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 <bin/lttng-consumerd/health-consumerd.h>
34 #include <common/common.h>
35 #include <common/utils.h>
36 #include <common/compat/poll.h>
37 #include <common/compat/endian.h>
38 #include <common/index/index.h>
39 #include <common/kernel-ctl/kernel-ctl.h>
40 #include <common/sessiond-comm/relayd.h>
41 #include <common/sessiond-comm/sessiond-comm.h>
42 #include <common/kernel-consumer/kernel-consumer.h>
43 #include <common/relayd/relayd.h>
44 #include <common/ust-consumer/ust-consumer.h>
45 #include <common/consumer-timer.h>
48 #include "consumer-stream.h"
49 #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
);
288 switch (consumer_data
.type
) {
289 case LTTNG_CONSUMER_KERNEL
:
291 case LTTNG_CONSUMER32_UST
:
292 case LTTNG_CONSUMER64_UST
:
293 lttng_ustconsumer_free_channel(channel
);
296 ERR("Unknown consumer_data type");
303 * RCU protected relayd socket pair free.
305 static void free_relayd_rcu(struct rcu_head
*head
)
307 struct lttng_ht_node_u64
*node
=
308 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
309 struct consumer_relayd_sock_pair
*relayd
=
310 caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
313 * Close all sockets. This is done in the call RCU since we don't want the
314 * socket fds to be reassigned thus potentially creating bad state of the
317 * We do not have to lock the control socket mutex here since at this stage
318 * there is no one referencing to this relayd object.
320 (void) relayd_close(&relayd
->control_sock
);
321 (void) relayd_close(&relayd
->data_sock
);
327 * Destroy and free relayd socket pair object.
329 void consumer_destroy_relayd(struct consumer_relayd_sock_pair
*relayd
)
332 struct lttng_ht_iter iter
;
334 if (relayd
== NULL
) {
338 DBG("Consumer destroy and close relayd socket pair");
340 iter
.iter
.node
= &relayd
->node
.node
;
341 ret
= lttng_ht_del(consumer_data
.relayd_ht
, &iter
);
343 /* We assume the relayd is being or is destroyed */
347 /* RCU free() call */
348 call_rcu(&relayd
->node
.head
, free_relayd_rcu
);
352 * Remove a channel from the global list protected by a mutex. This function is
353 * also responsible for freeing its data structures.
355 void consumer_del_channel(struct lttng_consumer_channel
*channel
)
358 struct lttng_ht_iter iter
;
360 DBG("Consumer delete channel key %" PRIu64
, channel
->key
);
362 pthread_mutex_lock(&consumer_data
.lock
);
363 pthread_mutex_lock(&channel
->lock
);
365 /* Destroy streams that might have been left in the stream list. */
366 clean_channel_stream_list(channel
);
368 if (channel
->live_timer_enabled
== 1) {
369 consumer_timer_live_stop(channel
);
372 switch (consumer_data
.type
) {
373 case LTTNG_CONSUMER_KERNEL
:
375 case LTTNG_CONSUMER32_UST
:
376 case LTTNG_CONSUMER64_UST
:
377 lttng_ustconsumer_del_channel(channel
);
380 ERR("Unknown consumer_data type");
386 iter
.iter
.node
= &channel
->node
.node
;
387 ret
= lttng_ht_del(consumer_data
.channel_ht
, &iter
);
391 call_rcu(&channel
->node
.head
, free_channel_rcu
);
393 pthread_mutex_unlock(&channel
->lock
);
394 pthread_mutex_unlock(&consumer_data
.lock
);
398 * Iterate over the relayd hash table and destroy each element. Finally,
399 * destroy the whole hash table.
401 static void cleanup_relayd_ht(void)
403 struct lttng_ht_iter iter
;
404 struct consumer_relayd_sock_pair
*relayd
;
408 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
410 consumer_destroy_relayd(relayd
);
415 lttng_ht_destroy(consumer_data
.relayd_ht
);
419 * Update the end point status of all streams having the given network sequence
420 * index (relayd index).
422 * It's atomically set without having the stream mutex locked which is fine
423 * because we handle the write/read race with a pipe wakeup for each thread.
425 static void update_endpoint_status_by_netidx(uint64_t net_seq_idx
,
426 enum consumer_endpoint_status status
)
428 struct lttng_ht_iter iter
;
429 struct lttng_consumer_stream
*stream
;
431 DBG("Consumer set delete flag on stream by idx %" PRIu64
, net_seq_idx
);
435 /* Let's begin with metadata */
436 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
437 if (stream
->net_seq_idx
== net_seq_idx
) {
438 uatomic_set(&stream
->endpoint_status
, status
);
439 DBG("Delete flag set to metadata stream %d", stream
->wait_fd
);
443 /* Follow up by the data streams */
444 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
445 if (stream
->net_seq_idx
== net_seq_idx
) {
446 uatomic_set(&stream
->endpoint_status
, status
);
447 DBG("Delete flag set to data stream %d", stream
->wait_fd
);
454 * Cleanup a relayd object by flagging every associated streams for deletion,
455 * destroying the object meaning removing it from the relayd hash table,
456 * closing the sockets and freeing the memory in a RCU call.
458 * If a local data context is available, notify the threads that the streams'
459 * state have changed.
461 static void cleanup_relayd(struct consumer_relayd_sock_pair
*relayd
,
462 struct lttng_consumer_local_data
*ctx
)
468 DBG("Cleaning up relayd sockets");
470 /* Save the net sequence index before destroying the object */
471 netidx
= relayd
->net_seq_idx
;
474 * Delete the relayd from the relayd hash table, close the sockets and free
475 * the object in a RCU call.
477 consumer_destroy_relayd(relayd
);
479 /* Set inactive endpoint to all streams */
480 update_endpoint_status_by_netidx(netidx
, CONSUMER_ENDPOINT_INACTIVE
);
483 * With a local data context, notify the threads that the streams' state
484 * have changed. The write() action on the pipe acts as an "implicit"
485 * memory barrier ordering the updates of the end point status from the
486 * read of this status which happens AFTER receiving this notify.
489 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
490 notify_thread_lttng_pipe(ctx
->consumer_metadata_pipe
);
495 * Flag a relayd socket pair for destruction. Destroy it if the refcount
498 * RCU read side lock MUST be aquired before calling this function.
500 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair
*relayd
)
504 /* Set destroy flag for this object */
505 uatomic_set(&relayd
->destroy_flag
, 1);
507 /* Destroy the relayd if refcount is 0 */
508 if (uatomic_read(&relayd
->refcount
) == 0) {
509 consumer_destroy_relayd(relayd
);
514 * Completly destroy stream from every visiable data structure and the given
517 * One this call returns, the stream object is not longer usable nor visible.
519 void consumer_del_stream(struct lttng_consumer_stream
*stream
,
522 consumer_stream_destroy(stream
, ht
);
526 * XXX naming of del vs destroy is all mixed up.
528 void consumer_del_stream_for_data(struct lttng_consumer_stream
*stream
)
530 consumer_stream_destroy(stream
, data_ht
);
533 void consumer_del_stream_for_metadata(struct lttng_consumer_stream
*stream
)
535 consumer_stream_destroy(stream
, metadata_ht
);
538 struct lttng_consumer_stream
*consumer_allocate_stream(uint64_t channel_key
,
540 enum lttng_consumer_stream_state state
,
541 const char *channel_name
,
548 enum consumer_channel_type type
,
549 unsigned int monitor
)
552 struct lttng_consumer_stream
*stream
;
554 stream
= zmalloc(sizeof(*stream
));
555 if (stream
== NULL
) {
556 PERROR("malloc struct lttng_consumer_stream");
563 stream
->key
= stream_key
;
565 stream
->out_fd_offset
= 0;
566 stream
->output_written
= 0;
567 stream
->state
= state
;
570 stream
->net_seq_idx
= relayd_id
;
571 stream
->session_id
= session_id
;
572 stream
->monitor
= monitor
;
573 stream
->endpoint_status
= CONSUMER_ENDPOINT_ACTIVE
;
574 stream
->index_fd
= -1;
575 pthread_mutex_init(&stream
->lock
, NULL
);
576 pthread_mutex_init(&stream
->metadata_timer_lock
, NULL
);
578 /* If channel is the metadata, flag this stream as metadata. */
579 if (type
== CONSUMER_CHANNEL_TYPE_METADATA
) {
580 stream
->metadata_flag
= 1;
581 /* Metadata is flat out. */
582 strncpy(stream
->name
, DEFAULT_METADATA_NAME
, sizeof(stream
->name
));
583 /* Live rendez-vous point. */
584 pthread_cond_init(&stream
->metadata_rdv
, NULL
);
585 pthread_mutex_init(&stream
->metadata_rdv_lock
, NULL
);
587 /* Format stream name to <channel_name>_<cpu_number> */
588 ret
= snprintf(stream
->name
, sizeof(stream
->name
), "%s_%d",
591 PERROR("snprintf stream name");
596 /* Key is always the wait_fd for streams. */
597 lttng_ht_node_init_u64(&stream
->node
, stream
->key
);
599 /* Init node per channel id key */
600 lttng_ht_node_init_u64(&stream
->node_channel_id
, channel_key
);
602 /* Init session id node with the stream session id */
603 lttng_ht_node_init_u64(&stream
->node_session_id
, stream
->session_id
);
605 DBG3("Allocated stream %s (key %" PRIu64
", chan_key %" PRIu64
606 " relayd_id %" PRIu64
", session_id %" PRIu64
,
607 stream
->name
, stream
->key
, channel_key
,
608 stream
->net_seq_idx
, stream
->session_id
);
624 * Add a stream to the global list protected by a mutex.
626 int consumer_add_data_stream(struct lttng_consumer_stream
*stream
)
628 struct lttng_ht
*ht
= data_ht
;
634 DBG3("Adding consumer stream %" PRIu64
, stream
->key
);
636 pthread_mutex_lock(&consumer_data
.lock
);
637 pthread_mutex_lock(&stream
->chan
->lock
);
638 pthread_mutex_lock(&stream
->chan
->timer_lock
);
639 pthread_mutex_lock(&stream
->lock
);
642 /* Steal stream identifier to avoid having streams with the same key */
643 steal_stream_key(stream
->key
, ht
);
645 lttng_ht_add_unique_u64(ht
, &stream
->node
);
647 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
648 &stream
->node_channel_id
);
651 * Add stream to the stream_list_ht of the consumer data. No need to steal
652 * the key since the HT does not use it and we allow to add redundant keys
655 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
658 * When nb_init_stream_left reaches 0, we don't need to trigger any action
659 * in terms of destroying the associated channel, because the action that
660 * causes the count to become 0 also causes a stream to be added. The
661 * channel deletion will thus be triggered by the following removal of this
664 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
665 /* Increment refcount before decrementing nb_init_stream_left */
667 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
670 /* Update consumer data once the node is inserted. */
671 consumer_data
.stream_count
++;
672 consumer_data
.need_update
= 1;
675 pthread_mutex_unlock(&stream
->lock
);
676 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
677 pthread_mutex_unlock(&stream
->chan
->lock
);
678 pthread_mutex_unlock(&consumer_data
.lock
);
683 void consumer_del_data_stream(struct lttng_consumer_stream
*stream
)
685 consumer_del_stream(stream
, data_ht
);
689 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
690 * be acquired before calling this.
692 static int add_relayd(struct consumer_relayd_sock_pair
*relayd
)
695 struct lttng_ht_node_u64
*node
;
696 struct lttng_ht_iter iter
;
700 lttng_ht_lookup(consumer_data
.relayd_ht
,
701 &relayd
->net_seq_idx
, &iter
);
702 node
= lttng_ht_iter_get_node_u64(&iter
);
706 lttng_ht_add_unique_u64(consumer_data
.relayd_ht
, &relayd
->node
);
713 * Allocate and return a consumer relayd socket.
715 struct consumer_relayd_sock_pair
*consumer_allocate_relayd_sock_pair(
716 uint64_t net_seq_idx
)
718 struct consumer_relayd_sock_pair
*obj
= NULL
;
720 /* net sequence index of -1 is a failure */
721 if (net_seq_idx
== (uint64_t) -1ULL) {
725 obj
= zmalloc(sizeof(struct consumer_relayd_sock_pair
));
727 PERROR("zmalloc relayd sock");
731 obj
->net_seq_idx
= net_seq_idx
;
733 obj
->destroy_flag
= 0;
734 obj
->control_sock
.sock
.fd
= -1;
735 obj
->data_sock
.sock
.fd
= -1;
736 lttng_ht_node_init_u64(&obj
->node
, obj
->net_seq_idx
);
737 pthread_mutex_init(&obj
->ctrl_sock_mutex
, NULL
);
744 * Find a relayd socket pair in the global consumer data.
746 * Return the object if found else NULL.
747 * RCU read-side lock must be held across this call and while using the
750 struct consumer_relayd_sock_pair
*consumer_find_relayd(uint64_t key
)
752 struct lttng_ht_iter iter
;
753 struct lttng_ht_node_u64
*node
;
754 struct consumer_relayd_sock_pair
*relayd
= NULL
;
756 /* Negative keys are lookup failures */
757 if (key
== (uint64_t) -1ULL) {
761 lttng_ht_lookup(consumer_data
.relayd_ht
, &key
,
763 node
= lttng_ht_iter_get_node_u64(&iter
);
765 relayd
= caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
773 * Find a relayd and send the stream
775 * Returns 0 on success, < 0 on error
777 int consumer_send_relayd_stream(struct lttng_consumer_stream
*stream
,
781 struct consumer_relayd_sock_pair
*relayd
;
784 assert(stream
->net_seq_idx
!= -1ULL);
787 /* The stream is not metadata. Get relayd reference if exists. */
789 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
790 if (relayd
!= NULL
) {
791 /* Add stream on the relayd */
792 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
793 ret
= relayd_add_stream(&relayd
->control_sock
, stream
->name
,
794 path
, &stream
->relayd_stream_id
,
795 stream
->chan
->tracefile_size
, stream
->chan
->tracefile_count
);
796 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
801 uatomic_inc(&relayd
->refcount
);
802 stream
->sent_to_relayd
= 1;
804 ERR("Stream %" PRIu64
" relayd ID %" PRIu64
" unknown. Can't send it.",
805 stream
->key
, stream
->net_seq_idx
);
810 DBG("Stream %s with key %" PRIu64
" sent to relayd id %" PRIu64
,
811 stream
->name
, stream
->key
, stream
->net_seq_idx
);
819 * Find a relayd and send the streams sent message
821 * Returns 0 on success, < 0 on error
823 int consumer_send_relayd_streams_sent(uint64_t net_seq_idx
)
826 struct consumer_relayd_sock_pair
*relayd
;
828 assert(net_seq_idx
!= -1ULL);
830 /* The stream is not metadata. Get relayd reference if exists. */
832 relayd
= consumer_find_relayd(net_seq_idx
);
833 if (relayd
!= NULL
) {
834 /* Add stream on the relayd */
835 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
836 ret
= relayd_streams_sent(&relayd
->control_sock
);
837 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
842 ERR("Relayd ID %" PRIu64
" unknown. Can't send streams_sent.",
849 DBG("All streams sent relayd id %" PRIu64
, net_seq_idx
);
857 * Find a relayd and close the stream
859 void close_relayd_stream(struct lttng_consumer_stream
*stream
)
861 struct consumer_relayd_sock_pair
*relayd
;
863 /* The stream is not metadata. Get relayd reference if exists. */
865 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
867 consumer_stream_relayd_close(stream
, relayd
);
873 * Handle stream for relayd transmission if the stream applies for network
874 * streaming where the net sequence index is set.
876 * Return destination file descriptor or negative value on error.
878 static int write_relayd_stream_header(struct lttng_consumer_stream
*stream
,
879 size_t data_size
, unsigned long padding
,
880 struct consumer_relayd_sock_pair
*relayd
)
883 struct lttcomm_relayd_data_hdr data_hdr
;
889 /* Reset data header */
890 memset(&data_hdr
, 0, sizeof(data_hdr
));
892 if (stream
->metadata_flag
) {
893 /* Caller MUST acquire the relayd control socket lock */
894 ret
= relayd_send_metadata(&relayd
->control_sock
, data_size
);
899 /* Metadata are always sent on the control socket. */
900 outfd
= relayd
->control_sock
.sock
.fd
;
902 /* Set header with stream information */
903 data_hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
904 data_hdr
.data_size
= htobe32(data_size
);
905 data_hdr
.padding_size
= htobe32(padding
);
907 * Note that net_seq_num below is assigned with the *current* value of
908 * next_net_seq_num and only after that the next_net_seq_num will be
909 * increment. This is why when issuing a command on the relayd using
910 * this next value, 1 should always be substracted in order to compare
911 * the last seen sequence number on the relayd side to the last sent.
913 data_hdr
.net_seq_num
= htobe64(stream
->next_net_seq_num
);
914 /* Other fields are zeroed previously */
916 ret
= relayd_send_data_hdr(&relayd
->data_sock
, &data_hdr
,
922 ++stream
->next_net_seq_num
;
924 /* Set to go on data socket */
925 outfd
= relayd
->data_sock
.sock
.fd
;
933 * Allocate and return a new lttng_consumer_channel object using the given key
934 * to initialize the hash table node.
936 * On error, return NULL.
938 struct lttng_consumer_channel
*consumer_allocate_channel(uint64_t key
,
940 const char *pathname
,
945 enum lttng_event_output output
,
946 uint64_t tracefile_size
,
947 uint64_t tracefile_count
,
948 uint64_t session_id_per_pid
,
949 unsigned int monitor
,
950 unsigned int live_timer_interval
)
952 struct lttng_consumer_channel
*channel
;
954 channel
= zmalloc(sizeof(*channel
));
955 if (channel
== NULL
) {
956 PERROR("malloc struct lttng_consumer_channel");
961 channel
->refcount
= 0;
962 channel
->session_id
= session_id
;
963 channel
->session_id_per_pid
= session_id_per_pid
;
966 channel
->relayd_id
= relayd_id
;
967 channel
->tracefile_size
= tracefile_size
;
968 channel
->tracefile_count
= tracefile_count
;
969 channel
->monitor
= monitor
;
970 channel
->live_timer_interval
= live_timer_interval
;
971 pthread_mutex_init(&channel
->lock
, NULL
);
972 pthread_mutex_init(&channel
->timer_lock
, NULL
);
975 case LTTNG_EVENT_SPLICE
:
976 channel
->output
= CONSUMER_CHANNEL_SPLICE
;
978 case LTTNG_EVENT_MMAP
:
979 channel
->output
= CONSUMER_CHANNEL_MMAP
;
989 * In monitor mode, the streams associated with the channel will be put in
990 * a special list ONLY owned by this channel. So, the refcount is set to 1
991 * here meaning that the channel itself has streams that are referenced.
993 * On a channel deletion, once the channel is no longer visible, the
994 * refcount is decremented and checked for a zero value to delete it. With
995 * streams in no monitor mode, it will now be safe to destroy the channel.
997 if (!channel
->monitor
) {
998 channel
->refcount
= 1;
1001 strncpy(channel
->pathname
, pathname
, sizeof(channel
->pathname
));
1002 channel
->pathname
[sizeof(channel
->pathname
) - 1] = '\0';
1004 strncpy(channel
->name
, name
, sizeof(channel
->name
));
1005 channel
->name
[sizeof(channel
->name
) - 1] = '\0';
1007 lttng_ht_node_init_u64(&channel
->node
, channel
->key
);
1009 channel
->wait_fd
= -1;
1011 CDS_INIT_LIST_HEAD(&channel
->streams
.head
);
1013 DBG("Allocated channel (key %" PRIu64
")", channel
->key
)
1020 * Add a channel to the global list protected by a mutex.
1022 * Always return 0 indicating success.
1024 int consumer_add_channel(struct lttng_consumer_channel
*channel
,
1025 struct lttng_consumer_local_data
*ctx
)
1027 pthread_mutex_lock(&consumer_data
.lock
);
1028 pthread_mutex_lock(&channel
->lock
);
1029 pthread_mutex_lock(&channel
->timer_lock
);
1032 * This gives us a guarantee that the channel we are about to add to the
1033 * channel hash table will be unique. See this function comment on the why
1034 * we need to steel the channel key at this stage.
1036 steal_channel_key(channel
->key
);
1039 lttng_ht_add_unique_u64(consumer_data
.channel_ht
, &channel
->node
);
1042 pthread_mutex_unlock(&channel
->timer_lock
);
1043 pthread_mutex_unlock(&channel
->lock
);
1044 pthread_mutex_unlock(&consumer_data
.lock
);
1046 if (channel
->wait_fd
!= -1 && channel
->type
== CONSUMER_CHANNEL_TYPE_DATA
) {
1047 notify_channel_pipe(ctx
, channel
, -1, CONSUMER_CHANNEL_ADD
);
1054 * Allocate the pollfd structure and the local view of the out fds to avoid
1055 * doing a lookup in the linked list and concurrency issues when writing is
1056 * needed. Called with consumer_data.lock held.
1058 * Returns the number of fds in the structures.
1060 static int update_poll_array(struct lttng_consumer_local_data
*ctx
,
1061 struct pollfd
**pollfd
, struct lttng_consumer_stream
**local_stream
,
1062 struct lttng_ht
*ht
)
1065 struct lttng_ht_iter iter
;
1066 struct lttng_consumer_stream
*stream
;
1071 assert(local_stream
);
1073 DBG("Updating poll fd array");
1075 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1077 * Only active streams with an active end point can be added to the
1078 * poll set and local stream storage of the thread.
1080 * There is a potential race here for endpoint_status to be updated
1081 * just after the check. However, this is OK since the stream(s) will
1082 * be deleted once the thread is notified that the end point state has
1083 * changed where this function will be called back again.
1085 if (stream
->state
!= LTTNG_CONSUMER_ACTIVE_STREAM
||
1086 stream
->endpoint_status
== CONSUMER_ENDPOINT_INACTIVE
) {
1090 * This clobbers way too much the debug output. Uncomment that if you
1091 * need it for debugging purposes.
1093 * DBG("Active FD %d", stream->wait_fd);
1095 (*pollfd
)[i
].fd
= stream
->wait_fd
;
1096 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1097 local_stream
[i
] = stream
;
1103 * Insert the consumer_data_pipe at the end of the array and don't
1104 * increment i so nb_fd is the number of real FD.
1106 (*pollfd
)[i
].fd
= lttng_pipe_get_readfd(ctx
->consumer_data_pipe
);
1107 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1109 (*pollfd
)[i
+ 1].fd
= lttng_pipe_get_readfd(ctx
->consumer_wakeup_pipe
);
1110 (*pollfd
)[i
+ 1].events
= POLLIN
| POLLPRI
;
1115 * Poll on the should_quit pipe and the command socket return -1 on
1116 * error, 1 if should exit, 0 if data is available on the command socket
1118 int lttng_consumer_poll_socket(struct pollfd
*consumer_sockpoll
)
1123 num_rdy
= poll(consumer_sockpoll
, 2, -1);
1124 if (num_rdy
== -1) {
1126 * Restart interrupted system call.
1128 if (errno
== EINTR
) {
1131 PERROR("Poll error");
1134 if (consumer_sockpoll
[0].revents
& (POLLIN
| POLLPRI
)) {
1135 DBG("consumer_should_quit wake up");
1142 * Set the error socket.
1144 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data
*ctx
,
1147 ctx
->consumer_error_socket
= sock
;
1151 * Set the command socket path.
1153 void lttng_consumer_set_command_sock_path(
1154 struct lttng_consumer_local_data
*ctx
, char *sock
)
1156 ctx
->consumer_command_sock_path
= sock
;
1160 * Send return code to the session daemon.
1161 * If the socket is not defined, we return 0, it is not a fatal error
1163 int lttng_consumer_send_error(struct lttng_consumer_local_data
*ctx
, int cmd
)
1165 if (ctx
->consumer_error_socket
> 0) {
1166 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
1167 sizeof(enum lttcomm_sessiond_command
));
1174 * Close all the tracefiles and stream fds and MUST be called when all
1175 * instances are destroyed i.e. when all threads were joined and are ended.
1177 void lttng_consumer_cleanup(void)
1179 struct lttng_ht_iter iter
;
1180 struct lttng_consumer_channel
*channel
;
1184 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
, channel
,
1186 consumer_del_channel(channel
);
1191 lttng_ht_destroy(consumer_data
.channel_ht
);
1193 cleanup_relayd_ht();
1195 lttng_ht_destroy(consumer_data
.stream_per_chan_id_ht
);
1198 * This HT contains streams that are freed by either the metadata thread or
1199 * the data thread so we do *nothing* on the hash table and simply destroy
1202 lttng_ht_destroy(consumer_data
.stream_list_ht
);
1206 * Called from signal handler.
1208 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
1213 ret
= lttng_write(ctx
->consumer_should_quit
[1], "4", 1);
1215 PERROR("write consumer quit");
1218 DBG("Consumer flag that it should quit");
1221 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1225 int outfd
= stream
->out_fd
;
1228 * This does a blocking write-and-wait on any page that belongs to the
1229 * subbuffer prior to the one we just wrote.
1230 * Don't care about error values, as these are just hints and ways to
1231 * limit the amount of page cache used.
1233 if (orig_offset
< stream
->max_sb_size
) {
1236 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1237 stream
->max_sb_size
,
1238 SYNC_FILE_RANGE_WAIT_BEFORE
1239 | SYNC_FILE_RANGE_WRITE
1240 | SYNC_FILE_RANGE_WAIT_AFTER
);
1242 * Give hints to the kernel about how we access the file:
1243 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1246 * We need to call fadvise again after the file grows because the
1247 * kernel does not seem to apply fadvise to non-existing parts of the
1250 * Call fadvise _after_ having waited for the page writeback to
1251 * complete because the dirty page writeback semantic is not well
1252 * defined. So it can be expected to lead to lower throughput in
1255 ret
= posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1256 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1257 if (ret
&& ret
!= -ENOSYS
) {
1259 PERROR("posix_fadvise on fd %i", outfd
);
1264 * Initialise the necessary environnement :
1265 * - create a new context
1266 * - create the poll_pipe
1267 * - create the should_quit pipe (for signal handler)
1268 * - create the thread pipe (for splice)
1270 * Takes a function pointer as argument, this function is called when data is
1271 * available on a buffer. This function is responsible to do the
1272 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1273 * buffer configuration and then kernctl_put_next_subbuf at the end.
1275 * Returns a pointer to the new context or NULL on error.
1277 struct lttng_consumer_local_data
*lttng_consumer_create(
1278 enum lttng_consumer_type type
,
1279 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1280 struct lttng_consumer_local_data
*ctx
),
1281 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1282 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1283 int (*update_stream
)(uint64_t stream_key
, uint32_t state
))
1286 struct lttng_consumer_local_data
*ctx
;
1288 assert(consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1289 consumer_data
.type
== type
);
1290 consumer_data
.type
= type
;
1292 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
1294 PERROR("allocating context");
1298 ctx
->consumer_error_socket
= -1;
1299 ctx
->consumer_metadata_socket
= -1;
1300 pthread_mutex_init(&ctx
->metadata_socket_lock
, NULL
);
1301 /* assign the callbacks */
1302 ctx
->on_buffer_ready
= buffer_ready
;
1303 ctx
->on_recv_channel
= recv_channel
;
1304 ctx
->on_recv_stream
= recv_stream
;
1305 ctx
->on_update_stream
= update_stream
;
1307 ctx
->consumer_data_pipe
= lttng_pipe_open(0);
1308 if (!ctx
->consumer_data_pipe
) {
1309 goto error_poll_pipe
;
1312 ctx
->consumer_wakeup_pipe
= lttng_pipe_open(0);
1313 if (!ctx
->consumer_wakeup_pipe
) {
1314 goto error_wakeup_pipe
;
1317 ret
= pipe(ctx
->consumer_should_quit
);
1319 PERROR("Error creating recv pipe");
1320 goto error_quit_pipe
;
1323 ret
= pipe(ctx
->consumer_channel_pipe
);
1325 PERROR("Error creating channel pipe");
1326 goto error_channel_pipe
;
1329 ctx
->consumer_metadata_pipe
= lttng_pipe_open(0);
1330 if (!ctx
->consumer_metadata_pipe
) {
1331 goto error_metadata_pipe
;
1336 error_metadata_pipe
:
1337 utils_close_pipe(ctx
->consumer_channel_pipe
);
1339 utils_close_pipe(ctx
->consumer_should_quit
);
1341 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1343 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1351 * Iterate over all streams of the hashtable and free them properly.
1353 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1355 struct lttng_ht_iter iter
;
1356 struct lttng_consumer_stream
*stream
;
1363 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1365 * Ignore return value since we are currently cleaning up so any error
1368 (void) consumer_del_stream(stream
, ht
);
1372 lttng_ht_destroy(ht
);
1376 * Iterate over all streams of the metadata hashtable and free them
1379 static void destroy_metadata_stream_ht(struct lttng_ht
*ht
)
1381 struct lttng_ht_iter iter
;
1382 struct lttng_consumer_stream
*stream
;
1389 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1391 * Ignore return value since we are currently cleaning up so any error
1394 (void) consumer_del_metadata_stream(stream
, ht
);
1398 lttng_ht_destroy(ht
);
1402 * Close all fds associated with the instance and free the context.
1404 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1408 DBG("Consumer destroying it. Closing everything.");
1414 destroy_data_stream_ht(data_ht
);
1415 destroy_metadata_stream_ht(metadata_ht
);
1417 ret
= close(ctx
->consumer_error_socket
);
1421 ret
= close(ctx
->consumer_metadata_socket
);
1425 utils_close_pipe(ctx
->consumer_channel_pipe
);
1426 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1427 lttng_pipe_destroy(ctx
->consumer_metadata_pipe
);
1428 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1429 utils_close_pipe(ctx
->consumer_should_quit
);
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
= close(stream
->index_fd
);
1596 PERROR("Closing index");
1599 stream
->index_fd
= -1;
1600 ret
= index_create_file(stream
->chan
->pathname
,
1601 stream
->name
, stream
->uid
, stream
->gid
,
1602 stream
->chan
->tracefile_size
,
1603 stream
->tracefile_count_current
);
1607 stream
->index_fd
= ret
;
1610 /* Reset current size because we just perform a rotation. */
1611 stream
->tracefile_size_current
= 0;
1612 stream
->out_fd_offset
= 0;
1615 stream
->tracefile_size_current
+= len
;
1617 index
->offset
= htobe64(stream
->out_fd_offset
);
1622 * This call guarantee that len or less is returned. It's impossible to
1623 * receive a ret value that is bigger than len.
1625 ret
= lttng_write(outfd
, mmap_base
+ mmap_offset
, len
);
1626 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, len
);
1627 if (ret
< 0 || ((size_t) ret
!= len
)) {
1629 * Report error to caller if nothing was written else at least send the
1637 /* Socket operation failed. We consider the relayd dead */
1638 if (errno
== EPIPE
|| errno
== EINVAL
|| errno
== EBADF
) {
1640 * This is possible if the fd is closed on the other side
1641 * (outfd) or any write problem. It can be verbose a bit for a
1642 * normal execution if for instance the relayd is stopped
1643 * abruptly. This can happen so set this to a DBG statement.
1645 DBG("Consumer mmap write detected relayd hang up");
1647 /* Unhandled error, print it and stop function right now. */
1648 PERROR("Error in write mmap (ret %zd != len %lu)", ret
, len
);
1652 stream
->output_written
+= ret
;
1654 /* This call is useless on a socket so better save a syscall. */
1656 /* This won't block, but will start writeout asynchronously */
1657 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, len
,
1658 SYNC_FILE_RANGE_WRITE
);
1659 stream
->out_fd_offset
+= len
;
1660 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1665 * This is a special case that the relayd has closed its socket. Let's
1666 * cleanup the relayd object and all associated streams.
1668 if (relayd
&& relayd_hang_up
) {
1669 cleanup_relayd(relayd
, ctx
);
1673 /* Unlock only if ctrl socket used */
1674 if (relayd
&& stream
->metadata_flag
) {
1675 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1683 * Splice the data from the ring buffer to the tracefile.
1685 * It must be called with the stream lock held.
1687 * Returns the number of bytes spliced.
1689 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1690 struct lttng_consumer_local_data
*ctx
,
1691 struct lttng_consumer_stream
*stream
, unsigned long len
,
1692 unsigned long padding
,
1693 struct ctf_packet_index
*index
)
1695 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1697 off_t orig_offset
= stream
->out_fd_offset
;
1698 int fd
= stream
->wait_fd
;
1699 /* Default is on the disk */
1700 int outfd
= stream
->out_fd
;
1701 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1703 unsigned int relayd_hang_up
= 0;
1705 switch (consumer_data
.type
) {
1706 case LTTNG_CONSUMER_KERNEL
:
1708 case LTTNG_CONSUMER32_UST
:
1709 case LTTNG_CONSUMER64_UST
:
1710 /* Not supported for user space tracing */
1713 ERR("Unknown consumer_data type");
1717 /* RCU lock for the relayd pointer */
1720 /* Flag that the current stream if set for network streaming. */
1721 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1722 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1723 if (relayd
== NULL
) {
1728 splice_pipe
= stream
->splice_pipe
;
1730 /* Write metadata stream id before payload */
1732 unsigned long total_len
= len
;
1734 if (stream
->metadata_flag
) {
1736 * Lock the control socket for the complete duration of the function
1737 * since from this point on we will use the socket.
1739 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1741 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
, relayd
,
1749 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1752 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1758 /* Use the returned socket. */
1761 /* No streaming, we have to set the len with the full padding */
1765 * Check if we need to change the tracefile before writing the packet.
1767 if (stream
->chan
->tracefile_size
> 0 &&
1768 (stream
->tracefile_size_current
+ len
) >
1769 stream
->chan
->tracefile_size
) {
1770 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1771 stream
->name
, stream
->chan
->tracefile_size
,
1772 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1773 stream
->out_fd
, &(stream
->tracefile_count_current
),
1777 ERR("Rotating output file");
1780 outfd
= stream
->out_fd
;
1782 if (stream
->index_fd
>= 0) {
1783 ret
= close(stream
->index_fd
);
1785 PERROR("Closing index");
1788 stream
->index_fd
= -1;
1789 ret
= index_create_file(stream
->chan
->pathname
,
1790 stream
->name
, stream
->uid
, stream
->gid
,
1791 stream
->chan
->tracefile_size
,
1792 stream
->tracefile_count_current
);
1797 stream
->index_fd
= ret
;
1800 /* Reset current size because we just perform a rotation. */
1801 stream
->tracefile_size_current
= 0;
1802 stream
->out_fd_offset
= 0;
1805 stream
->tracefile_size_current
+= len
;
1806 index
->offset
= htobe64(stream
->out_fd_offset
);
1810 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1811 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1812 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1813 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1814 DBG("splice chan to pipe, ret %zd", ret_splice
);
1815 if (ret_splice
< 0) {
1818 PERROR("Error in relay splice");
1822 /* Handle stream on the relayd if the output is on the network */
1823 if (relayd
&& stream
->metadata_flag
) {
1824 size_t metadata_payload_size
=
1825 sizeof(struct lttcomm_relayd_metadata_payload
);
1827 /* Update counter to fit the spliced data */
1828 ret_splice
+= metadata_payload_size
;
1829 len
+= metadata_payload_size
;
1831 * We do this so the return value can match the len passed as
1832 * argument to this function.
1834 written
-= metadata_payload_size
;
1837 /* Splice data out */
1838 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1839 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1840 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
1842 if (ret_splice
< 0) {
1847 } else if (ret_splice
> len
) {
1849 * We don't expect this code path to be executed but you never know
1850 * so this is an extra protection agains a buggy splice().
1853 written
+= ret_splice
;
1854 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice
,
1858 /* All good, update current len and continue. */
1862 /* This call is useless on a socket so better save a syscall. */
1864 /* This won't block, but will start writeout asynchronously */
1865 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1866 SYNC_FILE_RANGE_WRITE
);
1867 stream
->out_fd_offset
+= ret_splice
;
1869 stream
->output_written
+= ret_splice
;
1870 written
+= ret_splice
;
1873 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1879 * This is a special case that the relayd has closed its socket. Let's
1880 * cleanup the relayd object and all associated streams.
1882 if (relayd
&& relayd_hang_up
) {
1883 cleanup_relayd(relayd
, ctx
);
1884 /* Skip splice error so the consumer does not fail */
1889 /* send the appropriate error description to sessiond */
1892 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1895 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1898 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
1903 if (relayd
&& stream
->metadata_flag
) {
1904 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1912 * Take a snapshot for a specific fd
1914 * Returns 0 on success, < 0 on error
1916 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
1918 switch (consumer_data
.type
) {
1919 case LTTNG_CONSUMER_KERNEL
:
1920 return lttng_kconsumer_take_snapshot(stream
);
1921 case LTTNG_CONSUMER32_UST
:
1922 case LTTNG_CONSUMER64_UST
:
1923 return lttng_ustconsumer_take_snapshot(stream
);
1925 ERR("Unknown consumer_data type");
1932 * Get the produced position
1934 * Returns 0 on success, < 0 on error
1936 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
1939 switch (consumer_data
.type
) {
1940 case LTTNG_CONSUMER_KERNEL
:
1941 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
1942 case LTTNG_CONSUMER32_UST
:
1943 case LTTNG_CONSUMER64_UST
:
1944 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
1946 ERR("Unknown consumer_data type");
1952 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
1953 int sock
, struct pollfd
*consumer_sockpoll
)
1955 switch (consumer_data
.type
) {
1956 case LTTNG_CONSUMER_KERNEL
:
1957 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1958 case LTTNG_CONSUMER32_UST
:
1959 case LTTNG_CONSUMER64_UST
:
1960 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1962 ERR("Unknown consumer_data type");
1968 void lttng_consumer_close_all_metadata(void)
1970 switch (consumer_data
.type
) {
1971 case LTTNG_CONSUMER_KERNEL
:
1973 * The Kernel consumer has a different metadata scheme so we don't
1974 * close anything because the stream will be closed by the session
1978 case LTTNG_CONSUMER32_UST
:
1979 case LTTNG_CONSUMER64_UST
:
1981 * Close all metadata streams. The metadata hash table is passed and
1982 * this call iterates over it by closing all wakeup fd. This is safe
1983 * because at this point we are sure that the metadata producer is
1984 * either dead or blocked.
1986 lttng_ustconsumer_close_all_metadata(metadata_ht
);
1989 ERR("Unknown consumer_data type");
1995 * Clean up a metadata stream and free its memory.
1997 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
1998 struct lttng_ht
*ht
)
2000 struct lttng_consumer_channel
*free_chan
= NULL
;
2004 * This call should NEVER receive regular stream. It must always be
2005 * metadata stream and this is crucial for data structure synchronization.
2007 assert(stream
->metadata_flag
);
2009 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
2011 pthread_mutex_lock(&consumer_data
.lock
);
2012 pthread_mutex_lock(&stream
->chan
->lock
);
2013 pthread_mutex_lock(&stream
->lock
);
2015 /* Remove any reference to that stream. */
2016 consumer_stream_delete(stream
, ht
);
2018 /* Close down everything including the relayd if one. */
2019 consumer_stream_close(stream
);
2020 /* Destroy tracer buffers of the stream. */
2021 consumer_stream_destroy_buffers(stream
);
2023 /* Atomically decrement channel refcount since other threads can use it. */
2024 if (!uatomic_sub_return(&stream
->chan
->refcount
, 1)
2025 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
2026 /* Go for channel deletion! */
2027 free_chan
= stream
->chan
;
2031 * Nullify the stream reference so it is not used after deletion. The
2032 * channel lock MUST be acquired before being able to check for a NULL
2035 stream
->chan
->metadata_stream
= NULL
;
2037 pthread_mutex_unlock(&stream
->lock
);
2038 pthread_mutex_unlock(&stream
->chan
->lock
);
2039 pthread_mutex_unlock(&consumer_data
.lock
);
2042 consumer_del_channel(free_chan
);
2045 consumer_stream_free(stream
);
2049 * Action done with the metadata stream when adding it to the consumer internal
2050 * data structures to handle it.
2052 int consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
2054 struct lttng_ht
*ht
= metadata_ht
;
2056 struct lttng_ht_iter iter
;
2057 struct lttng_ht_node_u64
*node
;
2062 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2064 pthread_mutex_lock(&consumer_data
.lock
);
2065 pthread_mutex_lock(&stream
->chan
->lock
);
2066 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2067 pthread_mutex_lock(&stream
->lock
);
2070 * From here, refcounts are updated so be _careful_ when returning an error
2077 * Lookup the stream just to make sure it does not exist in our internal
2078 * state. This should NEVER happen.
2080 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2081 node
= lttng_ht_iter_get_node_u64(&iter
);
2085 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2086 * in terms of destroying the associated channel, because the action that
2087 * causes the count to become 0 also causes a stream to be added. The
2088 * channel deletion will thus be triggered by the following removal of this
2091 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2092 /* Increment refcount before decrementing nb_init_stream_left */
2094 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2097 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2099 lttng_ht_add_unique_u64(consumer_data
.stream_per_chan_id_ht
,
2100 &stream
->node_channel_id
);
2103 * Add stream to the stream_list_ht of the consumer data. No need to steal
2104 * the key since the HT does not use it and we allow to add redundant keys
2107 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2111 pthread_mutex_unlock(&stream
->lock
);
2112 pthread_mutex_unlock(&stream
->chan
->lock
);
2113 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2114 pthread_mutex_unlock(&consumer_data
.lock
);
2119 * Delete data stream that are flagged for deletion (endpoint_status).
2121 static void validate_endpoint_status_data_stream(void)
2123 struct lttng_ht_iter iter
;
2124 struct lttng_consumer_stream
*stream
;
2126 DBG("Consumer delete flagged data stream");
2129 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2130 /* Validate delete flag of the stream */
2131 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2134 /* Delete it right now */
2135 consumer_del_stream(stream
, data_ht
);
2141 * Delete metadata stream that are flagged for deletion (endpoint_status).
2143 static void validate_endpoint_status_metadata_stream(
2144 struct lttng_poll_event
*pollset
)
2146 struct lttng_ht_iter iter
;
2147 struct lttng_consumer_stream
*stream
;
2149 DBG("Consumer delete flagged metadata stream");
2154 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2155 /* Validate delete flag of the stream */
2156 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2160 * Remove from pollset so the metadata thread can continue without
2161 * blocking on a deleted stream.
2163 lttng_poll_del(pollset
, stream
->wait_fd
);
2165 /* Delete it right now */
2166 consumer_del_metadata_stream(stream
, metadata_ht
);
2172 * Thread polls on metadata file descriptor and write them on disk or on the
2175 void *consumer_thread_metadata_poll(void *data
)
2177 int ret
, i
, pollfd
, err
= -1;
2178 uint32_t revents
, nb_fd
;
2179 struct lttng_consumer_stream
*stream
= NULL
;
2180 struct lttng_ht_iter iter
;
2181 struct lttng_ht_node_u64
*node
;
2182 struct lttng_poll_event events
;
2183 struct lttng_consumer_local_data
*ctx
= data
;
2186 rcu_register_thread();
2188 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2190 if (testpoint(consumerd_thread_metadata
)) {
2191 goto error_testpoint
;
2194 health_code_update();
2196 DBG("Thread metadata poll started");
2198 /* Size is set to 1 for the consumer_metadata pipe */
2199 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2201 ERR("Poll set creation failed");
2205 ret
= lttng_poll_add(&events
,
2206 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2212 DBG("Metadata main loop started");
2216 health_code_update();
2217 health_poll_entry();
2218 DBG("Metadata poll wait");
2219 ret
= lttng_poll_wait(&events
, -1);
2220 DBG("Metadata poll return from wait with %d fd(s)",
2221 LTTNG_POLL_GETNB(&events
));
2223 DBG("Metadata event caught in thread");
2225 if (errno
== EINTR
) {
2226 ERR("Poll EINTR caught");
2229 if (LTTNG_POLL_GETNB(&events
) == 0) {
2230 err
= 0; /* All is OK */
2237 /* From here, the event is a metadata wait fd */
2238 for (i
= 0; i
< nb_fd
; i
++) {
2239 health_code_update();
2241 revents
= LTTNG_POLL_GETEV(&events
, i
);
2242 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2245 /* No activity for this FD (poll implementation). */
2249 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2250 if (revents
& LPOLLIN
) {
2253 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2254 &stream
, sizeof(stream
));
2255 if (pipe_len
< sizeof(stream
)) {
2257 PERROR("read metadata stream");
2260 * Remove the pipe from the poll set and continue the loop
2261 * since their might be data to consume.
2263 lttng_poll_del(&events
,
2264 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2265 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2269 /* A NULL stream means that the state has changed. */
2270 if (stream
== NULL
) {
2271 /* Check for deleted streams. */
2272 validate_endpoint_status_metadata_stream(&events
);
2276 DBG("Adding metadata stream %d to poll set",
2279 /* Add metadata stream to the global poll events list */
2280 lttng_poll_add(&events
, stream
->wait_fd
,
2281 LPOLLIN
| LPOLLPRI
| LPOLLHUP
);
2282 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2283 DBG("Metadata thread pipe hung up");
2285 * Remove the pipe from the poll set and continue the loop
2286 * since their might be data to consume.
2288 lttng_poll_del(&events
,
2289 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2290 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2293 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2297 /* Handle other stream */
2303 uint64_t tmp_id
= (uint64_t) pollfd
;
2305 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2307 node
= lttng_ht_iter_get_node_u64(&iter
);
2310 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2313 if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2314 /* Get the data out of the metadata file descriptor */
2315 DBG("Metadata available on fd %d", pollfd
);
2316 assert(stream
->wait_fd
== pollfd
);
2319 health_code_update();
2321 len
= ctx
->on_buffer_ready(stream
, ctx
);
2323 * We don't check the return value here since if we get
2324 * a negative len, it means an error occured thus we
2325 * simply remove it from the poll set and free the
2330 /* It's ok to have an unavailable sub-buffer */
2331 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2332 /* Clean up stream from consumer and free it. */
2333 lttng_poll_del(&events
, stream
->wait_fd
);
2334 consumer_del_metadata_stream(stream
, metadata_ht
);
2336 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2337 DBG("Metadata fd %d is hup|err.", pollfd
);
2338 if (!stream
->hangup_flush_done
2339 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2340 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2341 DBG("Attempting to flush and consume the UST buffers");
2342 lttng_ustconsumer_on_stream_hangup(stream
);
2344 /* We just flushed the stream now read it. */
2346 health_code_update();
2348 len
= ctx
->on_buffer_ready(stream
, ctx
);
2350 * We don't check the return value here since if we get
2351 * a negative len, it means an error occured thus we
2352 * simply remove it from the poll set and free the
2358 lttng_poll_del(&events
, stream
->wait_fd
);
2360 * This call update the channel states, closes file descriptors
2361 * and securely free the stream.
2363 consumer_del_metadata_stream(stream
, metadata_ht
);
2365 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2369 /* Release RCU lock for the stream looked up */
2377 DBG("Metadata poll thread exiting");
2379 lttng_poll_clean(&events
);
2384 ERR("Health error occurred in %s", __func__
);
2386 health_unregister(health_consumerd
);
2387 rcu_unregister_thread();
2392 * This thread polls the fds in the set to consume the data and write
2393 * it to tracefile if necessary.
2395 void *consumer_thread_data_poll(void *data
)
2397 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2398 struct pollfd
*pollfd
= NULL
;
2399 /* local view of the streams */
2400 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2401 /* local view of consumer_data.fds_count */
2403 struct lttng_consumer_local_data
*ctx
= data
;
2406 rcu_register_thread();
2408 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2410 if (testpoint(consumerd_thread_data
)) {
2411 goto error_testpoint
;
2414 health_code_update();
2416 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
*));
2417 if (local_stream
== NULL
) {
2418 PERROR("local_stream malloc");
2423 health_code_update();
2429 * the fds set has been updated, we need to update our
2430 * local array as well
2432 pthread_mutex_lock(&consumer_data
.lock
);
2433 if (consumer_data
.need_update
) {
2438 local_stream
= NULL
;
2441 * Allocate for all fds +1 for the consumer_data_pipe and +1 for
2444 pollfd
= zmalloc((consumer_data
.stream_count
+ 2) * sizeof(struct pollfd
));
2445 if (pollfd
== NULL
) {
2446 PERROR("pollfd malloc");
2447 pthread_mutex_unlock(&consumer_data
.lock
);
2451 local_stream
= zmalloc((consumer_data
.stream_count
+ 2) *
2452 sizeof(struct lttng_consumer_stream
*));
2453 if (local_stream
== NULL
) {
2454 PERROR("local_stream malloc");
2455 pthread_mutex_unlock(&consumer_data
.lock
);
2458 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2461 ERR("Error in allocating pollfd or local_outfds");
2462 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2463 pthread_mutex_unlock(&consumer_data
.lock
);
2467 consumer_data
.need_update
= 0;
2469 pthread_mutex_unlock(&consumer_data
.lock
);
2471 /* No FDs and consumer_quit, consumer_cleanup the thread */
2472 if (nb_fd
== 0 && consumer_quit
== 1) {
2473 err
= 0; /* All is OK */
2476 /* poll on the array of fds */
2478 DBG("polling on %d fd", nb_fd
+ 2);
2479 health_poll_entry();
2480 num_rdy
= poll(pollfd
, nb_fd
+ 2, -1);
2482 DBG("poll num_rdy : %d", num_rdy
);
2483 if (num_rdy
== -1) {
2485 * Restart interrupted system call.
2487 if (errno
== EINTR
) {
2490 PERROR("Poll error");
2491 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2493 } else if (num_rdy
== 0) {
2494 DBG("Polling thread timed out");
2499 * If the consumer_data_pipe triggered poll go directly to the
2500 * beginning of the loop to update the array. We want to prioritize
2501 * array update over low-priority reads.
2503 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2504 ssize_t pipe_readlen
;
2506 DBG("consumer_data_pipe wake up");
2507 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2508 &new_stream
, sizeof(new_stream
));
2509 if (pipe_readlen
< sizeof(new_stream
)) {
2510 PERROR("Consumer data pipe");
2511 /* Continue so we can at least handle the current stream(s). */
2516 * If the stream is NULL, just ignore it. It's also possible that
2517 * the sessiond poll thread changed the consumer_quit state and is
2518 * waking us up to test it.
2520 if (new_stream
== NULL
) {
2521 validate_endpoint_status_data_stream();
2525 /* Continue to update the local streams and handle prio ones */
2529 /* Handle wakeup pipe. */
2530 if (pollfd
[nb_fd
+ 1].revents
& (POLLIN
| POLLPRI
)) {
2532 ssize_t pipe_readlen
;
2534 pipe_readlen
= lttng_pipe_read(ctx
->consumer_wakeup_pipe
, &dummy
,
2536 if (pipe_readlen
< 0) {
2537 PERROR("Consumer data wakeup pipe");
2539 /* We've been awakened to handle stream(s). */
2540 ctx
->has_wakeup
= 0;
2543 /* Take care of high priority channels first. */
2544 for (i
= 0; i
< nb_fd
; i
++) {
2545 health_code_update();
2547 if (local_stream
[i
] == NULL
) {
2550 if (pollfd
[i
].revents
& POLLPRI
) {
2551 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2553 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2554 /* it's ok to have an unavailable sub-buffer */
2555 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2556 /* Clean the stream and free it. */
2557 consumer_del_stream(local_stream
[i
], data_ht
);
2558 local_stream
[i
] = NULL
;
2559 } else if (len
> 0) {
2560 local_stream
[i
]->data_read
= 1;
2566 * If we read high prio channel in this loop, try again
2567 * for more high prio data.
2573 /* Take care of low priority channels. */
2574 for (i
= 0; i
< nb_fd
; i
++) {
2575 health_code_update();
2577 if (local_stream
[i
] == NULL
) {
2580 if ((pollfd
[i
].revents
& POLLIN
) ||
2581 local_stream
[i
]->hangup_flush_done
||
2582 local_stream
[i
]->has_data
) {
2583 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2584 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2585 /* it's ok to have an unavailable sub-buffer */
2586 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2587 /* Clean the stream and free it. */
2588 consumer_del_stream(local_stream
[i
], data_ht
);
2589 local_stream
[i
] = NULL
;
2590 } else if (len
> 0) {
2591 local_stream
[i
]->data_read
= 1;
2596 /* Handle hangup and errors */
2597 for (i
= 0; i
< nb_fd
; i
++) {
2598 health_code_update();
2600 if (local_stream
[i
] == NULL
) {
2603 if (!local_stream
[i
]->hangup_flush_done
2604 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2605 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2606 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2607 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2609 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2610 /* Attempt read again, for the data we just flushed. */
2611 local_stream
[i
]->data_read
= 1;
2614 * If the poll flag is HUP/ERR/NVAL and we have
2615 * read no data in this pass, we can remove the
2616 * stream from its hash table.
2618 if ((pollfd
[i
].revents
& POLLHUP
)) {
2619 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2620 if (!local_stream
[i
]->data_read
) {
2621 consumer_del_stream(local_stream
[i
], data_ht
);
2622 local_stream
[i
] = NULL
;
2625 } else if (pollfd
[i
].revents
& POLLERR
) {
2626 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2627 if (!local_stream
[i
]->data_read
) {
2628 consumer_del_stream(local_stream
[i
], data_ht
);
2629 local_stream
[i
] = NULL
;
2632 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2633 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2634 if (!local_stream
[i
]->data_read
) {
2635 consumer_del_stream(local_stream
[i
], data_ht
);
2636 local_stream
[i
] = NULL
;
2640 if (local_stream
[i
] != NULL
) {
2641 local_stream
[i
]->data_read
= 0;
2648 DBG("polling thread exiting");
2653 * Close the write side of the pipe so epoll_wait() in
2654 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2655 * read side of the pipe. If we close them both, epoll_wait strangely does
2656 * not return and could create a endless wait period if the pipe is the
2657 * only tracked fd in the poll set. The thread will take care of closing
2660 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2665 ERR("Health error occurred in %s", __func__
);
2667 health_unregister(health_consumerd
);
2669 rcu_unregister_thread();
2674 * Close wake-up end of each stream belonging to the channel. This will
2675 * allow the poll() on the stream read-side to detect when the
2676 * write-side (application) finally closes them.
2679 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2681 struct lttng_ht
*ht
;
2682 struct lttng_consumer_stream
*stream
;
2683 struct lttng_ht_iter iter
;
2685 ht
= consumer_data
.stream_per_chan_id_ht
;
2688 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2689 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2690 ht
->match_fct
, &channel
->key
,
2691 &iter
.iter
, stream
, node_channel_id
.node
) {
2693 * Protect against teardown with mutex.
2695 pthread_mutex_lock(&stream
->lock
);
2696 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2699 switch (consumer_data
.type
) {
2700 case LTTNG_CONSUMER_KERNEL
:
2702 case LTTNG_CONSUMER32_UST
:
2703 case LTTNG_CONSUMER64_UST
:
2704 if (stream
->metadata_flag
) {
2705 /* Safe and protected by the stream lock. */
2706 lttng_ustconsumer_close_metadata(stream
->chan
);
2709 * Note: a mutex is taken internally within
2710 * liblttng-ust-ctl to protect timer wakeup_fd
2711 * use from concurrent close.
2713 lttng_ustconsumer_close_stream_wakeup(stream
);
2717 ERR("Unknown consumer_data type");
2721 pthread_mutex_unlock(&stream
->lock
);
2726 static void destroy_channel_ht(struct lttng_ht
*ht
)
2728 struct lttng_ht_iter iter
;
2729 struct lttng_consumer_channel
*channel
;
2737 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2738 ret
= lttng_ht_del(ht
, &iter
);
2743 lttng_ht_destroy(ht
);
2747 * This thread polls the channel fds to detect when they are being
2748 * closed. It closes all related streams if the channel is detected as
2749 * closed. It is currently only used as a shim layer for UST because the
2750 * consumerd needs to keep the per-stream wakeup end of pipes open for
2753 void *consumer_thread_channel_poll(void *data
)
2755 int ret
, i
, pollfd
, err
= -1;
2756 uint32_t revents
, nb_fd
;
2757 struct lttng_consumer_channel
*chan
= NULL
;
2758 struct lttng_ht_iter iter
;
2759 struct lttng_ht_node_u64
*node
;
2760 struct lttng_poll_event events
;
2761 struct lttng_consumer_local_data
*ctx
= data
;
2762 struct lttng_ht
*channel_ht
;
2764 rcu_register_thread();
2766 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
2768 if (testpoint(consumerd_thread_channel
)) {
2769 goto error_testpoint
;
2772 health_code_update();
2774 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2776 /* ENOMEM at this point. Better to bail out. */
2780 DBG("Thread channel poll started");
2782 /* Size is set to 1 for the consumer_channel pipe */
2783 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2785 ERR("Poll set creation failed");
2789 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2795 DBG("Channel main loop started");
2799 health_code_update();
2800 DBG("Channel poll wait");
2801 health_poll_entry();
2802 ret
= lttng_poll_wait(&events
, -1);
2803 DBG("Channel poll return from wait with %d fd(s)",
2804 LTTNG_POLL_GETNB(&events
));
2806 DBG("Channel event caught in thread");
2808 if (errno
== EINTR
) {
2809 ERR("Poll EINTR caught");
2812 if (LTTNG_POLL_GETNB(&events
) == 0) {
2813 err
= 0; /* All is OK */
2820 /* From here, the event is a channel wait fd */
2821 for (i
= 0; i
< nb_fd
; i
++) {
2822 health_code_update();
2824 revents
= LTTNG_POLL_GETEV(&events
, i
);
2825 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2828 /* No activity for this FD (poll implementation). */
2832 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2833 if (revents
& LPOLLIN
) {
2834 enum consumer_channel_action action
;
2837 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
2840 ERR("Error reading channel pipe");
2842 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2847 case CONSUMER_CHANNEL_ADD
:
2848 DBG("Adding channel %d to poll set",
2851 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
2854 lttng_ht_add_unique_u64(channel_ht
,
2855 &chan
->wait_fd_node
);
2857 /* Add channel to the global poll events list */
2858 lttng_poll_add(&events
, chan
->wait_fd
,
2859 LPOLLERR
| LPOLLHUP
);
2861 case CONSUMER_CHANNEL_DEL
:
2864 * This command should never be called if the channel
2865 * has streams monitored by either the data or metadata
2866 * thread. The consumer only notify this thread with a
2867 * channel del. command if it receives a destroy
2868 * channel command from the session daemon that send it
2869 * if a command prior to the GET_CHANNEL failed.
2873 chan
= consumer_find_channel(key
);
2876 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
2879 lttng_poll_del(&events
, chan
->wait_fd
);
2880 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
2881 ret
= lttng_ht_del(channel_ht
, &iter
);
2884 switch (consumer_data
.type
) {
2885 case LTTNG_CONSUMER_KERNEL
:
2887 case LTTNG_CONSUMER32_UST
:
2888 case LTTNG_CONSUMER64_UST
:
2889 health_code_update();
2890 /* Destroy streams that might have been left in the stream list. */
2891 clean_channel_stream_list(chan
);
2894 ERR("Unknown consumer_data type");
2899 * Release our own refcount. Force channel deletion even if
2900 * streams were not initialized.
2902 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
2903 consumer_del_channel(chan
);
2908 case CONSUMER_CHANNEL_QUIT
:
2910 * Remove the pipe from the poll set and continue the loop
2911 * since their might be data to consume.
2913 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2916 ERR("Unknown action");
2919 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2920 DBG("Channel thread pipe hung up");
2922 * Remove the pipe from the poll set and continue the loop
2923 * since their might be data to consume.
2925 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2928 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2932 /* Handle other stream */
2938 uint64_t tmp_id
= (uint64_t) pollfd
;
2940 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
2942 node
= lttng_ht_iter_get_node_u64(&iter
);
2945 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
2948 /* Check for error event */
2949 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2950 DBG("Channel fd %d is hup|err.", pollfd
);
2952 lttng_poll_del(&events
, chan
->wait_fd
);
2953 ret
= lttng_ht_del(channel_ht
, &iter
);
2957 * This will close the wait fd for each stream associated to
2958 * this channel AND monitored by the data/metadata thread thus
2959 * will be clean by the right thread.
2961 consumer_close_channel_streams(chan
);
2963 /* Release our own refcount */
2964 if (!uatomic_sub_return(&chan
->refcount
, 1)
2965 && !uatomic_read(&chan
->nb_init_stream_left
)) {
2966 consumer_del_channel(chan
);
2969 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2974 /* Release RCU lock for the channel looked up */
2982 lttng_poll_clean(&events
);
2984 destroy_channel_ht(channel_ht
);
2987 DBG("Channel poll thread exiting");
2990 ERR("Health error occurred in %s", __func__
);
2992 health_unregister(health_consumerd
);
2993 rcu_unregister_thread();
2997 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
2998 struct pollfd
*sockpoll
, int client_socket
)
3005 ret
= lttng_consumer_poll_socket(sockpoll
);
3009 DBG("Metadata connection on client_socket");
3011 /* Blocking call, waiting for transmission */
3012 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
3013 if (ctx
->consumer_metadata_socket
< 0) {
3014 WARN("On accept metadata");
3025 * This thread listens on the consumerd socket and receives the file
3026 * descriptors from the session daemon.
3028 void *consumer_thread_sessiond_poll(void *data
)
3030 int sock
= -1, client_socket
, ret
, err
= -1;
3032 * structure to poll for incoming data on communication socket avoids
3033 * making blocking sockets.
3035 struct pollfd consumer_sockpoll
[2];
3036 struct lttng_consumer_local_data
*ctx
= data
;
3038 rcu_register_thread();
3040 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
3042 if (testpoint(consumerd_thread_sessiond
)) {
3043 goto error_testpoint
;
3046 health_code_update();
3048 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
3049 unlink(ctx
->consumer_command_sock_path
);
3050 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
3051 if (client_socket
< 0) {
3052 ERR("Cannot create command socket");
3056 ret
= lttcomm_listen_unix_sock(client_socket
);
3061 DBG("Sending ready command to lttng-sessiond");
3062 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3063 /* return < 0 on error, but == 0 is not fatal */
3065 ERR("Error sending ready command to lttng-sessiond");
3069 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3070 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3071 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3072 consumer_sockpoll
[1].fd
= client_socket
;
3073 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3075 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3083 DBG("Connection on client_socket");
3085 /* Blocking call, waiting for transmission */
3086 sock
= lttcomm_accept_unix_sock(client_socket
);
3093 * Setup metadata socket which is the second socket connection on the
3094 * command unix socket.
3096 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3105 /* This socket is not useful anymore. */
3106 ret
= close(client_socket
);
3108 PERROR("close client_socket");
3112 /* update the polling structure to poll on the established socket */
3113 consumer_sockpoll
[1].fd
= sock
;
3114 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3117 health_code_update();
3119 health_poll_entry();
3120 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3129 DBG("Incoming command on sock");
3130 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3133 * This could simply be a session daemon quitting. Don't output
3136 DBG("Communication interrupted on command socket");
3140 if (consumer_quit
) {
3141 DBG("consumer_thread_receive_fds received quit from signal");
3142 err
= 0; /* All is OK */
3145 DBG("received command on sock");
3151 DBG("Consumer thread sessiond poll exiting");
3154 * Close metadata streams since the producer is the session daemon which
3157 * NOTE: for now, this only applies to the UST tracer.
3159 lttng_consumer_close_all_metadata();
3162 * when all fds have hung up, the polling thread
3168 * Notify the data poll thread to poll back again and test the
3169 * consumer_quit state that we just set so to quit gracefully.
3171 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3173 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3175 notify_health_quit_pipe(health_quit_pipe
);
3177 /* Cleaning up possibly open sockets. */
3181 PERROR("close sock sessiond poll");
3184 if (client_socket
>= 0) {
3185 ret
= close(client_socket
);
3187 PERROR("close client_socket sessiond poll");
3194 ERR("Health error occurred in %s", __func__
);
3196 health_unregister(health_consumerd
);
3198 rcu_unregister_thread();
3202 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3203 struct lttng_consumer_local_data
*ctx
)
3207 pthread_mutex_lock(&stream
->lock
);
3208 if (stream
->metadata_flag
) {
3209 pthread_mutex_lock(&stream
->metadata_rdv_lock
);
3212 switch (consumer_data
.type
) {
3213 case LTTNG_CONSUMER_KERNEL
:
3214 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
3216 case LTTNG_CONSUMER32_UST
:
3217 case LTTNG_CONSUMER64_UST
:
3218 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
3221 ERR("Unknown consumer_data type");
3227 if (stream
->metadata_flag
) {
3228 pthread_cond_broadcast(&stream
->metadata_rdv
);
3229 pthread_mutex_unlock(&stream
->metadata_rdv_lock
);
3231 pthread_mutex_unlock(&stream
->lock
);
3235 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3237 switch (consumer_data
.type
) {
3238 case LTTNG_CONSUMER_KERNEL
:
3239 return lttng_kconsumer_on_recv_stream(stream
);
3240 case LTTNG_CONSUMER32_UST
:
3241 case LTTNG_CONSUMER64_UST
:
3242 return lttng_ustconsumer_on_recv_stream(stream
);
3244 ERR("Unknown consumer_data type");
3251 * Allocate and set consumer data hash tables.
3253 int lttng_consumer_init(void)
3255 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3256 if (!consumer_data
.channel_ht
) {
3260 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3261 if (!consumer_data
.relayd_ht
) {
3265 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3266 if (!consumer_data
.stream_list_ht
) {
3270 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3271 if (!consumer_data
.stream_per_chan_id_ht
) {
3275 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3280 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3292 * Process the ADD_RELAYD command receive by a consumer.
3294 * This will create a relayd socket pair and add it to the relayd hash table.
3295 * The caller MUST acquire a RCU read side lock before calling it.
3297 int consumer_add_relayd_socket(uint64_t net_seq_idx
, int sock_type
,
3298 struct lttng_consumer_local_data
*ctx
, int sock
,
3299 struct pollfd
*consumer_sockpoll
,
3300 struct lttcomm_relayd_sock
*relayd_sock
, uint64_t sessiond_id
,
3301 uint64_t relayd_session_id
)
3303 int fd
= -1, ret
= -1, relayd_created
= 0;
3304 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3305 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3308 assert(relayd_sock
);
3310 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", net_seq_idx
);
3312 /* Get relayd reference if exists. */
3313 relayd
= consumer_find_relayd(net_seq_idx
);
3314 if (relayd
== NULL
) {
3315 assert(sock_type
== LTTNG_STREAM_CONTROL
);
3316 /* Not found. Allocate one. */
3317 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3318 if (relayd
== NULL
) {
3320 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3323 relayd
->sessiond_session_id
= sessiond_id
;
3328 * This code path MUST continue to the consumer send status message to
3329 * we can notify the session daemon and continue our work without
3330 * killing everything.
3334 * relayd key should never be found for control socket.
3336 assert(sock_type
!= LTTNG_STREAM_CONTROL
);
3339 /* First send a status message before receiving the fds. */
3340 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3342 /* Somehow, the session daemon is not responding anymore. */
3343 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3344 goto error_nosignal
;
3347 /* Poll on consumer socket. */
3348 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3350 /* Needing to exit in the middle of a command: error. */
3351 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3353 goto error_nosignal
;
3356 /* Get relayd socket from session daemon */
3357 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3358 if (ret
!= sizeof(fd
)) {
3360 fd
= -1; /* Just in case it gets set with an invalid value. */
3363 * Failing to receive FDs might indicate a major problem such as
3364 * reaching a fd limit during the receive where the kernel returns a
3365 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3366 * don't take any chances and stop everything.
3368 * XXX: Feature request #558 will fix that and avoid this possible
3369 * issue when reaching the fd limit.
3371 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3372 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3376 /* Copy socket information and received FD */
3377 switch (sock_type
) {
3378 case LTTNG_STREAM_CONTROL
:
3379 /* Copy received lttcomm socket */
3380 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3381 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3382 /* Handle create_sock error. */
3384 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3388 * Close the socket created internally by
3389 * lttcomm_create_sock, so we can replace it by the one
3390 * received from sessiond.
3392 if (close(relayd
->control_sock
.sock
.fd
)) {
3396 /* Assign new file descriptor */
3397 relayd
->control_sock
.sock
.fd
= fd
;
3398 fd
= -1; /* For error path */
3399 /* Assign version values. */
3400 relayd
->control_sock
.major
= relayd_sock
->major
;
3401 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3403 relayd
->relayd_session_id
= relayd_session_id
;
3406 case LTTNG_STREAM_DATA
:
3407 /* Copy received lttcomm socket */
3408 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3409 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3410 /* Handle create_sock error. */
3412 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3416 * Close the socket created internally by
3417 * lttcomm_create_sock, so we can replace it by the one
3418 * received from sessiond.
3420 if (close(relayd
->data_sock
.sock
.fd
)) {
3424 /* Assign new file descriptor */
3425 relayd
->data_sock
.sock
.fd
= fd
;
3426 fd
= -1; /* for eventual error paths */
3427 /* Assign version values. */
3428 relayd
->data_sock
.major
= relayd_sock
->major
;
3429 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3432 ERR("Unknown relayd socket type (%d)", sock_type
);
3434 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3438 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3439 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3440 relayd
->net_seq_idx
, fd
);
3442 /* We successfully added the socket. Send status back. */
3443 ret
= consumer_send_status_msg(sock
, ret_code
);
3445 /* Somehow, the session daemon is not responding anymore. */
3446 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3447 goto error_nosignal
;
3451 * Add relayd socket pair to consumer data hashtable. If object already
3452 * exists or on error, the function gracefully returns.
3460 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3461 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3465 /* Close received socket if valid. */
3468 PERROR("close received socket");
3472 if (relayd_created
) {
3480 * Try to lock the stream mutex.
3482 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
3484 static int stream_try_lock(struct lttng_consumer_stream
*stream
)
3491 * Try to lock the stream mutex. On failure, we know that the stream is
3492 * being used else where hence there is data still being extracted.
3494 ret
= pthread_mutex_trylock(&stream
->lock
);
3496 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
3508 * Search for a relayd associated to the session id and return the reference.
3510 * A rcu read side lock MUST be acquire before calling this function and locked
3511 * until the relayd object is no longer necessary.
3513 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3515 struct lttng_ht_iter iter
;
3516 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3518 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3519 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3522 * Check by sessiond id which is unique here where the relayd session
3523 * id might not be when having multiple relayd.
3525 if (relayd
->sessiond_session_id
== id
) {
3526 /* Found the relayd. There can be only one per id. */
3538 * Check if for a given session id there is still data needed to be extract
3541 * Return 1 if data is pending or else 0 meaning ready to be read.
3543 int consumer_data_pending(uint64_t id
)
3546 struct lttng_ht_iter iter
;
3547 struct lttng_ht
*ht
;
3548 struct lttng_consumer_stream
*stream
;
3549 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3550 int (*data_pending
)(struct lttng_consumer_stream
*);
3552 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3555 pthread_mutex_lock(&consumer_data
.lock
);
3557 switch (consumer_data
.type
) {
3558 case LTTNG_CONSUMER_KERNEL
:
3559 data_pending
= lttng_kconsumer_data_pending
;
3561 case LTTNG_CONSUMER32_UST
:
3562 case LTTNG_CONSUMER64_UST
:
3563 data_pending
= lttng_ustconsumer_data_pending
;
3566 ERR("Unknown consumer data type");
3570 /* Ease our life a bit */
3571 ht
= consumer_data
.stream_list_ht
;
3573 relayd
= find_relayd_by_session_id(id
);
3575 /* Send init command for data pending. */
3576 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3577 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3578 relayd
->relayd_session_id
);
3579 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3581 /* Communication error thus the relayd so no data pending. */
3582 goto data_not_pending
;
3586 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3587 ht
->hash_fct(&id
, lttng_ht_seed
),
3589 &iter
.iter
, stream
, node_session_id
.node
) {
3590 /* If this call fails, the stream is being used hence data pending. */
3591 ret
= stream_try_lock(stream
);
3597 * A removed node from the hash table indicates that the stream has
3598 * been deleted thus having a guarantee that the buffers are closed
3599 * on the consumer side. However, data can still be transmitted
3600 * over the network so don't skip the relayd check.
3602 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3604 /* Check the stream if there is data in the buffers. */
3605 ret
= data_pending(stream
);
3607 pthread_mutex_unlock(&stream
->lock
);
3614 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3615 if (stream
->metadata_flag
) {
3616 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3617 stream
->relayd_stream_id
);
3619 ret
= relayd_data_pending(&relayd
->control_sock
,
3620 stream
->relayd_stream_id
,
3621 stream
->next_net_seq_num
- 1);
3623 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3625 pthread_mutex_unlock(&stream
->lock
);
3629 pthread_mutex_unlock(&stream
->lock
);
3633 unsigned int is_data_inflight
= 0;
3635 /* Send init command for data pending. */
3636 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3637 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3638 relayd
->relayd_session_id
, &is_data_inflight
);
3639 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3641 goto data_not_pending
;
3643 if (is_data_inflight
) {
3649 * Finding _no_ node in the hash table and no inflight data means that the
3650 * stream(s) have been removed thus data is guaranteed to be available for
3651 * analysis from the trace files.
3655 /* Data is available to be read by a viewer. */
3656 pthread_mutex_unlock(&consumer_data
.lock
);
3661 /* Data is still being extracted from buffers. */
3662 pthread_mutex_unlock(&consumer_data
.lock
);
3668 * Send a ret code status message to the sessiond daemon.
3670 * Return the sendmsg() return value.
3672 int consumer_send_status_msg(int sock
, int ret_code
)
3674 struct lttcomm_consumer_status_msg msg
;
3676 memset(&msg
, 0, sizeof(msg
));
3677 msg
.ret_code
= ret_code
;
3679 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3683 * Send a channel status message to the sessiond daemon.
3685 * Return the sendmsg() return value.
3687 int consumer_send_status_channel(int sock
,
3688 struct lttng_consumer_channel
*channel
)
3690 struct lttcomm_consumer_status_channel msg
;
3694 memset(&msg
, 0, sizeof(msg
));
3696 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3698 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3699 msg
.key
= channel
->key
;
3700 msg
.stream_count
= channel
->streams
.count
;
3703 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3706 unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos
,
3707 unsigned long produced_pos
, uint64_t nb_packets_per_stream
,
3708 uint64_t max_sb_size
)
3710 unsigned long start_pos
;
3712 if (!nb_packets_per_stream
) {
3713 return consumed_pos
; /* Grab everything */
3715 start_pos
= produced_pos
- offset_align_floor(produced_pos
, max_sb_size
);
3716 start_pos
-= max_sb_size
* nb_packets_per_stream
;
3717 if ((long) (start_pos
- consumed_pos
) < 0) {
3718 return consumed_pos
; /* Grab everything */