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.
28 #include <sys/socket.h>
29 #include <sys/types.h>
34 #include <bin/lttng-consumerd/health-consumerd.h>
35 #include <common/common.h>
36 #include <common/utils.h>
37 #include <common/compat/poll.h>
38 #include <common/compat/endian.h>
39 #include <common/index/index.h>
40 #include <common/kernel-ctl/kernel-ctl.h>
41 #include <common/sessiond-comm/relayd.h>
42 #include <common/sessiond-comm/sessiond-comm.h>
43 #include <common/kernel-consumer/kernel-consumer.h>
44 #include <common/relayd/relayd.h>
45 #include <common/ust-consumer/ust-consumer.h>
46 #include <common/consumer-timer.h>
49 #include "consumer-stream.h"
50 #include "consumer-testpoint.h"
52 struct lttng_consumer_global_data consumer_data
= {
55 .type
= LTTNG_CONSUMER_UNKNOWN
,
58 enum consumer_channel_action
{
61 CONSUMER_CHANNEL_QUIT
,
64 struct consumer_channel_msg
{
65 enum consumer_channel_action action
;
66 struct lttng_consumer_channel
*chan
; /* add */
67 uint64_t key
; /* del */
71 * Flag to inform the polling thread to quit when all fd hung up. Updated by
72 * the consumer_thread_receive_fds when it notices that all fds has hung up.
73 * Also updated by the signal handler (consumer_should_exit()). Read by the
76 volatile int consumer_quit
;
79 * Global hash table containing respectively metadata and data streams. The
80 * stream element in this ht should only be updated by the metadata poll thread
81 * for the metadata and the data poll thread for the data.
83 static struct lttng_ht
*metadata_ht
;
84 static struct lttng_ht
*data_ht
;
87 * Notify a thread lttng pipe to poll back again. This usually means that some
88 * global state has changed so we just send back the thread in a poll wait
91 static void notify_thread_lttng_pipe(struct lttng_pipe
*pipe
)
93 struct lttng_consumer_stream
*null_stream
= NULL
;
97 (void) lttng_pipe_write(pipe
, &null_stream
, sizeof(null_stream
));
100 static void notify_health_quit_pipe(int *pipe
)
104 ret
= lttng_write(pipe
[1], "4", 1);
106 PERROR("write consumer health quit");
110 static void notify_channel_pipe(struct lttng_consumer_local_data
*ctx
,
111 struct lttng_consumer_channel
*chan
,
113 enum consumer_channel_action action
)
115 struct consumer_channel_msg msg
;
118 memset(&msg
, 0, sizeof(msg
));
123 ret
= lttng_write(ctx
->consumer_channel_pipe
[1], &msg
, sizeof(msg
));
124 if (ret
< sizeof(msg
)) {
125 PERROR("notify_channel_pipe write error");
129 void notify_thread_del_channel(struct lttng_consumer_local_data
*ctx
,
132 notify_channel_pipe(ctx
, NULL
, key
, CONSUMER_CHANNEL_DEL
);
135 static int read_channel_pipe(struct lttng_consumer_local_data
*ctx
,
136 struct lttng_consumer_channel
**chan
,
138 enum consumer_channel_action
*action
)
140 struct consumer_channel_msg msg
;
143 ret
= lttng_read(ctx
->consumer_channel_pipe
[0], &msg
, sizeof(msg
));
144 if (ret
< sizeof(msg
)) {
148 *action
= msg
.action
;
156 * Cleanup the stream list of a channel. Those streams are not yet globally
159 static void clean_channel_stream_list(struct lttng_consumer_channel
*channel
)
161 struct lttng_consumer_stream
*stream
, *stmp
;
165 /* Delete streams that might have been left in the stream list. */
166 cds_list_for_each_entry_safe(stream
, stmp
, &channel
->streams
.head
,
168 cds_list_del(&stream
->send_node
);
170 * Once a stream is added to this list, the buffers were created so we
171 * have a guarantee that this call will succeed. Setting the monitor
172 * mode to 0 so we don't lock nor try to delete the stream from the
176 consumer_stream_destroy(stream
, NULL
);
181 * Find a stream. The consumer_data.lock must be locked during this
184 static struct lttng_consumer_stream
*find_stream(uint64_t key
,
187 struct lttng_ht_iter iter
;
188 struct lttng_ht_node_u64
*node
;
189 struct lttng_consumer_stream
*stream
= NULL
;
193 /* -1ULL keys are lookup failures */
194 if (key
== (uint64_t) -1ULL) {
200 lttng_ht_lookup(ht
, &key
, &iter
);
201 node
= lttng_ht_iter_get_node_u64(&iter
);
203 stream
= caa_container_of(node
, struct lttng_consumer_stream
, node
);
211 static void steal_stream_key(uint64_t key
, struct lttng_ht
*ht
)
213 struct lttng_consumer_stream
*stream
;
216 stream
= find_stream(key
, ht
);
218 stream
->key
= (uint64_t) -1ULL;
220 * We don't want the lookup to match, but we still need
221 * to iterate on this stream when iterating over the hash table. Just
222 * change the node key.
224 stream
->node
.key
= (uint64_t) -1ULL;
230 * Return a channel object for the given key.
232 * RCU read side lock MUST be acquired before calling this function and
233 * protects the channel ptr.
235 struct lttng_consumer_channel
*consumer_find_channel(uint64_t key
)
237 struct lttng_ht_iter iter
;
238 struct lttng_ht_node_u64
*node
;
239 struct lttng_consumer_channel
*channel
= NULL
;
241 /* -1ULL keys are lookup failures */
242 if (key
== (uint64_t) -1ULL) {
246 lttng_ht_lookup(consumer_data
.channel_ht
, &key
, &iter
);
247 node
= lttng_ht_iter_get_node_u64(&iter
);
249 channel
= caa_container_of(node
, struct lttng_consumer_channel
, node
);
256 * There is a possibility that the consumer does not have enough time between
257 * the close of the channel on the session daemon and the cleanup in here thus
258 * once we have a channel add with an existing key, we know for sure that this
259 * channel will eventually get cleaned up by all streams being closed.
261 * This function just nullifies the already existing channel key.
263 static void steal_channel_key(uint64_t key
)
265 struct lttng_consumer_channel
*channel
;
268 channel
= consumer_find_channel(key
);
270 channel
->key
= (uint64_t) -1ULL;
272 * We don't want the lookup to match, but we still need to iterate on
273 * this channel when iterating over the hash table. Just change the
276 channel
->node
.key
= (uint64_t) -1ULL;
281 static void free_channel_rcu(struct rcu_head
*head
)
283 struct lttng_ht_node_u64
*node
=
284 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
285 struct lttng_consumer_channel
*channel
=
286 caa_container_of(node
, struct lttng_consumer_channel
, node
);
292 * RCU protected relayd socket pair free.
294 static void free_relayd_rcu(struct rcu_head
*head
)
296 struct lttng_ht_node_u64
*node
=
297 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
298 struct consumer_relayd_sock_pair
*relayd
=
299 caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
302 * Close all sockets. This is done in the call RCU since we don't want the
303 * socket fds to be reassigned thus potentially creating bad state of the
306 * We do not have to lock the control socket mutex here since at this stage
307 * there is no one referencing to this relayd object.
309 (void) relayd_close(&relayd
->control_sock
);
310 (void) relayd_close(&relayd
->data_sock
);
316 * Destroy and free relayd socket pair object.
318 void consumer_destroy_relayd(struct consumer_relayd_sock_pair
*relayd
)
321 struct lttng_ht_iter iter
;
323 if (relayd
== NULL
) {
327 DBG("Consumer destroy and close relayd socket pair");
329 iter
.iter
.node
= &relayd
->node
.node
;
330 ret
= lttng_ht_del(consumer_data
.relayd_ht
, &iter
);
332 /* We assume the relayd is being or is destroyed */
336 /* RCU free() call */
337 call_rcu(&relayd
->node
.head
, free_relayd_rcu
);
341 * Remove a channel from the global list protected by a mutex. This function is
342 * also responsible for freeing its data structures.
344 void consumer_del_channel(struct lttng_consumer_channel
*channel
)
347 struct lttng_ht_iter iter
;
349 DBG("Consumer delete channel key %" PRIu64
, channel
->key
);
351 pthread_mutex_lock(&consumer_data
.lock
);
352 pthread_mutex_lock(&channel
->lock
);
354 /* Destroy streams that might have been left in the stream list. */
355 clean_channel_stream_list(channel
);
357 if (channel
->live_timer_enabled
== 1) {
358 consumer_timer_live_stop(channel
);
361 switch (consumer_data
.type
) {
362 case LTTNG_CONSUMER_KERNEL
:
364 case LTTNG_CONSUMER32_UST
:
365 case LTTNG_CONSUMER64_UST
:
366 lttng_ustconsumer_del_channel(channel
);
369 ERR("Unknown consumer_data type");
375 iter
.iter
.node
= &channel
->node
.node
;
376 ret
= lttng_ht_del(consumer_data
.channel_ht
, &iter
);
380 call_rcu(&channel
->node
.head
, free_channel_rcu
);
382 pthread_mutex_unlock(&channel
->lock
);
383 pthread_mutex_unlock(&consumer_data
.lock
);
387 * Iterate over the relayd hash table and destroy each element. Finally,
388 * destroy the whole hash table.
390 static void cleanup_relayd_ht(void)
392 struct lttng_ht_iter iter
;
393 struct consumer_relayd_sock_pair
*relayd
;
397 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
399 consumer_destroy_relayd(relayd
);
404 lttng_ht_destroy(consumer_data
.relayd_ht
);
408 * Update the end point status of all streams having the given network sequence
409 * index (relayd index).
411 * It's atomically set without having the stream mutex locked which is fine
412 * because we handle the write/read race with a pipe wakeup for each thread.
414 static void update_endpoint_status_by_netidx(uint64_t net_seq_idx
,
415 enum consumer_endpoint_status status
)
417 struct lttng_ht_iter iter
;
418 struct lttng_consumer_stream
*stream
;
420 DBG("Consumer set delete flag on stream by idx %" PRIu64
, net_seq_idx
);
424 /* Let's begin with metadata */
425 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
426 if (stream
->net_seq_idx
== net_seq_idx
) {
427 uatomic_set(&stream
->endpoint_status
, status
);
428 DBG("Delete flag set to metadata stream %d", stream
->wait_fd
);
432 /* Follow up by the data streams */
433 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
434 if (stream
->net_seq_idx
== net_seq_idx
) {
435 uatomic_set(&stream
->endpoint_status
, status
);
436 DBG("Delete flag set to data stream %d", stream
->wait_fd
);
443 * Cleanup a relayd object by flagging every associated streams for deletion,
444 * destroying the object meaning removing it from the relayd hash table,
445 * closing the sockets and freeing the memory in a RCU call.
447 * If a local data context is available, notify the threads that the streams'
448 * state have changed.
450 static void cleanup_relayd(struct consumer_relayd_sock_pair
*relayd
,
451 struct lttng_consumer_local_data
*ctx
)
457 DBG("Cleaning up relayd sockets");
459 /* Save the net sequence index before destroying the object */
460 netidx
= relayd
->net_seq_idx
;
463 * Delete the relayd from the relayd hash table, close the sockets and free
464 * the object in a RCU call.
466 consumer_destroy_relayd(relayd
);
468 /* Set inactive endpoint to all streams */
469 update_endpoint_status_by_netidx(netidx
, CONSUMER_ENDPOINT_INACTIVE
);
472 * With a local data context, notify the threads that the streams' state
473 * have changed. The write() action on the pipe acts as an "implicit"
474 * memory barrier ordering the updates of the end point status from the
475 * read of this status which happens AFTER receiving this notify.
478 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
479 notify_thread_lttng_pipe(ctx
->consumer_metadata_pipe
);
484 * Flag a relayd socket pair for destruction. Destroy it if the refcount
487 * RCU read side lock MUST be aquired before calling this function.
489 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair
*relayd
)
493 /* Set destroy flag for this object */
494 uatomic_set(&relayd
->destroy_flag
, 1);
496 /* Destroy the relayd if refcount is 0 */
497 if (uatomic_read(&relayd
->refcount
) == 0) {
498 consumer_destroy_relayd(relayd
);
503 * Completly destroy stream from every visiable data structure and the given
506 * One this call returns, the stream object is not longer usable nor visible.
508 void consumer_del_stream(struct lttng_consumer_stream
*stream
,
511 consumer_stream_destroy(stream
, ht
);
515 * XXX naming of del vs destroy is all mixed up.
517 void consumer_del_stream_for_data(struct lttng_consumer_stream
*stream
)
519 consumer_stream_destroy(stream
, data_ht
);
522 void consumer_del_stream_for_metadata(struct lttng_consumer_stream
*stream
)
524 consumer_stream_destroy(stream
, metadata_ht
);
527 struct lttng_consumer_stream
*consumer_allocate_stream(uint64_t channel_key
,
529 enum lttng_consumer_stream_state state
,
530 const char *channel_name
,
537 enum consumer_channel_type type
,
538 unsigned int monitor
)
541 struct lttng_consumer_stream
*stream
;
543 stream
= zmalloc(sizeof(*stream
));
544 if (stream
== NULL
) {
545 PERROR("malloc struct lttng_consumer_stream");
552 stream
->key
= stream_key
;
554 stream
->out_fd_offset
= 0;
555 stream
->output_written
= 0;
556 stream
->state
= state
;
559 stream
->net_seq_idx
= relayd_id
;
560 stream
->session_id
= session_id
;
561 stream
->monitor
= monitor
;
562 stream
->endpoint_status
= CONSUMER_ENDPOINT_ACTIVE
;
563 stream
->index_fd
= -1;
564 pthread_mutex_init(&stream
->lock
, NULL
);
566 /* If channel is the metadata, flag this stream as metadata. */
567 if (type
== CONSUMER_CHANNEL_TYPE_METADATA
) {
568 stream
->metadata_flag
= 1;
569 /* Metadata is flat out. */
570 strncpy(stream
->name
, DEFAULT_METADATA_NAME
, sizeof(stream
->name
));
571 /* Live rendez-vous point. */
572 pthread_cond_init(&stream
->metadata_rdv
, NULL
);
573 pthread_mutex_init(&stream
->metadata_rdv_lock
, NULL
);
575 /* Format stream name to <channel_name>_<cpu_number> */
576 ret
= snprintf(stream
->name
, sizeof(stream
->name
), "%s_%d",
579 PERROR("snprintf stream name");
584 /* Key is always the wait_fd for streams. */
585 lttng_ht_node_init_u64(&stream
->node
, stream
->key
);
587 /* Init node per channel id key */
588 lttng_ht_node_init_u64(&stream
->node_channel_id
, channel_key
);
590 /* Init session id node with the stream session id */
591 lttng_ht_node_init_u64(&stream
->node_session_id
, stream
->session_id
);
593 DBG3("Allocated stream %s (key %" PRIu64
", chan_key %" PRIu64
594 " relayd_id %" PRIu64
", session_id %" PRIu64
,
595 stream
->name
, stream
->key
, channel_key
,
596 stream
->net_seq_idx
, stream
->session_id
);
612 * Add a stream to the global list protected by a mutex.
614 int consumer_add_data_stream(struct lttng_consumer_stream
*stream
)
616 struct lttng_ht
*ht
= data_ht
;
622 DBG3("Adding consumer stream %" PRIu64
, stream
->key
);
624 pthread_mutex_lock(&consumer_data
.lock
);
625 pthread_mutex_lock(&stream
->chan
->lock
);
626 pthread_mutex_lock(&stream
->chan
->timer_lock
);
627 pthread_mutex_lock(&stream
->lock
);
630 /* Steal stream identifier to avoid having streams with the same key */
631 steal_stream_key(stream
->key
, ht
);
633 lttng_ht_add_unique_u64(ht
, &stream
->node
);
635 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
636 &stream
->node_channel_id
);
639 * Add stream to the stream_list_ht of the consumer data. No need to steal
640 * the key since the HT does not use it and we allow to add redundant keys
643 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
646 * When nb_init_stream_left reaches 0, we don't need to trigger any action
647 * in terms of destroying the associated channel, because the action that
648 * causes the count to become 0 also causes a stream to be added. The
649 * channel deletion will thus be triggered by the following removal of this
652 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
653 /* Increment refcount before decrementing nb_init_stream_left */
655 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
658 /* Update consumer data once the node is inserted. */
659 consumer_data
.stream_count
++;
660 consumer_data
.need_update
= 1;
663 pthread_mutex_unlock(&stream
->lock
);
664 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
665 pthread_mutex_unlock(&stream
->chan
->lock
);
666 pthread_mutex_unlock(&consumer_data
.lock
);
671 void consumer_del_data_stream(struct lttng_consumer_stream
*stream
)
673 consumer_del_stream(stream
, data_ht
);
677 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
678 * be acquired before calling this.
680 static int add_relayd(struct consumer_relayd_sock_pair
*relayd
)
683 struct lttng_ht_node_u64
*node
;
684 struct lttng_ht_iter iter
;
688 lttng_ht_lookup(consumer_data
.relayd_ht
,
689 &relayd
->net_seq_idx
, &iter
);
690 node
= lttng_ht_iter_get_node_u64(&iter
);
694 lttng_ht_add_unique_u64(consumer_data
.relayd_ht
, &relayd
->node
);
701 * Allocate and return a consumer relayd socket.
703 struct consumer_relayd_sock_pair
*consumer_allocate_relayd_sock_pair(
704 uint64_t net_seq_idx
)
706 struct consumer_relayd_sock_pair
*obj
= NULL
;
708 /* net sequence index of -1 is a failure */
709 if (net_seq_idx
== (uint64_t) -1ULL) {
713 obj
= zmalloc(sizeof(struct consumer_relayd_sock_pair
));
715 PERROR("zmalloc relayd sock");
719 obj
->net_seq_idx
= net_seq_idx
;
721 obj
->destroy_flag
= 0;
722 obj
->control_sock
.sock
.fd
= -1;
723 obj
->data_sock
.sock
.fd
= -1;
724 lttng_ht_node_init_u64(&obj
->node
, obj
->net_seq_idx
);
725 pthread_mutex_init(&obj
->ctrl_sock_mutex
, NULL
);
732 * Find a relayd socket pair in the global consumer data.
734 * Return the object if found else NULL.
735 * RCU read-side lock must be held across this call and while using the
738 struct consumer_relayd_sock_pair
*consumer_find_relayd(uint64_t key
)
740 struct lttng_ht_iter iter
;
741 struct lttng_ht_node_u64
*node
;
742 struct consumer_relayd_sock_pair
*relayd
= NULL
;
744 /* Negative keys are lookup failures */
745 if (key
== (uint64_t) -1ULL) {
749 lttng_ht_lookup(consumer_data
.relayd_ht
, &key
,
751 node
= lttng_ht_iter_get_node_u64(&iter
);
753 relayd
= caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
761 * Find a relayd and send the stream
763 * Returns 0 on success, < 0 on error
765 int consumer_send_relayd_stream(struct lttng_consumer_stream
*stream
,
769 struct consumer_relayd_sock_pair
*relayd
;
772 assert(stream
->net_seq_idx
!= -1ULL);
775 /* The stream is not metadata. Get relayd reference if exists. */
777 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
778 if (relayd
!= NULL
) {
779 /* Add stream on the relayd */
780 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
781 ret
= relayd_add_stream(&relayd
->control_sock
, stream
->name
,
782 path
, &stream
->relayd_stream_id
,
783 stream
->chan
->tracefile_size
, stream
->chan
->tracefile_count
);
784 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
789 uatomic_inc(&relayd
->refcount
);
790 stream
->sent_to_relayd
= 1;
792 ERR("Stream %" PRIu64
" relayd ID %" PRIu64
" unknown. Can't send it.",
793 stream
->key
, stream
->net_seq_idx
);
798 DBG("Stream %s with key %" PRIu64
" sent to relayd id %" PRIu64
,
799 stream
->name
, stream
->key
, stream
->net_seq_idx
);
807 * Find a relayd and send the streams sent message
809 * Returns 0 on success, < 0 on error
811 int consumer_send_relayd_streams_sent(uint64_t net_seq_idx
)
814 struct consumer_relayd_sock_pair
*relayd
;
816 assert(net_seq_idx
!= -1ULL);
818 /* The stream is not metadata. Get relayd reference if exists. */
820 relayd
= consumer_find_relayd(net_seq_idx
);
821 if (relayd
!= NULL
) {
822 /* Add stream on the relayd */
823 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
824 ret
= relayd_streams_sent(&relayd
->control_sock
);
825 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
830 ERR("Relayd ID %" PRIu64
" unknown. Can't send streams_sent.",
837 DBG("All streams sent relayd id %" PRIu64
, net_seq_idx
);
845 * Find a relayd and close the stream
847 void close_relayd_stream(struct lttng_consumer_stream
*stream
)
849 struct consumer_relayd_sock_pair
*relayd
;
851 /* The stream is not metadata. Get relayd reference if exists. */
853 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
855 consumer_stream_relayd_close(stream
, relayd
);
861 * Handle stream for relayd transmission if the stream applies for network
862 * streaming where the net sequence index is set.
864 * Return destination file descriptor or negative value on error.
866 static int write_relayd_stream_header(struct lttng_consumer_stream
*stream
,
867 size_t data_size
, unsigned long padding
,
868 struct consumer_relayd_sock_pair
*relayd
)
871 struct lttcomm_relayd_data_hdr data_hdr
;
877 /* Reset data header */
878 memset(&data_hdr
, 0, sizeof(data_hdr
));
880 if (stream
->metadata_flag
) {
881 /* Caller MUST acquire the relayd control socket lock */
882 ret
= relayd_send_metadata(&relayd
->control_sock
, data_size
);
887 /* Metadata are always sent on the control socket. */
888 outfd
= relayd
->control_sock
.sock
.fd
;
890 /* Set header with stream information */
891 data_hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
892 data_hdr
.data_size
= htobe32(data_size
);
893 data_hdr
.padding_size
= htobe32(padding
);
895 * Note that net_seq_num below is assigned with the *current* value of
896 * next_net_seq_num and only after that the next_net_seq_num will be
897 * increment. This is why when issuing a command on the relayd using
898 * this next value, 1 should always be substracted in order to compare
899 * the last seen sequence number on the relayd side to the last sent.
901 data_hdr
.net_seq_num
= htobe64(stream
->next_net_seq_num
);
902 /* Other fields are zeroed previously */
904 ret
= relayd_send_data_hdr(&relayd
->data_sock
, &data_hdr
,
910 ++stream
->next_net_seq_num
;
912 /* Set to go on data socket */
913 outfd
= relayd
->data_sock
.sock
.fd
;
921 * Allocate and return a new lttng_consumer_channel object using the given key
922 * to initialize the hash table node.
924 * On error, return NULL.
926 struct lttng_consumer_channel
*consumer_allocate_channel(uint64_t key
,
928 const char *pathname
,
933 enum lttng_event_output output
,
934 uint64_t tracefile_size
,
935 uint64_t tracefile_count
,
936 uint64_t session_id_per_pid
,
937 unsigned int monitor
,
938 unsigned int live_timer_interval
)
940 struct lttng_consumer_channel
*channel
;
942 channel
= zmalloc(sizeof(*channel
));
943 if (channel
== NULL
) {
944 PERROR("malloc struct lttng_consumer_channel");
949 channel
->refcount
= 0;
950 channel
->session_id
= session_id
;
951 channel
->session_id_per_pid
= session_id_per_pid
;
954 channel
->relayd_id
= relayd_id
;
955 channel
->tracefile_size
= tracefile_size
;
956 channel
->tracefile_count
= tracefile_count
;
957 channel
->monitor
= monitor
;
958 channel
->live_timer_interval
= live_timer_interval
;
959 pthread_mutex_init(&channel
->lock
, NULL
);
960 pthread_mutex_init(&channel
->timer_lock
, NULL
);
963 case LTTNG_EVENT_SPLICE
:
964 channel
->output
= CONSUMER_CHANNEL_SPLICE
;
966 case LTTNG_EVENT_MMAP
:
967 channel
->output
= CONSUMER_CHANNEL_MMAP
;
977 * In monitor mode, the streams associated with the channel will be put in
978 * a special list ONLY owned by this channel. So, the refcount is set to 1
979 * here meaning that the channel itself has streams that are referenced.
981 * On a channel deletion, once the channel is no longer visible, the
982 * refcount is decremented and checked for a zero value to delete it. With
983 * streams in no monitor mode, it will now be safe to destroy the channel.
985 if (!channel
->monitor
) {
986 channel
->refcount
= 1;
989 strncpy(channel
->pathname
, pathname
, sizeof(channel
->pathname
));
990 channel
->pathname
[sizeof(channel
->pathname
) - 1] = '\0';
992 strncpy(channel
->name
, name
, sizeof(channel
->name
));
993 channel
->name
[sizeof(channel
->name
) - 1] = '\0';
995 lttng_ht_node_init_u64(&channel
->node
, channel
->key
);
997 channel
->wait_fd
= -1;
999 CDS_INIT_LIST_HEAD(&channel
->streams
.head
);
1001 DBG("Allocated channel (key %" PRIu64
")", channel
->key
)
1008 * Add a channel to the global list protected by a mutex.
1010 * Always return 0 indicating success.
1012 int consumer_add_channel(struct lttng_consumer_channel
*channel
,
1013 struct lttng_consumer_local_data
*ctx
)
1015 pthread_mutex_lock(&consumer_data
.lock
);
1016 pthread_mutex_lock(&channel
->lock
);
1017 pthread_mutex_lock(&channel
->timer_lock
);
1020 * This gives us a guarantee that the channel we are about to add to the
1021 * channel hash table will be unique. See this function comment on the why
1022 * we need to steel the channel key at this stage.
1024 steal_channel_key(channel
->key
);
1027 lttng_ht_add_unique_u64(consumer_data
.channel_ht
, &channel
->node
);
1030 pthread_mutex_unlock(&channel
->timer_lock
);
1031 pthread_mutex_unlock(&channel
->lock
);
1032 pthread_mutex_unlock(&consumer_data
.lock
);
1034 if (channel
->wait_fd
!= -1 && channel
->type
== CONSUMER_CHANNEL_TYPE_DATA
) {
1035 notify_channel_pipe(ctx
, channel
, -1, CONSUMER_CHANNEL_ADD
);
1042 * Allocate the pollfd structure and the local view of the out fds to avoid
1043 * doing a lookup in the linked list and concurrency issues when writing is
1044 * needed. Called with consumer_data.lock held.
1046 * Returns the number of fds in the structures.
1048 static int update_poll_array(struct lttng_consumer_local_data
*ctx
,
1049 struct pollfd
**pollfd
, struct lttng_consumer_stream
**local_stream
,
1050 struct lttng_ht
*ht
)
1053 struct lttng_ht_iter iter
;
1054 struct lttng_consumer_stream
*stream
;
1059 assert(local_stream
);
1061 DBG("Updating poll fd array");
1063 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1065 * Only active streams with an active end point can be added to the
1066 * poll set and local stream storage of the thread.
1068 * There is a potential race here for endpoint_status to be updated
1069 * just after the check. However, this is OK since the stream(s) will
1070 * be deleted once the thread is notified that the end point state has
1071 * changed where this function will be called back again.
1073 if (stream
->state
!= LTTNG_CONSUMER_ACTIVE_STREAM
||
1074 stream
->endpoint_status
== CONSUMER_ENDPOINT_INACTIVE
) {
1078 * This clobbers way too much the debug output. Uncomment that if you
1079 * need it for debugging purposes.
1081 * DBG("Active FD %d", stream->wait_fd);
1083 (*pollfd
)[i
].fd
= stream
->wait_fd
;
1084 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1085 local_stream
[i
] = stream
;
1091 * Insert the consumer_data_pipe at the end of the array and don't
1092 * increment i so nb_fd is the number of real FD.
1094 (*pollfd
)[i
].fd
= lttng_pipe_get_readfd(ctx
->consumer_data_pipe
);
1095 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1097 (*pollfd
)[i
+ 1].fd
= lttng_pipe_get_readfd(ctx
->consumer_wakeup_pipe
);
1098 (*pollfd
)[i
+ 1].events
= POLLIN
| POLLPRI
;
1103 * Poll on the should_quit pipe and the command socket return -1 on
1104 * error, 1 if should exit, 0 if data is available on the command socket
1106 int lttng_consumer_poll_socket(struct pollfd
*consumer_sockpoll
)
1111 num_rdy
= poll(consumer_sockpoll
, 2, -1);
1112 if (num_rdy
== -1) {
1114 * Restart interrupted system call.
1116 if (errno
== EINTR
) {
1119 PERROR("Poll error");
1122 if (consumer_sockpoll
[0].revents
& (POLLIN
| POLLPRI
)) {
1123 DBG("consumer_should_quit wake up");
1130 * Set the error socket.
1132 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data
*ctx
,
1135 ctx
->consumer_error_socket
= sock
;
1139 * Set the command socket path.
1141 void lttng_consumer_set_command_sock_path(
1142 struct lttng_consumer_local_data
*ctx
, char *sock
)
1144 ctx
->consumer_command_sock_path
= sock
;
1148 * Send return code to the session daemon.
1149 * If the socket is not defined, we return 0, it is not a fatal error
1151 int lttng_consumer_send_error(struct lttng_consumer_local_data
*ctx
, int cmd
)
1153 if (ctx
->consumer_error_socket
> 0) {
1154 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
1155 sizeof(enum lttcomm_sessiond_command
));
1162 * Close all the tracefiles and stream fds and MUST be called when all
1163 * instances are destroyed i.e. when all threads were joined and are ended.
1165 void lttng_consumer_cleanup(void)
1167 struct lttng_ht_iter iter
;
1168 struct lttng_consumer_channel
*channel
;
1172 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
, channel
,
1174 consumer_del_channel(channel
);
1179 lttng_ht_destroy(consumer_data
.channel_ht
);
1181 cleanup_relayd_ht();
1183 lttng_ht_destroy(consumer_data
.stream_per_chan_id_ht
);
1186 * This HT contains streams that are freed by either the metadata thread or
1187 * the data thread so we do *nothing* on the hash table and simply destroy
1190 lttng_ht_destroy(consumer_data
.stream_list_ht
);
1194 * Called from signal handler.
1196 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
1201 ret
= lttng_write(ctx
->consumer_should_quit
[1], "4", 1);
1203 PERROR("write consumer quit");
1206 DBG("Consumer flag that it should quit");
1209 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1212 int outfd
= stream
->out_fd
;
1215 * This does a blocking write-and-wait on any page that belongs to the
1216 * subbuffer prior to the one we just wrote.
1217 * Don't care about error values, as these are just hints and ways to
1218 * limit the amount of page cache used.
1220 if (orig_offset
< stream
->max_sb_size
) {
1223 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1224 stream
->max_sb_size
,
1225 SYNC_FILE_RANGE_WAIT_BEFORE
1226 | SYNC_FILE_RANGE_WRITE
1227 | SYNC_FILE_RANGE_WAIT_AFTER
);
1229 * Give hints to the kernel about how we access the file:
1230 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1233 * We need to call fadvise again after the file grows because the
1234 * kernel does not seem to apply fadvise to non-existing parts of the
1237 * Call fadvise _after_ having waited for the page writeback to
1238 * complete because the dirty page writeback semantic is not well
1239 * defined. So it can be expected to lead to lower throughput in
1242 posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1243 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1247 * Initialise the necessary environnement :
1248 * - create a new context
1249 * - create the poll_pipe
1250 * - create the should_quit pipe (for signal handler)
1251 * - create the thread pipe (for splice)
1253 * Takes a function pointer as argument, this function is called when data is
1254 * available on a buffer. This function is responsible to do the
1255 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1256 * buffer configuration and then kernctl_put_next_subbuf at the end.
1258 * Returns a pointer to the new context or NULL on error.
1260 struct lttng_consumer_local_data
*lttng_consumer_create(
1261 enum lttng_consumer_type type
,
1262 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1263 struct lttng_consumer_local_data
*ctx
),
1264 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1265 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1266 int (*update_stream
)(uint64_t stream_key
, uint32_t state
))
1269 struct lttng_consumer_local_data
*ctx
;
1271 assert(consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1272 consumer_data
.type
== type
);
1273 consumer_data
.type
= type
;
1275 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
1277 PERROR("allocating context");
1281 ctx
->consumer_error_socket
= -1;
1282 ctx
->consumer_metadata_socket
= -1;
1283 pthread_mutex_init(&ctx
->metadata_socket_lock
, NULL
);
1284 /* assign the callbacks */
1285 ctx
->on_buffer_ready
= buffer_ready
;
1286 ctx
->on_recv_channel
= recv_channel
;
1287 ctx
->on_recv_stream
= recv_stream
;
1288 ctx
->on_update_stream
= update_stream
;
1290 ctx
->consumer_data_pipe
= lttng_pipe_open(0);
1291 if (!ctx
->consumer_data_pipe
) {
1292 goto error_poll_pipe
;
1295 ctx
->consumer_wakeup_pipe
= lttng_pipe_open(0);
1296 if (!ctx
->consumer_wakeup_pipe
) {
1297 goto error_wakeup_pipe
;
1300 ret
= pipe(ctx
->consumer_should_quit
);
1302 PERROR("Error creating recv pipe");
1303 goto error_quit_pipe
;
1306 ret
= pipe(ctx
->consumer_thread_pipe
);
1308 PERROR("Error creating thread pipe");
1309 goto error_thread_pipe
;
1312 ret
= pipe(ctx
->consumer_channel_pipe
);
1314 PERROR("Error creating channel pipe");
1315 goto error_channel_pipe
;
1318 ctx
->consumer_metadata_pipe
= lttng_pipe_open(0);
1319 if (!ctx
->consumer_metadata_pipe
) {
1320 goto error_metadata_pipe
;
1323 ret
= utils_create_pipe(ctx
->consumer_splice_metadata_pipe
);
1325 goto error_splice_pipe
;
1331 lttng_pipe_destroy(ctx
->consumer_metadata_pipe
);
1332 error_metadata_pipe
:
1333 utils_close_pipe(ctx
->consumer_channel_pipe
);
1335 utils_close_pipe(ctx
->consumer_thread_pipe
);
1337 utils_close_pipe(ctx
->consumer_should_quit
);
1339 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1341 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1349 * Iterate over all streams of the hashtable and free them properly.
1351 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1353 struct lttng_ht_iter iter
;
1354 struct lttng_consumer_stream
*stream
;
1361 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1363 * Ignore return value since we are currently cleaning up so any error
1366 (void) consumer_del_stream(stream
, ht
);
1370 lttng_ht_destroy(ht
);
1374 * Iterate over all streams of the metadata hashtable and free them
1377 static void destroy_metadata_stream_ht(struct lttng_ht
*ht
)
1379 struct lttng_ht_iter iter
;
1380 struct lttng_consumer_stream
*stream
;
1387 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1389 * Ignore return value since we are currently cleaning up so any error
1392 (void) consumer_del_metadata_stream(stream
, ht
);
1396 lttng_ht_destroy(ht
);
1400 * Close all fds associated with the instance and free the context.
1402 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1406 DBG("Consumer destroying it. Closing everything.");
1412 destroy_data_stream_ht(data_ht
);
1413 destroy_metadata_stream_ht(metadata_ht
);
1415 ret
= close(ctx
->consumer_error_socket
);
1419 ret
= close(ctx
->consumer_metadata_socket
);
1423 utils_close_pipe(ctx
->consumer_thread_pipe
);
1424 utils_close_pipe(ctx
->consumer_channel_pipe
);
1425 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1426 lttng_pipe_destroy(ctx
->consumer_metadata_pipe
);
1427 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1428 utils_close_pipe(ctx
->consumer_should_quit
);
1429 utils_close_pipe(ctx
->consumer_splice_metadata_pipe
);
1431 unlink(ctx
->consumer_command_sock_path
);
1436 * Write the metadata stream id on the specified file descriptor.
1438 static int write_relayd_metadata_id(int fd
,
1439 struct lttng_consumer_stream
*stream
,
1440 struct consumer_relayd_sock_pair
*relayd
, unsigned long padding
)
1443 struct lttcomm_relayd_metadata_payload hdr
;
1445 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1446 hdr
.padding_size
= htobe32(padding
);
1447 ret
= lttng_write(fd
, (void *) &hdr
, sizeof(hdr
));
1448 if (ret
< sizeof(hdr
)) {
1450 * This error means that the fd's end is closed so ignore the perror
1451 * not to clubber the error output since this can happen in a normal
1454 if (errno
!= EPIPE
) {
1455 PERROR("write metadata stream id");
1457 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1459 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1460 * handle writting the missing part so report that as an error and
1461 * don't lie to the caller.
1466 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1467 stream
->relayd_stream_id
, padding
);
1474 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1475 * core function for writing trace buffers to either the local filesystem or
1478 * It must be called with the stream lock held.
1480 * Careful review MUST be put if any changes occur!
1482 * Returns the number of bytes written
1484 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1485 struct lttng_consumer_local_data
*ctx
,
1486 struct lttng_consumer_stream
*stream
, unsigned long len
,
1487 unsigned long padding
,
1488 struct ctf_packet_index
*index
)
1490 unsigned long mmap_offset
;
1493 off_t orig_offset
= stream
->out_fd_offset
;
1494 /* Default is on the disk */
1495 int outfd
= stream
->out_fd
;
1496 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1497 unsigned int relayd_hang_up
= 0;
1499 /* RCU lock for the relayd pointer */
1502 /* Flag that the current stream if set for network streaming. */
1503 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1504 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1505 if (relayd
== NULL
) {
1511 /* get the offset inside the fd to mmap */
1512 switch (consumer_data
.type
) {
1513 case LTTNG_CONSUMER_KERNEL
:
1514 mmap_base
= stream
->mmap_base
;
1515 ret
= kernctl_get_mmap_read_offset(stream
->wait_fd
, &mmap_offset
);
1518 PERROR("tracer ctl get_mmap_read_offset");
1522 case LTTNG_CONSUMER32_UST
:
1523 case LTTNG_CONSUMER64_UST
:
1524 mmap_base
= lttng_ustctl_get_mmap_base(stream
);
1526 ERR("read mmap get mmap base for stream %s", stream
->name
);
1530 ret
= lttng_ustctl_get_mmap_read_offset(stream
, &mmap_offset
);
1532 PERROR("tracer ctl get_mmap_read_offset");
1538 ERR("Unknown consumer_data type");
1542 /* Handle stream on the relayd if the output is on the network */
1544 unsigned long netlen
= len
;
1547 * Lock the control socket for the complete duration of the function
1548 * since from this point on we will use the socket.
1550 if (stream
->metadata_flag
) {
1551 /* Metadata requires the control socket. */
1552 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1553 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1556 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1561 /* Use the returned socket. */
1564 /* Write metadata stream id before payload */
1565 if (stream
->metadata_flag
) {
1566 ret
= write_relayd_metadata_id(outfd
, stream
, relayd
, padding
);
1573 /* No streaming, we have to set the len with the full padding */
1577 * Check if we need to change the tracefile before writing the packet.
1579 if (stream
->chan
->tracefile_size
> 0 &&
1580 (stream
->tracefile_size_current
+ len
) >
1581 stream
->chan
->tracefile_size
) {
1582 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1583 stream
->name
, stream
->chan
->tracefile_size
,
1584 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1585 stream
->out_fd
, &(stream
->tracefile_count_current
),
1588 ERR("Rotating output file");
1591 outfd
= stream
->out_fd
;
1593 if (stream
->index_fd
>= 0) {
1594 ret
= index_create_file(stream
->chan
->pathname
,
1595 stream
->name
, stream
->uid
, stream
->gid
,
1596 stream
->chan
->tracefile_size
,
1597 stream
->tracefile_count_current
);
1601 stream
->index_fd
= ret
;
1604 /* Reset current size because we just perform a rotation. */
1605 stream
->tracefile_size_current
= 0;
1606 stream
->out_fd_offset
= 0;
1609 stream
->tracefile_size_current
+= len
;
1611 index
->offset
= htobe64(stream
->out_fd_offset
);
1616 * This call guarantee that len or less is returned. It's impossible to
1617 * receive a ret value that is bigger than len.
1619 ret
= lttng_write(outfd
, mmap_base
+ mmap_offset
, len
);
1620 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, len
);
1621 if (ret
< 0 || ((size_t) ret
!= len
)) {
1623 * Report error to caller if nothing was written else at least send the
1631 /* Socket operation failed. We consider the relayd dead */
1632 if (errno
== EPIPE
|| errno
== EINVAL
|| errno
== EBADF
) {
1634 * This is possible if the fd is closed on the other side
1635 * (outfd) or any write problem. It can be verbose a bit for a
1636 * normal execution if for instance the relayd is stopped
1637 * abruptly. This can happen so set this to a DBG statement.
1639 DBG("Consumer mmap write detected relayd hang up");
1641 /* Unhandled error, print it and stop function right now. */
1642 PERROR("Error in write mmap (ret %zd != len %lu)", ret
, len
);
1646 stream
->output_written
+= ret
;
1648 /* This call is useless on a socket so better save a syscall. */
1650 /* This won't block, but will start writeout asynchronously */
1651 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, len
,
1652 SYNC_FILE_RANGE_WRITE
);
1653 stream
->out_fd_offset
+= len
;
1655 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1659 * This is a special case that the relayd has closed its socket. Let's
1660 * cleanup the relayd object and all associated streams.
1662 if (relayd
&& relayd_hang_up
) {
1663 cleanup_relayd(relayd
, ctx
);
1667 /* Unlock only if ctrl socket used */
1668 if (relayd
&& stream
->metadata_flag
) {
1669 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1677 * Splice the data from the ring buffer to the tracefile.
1679 * It must be called with the stream lock held.
1681 * Returns the number of bytes spliced.
1683 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1684 struct lttng_consumer_local_data
*ctx
,
1685 struct lttng_consumer_stream
*stream
, unsigned long len
,
1686 unsigned long padding
,
1687 struct ctf_packet_index
*index
)
1689 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1691 off_t orig_offset
= stream
->out_fd_offset
;
1692 int fd
= stream
->wait_fd
;
1693 /* Default is on the disk */
1694 int outfd
= stream
->out_fd
;
1695 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1697 unsigned int relayd_hang_up
= 0;
1699 switch (consumer_data
.type
) {
1700 case LTTNG_CONSUMER_KERNEL
:
1702 case LTTNG_CONSUMER32_UST
:
1703 case LTTNG_CONSUMER64_UST
:
1704 /* Not supported for user space tracing */
1707 ERR("Unknown consumer_data type");
1711 /* RCU lock for the relayd pointer */
1714 /* Flag that the current stream if set for network streaming. */
1715 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1716 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1717 if (relayd
== NULL
) {
1724 * Choose right pipe for splice. Metadata and trace data are handled by
1725 * different threads hence the use of two pipes in order not to race or
1726 * corrupt the written data.
1728 if (stream
->metadata_flag
) {
1729 splice_pipe
= ctx
->consumer_splice_metadata_pipe
;
1731 splice_pipe
= ctx
->consumer_thread_pipe
;
1734 /* Write metadata stream id before payload */
1736 unsigned long total_len
= len
;
1738 if (stream
->metadata_flag
) {
1740 * Lock the control socket for the complete duration of the function
1741 * since from this point on we will use the socket.
1743 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1745 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
, relayd
,
1753 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1756 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1762 /* Use the returned socket. */
1765 /* No streaming, we have to set the len with the full padding */
1769 * Check if we need to change the tracefile before writing the packet.
1771 if (stream
->chan
->tracefile_size
> 0 &&
1772 (stream
->tracefile_size_current
+ len
) >
1773 stream
->chan
->tracefile_size
) {
1774 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1775 stream
->name
, stream
->chan
->tracefile_size
,
1776 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1777 stream
->out_fd
, &(stream
->tracefile_count_current
),
1781 ERR("Rotating output file");
1784 outfd
= stream
->out_fd
;
1786 if (stream
->index_fd
>= 0) {
1787 ret
= index_create_file(stream
->chan
->pathname
,
1788 stream
->name
, stream
->uid
, stream
->gid
,
1789 stream
->chan
->tracefile_size
,
1790 stream
->tracefile_count_current
);
1795 stream
->index_fd
= ret
;
1798 /* Reset current size because we just perform a rotation. */
1799 stream
->tracefile_size_current
= 0;
1800 stream
->out_fd_offset
= 0;
1803 stream
->tracefile_size_current
+= len
;
1804 index
->offset
= htobe64(stream
->out_fd_offset
);
1808 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1809 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1810 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1811 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1812 DBG("splice chan to pipe, ret %zd", ret_splice
);
1813 if (ret_splice
< 0) {
1816 PERROR("Error in relay splice");
1820 /* Handle stream on the relayd if the output is on the network */
1821 if (relayd
&& stream
->metadata_flag
) {
1822 size_t metadata_payload_size
=
1823 sizeof(struct lttcomm_relayd_metadata_payload
);
1825 /* Update counter to fit the spliced data */
1826 ret_splice
+= metadata_payload_size
;
1827 len
+= metadata_payload_size
;
1829 * We do this so the return value can match the len passed as
1830 * argument to this function.
1832 written
-= metadata_payload_size
;
1835 /* Splice data out */
1836 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1837 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1838 DBG("Consumer splice pipe to file, ret %zd", ret_splice
);
1839 if (ret_splice
< 0) {
1844 } else if (ret_splice
> len
) {
1846 * We don't expect this code path to be executed but you never know
1847 * so this is an extra protection agains a buggy splice().
1850 written
+= ret_splice
;
1851 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice
,
1855 /* All good, update current len and continue. */
1859 /* This call is useless on a socket so better save a syscall. */
1861 /* This won't block, but will start writeout asynchronously */
1862 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1863 SYNC_FILE_RANGE_WRITE
);
1864 stream
->out_fd_offset
+= ret_splice
;
1866 stream
->output_written
+= ret_splice
;
1867 written
+= ret_splice
;
1869 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1874 * This is a special case that the relayd has closed its socket. Let's
1875 * cleanup the relayd object and all associated streams.
1877 if (relayd
&& relayd_hang_up
) {
1878 cleanup_relayd(relayd
, ctx
);
1879 /* Skip splice error so the consumer does not fail */
1884 /* send the appropriate error description to sessiond */
1887 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1890 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1893 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
1898 if (relayd
&& stream
->metadata_flag
) {
1899 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1907 * Take a snapshot for a specific fd
1909 * Returns 0 on success, < 0 on error
1911 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
1913 switch (consumer_data
.type
) {
1914 case LTTNG_CONSUMER_KERNEL
:
1915 return lttng_kconsumer_take_snapshot(stream
);
1916 case LTTNG_CONSUMER32_UST
:
1917 case LTTNG_CONSUMER64_UST
:
1918 return lttng_ustconsumer_take_snapshot(stream
);
1920 ERR("Unknown consumer_data type");
1927 * Get the produced position
1929 * Returns 0 on success, < 0 on error
1931 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
1934 switch (consumer_data
.type
) {
1935 case LTTNG_CONSUMER_KERNEL
:
1936 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
1937 case LTTNG_CONSUMER32_UST
:
1938 case LTTNG_CONSUMER64_UST
:
1939 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
1941 ERR("Unknown consumer_data type");
1947 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
1948 int sock
, struct pollfd
*consumer_sockpoll
)
1950 switch (consumer_data
.type
) {
1951 case LTTNG_CONSUMER_KERNEL
:
1952 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1953 case LTTNG_CONSUMER32_UST
:
1954 case LTTNG_CONSUMER64_UST
:
1955 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1957 ERR("Unknown consumer_data type");
1963 void lttng_consumer_close_all_metadata(void)
1965 switch (consumer_data
.type
) {
1966 case LTTNG_CONSUMER_KERNEL
:
1968 * The Kernel consumer has a different metadata scheme so we don't
1969 * close anything because the stream will be closed by the session
1973 case LTTNG_CONSUMER32_UST
:
1974 case LTTNG_CONSUMER64_UST
:
1976 * Close all metadata streams. The metadata hash table is passed and
1977 * this call iterates over it by closing all wakeup fd. This is safe
1978 * because at this point we are sure that the metadata producer is
1979 * either dead or blocked.
1981 lttng_ustconsumer_close_all_metadata(metadata_ht
);
1984 ERR("Unknown consumer_data type");
1990 * Clean up a metadata stream and free its memory.
1992 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
1993 struct lttng_ht
*ht
)
1995 struct lttng_consumer_channel
*free_chan
= NULL
;
1999 * This call should NEVER receive regular stream. It must always be
2000 * metadata stream and this is crucial for data structure synchronization.
2002 assert(stream
->metadata_flag
);
2004 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
2006 pthread_mutex_lock(&consumer_data
.lock
);
2007 pthread_mutex_lock(&stream
->chan
->lock
);
2008 pthread_mutex_lock(&stream
->lock
);
2010 /* Remove any reference to that stream. */
2011 consumer_stream_delete(stream
, ht
);
2013 /* Close down everything including the relayd if one. */
2014 consumer_stream_close(stream
);
2015 /* Destroy tracer buffers of the stream. */
2016 consumer_stream_destroy_buffers(stream
);
2018 /* Atomically decrement channel refcount since other threads can use it. */
2019 if (!uatomic_sub_return(&stream
->chan
->refcount
, 1)
2020 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
2021 /* Go for channel deletion! */
2022 free_chan
= stream
->chan
;
2026 * Nullify the stream reference so it is not used after deletion. The
2027 * channel lock MUST be acquired before being able to check for a NULL
2030 stream
->chan
->metadata_stream
= NULL
;
2032 pthread_mutex_unlock(&stream
->lock
);
2033 pthread_mutex_unlock(&stream
->chan
->lock
);
2034 pthread_mutex_unlock(&consumer_data
.lock
);
2037 consumer_del_channel(free_chan
);
2040 consumer_stream_free(stream
);
2044 * Action done with the metadata stream when adding it to the consumer internal
2045 * data structures to handle it.
2047 int consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
2049 struct lttng_ht
*ht
= metadata_ht
;
2051 struct lttng_ht_iter iter
;
2052 struct lttng_ht_node_u64
*node
;
2057 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2059 pthread_mutex_lock(&consumer_data
.lock
);
2060 pthread_mutex_lock(&stream
->chan
->lock
);
2061 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2062 pthread_mutex_lock(&stream
->lock
);
2065 * From here, refcounts are updated so be _careful_ when returning an error
2072 * Lookup the stream just to make sure it does not exist in our internal
2073 * state. This should NEVER happen.
2075 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2076 node
= lttng_ht_iter_get_node_u64(&iter
);
2080 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2081 * in terms of destroying the associated channel, because the action that
2082 * causes the count to become 0 also causes a stream to be added. The
2083 * channel deletion will thus be triggered by the following removal of this
2086 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2087 /* Increment refcount before decrementing nb_init_stream_left */
2089 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2092 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2094 lttng_ht_add_unique_u64(consumer_data
.stream_per_chan_id_ht
,
2095 &stream
->node_channel_id
);
2098 * Add stream to the stream_list_ht of the consumer data. No need to steal
2099 * the key since the HT does not use it and we allow to add redundant keys
2102 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2106 pthread_mutex_unlock(&stream
->lock
);
2107 pthread_mutex_unlock(&stream
->chan
->lock
);
2108 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2109 pthread_mutex_unlock(&consumer_data
.lock
);
2114 * Delete data stream that are flagged for deletion (endpoint_status).
2116 static void validate_endpoint_status_data_stream(void)
2118 struct lttng_ht_iter iter
;
2119 struct lttng_consumer_stream
*stream
;
2121 DBG("Consumer delete flagged data stream");
2124 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2125 /* Validate delete flag of the stream */
2126 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2129 /* Delete it right now */
2130 consumer_del_stream(stream
, data_ht
);
2136 * Delete metadata stream that are flagged for deletion (endpoint_status).
2138 static void validate_endpoint_status_metadata_stream(
2139 struct lttng_poll_event
*pollset
)
2141 struct lttng_ht_iter iter
;
2142 struct lttng_consumer_stream
*stream
;
2144 DBG("Consumer delete flagged metadata stream");
2149 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2150 /* Validate delete flag of the stream */
2151 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2155 * Remove from pollset so the metadata thread can continue without
2156 * blocking on a deleted stream.
2158 lttng_poll_del(pollset
, stream
->wait_fd
);
2160 /* Delete it right now */
2161 consumer_del_metadata_stream(stream
, metadata_ht
);
2167 * Thread polls on metadata file descriptor and write them on disk or on the
2170 void *consumer_thread_metadata_poll(void *data
)
2172 int ret
, i
, pollfd
, err
= -1;
2173 uint32_t revents
, nb_fd
;
2174 struct lttng_consumer_stream
*stream
= NULL
;
2175 struct lttng_ht_iter iter
;
2176 struct lttng_ht_node_u64
*node
;
2177 struct lttng_poll_event events
;
2178 struct lttng_consumer_local_data
*ctx
= data
;
2181 rcu_register_thread();
2183 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2185 if (testpoint(consumerd_thread_metadata
)) {
2186 goto error_testpoint
;
2189 health_code_update();
2191 DBG("Thread metadata poll started");
2193 /* Size is set to 1 for the consumer_metadata pipe */
2194 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2196 ERR("Poll set creation failed");
2200 ret
= lttng_poll_add(&events
,
2201 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2207 DBG("Metadata main loop started");
2210 health_code_update();
2212 /* Only the metadata pipe is set */
2213 if (LTTNG_POLL_GETNB(&events
) == 0 && consumer_quit
== 1) {
2214 err
= 0; /* All is OK */
2219 DBG("Metadata poll wait with %d fd(s)", LTTNG_POLL_GETNB(&events
));
2220 health_poll_entry();
2221 ret
= lttng_poll_wait(&events
, -1);
2223 DBG("Metadata event catched in thread");
2225 if (errno
== EINTR
) {
2226 ERR("Poll EINTR catched");
2234 /* From here, the event is a metadata wait fd */
2235 for (i
= 0; i
< nb_fd
; i
++) {
2236 health_code_update();
2238 revents
= LTTNG_POLL_GETEV(&events
, i
);
2239 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2241 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2242 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2243 DBG("Metadata thread pipe hung up");
2245 * Remove the pipe from the poll set and continue the loop
2246 * since their might be data to consume.
2248 lttng_poll_del(&events
,
2249 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2250 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2252 } else if (revents
& LPOLLIN
) {
2255 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2256 &stream
, sizeof(stream
));
2257 if (pipe_len
< sizeof(stream
)) {
2258 PERROR("read metadata stream");
2260 * Continue here to handle the rest of the streams.
2265 /* A NULL stream means that the state has changed. */
2266 if (stream
== NULL
) {
2267 /* Check for deleted streams. */
2268 validate_endpoint_status_metadata_stream(&events
);
2272 DBG("Adding metadata stream %d to poll set",
2275 /* Add metadata stream to the global poll events list */
2276 lttng_poll_add(&events
, stream
->wait_fd
,
2277 LPOLLIN
| LPOLLPRI
| LPOLLHUP
);
2280 /* Handle other stream */
2286 uint64_t tmp_id
= (uint64_t) pollfd
;
2288 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2290 node
= lttng_ht_iter_get_node_u64(&iter
);
2293 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2296 /* Check for error event */
2297 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2298 DBG("Metadata fd %d is hup|err.", pollfd
);
2299 if (!stream
->hangup_flush_done
2300 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2301 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2302 DBG("Attempting to flush and consume the UST buffers");
2303 lttng_ustconsumer_on_stream_hangup(stream
);
2305 /* We just flushed the stream now read it. */
2307 health_code_update();
2309 len
= ctx
->on_buffer_ready(stream
, ctx
);
2311 * We don't check the return value here since if we get
2312 * a negative len, it means an error occured thus we
2313 * simply remove it from the poll set and free the
2319 lttng_poll_del(&events
, stream
->wait_fd
);
2321 * This call update the channel states, closes file descriptors
2322 * and securely free the stream.
2324 consumer_del_metadata_stream(stream
, metadata_ht
);
2325 } else if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2326 /* Get the data out of the metadata file descriptor */
2327 DBG("Metadata available on fd %d", pollfd
);
2328 assert(stream
->wait_fd
== pollfd
);
2331 health_code_update();
2333 len
= ctx
->on_buffer_ready(stream
, ctx
);
2335 * We don't check the return value here since if we get
2336 * a negative len, it means an error occured thus we
2337 * simply remove it from the poll set and free the
2342 /* It's ok to have an unavailable sub-buffer */
2343 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2344 /* Clean up stream from consumer and free it. */
2345 lttng_poll_del(&events
, stream
->wait_fd
);
2346 consumer_del_metadata_stream(stream
, metadata_ht
);
2350 /* Release RCU lock for the stream looked up */
2359 DBG("Metadata poll thread exiting");
2361 lttng_poll_clean(&events
);
2366 ERR("Health error occurred in %s", __func__
);
2368 health_unregister(health_consumerd
);
2369 rcu_unregister_thread();
2374 * This thread polls the fds in the set to consume the data and write
2375 * it to tracefile if necessary.
2377 void *consumer_thread_data_poll(void *data
)
2379 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2380 struct pollfd
*pollfd
= NULL
;
2381 /* local view of the streams */
2382 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2383 /* local view of consumer_data.fds_count */
2385 struct lttng_consumer_local_data
*ctx
= data
;
2388 rcu_register_thread();
2390 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2392 if (testpoint(consumerd_thread_data
)) {
2393 goto error_testpoint
;
2396 health_code_update();
2398 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
*));
2399 if (local_stream
== NULL
) {
2400 PERROR("local_stream malloc");
2405 health_code_update();
2411 * the fds set has been updated, we need to update our
2412 * local array as well
2414 pthread_mutex_lock(&consumer_data
.lock
);
2415 if (consumer_data
.need_update
) {
2420 local_stream
= NULL
;
2423 * Allocate for all fds +1 for the consumer_data_pipe and +1 for
2426 pollfd
= zmalloc((consumer_data
.stream_count
+ 2) * sizeof(struct pollfd
));
2427 if (pollfd
== NULL
) {
2428 PERROR("pollfd malloc");
2429 pthread_mutex_unlock(&consumer_data
.lock
);
2433 local_stream
= zmalloc((consumer_data
.stream_count
+ 2) *
2434 sizeof(struct lttng_consumer_stream
*));
2435 if (local_stream
== NULL
) {
2436 PERROR("local_stream malloc");
2437 pthread_mutex_unlock(&consumer_data
.lock
);
2440 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2443 ERR("Error in allocating pollfd or local_outfds");
2444 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2445 pthread_mutex_unlock(&consumer_data
.lock
);
2449 consumer_data
.need_update
= 0;
2451 pthread_mutex_unlock(&consumer_data
.lock
);
2453 /* No FDs and consumer_quit, consumer_cleanup the thread */
2454 if (nb_fd
== 0 && consumer_quit
== 1) {
2455 err
= 0; /* All is OK */
2458 /* poll on the array of fds */
2460 DBG("polling on %d fd", nb_fd
+ 2);
2461 health_poll_entry();
2462 num_rdy
= poll(pollfd
, nb_fd
+ 2, -1);
2464 DBG("poll num_rdy : %d", num_rdy
);
2465 if (num_rdy
== -1) {
2467 * Restart interrupted system call.
2469 if (errno
== EINTR
) {
2472 PERROR("Poll error");
2473 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2475 } else if (num_rdy
== 0) {
2476 DBG("Polling thread timed out");
2481 * If the consumer_data_pipe triggered poll go directly to the
2482 * beginning of the loop to update the array. We want to prioritize
2483 * array update over low-priority reads.
2485 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2486 ssize_t pipe_readlen
;
2488 DBG("consumer_data_pipe wake up");
2489 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2490 &new_stream
, sizeof(new_stream
));
2491 if (pipe_readlen
< sizeof(new_stream
)) {
2492 PERROR("Consumer data pipe");
2493 /* Continue so we can at least handle the current stream(s). */
2498 * If the stream is NULL, just ignore it. It's also possible that
2499 * the sessiond poll thread changed the consumer_quit state and is
2500 * waking us up to test it.
2502 if (new_stream
== NULL
) {
2503 validate_endpoint_status_data_stream();
2507 /* Continue to update the local streams and handle prio ones */
2511 /* Handle wakeup pipe. */
2512 if (pollfd
[nb_fd
+ 1].revents
& (POLLIN
| POLLPRI
)) {
2514 ssize_t pipe_readlen
;
2516 pipe_readlen
= lttng_pipe_read(ctx
->consumer_wakeup_pipe
, &dummy
,
2518 if (pipe_readlen
< 0) {
2519 PERROR("Consumer data wakeup pipe");
2521 /* We've been awakened to handle stream(s). */
2522 ctx
->has_wakeup
= 0;
2525 /* Take care of high priority channels first. */
2526 for (i
= 0; i
< nb_fd
; i
++) {
2527 health_code_update();
2529 if (local_stream
[i
] == NULL
) {
2532 if (pollfd
[i
].revents
& POLLPRI
) {
2533 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2535 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2536 /* it's ok to have an unavailable sub-buffer */
2537 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2538 /* Clean the stream and free it. */
2539 consumer_del_stream(local_stream
[i
], data_ht
);
2540 local_stream
[i
] = NULL
;
2541 } else if (len
> 0) {
2542 local_stream
[i
]->data_read
= 1;
2548 * If we read high prio channel in this loop, try again
2549 * for more high prio data.
2555 /* Take care of low priority channels. */
2556 for (i
= 0; i
< nb_fd
; i
++) {
2557 health_code_update();
2559 if (local_stream
[i
] == NULL
) {
2562 if ((pollfd
[i
].revents
& POLLIN
) ||
2563 local_stream
[i
]->hangup_flush_done
||
2564 local_stream
[i
]->has_data
) {
2565 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2566 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2567 /* it's ok to have an unavailable sub-buffer */
2568 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2569 /* Clean the stream and free it. */
2570 consumer_del_stream(local_stream
[i
], data_ht
);
2571 local_stream
[i
] = NULL
;
2572 } else if (len
> 0) {
2573 local_stream
[i
]->data_read
= 1;
2578 /* Handle hangup and errors */
2579 for (i
= 0; i
< nb_fd
; i
++) {
2580 health_code_update();
2582 if (local_stream
[i
] == NULL
) {
2585 if (!local_stream
[i
]->hangup_flush_done
2586 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2587 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2588 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2589 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2591 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2592 /* Attempt read again, for the data we just flushed. */
2593 local_stream
[i
]->data_read
= 1;
2596 * If the poll flag is HUP/ERR/NVAL and we have
2597 * read no data in this pass, we can remove the
2598 * stream from its hash table.
2600 if ((pollfd
[i
].revents
& POLLHUP
)) {
2601 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2602 if (!local_stream
[i
]->data_read
) {
2603 consumer_del_stream(local_stream
[i
], data_ht
);
2604 local_stream
[i
] = NULL
;
2607 } else if (pollfd
[i
].revents
& POLLERR
) {
2608 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2609 if (!local_stream
[i
]->data_read
) {
2610 consumer_del_stream(local_stream
[i
], data_ht
);
2611 local_stream
[i
] = NULL
;
2614 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2615 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2616 if (!local_stream
[i
]->data_read
) {
2617 consumer_del_stream(local_stream
[i
], data_ht
);
2618 local_stream
[i
] = NULL
;
2622 if (local_stream
[i
] != NULL
) {
2623 local_stream
[i
]->data_read
= 0;
2630 DBG("polling thread exiting");
2635 * Close the write side of the pipe so epoll_wait() in
2636 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2637 * read side of the pipe. If we close them both, epoll_wait strangely does
2638 * not return and could create a endless wait period if the pipe is the
2639 * only tracked fd in the poll set. The thread will take care of closing
2642 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2647 ERR("Health error occurred in %s", __func__
);
2649 health_unregister(health_consumerd
);
2651 rcu_unregister_thread();
2656 * Close wake-up end of each stream belonging to the channel. This will
2657 * allow the poll() on the stream read-side to detect when the
2658 * write-side (application) finally closes them.
2661 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2663 struct lttng_ht
*ht
;
2664 struct lttng_consumer_stream
*stream
;
2665 struct lttng_ht_iter iter
;
2667 ht
= consumer_data
.stream_per_chan_id_ht
;
2670 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2671 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2672 ht
->match_fct
, &channel
->key
,
2673 &iter
.iter
, stream
, node_channel_id
.node
) {
2675 * Protect against teardown with mutex.
2677 pthread_mutex_lock(&stream
->lock
);
2678 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2681 switch (consumer_data
.type
) {
2682 case LTTNG_CONSUMER_KERNEL
:
2684 case LTTNG_CONSUMER32_UST
:
2685 case LTTNG_CONSUMER64_UST
:
2686 if (stream
->metadata_flag
) {
2687 /* Safe and protected by the stream lock. */
2688 lttng_ustconsumer_close_metadata(stream
->chan
);
2691 * Note: a mutex is taken internally within
2692 * liblttng-ust-ctl to protect timer wakeup_fd
2693 * use from concurrent close.
2695 lttng_ustconsumer_close_stream_wakeup(stream
);
2699 ERR("Unknown consumer_data type");
2703 pthread_mutex_unlock(&stream
->lock
);
2708 static void destroy_channel_ht(struct lttng_ht
*ht
)
2710 struct lttng_ht_iter iter
;
2711 struct lttng_consumer_channel
*channel
;
2719 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2720 ret
= lttng_ht_del(ht
, &iter
);
2725 lttng_ht_destroy(ht
);
2729 * This thread polls the channel fds to detect when they are being
2730 * closed. It closes all related streams if the channel is detected as
2731 * closed. It is currently only used as a shim layer for UST because the
2732 * consumerd needs to keep the per-stream wakeup end of pipes open for
2735 void *consumer_thread_channel_poll(void *data
)
2737 int ret
, i
, pollfd
, err
= -1;
2738 uint32_t revents
, nb_fd
;
2739 struct lttng_consumer_channel
*chan
= NULL
;
2740 struct lttng_ht_iter iter
;
2741 struct lttng_ht_node_u64
*node
;
2742 struct lttng_poll_event events
;
2743 struct lttng_consumer_local_data
*ctx
= data
;
2744 struct lttng_ht
*channel_ht
;
2746 rcu_register_thread();
2748 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
2750 if (testpoint(consumerd_thread_channel
)) {
2751 goto error_testpoint
;
2754 health_code_update();
2756 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2758 /* ENOMEM at this point. Better to bail out. */
2762 DBG("Thread channel poll started");
2764 /* Size is set to 1 for the consumer_channel pipe */
2765 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2767 ERR("Poll set creation failed");
2771 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2777 DBG("Channel main loop started");
2780 health_code_update();
2782 /* Only the channel pipe is set */
2783 if (LTTNG_POLL_GETNB(&events
) == 0 && consumer_quit
== 1) {
2784 err
= 0; /* All is OK */
2789 DBG("Channel poll wait with %d fd(s)", LTTNG_POLL_GETNB(&events
));
2790 health_poll_entry();
2791 ret
= lttng_poll_wait(&events
, -1);
2793 DBG("Channel event catched in thread");
2795 if (errno
== EINTR
) {
2796 ERR("Poll EINTR catched");
2804 /* From here, the event is a channel wait fd */
2805 for (i
= 0; i
< nb_fd
; i
++) {
2806 health_code_update();
2808 revents
= LTTNG_POLL_GETEV(&events
, i
);
2809 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2811 /* Just don't waste time if no returned events for the fd */
2815 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2816 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2817 DBG("Channel thread pipe hung up");
2819 * Remove the pipe from the poll set and continue the loop
2820 * since their might be data to consume.
2822 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2824 } else if (revents
& LPOLLIN
) {
2825 enum consumer_channel_action action
;
2828 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
2830 ERR("Error reading channel pipe");
2835 case CONSUMER_CHANNEL_ADD
:
2836 DBG("Adding channel %d to poll set",
2839 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
2842 lttng_ht_add_unique_u64(channel_ht
,
2843 &chan
->wait_fd_node
);
2845 /* Add channel to the global poll events list */
2846 lttng_poll_add(&events
, chan
->wait_fd
,
2847 LPOLLIN
| LPOLLPRI
);
2849 case CONSUMER_CHANNEL_DEL
:
2852 * This command should never be called if the channel
2853 * has streams monitored by either the data or metadata
2854 * thread. The consumer only notify this thread with a
2855 * channel del. command if it receives a destroy
2856 * channel command from the session daemon that send it
2857 * if a command prior to the GET_CHANNEL failed.
2861 chan
= consumer_find_channel(key
);
2864 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
2867 lttng_poll_del(&events
, chan
->wait_fd
);
2868 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
2869 ret
= lttng_ht_del(channel_ht
, &iter
);
2872 switch (consumer_data
.type
) {
2873 case LTTNG_CONSUMER_KERNEL
:
2875 case LTTNG_CONSUMER32_UST
:
2876 case LTTNG_CONSUMER64_UST
:
2877 health_code_update();
2878 /* Destroy streams that might have been left in the stream list. */
2879 clean_channel_stream_list(chan
);
2882 ERR("Unknown consumer_data type");
2887 * Release our own refcount. Force channel deletion even if
2888 * streams were not initialized.
2890 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
2891 consumer_del_channel(chan
);
2896 case CONSUMER_CHANNEL_QUIT
:
2898 * Remove the pipe from the poll set and continue the loop
2899 * since their might be data to consume.
2901 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2904 ERR("Unknown action");
2909 /* Handle other stream */
2915 uint64_t tmp_id
= (uint64_t) pollfd
;
2917 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
2919 node
= lttng_ht_iter_get_node_u64(&iter
);
2922 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
2925 /* Check for error event */
2926 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2927 DBG("Channel fd %d is hup|err.", pollfd
);
2929 lttng_poll_del(&events
, chan
->wait_fd
);
2930 ret
= lttng_ht_del(channel_ht
, &iter
);
2934 * This will close the wait fd for each stream associated to
2935 * this channel AND monitored by the data/metadata thread thus
2936 * will be clean by the right thread.
2938 consumer_close_channel_streams(chan
);
2940 /* Release our own refcount */
2941 if (!uatomic_sub_return(&chan
->refcount
, 1)
2942 && !uatomic_read(&chan
->nb_init_stream_left
)) {
2943 consumer_del_channel(chan
);
2947 /* Release RCU lock for the channel looked up */
2955 lttng_poll_clean(&events
);
2957 destroy_channel_ht(channel_ht
);
2960 DBG("Channel poll thread exiting");
2963 ERR("Health error occurred in %s", __func__
);
2965 health_unregister(health_consumerd
);
2966 rcu_unregister_thread();
2970 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
2971 struct pollfd
*sockpoll
, int client_socket
)
2978 ret
= lttng_consumer_poll_socket(sockpoll
);
2982 DBG("Metadata connection on client_socket");
2984 /* Blocking call, waiting for transmission */
2985 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
2986 if (ctx
->consumer_metadata_socket
< 0) {
2987 WARN("On accept metadata");
2998 * This thread listens on the consumerd socket and receives the file
2999 * descriptors from the session daemon.
3001 void *consumer_thread_sessiond_poll(void *data
)
3003 int sock
= -1, client_socket
, ret
, err
= -1;
3005 * structure to poll for incoming data on communication socket avoids
3006 * making blocking sockets.
3008 struct pollfd consumer_sockpoll
[2];
3009 struct lttng_consumer_local_data
*ctx
= data
;
3011 rcu_register_thread();
3013 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
3015 if (testpoint(consumerd_thread_sessiond
)) {
3016 goto error_testpoint
;
3019 health_code_update();
3021 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
3022 unlink(ctx
->consumer_command_sock_path
);
3023 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
3024 if (client_socket
< 0) {
3025 ERR("Cannot create command socket");
3029 ret
= lttcomm_listen_unix_sock(client_socket
);
3034 DBG("Sending ready command to lttng-sessiond");
3035 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3036 /* return < 0 on error, but == 0 is not fatal */
3038 ERR("Error sending ready command to lttng-sessiond");
3042 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3043 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3044 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3045 consumer_sockpoll
[1].fd
= client_socket
;
3046 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3048 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3056 DBG("Connection on client_socket");
3058 /* Blocking call, waiting for transmission */
3059 sock
= lttcomm_accept_unix_sock(client_socket
);
3066 * Setup metadata socket which is the second socket connection on the
3067 * command unix socket.
3069 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3078 /* This socket is not useful anymore. */
3079 ret
= close(client_socket
);
3081 PERROR("close client_socket");
3085 /* update the polling structure to poll on the established socket */
3086 consumer_sockpoll
[1].fd
= sock
;
3087 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3090 health_code_update();
3092 health_poll_entry();
3093 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3102 DBG("Incoming command on sock");
3103 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3106 * This could simply be a session daemon quitting. Don't output
3109 DBG("Communication interrupted on command socket");
3113 if (consumer_quit
) {
3114 DBG("consumer_thread_receive_fds received quit from signal");
3115 err
= 0; /* All is OK */
3118 DBG("received command on sock");
3124 DBG("Consumer thread sessiond poll exiting");
3127 * Close metadata streams since the producer is the session daemon which
3130 * NOTE: for now, this only applies to the UST tracer.
3132 lttng_consumer_close_all_metadata();
3135 * when all fds have hung up, the polling thread
3141 * Notify the data poll thread to poll back again and test the
3142 * consumer_quit state that we just set so to quit gracefully.
3144 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3146 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3148 notify_health_quit_pipe(health_quit_pipe
);
3150 /* Cleaning up possibly open sockets. */
3154 PERROR("close sock sessiond poll");
3157 if (client_socket
>= 0) {
3158 ret
= close(client_socket
);
3160 PERROR("close client_socket sessiond poll");
3167 ERR("Health error occurred in %s", __func__
);
3169 health_unregister(health_consumerd
);
3171 rcu_unregister_thread();
3175 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3176 struct lttng_consumer_local_data
*ctx
)
3180 pthread_mutex_lock(&stream
->lock
);
3181 if (stream
->metadata_flag
) {
3182 pthread_mutex_lock(&stream
->metadata_rdv_lock
);
3185 switch (consumer_data
.type
) {
3186 case LTTNG_CONSUMER_KERNEL
:
3187 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
3189 case LTTNG_CONSUMER32_UST
:
3190 case LTTNG_CONSUMER64_UST
:
3191 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
3194 ERR("Unknown consumer_data type");
3200 if (stream
->metadata_flag
) {
3201 pthread_cond_broadcast(&stream
->metadata_rdv
);
3202 pthread_mutex_unlock(&stream
->metadata_rdv_lock
);
3204 pthread_mutex_unlock(&stream
->lock
);
3208 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3210 switch (consumer_data
.type
) {
3211 case LTTNG_CONSUMER_KERNEL
:
3212 return lttng_kconsumer_on_recv_stream(stream
);
3213 case LTTNG_CONSUMER32_UST
:
3214 case LTTNG_CONSUMER64_UST
:
3215 return lttng_ustconsumer_on_recv_stream(stream
);
3217 ERR("Unknown consumer_data type");
3224 * Allocate and set consumer data hash tables.
3226 int lttng_consumer_init(void)
3228 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3229 if (!consumer_data
.channel_ht
) {
3233 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3234 if (!consumer_data
.relayd_ht
) {
3238 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3239 if (!consumer_data
.stream_list_ht
) {
3243 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3244 if (!consumer_data
.stream_per_chan_id_ht
) {
3248 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3253 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3265 * Process the ADD_RELAYD command receive by a consumer.
3267 * This will create a relayd socket pair and add it to the relayd hash table.
3268 * The caller MUST acquire a RCU read side lock before calling it.
3270 int consumer_add_relayd_socket(uint64_t net_seq_idx
, int sock_type
,
3271 struct lttng_consumer_local_data
*ctx
, int sock
,
3272 struct pollfd
*consumer_sockpoll
,
3273 struct lttcomm_relayd_sock
*relayd_sock
, uint64_t sessiond_id
,
3274 uint64_t relayd_session_id
)
3276 int fd
= -1, ret
= -1, relayd_created
= 0;
3277 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3278 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3281 assert(relayd_sock
);
3283 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", net_seq_idx
);
3285 /* Get relayd reference if exists. */
3286 relayd
= consumer_find_relayd(net_seq_idx
);
3287 if (relayd
== NULL
) {
3288 assert(sock_type
== LTTNG_STREAM_CONTROL
);
3289 /* Not found. Allocate one. */
3290 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3291 if (relayd
== NULL
) {
3293 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3296 relayd
->sessiond_session_id
= sessiond_id
;
3301 * This code path MUST continue to the consumer send status message to
3302 * we can notify the session daemon and continue our work without
3303 * killing everything.
3307 * relayd key should never be found for control socket.
3309 assert(sock_type
!= LTTNG_STREAM_CONTROL
);
3312 /* First send a status message before receiving the fds. */
3313 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3315 /* Somehow, the session daemon is not responding anymore. */
3316 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3317 goto error_nosignal
;
3320 /* Poll on consumer socket. */
3321 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3323 /* Needing to exit in the middle of a command: error. */
3324 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3326 goto error_nosignal
;
3329 /* Get relayd socket from session daemon */
3330 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3331 if (ret
!= sizeof(fd
)) {
3333 fd
= -1; /* Just in case it gets set with an invalid value. */
3336 * Failing to receive FDs might indicate a major problem such as
3337 * reaching a fd limit during the receive where the kernel returns a
3338 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3339 * don't take any chances and stop everything.
3341 * XXX: Feature request #558 will fix that and avoid this possible
3342 * issue when reaching the fd limit.
3344 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3345 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3349 /* Copy socket information and received FD */
3350 switch (sock_type
) {
3351 case LTTNG_STREAM_CONTROL
:
3352 /* Copy received lttcomm socket */
3353 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3354 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3355 /* Handle create_sock error. */
3357 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3361 * Close the socket created internally by
3362 * lttcomm_create_sock, so we can replace it by the one
3363 * received from sessiond.
3365 if (close(relayd
->control_sock
.sock
.fd
)) {
3369 /* Assign new file descriptor */
3370 relayd
->control_sock
.sock
.fd
= fd
;
3371 fd
= -1; /* For error path */
3372 /* Assign version values. */
3373 relayd
->control_sock
.major
= relayd_sock
->major
;
3374 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3376 relayd
->relayd_session_id
= relayd_session_id
;
3379 case LTTNG_STREAM_DATA
:
3380 /* Copy received lttcomm socket */
3381 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3382 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3383 /* Handle create_sock error. */
3385 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3389 * Close the socket created internally by
3390 * lttcomm_create_sock, so we can replace it by the one
3391 * received from sessiond.
3393 if (close(relayd
->data_sock
.sock
.fd
)) {
3397 /* Assign new file descriptor */
3398 relayd
->data_sock
.sock
.fd
= fd
;
3399 fd
= -1; /* for eventual error paths */
3400 /* Assign version values. */
3401 relayd
->data_sock
.major
= relayd_sock
->major
;
3402 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3405 ERR("Unknown relayd socket type (%d)", sock_type
);
3407 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3411 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3412 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3413 relayd
->net_seq_idx
, fd
);
3415 /* We successfully added the socket. Send status back. */
3416 ret
= consumer_send_status_msg(sock
, ret_code
);
3418 /* Somehow, the session daemon is not responding anymore. */
3419 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3420 goto error_nosignal
;
3424 * Add relayd socket pair to consumer data hashtable. If object already
3425 * exists or on error, the function gracefully returns.
3433 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3434 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3438 /* Close received socket if valid. */
3441 PERROR("close received socket");
3445 if (relayd_created
) {
3453 * Try to lock the stream mutex.
3455 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
3457 static int stream_try_lock(struct lttng_consumer_stream
*stream
)
3464 * Try to lock the stream mutex. On failure, we know that the stream is
3465 * being used else where hence there is data still being extracted.
3467 ret
= pthread_mutex_trylock(&stream
->lock
);
3469 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
3481 * Search for a relayd associated to the session id and return the reference.
3483 * A rcu read side lock MUST be acquire before calling this function and locked
3484 * until the relayd object is no longer necessary.
3486 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3488 struct lttng_ht_iter iter
;
3489 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3491 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3492 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3495 * Check by sessiond id which is unique here where the relayd session
3496 * id might not be when having multiple relayd.
3498 if (relayd
->sessiond_session_id
== id
) {
3499 /* Found the relayd. There can be only one per id. */
3511 * Check if for a given session id there is still data needed to be extract
3514 * Return 1 if data is pending or else 0 meaning ready to be read.
3516 int consumer_data_pending(uint64_t id
)
3519 struct lttng_ht_iter iter
;
3520 struct lttng_ht
*ht
;
3521 struct lttng_consumer_stream
*stream
;
3522 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3523 int (*data_pending
)(struct lttng_consumer_stream
*);
3525 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3528 pthread_mutex_lock(&consumer_data
.lock
);
3530 switch (consumer_data
.type
) {
3531 case LTTNG_CONSUMER_KERNEL
:
3532 data_pending
= lttng_kconsumer_data_pending
;
3534 case LTTNG_CONSUMER32_UST
:
3535 case LTTNG_CONSUMER64_UST
:
3536 data_pending
= lttng_ustconsumer_data_pending
;
3539 ERR("Unknown consumer data type");
3543 /* Ease our life a bit */
3544 ht
= consumer_data
.stream_list_ht
;
3546 relayd
= find_relayd_by_session_id(id
);
3548 /* Send init command for data pending. */
3549 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3550 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3551 relayd
->relayd_session_id
);
3552 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3554 /* Communication error thus the relayd so no data pending. */
3555 goto data_not_pending
;
3559 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3560 ht
->hash_fct(&id
, lttng_ht_seed
),
3562 &iter
.iter
, stream
, node_session_id
.node
) {
3563 /* If this call fails, the stream is being used hence data pending. */
3564 ret
= stream_try_lock(stream
);
3570 * A removed node from the hash table indicates that the stream has
3571 * been deleted thus having a guarantee that the buffers are closed
3572 * on the consumer side. However, data can still be transmitted
3573 * over the network so don't skip the relayd check.
3575 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3578 * An empty output file is not valid. We need at least one packet
3579 * generated per stream, even if it contains no event, so it
3580 * contains at least one packet header.
3582 if (stream
->output_written
== 0) {
3583 pthread_mutex_unlock(&stream
->lock
);
3586 /* Check the stream if there is data in the buffers. */
3587 ret
= data_pending(stream
);
3589 pthread_mutex_unlock(&stream
->lock
);
3596 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3597 if (stream
->metadata_flag
) {
3598 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3599 stream
->relayd_stream_id
);
3601 ret
= relayd_data_pending(&relayd
->control_sock
,
3602 stream
->relayd_stream_id
,
3603 stream
->next_net_seq_num
- 1);
3605 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3607 pthread_mutex_unlock(&stream
->lock
);
3611 pthread_mutex_unlock(&stream
->lock
);
3615 unsigned int is_data_inflight
= 0;
3617 /* Send init command for data pending. */
3618 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3619 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3620 relayd
->relayd_session_id
, &is_data_inflight
);
3621 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3623 goto data_not_pending
;
3625 if (is_data_inflight
) {
3631 * Finding _no_ node in the hash table and no inflight data means that the
3632 * stream(s) have been removed thus data is guaranteed to be available for
3633 * analysis from the trace files.
3637 /* Data is available to be read by a viewer. */
3638 pthread_mutex_unlock(&consumer_data
.lock
);
3643 /* Data is still being extracted from buffers. */
3644 pthread_mutex_unlock(&consumer_data
.lock
);
3650 * Send a ret code status message to the sessiond daemon.
3652 * Return the sendmsg() return value.
3654 int consumer_send_status_msg(int sock
, int ret_code
)
3656 struct lttcomm_consumer_status_msg msg
;
3658 memset(&msg
, 0, sizeof(msg
));
3659 msg
.ret_code
= ret_code
;
3661 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3665 * Send a channel status message to the sessiond daemon.
3667 * Return the sendmsg() return value.
3669 int consumer_send_status_channel(int sock
,
3670 struct lttng_consumer_channel
*channel
)
3672 struct lttcomm_consumer_status_channel msg
;
3676 memset(&msg
, 0, sizeof(msg
));
3678 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3680 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3681 msg
.key
= channel
->key
;
3682 msg
.stream_count
= channel
->streams
.count
;
3685 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3689 * Using a maximum stream size with the produced and consumed position of a
3690 * stream, computes the new consumed position to be as close as possible to the
3691 * maximum possible stream size.
3693 * If maximum stream size is lower than the possible buffer size (produced -
3694 * consumed), the consumed_pos given is returned untouched else the new value
3697 unsigned long consumer_get_consumed_maxsize(unsigned long consumed_pos
,
3698 unsigned long produced_pos
, uint64_t max_stream_size
)
3700 if (max_stream_size
&& max_stream_size
< (produced_pos
- consumed_pos
)) {
3701 /* Offset from the produced position to get the latest buffers. */
3702 return produced_pos
- max_stream_size
;
3705 return consumed_pos
;