
2178 * Continue here to handle the rest of the streams.
2183 /* A NULL stream means that the state has changed. */
2184 if (stream
== NULL
) {
2185 /* Check for deleted streams. */
2186 validate_endpoint_status_metadata_stream(&events
);
2190 DBG("Adding metadata stream %d to poll set",
2193 ret
= add_metadata_stream(stream
, metadata_ht
);
2195 ERR("Unable to add metadata stream");
2196 /* Stream was not setup properly. Continuing. */
2197 consumer_del_metadata_stream(stream
, NULL
);
2201 /* Add metadata stream to the global poll events list */
2202 lttng_poll_add(&events
, stream
->wait_fd
,
2203 LPOLLIN
| LPOLLPRI
);
2206 /* Handle other stream */
2212 uint64_t tmp_id
= (uint64_t) pollfd
;
2214 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2216 node
= lttng_ht_iter_get_node_u64(&iter
);
2219 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2222 /* Check for error event */
2223 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2224 DBG("Metadata fd %d is hup|err.", pollfd
);
2225 if (!stream
->hangup_flush_done
2226 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2227 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2228 DBG("Attempting to flush and consume the UST buffers");
2229 lttng_ustconsumer_on_stream_hangup(stream
);
2231 /* We just flushed the stream now read it. */
2233 len
= ctx
->on_buffer_ready(stream
, ctx
);
2235 * We don't check the return value here since if we get
2236 * a negative len, it means an error occured thus we
2237 * simply remove it from the poll set and free the
2243 lttng_poll_del(&events
, stream
->wait_fd
);
2245 * This call update the channel states, closes file descriptors
2246 * and securely free the stream.
2248 consumer_del_metadata_stream(stream
, metadata_ht
);
2249 } else if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2250 /* Get the data out of the metadata file descriptor */
2251 DBG("Metadata available on fd %d", pollfd
);
2252 assert(stream
->wait_fd
== pollfd
);
2254 len
= ctx
->on_buffer_ready(stream
, ctx
);
2255 /* It's ok to have an unavailable sub-buffer */
2256 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2257 /* Clean up stream from consumer and free it. */
2258 lttng_poll_del(&events
, stream
->wait_fd
);
2259 consumer_del_metadata_stream(stream
, metadata_ht
);
2260 } else if (len
> 0) {
2261 stream
->data_read
= 1;
2265 /* Release RCU lock for the stream looked up */
2272 DBG("Metadata poll thread exiting");
2274 lttng_poll_clean(&events
);
2276 destroy_stream_ht(metadata_ht
);
2278 rcu_unregister_thread();
2283 * This thread polls the fds in the set to consume the data and write
2284 * it to tracefile if necessary.
2286 void *consumer_thread_data_poll(void *data
)
2288 int num_rdy
, num_hup
, high_prio
, ret
, i
;
2289 struct pollfd
*pollfd
= NULL
;
2290 /* local view of the streams */
2291 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2292 /* local view of consumer_data.fds_count */
2294 struct lttng_consumer_local_data
*ctx
= data
;
2297 rcu_register_thread();
2299 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2300 if (data_ht
== NULL
) {
2301 /* ENOMEM at this point. Better to bail out. */
2305 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
));
2312 * the fds set has been updated, we need to update our
2313 * local array as well
2315 pthread_mutex_lock(&consumer_data
.lock
);
2316 if (consumer_data
.need_update
) {
2321 local_stream
= NULL
;
2323 /* allocate for all fds + 1 for the consumer_data_pipe */
2324 pollfd
= zmalloc((consumer_data
.stream_count
+ 1) * sizeof(struct pollfd
));
2325 if (pollfd
== NULL
) {
2326 PERROR("pollfd malloc");
2327 pthread_mutex_unlock(&consumer_data
.lock
);
2331 /* allocate for all fds + 1 for the consumer_data_pipe */
2332 local_stream
= zmalloc((consumer_data
.stream_count
+ 1) *
2333 sizeof(struct lttng_consumer_stream
));
2334 if (local_stream
== NULL
) {
2335 PERROR("local_stream malloc");
2336 pthread_mutex_unlock(&consumer_data
.lock
);
2339 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2342 ERR("Error in allocating pollfd or local_outfds");
2343 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2344 pthread_mutex_unlock(&consumer_data
.lock
);
2348 consumer_data
.need_update
= 0;
2350 pthread_mutex_unlock(&consumer_data
.lock
);
2352 /* No FDs and consumer_quit, consumer_cleanup the thread */
2353 if (nb_fd
== 0 && consumer_quit
== 1) {
2356 /* poll on the array of fds */
2358 DBG("polling on %d fd", nb_fd
+ 1);
2359 num_rdy
= poll(pollfd
, nb_fd
+ 1, -1);
2360 DBG("poll num_rdy : %d", num_rdy
);
2361 if (num_rdy
== -1) {
2363 * Restart interrupted system call.
2365 if (errno
== EINTR
) {
2368 PERROR("Poll error");
2369 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2371 } else if (num_rdy
== 0) {
2372 DBG("Polling thread timed out");
2377 * If the consumer_data_pipe triggered poll go directly to the
2378 * beginning of the loop to update the array. We want to prioritize
2379 * array update over low-priority reads.
2381 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2382 ssize_t pipe_readlen
;
2384 DBG("consumer_data_pipe wake up");
2385 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2386 &new_stream
, sizeof(new_stream
));
2387 if (pipe_readlen
< 0) {
2388 ERR("Consumer data pipe ret %ld", pipe_readlen
);
2389 /* Continue so we can at least handle the current stream(s). */
2394 * If the stream is NULL, just ignore it. It's also possible that
2395 * the sessiond poll thread changed the consumer_quit state and is
2396 * waking us up to test it.
2398 if (new_stream
== NULL
) {
2399 validate_endpoint_status_data_stream();
2403 ret
= add_stream(new_stream
, data_ht
);
2405 ERR("Consumer add stream %" PRIu64
" failed. Continuing",
2408 * At this point, if the add_stream fails, it is not in the
2409 * hash table thus passing the NULL value here.
2411 consumer_del_stream(new_stream
, NULL
);
2414 /* Continue to update the local streams and handle prio ones */
2418 /* Take care of high priority channels first. */
2419 for (i
= 0; i
< nb_fd
; i
++) {
2420 if (local_stream
[i
] == NULL
) {
2423 if (pollfd
[i
].revents
& POLLPRI
) {
2424 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2426 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2427 /* it's ok to have an unavailable sub-buffer */
2428 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2429 /* Clean the stream and free it. */
2430 consumer_del_stream(local_stream
[i
], data_ht
);
2431 local_stream
[i
] = NULL
;
2432 } else if (len
> 0) {
2433 local_stream
[i
]->data_read
= 1;
2439 * If we read high prio channel in this loop, try again
2440 * for more high prio data.
2446 /* Take care of low priority channels. */
2447 for (i
= 0; i
< nb_fd
; i
++) {
2448 if (local_stream
[i
] == NULL
) {
2451 if ((pollfd
[i
].revents
& POLLIN
) ||
2452 local_stream
[i
]->hangup_flush_done
) {
2453 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2454 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2455 /* it's ok to have an unavailable sub-buffer */
2456 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2457 /* Clean the stream and free it. */
2458 consumer_del_stream(local_stream
[i
], data_ht
);
2459 local_stream
[i
] = NULL
;
2460 } else if (len
> 0) {
2461 local_stream
[i
]->data_read
= 1;
2466 /* Handle hangup and errors */
2467 for (i
= 0; i
< nb_fd
; i
++) {
2468 if (local_stream
[i
] == NULL
) {
2471 if (!local_stream
[i
]->hangup_flush_done
2472 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2473 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2474 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2475 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2477 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2478 /* Attempt read again, for the data we just flushed. */
2479 local_stream
[i
]->data_read
= 1;
2482 * If the poll flag is HUP/ERR/NVAL and we have
2483 * read no data in this pass, we can remove the
2484 * stream from its hash table.
2486 if ((pollfd
[i
].revents
& POLLHUP
)) {
2487 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2488 if (!local_stream
[i
]->data_read
) {
2489 consumer_del_stream(local_stream
[i
], data_ht
);
2490 local_stream
[i
] = NULL
;
2493 } else if (pollfd
[i
].revents
& POLLERR
) {
2494 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2495 if (!local_stream
[i
]->data_read
) {
2496 consumer_del_stream(local_stream
[i
], data_ht
);
2497 local_stream
[i
] = NULL
;
2500 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2501 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2502 if (!local_stream
[i
]->data_read
) {
2503 consumer_del_stream(local_stream
[i
], data_ht
);
2504 local_stream
[i
] = NULL
;
2508 if (local_stream
[i
] != NULL
) {
2509 local_stream
[i
]->data_read
= 0;
2514 DBG("polling thread exiting");
2519 * Close the write side of the pipe so epoll_wait() in
2520 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2521 * read side of the pipe. If we close them both, epoll_wait strangely does
2522 * not return and could create a endless wait period if the pipe is the
2523 * only tracked fd in the poll set. The thread will take care of closing
2526 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2528 destroy_data_stream_ht(data_ht
);
2530 rcu_unregister_thread();
2535 * Close wake-up end of each stream belonging to the channel. This will
2536 * allow the poll() on the stream read-side to detect when the
2537 * write-side (application) finally closes them.
2540 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2542 struct lttng_ht
*ht
;
2543 struct lttng_consumer_stream
*stream
;
2544 struct lttng_ht_iter iter
;
2546 ht
= consumer_data
.stream_per_chan_id_ht
;
2549 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2550 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2551 ht
->match_fct
, &channel
->key
,
2552 &iter
.iter
, stream
, node_channel_id
.node
) {
2554 * Protect against teardown with mutex.
2556 pthread_mutex_lock(&stream
->lock
);
2557 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2560 switch (consumer_data
.type
) {
2561 case LTTNG_CONSUMER_KERNEL
:
2563 case LTTNG_CONSUMER32_UST
:
2564 case LTTNG_CONSUMER64_UST
:
2566 * Note: a mutex is taken internally within
2567 * liblttng-ust-ctl to protect timer wakeup_fd
2568 * use from concurrent close.
2570 lttng_ustconsumer_close_stream_wakeup(stream
);
2573 ERR("Unknown consumer_data type");
2577 pthread_mutex_unlock(&stream
->lock
);
2582 static void destroy_channel_ht(struct lttng_ht
*ht
)
2584 struct lttng_ht_iter iter
;
2585 struct lttng_consumer_channel
*channel
;
2593 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2594 ret
= lttng_ht_del(ht
, &iter
);
2599 lttng_ht_destroy(ht
);
2603 * This thread polls the channel fds to detect when they are being
2604 * closed. It closes all related streams if the channel is detected as
2605 * closed. It is currently only used as a shim layer for UST because the
2606 * consumerd needs to keep the per-stream wakeup end of pipes open for
2609 void *consumer_thread_channel_poll(void *data
)
2612 uint32_t revents
, nb_fd
;
2613 struct lttng_consumer_channel
*chan
= NULL
;
2614 struct lttng_ht_iter iter
;
2615 struct lttng_ht_node_u64
*node
;
2616 struct lttng_poll_event events
;
2617 struct lttng_consumer_local_data
*ctx
= data
;
2618 struct lttng_ht
*channel_ht
;
2620 rcu_register_thread();
2622 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2624 /* ENOMEM at this point. Better to bail out. */
2628 DBG("Thread channel poll started");
2630 /* Size is set to 1 for the consumer_channel pipe */
2631 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2633 ERR("Poll set creation failed");
2637 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2643 DBG("Channel main loop started");
2646 /* Only the channel pipe is set */
2647 if (LTTNG_POLL_GETNB(&events
) == 0 && consumer_quit
== 1) {
2652 DBG("Channel poll wait with %d fd(s)", LTTNG_POLL_GETNB(&events
));
2653 ret
= lttng_poll_wait(&events
, -1);
2654 DBG("Channel event catched in thread");
2656 if (errno
== EINTR
) {
2657 ERR("Poll EINTR catched");
2665 /* From here, the event is a channel wait fd */
2666 for (i
= 0; i
< nb_fd
; i
++) {
2667 revents
= LTTNG_POLL_GETEV(&events
, i
);
2668 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2670 /* Just don't waste time if no returned events for the fd */
2674 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2675 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2676 DBG("Channel thread pipe hung up");
2678 * Remove the pipe from the poll set and continue the loop
2679 * since their might be data to consume.
2681 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2683 } else if (revents
& LPOLLIN
) {
2684 enum consumer_channel_action action
;
2687 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
2689 ERR("Error reading channel pipe");
2694 case CONSUMER_CHANNEL_ADD
:
2695 DBG("Adding channel %d to poll set",
2698 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
2700 lttng_ht_add_unique_u64(channel_ht
,
2701 &chan
->wait_fd_node
);
2702 /* Add channel to the global poll events list */
2703 lttng_poll_add(&events
, chan
->wait_fd
,
2704 LPOLLIN
| LPOLLPRI
);
2706 case CONSUMER_CHANNEL_DEL
:
2708 chan
= consumer_find_channel(key
);
2710 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
2713 lttng_poll_del(&events
, chan
->wait_fd
);
2714 ret
= lttng_ht_del(channel_ht
, &iter
);
2716 consumer_close_channel_streams(chan
);
2719 * Release our own refcount. Force channel deletion even if
2720 * streams were not initialized.
2722 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
2723 consumer_del_channel(chan
);
2727 case CONSUMER_CHANNEL_QUIT
:
2729 * Remove the pipe from the poll set and continue the loop
2730 * since their might be data to consume.
2732 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2735 ERR("Unknown action");
2740 /* Handle other stream */
2746 uint64_t tmp_id
= (uint64_t) pollfd
;
2748 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
2750 node
= lttng_ht_iter_get_node_u64(&iter
);
2753 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
2756 /* Check for error event */
2757 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2758 DBG("Channel fd %d is hup|err.", pollfd
);
2760 lttng_poll_del(&events
, chan
->wait_fd
);
2761 ret
= lttng_ht_del(channel_ht
, &iter
);
2763 consumer_close_channel_streams(chan
);
2765 /* Release our own refcount */
2766 if (!uatomic_sub_return(&chan
->refcount
, 1)
2767 && !uatomic_read(&chan
->nb_init_stream_left
)) {
2768 consumer_del_channel(chan
);
2772 /* Release RCU lock for the channel looked up */
2778 lttng_poll_clean(&events
);
2780 destroy_channel_ht(channel_ht
);
2782 DBG("Channel poll thread exiting");
2783 rcu_unregister_thread();
2787 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
2788 struct pollfd
*sockpoll
, int client_socket
)
2795 if (lttng_consumer_poll_socket(sockpoll
) < 0) {
2799 DBG("Metadata connection on client_socket");
2801 /* Blocking call, waiting for transmission */
2802 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
2803 if (ctx
->consumer_metadata_socket
< 0) {
2804 WARN("On accept metadata");
2815 * This thread listens on the consumerd socket and receives the file
2816 * descriptors from the session daemon.
2818 void *consumer_thread_sessiond_poll(void *data
)
2820 int sock
= -1, client_socket
, ret
;
2822 * structure to poll for incoming data on communication socket avoids
2823 * making blocking sockets.
2825 struct pollfd consumer_sockpoll
[2];
2826 struct lttng_consumer_local_data
*ctx
= data
;
2828 rcu_register_thread();
2830 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
2831 unlink(ctx
->consumer_command_sock_path
);
2832 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
2833 if (client_socket
< 0) {
2834 ERR("Cannot create command socket");
2838 ret
= lttcomm_listen_unix_sock(client_socket
);
2843 DBG("Sending ready command to lttng-sessiond");
2844 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
2845 /* return < 0 on error, but == 0 is not fatal */
2847 ERR("Error sending ready command to lttng-sessiond");
2851 ret
= fcntl(client_socket
, F_SETFL
, O_NONBLOCK
);
2853 PERROR("fcntl O_NONBLOCK");
2857 /* prepare the FDs to poll : to client socket and the should_quit pipe */
2858 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
2859 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
2860 consumer_sockpoll
[1].fd
= client_socket
;
2861 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
2863 if (lttng_consumer_poll_socket(consumer_sockpoll
) < 0) {
2866 DBG("Connection on client_socket");
2868 /* Blocking call, waiting for transmission */
2869 sock
= lttcomm_accept_unix_sock(client_socket
);
2874 ret
= fcntl(sock
, F_SETFL
, O_NONBLOCK
);
2876 PERROR("fcntl O_NONBLOCK");
2881 * Setup metadata socket which is the second socket connection on the
2882 * command unix socket.
2884 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
2889 /* This socket is not useful anymore. */
2890 ret
= close(client_socket
);
2892 PERROR("close client_socket");
2896 /* update the polling structure to poll on the established socket */
2897 consumer_sockpoll
[1].fd
= sock
;
2898 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
2901 if (lttng_consumer_poll_socket(consumer_sockpoll
) < 0) {
2904 DBG("Incoming command on sock");
2905 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2906 if (ret
== -ENOENT
) {
2907 DBG("Received STOP command");
2912 * This could simply be a session daemon quitting. Don't output
2915 DBG("Communication interrupted on command socket");
2918 if (consumer_quit
) {
2919 DBG("consumer_thread_receive_fds received quit from signal");
2922 DBG("received command on sock");
2925 DBG("Consumer thread sessiond poll exiting");
2928 * Close metadata streams since the producer is the session daemon which
2931 * NOTE: for now, this only applies to the UST tracer.
2933 lttng_consumer_close_metadata();
2936 * when all fds have hung up, the polling thread
2942 * Notify the data poll thread to poll back again and test the
2943 * consumer_quit state that we just set so to quit gracefully.
2945 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
2947 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
2949 /* Cleaning up possibly open sockets. */
2953 PERROR("close sock sessiond poll");
2956 if (client_socket
>= 0) {
2959 PERROR("close client_socket sessiond poll");
2963 rcu_unregister_thread();
2967 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
2968 struct lttng_consumer_local_data
*ctx
)
2972 pthread_mutex_lock(&stream
->lock
);
2974 switch (consumer_data
.type
) {
2975 case LTTNG_CONSUMER_KERNEL
:
2976 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
2978 case LTTNG_CONSUMER32_UST
:
2979 case LTTNG_CONSUMER64_UST
:
2980 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
2983 ERR("Unknown consumer_data type");
2989 pthread_mutex_unlock(&stream
->lock
);
2993 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
2995 switch (consumer_data
.type
) {
2996 case LTTNG_CONSUMER_KERNEL
:
2997 return lttng_kconsumer_on_recv_stream(stream
);
2998 case LTTNG_CONSUMER32_UST
:
2999 case LTTNG_CONSUMER64_UST
:
3000 return lttng_ustconsumer_on_recv_stream(stream
);
3002 ERR("Unknown consumer_data type");
3009 * Allocate and set consumer data hash tables.
3011 void lttng_consumer_init(void)
3013 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3014 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3015 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3016 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3020 * Process the ADD_RELAYD command receive by a consumer.
3022 * This will create a relayd socket pair and add it to the relayd hash table.
3023 * The caller MUST acquire a RCU read side lock before calling it.
3025 int consumer_add_relayd_socket(int net_seq_idx
, int sock_type
,
3026 struct lttng_consumer_local_data
*ctx
, int sock
,
3027 struct pollfd
*consumer_sockpoll
,
3028 struct lttcomm_relayd_sock
*relayd_sock
, unsigned int sessiond_id
)
3030 int fd
= -1, ret
= -1, relayd_created
= 0;
3031 enum lttng_error_code ret_code
= LTTNG_OK
;
3032 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3035 assert(relayd_sock
);
3037 DBG("Consumer adding relayd socket (idx: %d)", net_seq_idx
);
3039 /* First send a status message before receiving the fds. */
3040 ret
= consumer_send_status_msg(sock
, ret_code
);
3042 /* Somehow, the session daemon is not responding anymore. */
3046 /* Get relayd reference if exists. */
3047 relayd
= consumer_find_relayd(net_seq_idx
);
3048 if (relayd
== NULL
) {
3049 /* Not found. Allocate one. */
3050 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3051 if (relayd
== NULL
) {
3052 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_OUTFD_ERROR
);
3056 relayd
->sessiond_session_id
= (uint64_t) sessiond_id
;
3060 /* Poll on consumer socket. */
3061 if (lttng_consumer_poll_socket(consumer_sockpoll
) < 0) {
3066 /* Get relayd socket from session daemon */
3067 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3068 if (ret
!= sizeof(fd
)) {
3069 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3071 fd
= -1; /* Just in case it gets set with an invalid value. */
3075 /* We have the fds without error. Send status back. */
3076 ret
= consumer_send_status_msg(sock
, ret_code
);
3078 /* Somehow, the session daemon is not responding anymore. */
3082 /* Copy socket information and received FD */
3083 switch (sock_type
) {
3084 case LTTNG_STREAM_CONTROL
:
3085 /* Copy received lttcomm socket */
3086 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3087 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3088 /* Immediately try to close the created socket if valid. */
3089 if (relayd
->control_sock
.sock
.fd
>= 0) {
3090 if (close(relayd
->control_sock
.sock
.fd
)) {
3091 PERROR("close relayd control socket");
3094 /* Handle create_sock error. */
3099 /* Assign new file descriptor */
3100 relayd
->control_sock
.sock
.fd
= fd
;
3101 /* Assign version values. */
3102 relayd
->control_sock
.major
= relayd_sock
->major
;
3103 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3106 * Create a session on the relayd and store the returned id. Lock the
3107 * control socket mutex if the relayd was NOT created before.
3109 if (!relayd_created
) {
3110 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3112 ret
= relayd_create_session(&relayd
->control_sock
,
3113 &relayd
->relayd_session_id
);
3114 if (!relayd_created
) {
3115 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3119 * Close all sockets of a relayd object. It will be freed if it was
3120 * created at the error code path or else it will be garbage
3123 (void) relayd_close(&relayd
->control_sock
);
3124 (void) relayd_close(&relayd
->data_sock
);
3129 case LTTNG_STREAM_DATA
:
3130 /* Copy received lttcomm socket */
3131 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3132 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3133 /* Immediately try to close the created socket if valid. */
3134 if (relayd
->data_sock
.sock
.fd
>= 0) {
3135 if (close(relayd
->data_sock
.sock
.fd
)) {
3136 PERROR("close relayd data socket");
3139 /* Handle create_sock error. */
3144 /* Assign new file descriptor */
3145 relayd
->data_sock
.sock
.fd
= fd
;
3146 /* Assign version values. */
3147 relayd
->data_sock
.major
= relayd_sock
->major
;
3148 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3151 ERR("Unknown relayd socket type (%d)", sock_type
);
3156 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3157 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3158 relayd
->net_seq_idx
, fd
);
3161 * Add relayd socket pair to consumer data hashtable. If object already
3162 * exists or on error, the function gracefully returns.
3170 /* Close received socket if valid. */
3173 PERROR("close received socket");
3178 if (relayd_created
) {
3186 * Try to lock the stream mutex.
3188 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
3190 static int stream_try_lock(struct lttng_consumer_stream
*stream
)
3197 * Try to lock the stream mutex. On failure, we know that the stream is
3198 * being used else where hence there is data still being extracted.
3200 ret
= pthread_mutex_trylock(&stream
->lock
);
3202 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
3214 * Search for a relayd associated to the session id and return the reference.
3216 * A rcu read side lock MUST be acquire before calling this function and locked
3217 * until the relayd object is no longer necessary.
3219 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3221 struct lttng_ht_iter iter
;
3222 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3224 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3225 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3228 * Check by sessiond id which is unique here where the relayd session
3229 * id might not be when having multiple relayd.
3231 if (relayd
->sessiond_session_id
== id
) {
3232 /* Found the relayd. There can be only one per id. */
3244 * Check if for a given session id there is still data needed to be extract
3247 * Return 1 if data is pending or else 0 meaning ready to be read.
3249 int consumer_data_pending(uint64_t id
)
3252 struct lttng_ht_iter iter
;
3253 struct lttng_ht
*ht
;
3254 struct lttng_consumer_stream
*stream
;
3255 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3256 int (*data_pending
)(struct lttng_consumer_stream
*);
3258 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3261 pthread_mutex_lock(&consumer_data
.lock
);
3263 switch (consumer_data
.type
) {
3264 case LTTNG_CONSUMER_KERNEL
:
3265 data_pending
= lttng_kconsumer_data_pending
;
3267 case LTTNG_CONSUMER32_UST
:
3268 case LTTNG_CONSUMER64_UST
:
3269 data_pending
= lttng_ustconsumer_data_pending
;
3272 ERR("Unknown consumer data type");
3276 /* Ease our life a bit */
3277 ht
= consumer_data
.stream_list_ht
;
3279 relayd
= find_relayd_by_session_id(id
);
3281 /* Send init command for data pending. */
3282 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3283 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3284 relayd
->relayd_session_id
);
3285 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3287 /* Communication error thus the relayd so no data pending. */
3288 goto data_not_pending
;
3292 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3293 ht
->hash_fct(&id
, lttng_ht_seed
),
3295 &iter
.iter
, stream
, node_session_id
.node
) {
3296 /* If this call fails, the stream is being used hence data pending. */
3297 ret
= stream_try_lock(stream
);
3303 * A removed node from the hash table indicates that the stream has
3304 * been deleted thus having a guarantee that the buffers are closed
3305 * on the consumer side. However, data can still be transmitted
3306 * over the network so don't skip the relayd check.
3308 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3310 /* Check the stream if there is data in the buffers. */
3311 ret
= data_pending(stream
);
3313 pthread_mutex_unlock(&stream
->lock
);
3320 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3321 if (stream
->metadata_flag
) {
3322 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3323 stream
->relayd_stream_id
);
3325 ret
= relayd_data_pending(&relayd
->control_sock
,
3326 stream
->relayd_stream_id
,
3327 stream
->next_net_seq_num
- 1);
3329 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3331 pthread_mutex_unlock(&stream
->lock
);
3335 pthread_mutex_unlock(&stream
->lock
);
3339 unsigned int is_data_inflight
= 0;
3341 /* Send init command for data pending. */
3342 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3343 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3344 relayd
->relayd_session_id
, &is_data_inflight
);
3345 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3347 goto data_not_pending
;
3349 if (is_data_inflight
) {
3355 * Finding _no_ node in the hash table and no inflight data means that the
3356 * stream(s) have been removed thus data is guaranteed to be available for
3357 * analysis from the trace files.
3361 /* Data is available to be read by a viewer. */
3362 pthread_mutex_unlock(&consumer_data
.lock
);
3367 /* Data is still being extracted from buffers. */
3368 pthread_mutex_unlock(&consumer_data
.lock
);
3374 * Send a ret code status message to the sessiond daemon.
3376 * Return the sendmsg() return value.
3378 int consumer_send_status_msg(int sock
, int ret_code
)
3380 struct lttcomm_consumer_status_msg msg
;
3382 msg
.ret_code
= ret_code
;
3384 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3388 * Send a channel status message to the sessiond daemon.
3390 * Return the sendmsg() return value.
3392 int consumer_send_status_channel(int sock
,
3393 struct lttng_consumer_channel
*channel
)
3395 struct lttcomm_consumer_status_channel msg
;
3400 msg
.ret_code
= -LTTNG_ERR_UST_CHAN_FAIL
;
3402 msg
.ret_code
= LTTNG_OK
;
3403 msg
.key
= channel
->key
;
3404 msg
.stream_count
= channel
->streams
.count
;
3407 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));