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
);
1258 WARN("posix_fadvise() error (%i)", ret
);
1263 * Initialise the necessary environnement :
1264 * - create a new context
1265 * - create the poll_pipe
1266 * - create the should_quit pipe (for signal handler)
1267 * - create the thread pipe (for splice)
1269 * Takes a function pointer as argument, this function is called when data is
1270 * available on a buffer. This function is responsible to do the
1271 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1272 * buffer configuration and then kernctl_put_next_subbuf at the end.
1274 * Returns a pointer to the new context or NULL on error.
1276 struct lttng_consumer_local_data
*lttng_consumer_create(
1277 enum lttng_consumer_type type
,
1278 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1279 struct lttng_consumer_local_data
*ctx
),
1280 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1281 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1282 int (*update_stream
)(uint64_t stream_key
, uint32_t state
))
1285 struct lttng_consumer_local_data
*ctx
;
1287 assert(consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1288 consumer_data
.type
== type
);
1289 consumer_data
.type
= type
;
1291 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
1293 PERROR("allocating context");
1297 ctx
->consumer_error_socket
= -1;
1298 ctx
->consumer_metadata_socket
= -1;
1299 pthread_mutex_init(&ctx
->metadata_socket_lock
, NULL
);
1300 /* assign the callbacks */
1301 ctx
->on_buffer_ready
= buffer_ready
;
1302 ctx
->on_recv_channel
= recv_channel
;
1303 ctx
->on_recv_stream
= recv_stream
;
1304 ctx
->on_update_stream
= update_stream
;
1306 ctx
->consumer_data_pipe
= lttng_pipe_open(0);
1307 if (!ctx
->consumer_data_pipe
) {
1308 goto error_poll_pipe
;
1311 ctx
->consumer_wakeup_pipe
= lttng_pipe_open(0);
1312 if (!ctx
->consumer_wakeup_pipe
) {
1313 goto error_wakeup_pipe
;
1316 ret
= pipe(ctx
->consumer_should_quit
);
1318 PERROR("Error creating recv pipe");
1319 goto error_quit_pipe
;
1322 ret
= pipe(ctx
->consumer_channel_pipe
);
1324 PERROR("Error creating channel pipe");
1325 goto error_channel_pipe
;
1328 ctx
->consumer_metadata_pipe
= lttng_pipe_open(0);
1329 if (!ctx
->consumer_metadata_pipe
) {
1330 goto error_metadata_pipe
;
1335 error_metadata_pipe
:
1336 utils_close_pipe(ctx
->consumer_channel_pipe
);
1338 utils_close_pipe(ctx
->consumer_should_quit
);
1340 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1342 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1350 * Iterate over all streams of the hashtable and free them properly.
1352 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1354 struct lttng_ht_iter iter
;
1355 struct lttng_consumer_stream
*stream
;
1362 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1364 * Ignore return value since we are currently cleaning up so any error
1367 (void) consumer_del_stream(stream
, ht
);
1371 lttng_ht_destroy(ht
);
1375 * Iterate over all streams of the metadata hashtable and free them
1378 static void destroy_metadata_stream_ht(struct lttng_ht
*ht
)
1380 struct lttng_ht_iter iter
;
1381 struct lttng_consumer_stream
*stream
;
1388 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1390 * Ignore return value since we are currently cleaning up so any error
1393 (void) consumer_del_metadata_stream(stream
, ht
);
1397 lttng_ht_destroy(ht
);
1401 * Close all fds associated with the instance and free the context.
1403 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1407 DBG("Consumer destroying it. Closing everything.");
1413 destroy_data_stream_ht(data_ht
);
1414 destroy_metadata_stream_ht(metadata_ht
);
1416 ret
= close(ctx
->consumer_error_socket
);
1420 ret
= close(ctx
->consumer_metadata_socket
);
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
);
1430 unlink(ctx
->consumer_command_sock_path
);
1435 * Write the metadata stream id on the specified file descriptor.
1437 static int write_relayd_metadata_id(int fd
,
1438 struct lttng_consumer_stream
*stream
,
1439 struct consumer_relayd_sock_pair
*relayd
, unsigned long padding
)
1442 struct lttcomm_relayd_metadata_payload hdr
;
1444 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1445 hdr
.padding_size
= htobe32(padding
);
1446 ret
= lttng_write(fd
, (void *) &hdr
, sizeof(hdr
));
1447 if (ret
< sizeof(hdr
)) {
1449 * This error means that the fd's end is closed so ignore the perror
1450 * not to clubber the error output since this can happen in a normal
1453 if (errno
!= EPIPE
) {
1454 PERROR("write metadata stream id");
1456 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1458 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1459 * handle writting the missing part so report that as an error and
1460 * don't lie to the caller.
1465 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1466 stream
->relayd_stream_id
, padding
);
1473 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1474 * core function for writing trace buffers to either the local filesystem or
1477 * It must be called with the stream lock held.
1479 * Careful review MUST be put if any changes occur!
1481 * Returns the number of bytes written
1483 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1484 struct lttng_consumer_local_data
*ctx
,
1485 struct lttng_consumer_stream
*stream
, unsigned long len
,
1486 unsigned long padding
,
1487 struct ctf_packet_index
*index
)
1489 unsigned long mmap_offset
;
1492 off_t orig_offset
= stream
->out_fd_offset
;
1493 /* Default is on the disk */
1494 int outfd
= stream
->out_fd
;
1495 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1496 unsigned int relayd_hang_up
= 0;
1498 /* RCU lock for the relayd pointer */
1501 /* Flag that the current stream if set for network streaming. */
1502 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1503 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1504 if (relayd
== NULL
) {
1510 /* get the offset inside the fd to mmap */
1511 switch (consumer_data
.type
) {
1512 case LTTNG_CONSUMER_KERNEL
:
1513 mmap_base
= stream
->mmap_base
;
1514 ret
= kernctl_get_mmap_read_offset(stream
->wait_fd
, &mmap_offset
);
1517 PERROR("tracer ctl get_mmap_read_offset");
1521 case LTTNG_CONSUMER32_UST
:
1522 case LTTNG_CONSUMER64_UST
:
1523 mmap_base
= lttng_ustctl_get_mmap_base(stream
);
1525 ERR("read mmap get mmap base for stream %s", stream
->name
);
1529 ret
= lttng_ustctl_get_mmap_read_offset(stream
, &mmap_offset
);
1531 PERROR("tracer ctl get_mmap_read_offset");
1537 ERR("Unknown consumer_data type");
1541 /* Handle stream on the relayd if the output is on the network */
1543 unsigned long netlen
= len
;
1546 * Lock the control socket for the complete duration of the function
1547 * since from this point on we will use the socket.
1549 if (stream
->metadata_flag
) {
1550 /* Metadata requires the control socket. */
1551 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1552 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1555 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1560 /* Use the returned socket. */
1563 /* Write metadata stream id before payload */
1564 if (stream
->metadata_flag
) {
1565 ret
= write_relayd_metadata_id(outfd
, stream
, relayd
, padding
);
1572 /* No streaming, we have to set the len with the full padding */
1576 * Check if we need to change the tracefile before writing the packet.
1578 if (stream
->chan
->tracefile_size
> 0 &&
1579 (stream
->tracefile_size_current
+ len
) >
1580 stream
->chan
->tracefile_size
) {
1581 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1582 stream
->name
, stream
->chan
->tracefile_size
,
1583 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1584 stream
->out_fd
, &(stream
->tracefile_count_current
),
1587 ERR("Rotating output file");
1590 outfd
= stream
->out_fd
;
1592 if (stream
->index_fd
>= 0) {
1593 ret
= close(stream
->index_fd
);
1595 PERROR("Closing index");
1598 stream
->index_fd
= -1;
1599 ret
= index_create_file(stream
->chan
->pathname
,
1600 stream
->name
, stream
->uid
, stream
->gid
,
1601 stream
->chan
->tracefile_size
,
1602 stream
->tracefile_count_current
);
1606 stream
->index_fd
= ret
;
1609 /* Reset current size because we just perform a rotation. */
1610 stream
->tracefile_size_current
= 0;
1611 stream
->out_fd_offset
= 0;
1614 stream
->tracefile_size_current
+= len
;
1616 index
->offset
= htobe64(stream
->out_fd_offset
);
1621 * This call guarantee that len or less is returned. It's impossible to
1622 * receive a ret value that is bigger than len.
1624 ret
= lttng_write(outfd
, mmap_base
+ mmap_offset
, len
);
1625 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, len
);
1626 if (ret
< 0 || ((size_t) ret
!= len
)) {
1628 * Report error to caller if nothing was written else at least send the
1636 /* Socket operation failed. We consider the relayd dead */
1637 if (errno
== EPIPE
|| errno
== EINVAL
|| errno
== EBADF
) {
1639 * This is possible if the fd is closed on the other side
1640 * (outfd) or any write problem. It can be verbose a bit for a
1641 * normal execution if for instance the relayd is stopped
1642 * abruptly. This can happen so set this to a DBG statement.
1644 DBG("Consumer mmap write detected relayd hang up");
1646 /* Unhandled error, print it and stop function right now. */
1647 PERROR("Error in write mmap (ret %zd != len %lu)", ret
, len
);
1651 stream
->output_written
+= ret
;
1653 /* This call is useless on a socket so better save a syscall. */
1655 /* This won't block, but will start writeout asynchronously */
1656 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, len
,
1657 SYNC_FILE_RANGE_WRITE
);
1658 stream
->out_fd_offset
+= len
;
1660 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1664 * This is a special case that the relayd has closed its socket. Let's
1665 * cleanup the relayd object and all associated streams.
1667 if (relayd
&& relayd_hang_up
) {
1668 cleanup_relayd(relayd
, ctx
);
1672 /* Unlock only if ctrl socket used */
1673 if (relayd
&& stream
->metadata_flag
) {
1674 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1682 * Splice the data from the ring buffer to the tracefile.
1684 * It must be called with the stream lock held.
1686 * Returns the number of bytes spliced.
1688 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1689 struct lttng_consumer_local_data
*ctx
,
1690 struct lttng_consumer_stream
*stream
, unsigned long len
,
1691 unsigned long padding
,
1692 struct ctf_packet_index
*index
)
1694 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1696 off_t orig_offset
= stream
->out_fd_offset
;
1697 int fd
= stream
->wait_fd
;
1698 /* Default is on the disk */
1699 int outfd
= stream
->out_fd
;
1700 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1702 unsigned int relayd_hang_up
= 0;
1704 switch (consumer_data
.type
) {
1705 case LTTNG_CONSUMER_KERNEL
:
1707 case LTTNG_CONSUMER32_UST
:
1708 case LTTNG_CONSUMER64_UST
:
1709 /* Not supported for user space tracing */
1712 ERR("Unknown consumer_data type");
1716 /* RCU lock for the relayd pointer */
1719 /* Flag that the current stream if set for network streaming. */
1720 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1721 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1722 if (relayd
== NULL
) {
1727 splice_pipe
= stream
->splice_pipe
;
1729 /* Write metadata stream id before payload */
1731 unsigned long total_len
= len
;
1733 if (stream
->metadata_flag
) {
1735 * Lock the control socket for the complete duration of the function
1736 * since from this point on we will use the socket.
1738 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1740 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
, relayd
,
1748 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1751 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1757 /* Use the returned socket. */
1760 /* No streaming, we have to set the len with the full padding */
1764 * Check if we need to change the tracefile before writing the packet.
1766 if (stream
->chan
->tracefile_size
> 0 &&
1767 (stream
->tracefile_size_current
+ len
) >
1768 stream
->chan
->tracefile_size
) {
1769 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1770 stream
->name
, stream
->chan
->tracefile_size
,
1771 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1772 stream
->out_fd
, &(stream
->tracefile_count_current
),
1776 ERR("Rotating output file");
1779 outfd
= stream
->out_fd
;
1781 if (stream
->index_fd
>= 0) {
1782 ret
= close(stream
->index_fd
);
1784 PERROR("Closing index");
1787 stream
->index_fd
= -1;
1788 ret
= index_create_file(stream
->chan
->pathname
,
1789 stream
->name
, stream
->uid
, stream
->gid
,
1790 stream
->chan
->tracefile_size
,
1791 stream
->tracefile_count_current
);
1796 stream
->index_fd
= ret
;
1799 /* Reset current size because we just perform a rotation. */
1800 stream
->tracefile_size_current
= 0;
1801 stream
->out_fd_offset
= 0;
1804 stream
->tracefile_size_current
+= len
;
1805 index
->offset
= htobe64(stream
->out_fd_offset
);
1809 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1810 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1811 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1812 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1813 DBG("splice chan to pipe, ret %zd", ret_splice
);
1814 if (ret_splice
< 0) {
1817 PERROR("Error in relay splice");
1821 /* Handle stream on the relayd if the output is on the network */
1822 if (relayd
&& stream
->metadata_flag
) {
1823 size_t metadata_payload_size
=
1824 sizeof(struct lttcomm_relayd_metadata_payload
);
1826 /* Update counter to fit the spliced data */
1827 ret_splice
+= metadata_payload_size
;
1828 len
+= metadata_payload_size
;
1830 * We do this so the return value can match the len passed as
1831 * argument to this function.
1833 written
-= metadata_payload_size
;
1836 /* Splice data out */
1837 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1838 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1839 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
1841 if (ret_splice
< 0) {
1846 } else if (ret_splice
> len
) {
1848 * We don't expect this code path to be executed but you never know
1849 * so this is an extra protection agains a buggy splice().
1852 written
+= ret_splice
;
1853 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice
,
1857 /* All good, update current len and continue. */
1861 /* This call is useless on a socket so better save a syscall. */
1863 /* This won't block, but will start writeout asynchronously */
1864 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1865 SYNC_FILE_RANGE_WRITE
);
1866 stream
->out_fd_offset
+= ret_splice
;
1868 stream
->output_written
+= ret_splice
;
1869 written
+= ret_splice
;
1871 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1876 * This is a special case that the relayd has closed its socket. Let's
1877 * cleanup the relayd object and all associated streams.
1879 if (relayd
&& relayd_hang_up
) {
1880 cleanup_relayd(relayd
, ctx
);
1881 /* Skip splice error so the consumer does not fail */
1886 /* send the appropriate error description to sessiond */
1889 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1892 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1895 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
1900 if (relayd
&& stream
->metadata_flag
) {
1901 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1909 * Take a snapshot for a specific fd
1911 * Returns 0 on success, < 0 on error
1913 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
1915 switch (consumer_data
.type
) {
1916 case LTTNG_CONSUMER_KERNEL
:
1917 return lttng_kconsumer_take_snapshot(stream
);
1918 case LTTNG_CONSUMER32_UST
:
1919 case LTTNG_CONSUMER64_UST
:
1920 return lttng_ustconsumer_take_snapshot(stream
);
1922 ERR("Unknown consumer_data type");
1929 * Get the produced position
1931 * Returns 0 on success, < 0 on error
1933 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
1936 switch (consumer_data
.type
) {
1937 case LTTNG_CONSUMER_KERNEL
:
1938 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
1939 case LTTNG_CONSUMER32_UST
:
1940 case LTTNG_CONSUMER64_UST
:
1941 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
1943 ERR("Unknown consumer_data type");
1949 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
1950 int sock
, struct pollfd
*consumer_sockpoll
)
1952 switch (consumer_data
.type
) {
1953 case LTTNG_CONSUMER_KERNEL
:
1954 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1955 case LTTNG_CONSUMER32_UST
:
1956 case LTTNG_CONSUMER64_UST
:
1957 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1959 ERR("Unknown consumer_data type");
1965 void lttng_consumer_close_all_metadata(void)
1967 switch (consumer_data
.type
) {
1968 case LTTNG_CONSUMER_KERNEL
:
1970 * The Kernel consumer has a different metadata scheme so we don't
1971 * close anything because the stream will be closed by the session
1975 case LTTNG_CONSUMER32_UST
:
1976 case LTTNG_CONSUMER64_UST
:
1978 * Close all metadata streams. The metadata hash table is passed and
1979 * this call iterates over it by closing all wakeup fd. This is safe
1980 * because at this point we are sure that the metadata producer is
1981 * either dead or blocked.
1983 lttng_ustconsumer_close_all_metadata(metadata_ht
);
1986 ERR("Unknown consumer_data type");
1992 * Clean up a metadata stream and free its memory.
1994 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
1995 struct lttng_ht
*ht
)
1997 struct lttng_consumer_channel
*free_chan
= NULL
;
2001 * This call should NEVER receive regular stream. It must always be
2002 * metadata stream and this is crucial for data structure synchronization.
2004 assert(stream
->metadata_flag
);
2006 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
2008 pthread_mutex_lock(&consumer_data
.lock
);
2009 pthread_mutex_lock(&stream
->chan
->lock
);
2010 pthread_mutex_lock(&stream
->lock
);
2012 /* Remove any reference to that stream. */
2013 consumer_stream_delete(stream
, ht
);
2015 /* Close down everything including the relayd if one. */
2016 consumer_stream_close(stream
);
2017 /* Destroy tracer buffers of the stream. */
2018 consumer_stream_destroy_buffers(stream
);
2020 /* Atomically decrement channel refcount since other threads can use it. */
2021 if (!uatomic_sub_return(&stream
->chan
->refcount
, 1)
2022 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
2023 /* Go for channel deletion! */
2024 free_chan
= stream
->chan
;
2028 * Nullify the stream reference so it is not used after deletion. The
2029 * channel lock MUST be acquired before being able to check for a NULL
2032 stream
->chan
->metadata_stream
= NULL
;
2034 pthread_mutex_unlock(&stream
->lock
);
2035 pthread_mutex_unlock(&stream
->chan
->lock
);
2036 pthread_mutex_unlock(&consumer_data
.lock
);
2039 consumer_del_channel(free_chan
);
2042 consumer_stream_free(stream
);
2046 * Action done with the metadata stream when adding it to the consumer internal
2047 * data structures to handle it.
2049 int consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
2051 struct lttng_ht
*ht
= metadata_ht
;
2053 struct lttng_ht_iter iter
;
2054 struct lttng_ht_node_u64
*node
;
2059 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2061 pthread_mutex_lock(&consumer_data
.lock
);
2062 pthread_mutex_lock(&stream
->chan
->lock
);
2063 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2064 pthread_mutex_lock(&stream
->lock
);
2067 * From here, refcounts are updated so be _careful_ when returning an error
2074 * Lookup the stream just to make sure it does not exist in our internal
2075 * state. This should NEVER happen.
2077 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2078 node
= lttng_ht_iter_get_node_u64(&iter
);
2082 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2083 * in terms of destroying the associated channel, because the action that
2084 * causes the count to become 0 also causes a stream to be added. The
2085 * channel deletion will thus be triggered by the following removal of this
2088 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2089 /* Increment refcount before decrementing nb_init_stream_left */
2091 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2094 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2096 lttng_ht_add_unique_u64(consumer_data
.stream_per_chan_id_ht
,
2097 &stream
->node_channel_id
);
2100 * Add stream to the stream_list_ht of the consumer data. No need to steal
2101 * the key since the HT does not use it and we allow to add redundant keys
2104 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2108 pthread_mutex_unlock(&stream
->lock
);
2109 pthread_mutex_unlock(&stream
->chan
->lock
);
2110 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2111 pthread_mutex_unlock(&consumer_data
.lock
);
2116 * Delete data stream that are flagged for deletion (endpoint_status).
2118 static void validate_endpoint_status_data_stream(void)
2120 struct lttng_ht_iter iter
;
2121 struct lttng_consumer_stream
*stream
;
2123 DBG("Consumer delete flagged data stream");
2126 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2127 /* Validate delete flag of the stream */
2128 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2131 /* Delete it right now */
2132 consumer_del_stream(stream
, data_ht
);
2138 * Delete metadata stream that are flagged for deletion (endpoint_status).
2140 static void validate_endpoint_status_metadata_stream(
2141 struct lttng_poll_event
*pollset
)
2143 struct lttng_ht_iter iter
;
2144 struct lttng_consumer_stream
*stream
;
2146 DBG("Consumer delete flagged metadata stream");
2151 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2152 /* Validate delete flag of the stream */
2153 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2157 * Remove from pollset so the metadata thread can continue without
2158 * blocking on a deleted stream.
2160 lttng_poll_del(pollset
, stream
->wait_fd
);
2162 /* Delete it right now */
2163 consumer_del_metadata_stream(stream
, metadata_ht
);
2169 * Thread polls on metadata file descriptor and write them on disk or on the
2172 void *consumer_thread_metadata_poll(void *data
)
2174 int ret
, i
, pollfd
, err
= -1;
2175 uint32_t revents
, nb_fd
;
2176 struct lttng_consumer_stream
*stream
= NULL
;
2177 struct lttng_ht_iter iter
;
2178 struct lttng_ht_node_u64
*node
;
2179 struct lttng_poll_event events
;
2180 struct lttng_consumer_local_data
*ctx
= data
;
2183 rcu_register_thread();
2185 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2187 if (testpoint(consumerd_thread_metadata
)) {
2188 goto error_testpoint
;
2191 health_code_update();
2193 DBG("Thread metadata poll started");
2195 /* Size is set to 1 for the consumer_metadata pipe */
2196 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2198 ERR("Poll set creation failed");
2202 ret
= lttng_poll_add(&events
,
2203 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2209 DBG("Metadata main loop started");
2213 health_code_update();
2214 health_poll_entry();
2215 DBG("Metadata poll wait");
2216 ret
= lttng_poll_wait(&events
, -1);
2217 DBG("Metadata poll return from wait with %d fd(s)",
2218 LTTNG_POLL_GETNB(&events
));
2220 DBG("Metadata event caught in thread");
2222 if (errno
== EINTR
) {
2223 ERR("Poll EINTR caught");
2226 if (LTTNG_POLL_GETNB(&events
) == 0) {
2227 err
= 0; /* All is OK */
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
);
2242 /* No activity for this FD (poll implementation). */
2246 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2247 if (revents
& LPOLLIN
) {
2250 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2251 &stream
, sizeof(stream
));
2252 if (pipe_len
< sizeof(stream
)) {
2254 PERROR("read metadata stream");
2257 * Remove the pipe from the poll set and continue the loop
2258 * since their might be data to consume.
2260 lttng_poll_del(&events
,
2261 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2262 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2266 /* A NULL stream means that the state has changed. */
2267 if (stream
== NULL
) {
2268 /* Check for deleted streams. */
2269 validate_endpoint_status_metadata_stream(&events
);
2273 DBG("Adding metadata stream %d to poll set",
2276 /* Add metadata stream to the global poll events list */
2277 lttng_poll_add(&events
, stream
->wait_fd
,
2278 LPOLLIN
| LPOLLPRI
| LPOLLHUP
);
2279 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2280 DBG("Metadata thread pipe hung up");
2282 * Remove the pipe from the poll set and continue the loop
2283 * since their might be data to consume.
2285 lttng_poll_del(&events
,
2286 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2287 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2290 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2294 /* Handle other stream */
2300 uint64_t tmp_id
= (uint64_t) pollfd
;
2302 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2304 node
= lttng_ht_iter_get_node_u64(&iter
);
2307 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2310 if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2311 /* Get the data out of the metadata file descriptor */
2312 DBG("Metadata available on fd %d", pollfd
);
2313 assert(stream
->wait_fd
== pollfd
);
2316 health_code_update();
2318 len
= ctx
->on_buffer_ready(stream
, ctx
);
2320 * We don't check the return value here since if we get
2321 * a negative len, it means an error occured thus we
2322 * simply remove it from the poll set and free the
2327 /* It's ok to have an unavailable sub-buffer */
2328 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2329 /* Clean up stream from consumer and free it. */
2330 lttng_poll_del(&events
, stream
->wait_fd
);
2331 consumer_del_metadata_stream(stream
, metadata_ht
);
2333 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2334 DBG("Metadata fd %d is hup|err.", pollfd
);
2335 if (!stream
->hangup_flush_done
2336 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2337 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2338 DBG("Attempting to flush and consume the UST buffers");
2339 lttng_ustconsumer_on_stream_hangup(stream
);
2341 /* We just flushed the stream now read it. */
2343 health_code_update();
2345 len
= ctx
->on_buffer_ready(stream
, ctx
);
2347 * We don't check the return value here since if we get
2348 * a negative len, it means an error occured thus we
2349 * simply remove it from the poll set and free the
2355 lttng_poll_del(&events
, stream
->wait_fd
);
2357 * This call update the channel states, closes file descriptors
2358 * and securely free the stream.
2360 consumer_del_metadata_stream(stream
, metadata_ht
);
2362 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2366 /* Release RCU lock for the stream looked up */
2374 DBG("Metadata poll thread exiting");
2376 lttng_poll_clean(&events
);
2381 ERR("Health error occurred in %s", __func__
);
2383 health_unregister(health_consumerd
);
2384 rcu_unregister_thread();
2389 * This thread polls the fds in the set to consume the data and write
2390 * it to tracefile if necessary.
2392 void *consumer_thread_data_poll(void *data
)
2394 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2395 struct pollfd
*pollfd
= NULL
;
2396 /* local view of the streams */
2397 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2398 /* local view of consumer_data.fds_count */
2400 struct lttng_consumer_local_data
*ctx
= data
;
2403 rcu_register_thread();
2405 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2407 if (testpoint(consumerd_thread_data
)) {
2408 goto error_testpoint
;
2411 health_code_update();
2413 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
*));
2414 if (local_stream
== NULL
) {
2415 PERROR("local_stream malloc");
2420 health_code_update();
2426 * the fds set has been updated, we need to update our
2427 * local array as well
2429 pthread_mutex_lock(&consumer_data
.lock
);
2430 if (consumer_data
.need_update
) {
2435 local_stream
= NULL
;
2438 * Allocate for all fds +1 for the consumer_data_pipe and +1 for
2441 pollfd
= zmalloc((consumer_data
.stream_count
+ 2) * sizeof(struct pollfd
));
2442 if (pollfd
== NULL
) {
2443 PERROR("pollfd malloc");
2444 pthread_mutex_unlock(&consumer_data
.lock
);
2448 local_stream
= zmalloc((consumer_data
.stream_count
+ 2) *
2449 sizeof(struct lttng_consumer_stream
*));
2450 if (local_stream
== NULL
) {
2451 PERROR("local_stream malloc");
2452 pthread_mutex_unlock(&consumer_data
.lock
);
2455 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2458 ERR("Error in allocating pollfd or local_outfds");
2459 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2460 pthread_mutex_unlock(&consumer_data
.lock
);
2464 consumer_data
.need_update
= 0;
2466 pthread_mutex_unlock(&consumer_data
.lock
);
2468 /* No FDs and consumer_quit, consumer_cleanup the thread */
2469 if (nb_fd
== 0 && consumer_quit
== 1) {
2470 err
= 0; /* All is OK */
2473 /* poll on the array of fds */
2475 DBG("polling on %d fd", nb_fd
+ 2);
2476 health_poll_entry();
2477 num_rdy
= poll(pollfd
, nb_fd
+ 2, -1);
2479 DBG("poll num_rdy : %d", num_rdy
);
2480 if (num_rdy
== -1) {
2482 * Restart interrupted system call.
2484 if (errno
== EINTR
) {
2487 PERROR("Poll error");
2488 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2490 } else if (num_rdy
== 0) {
2491 DBG("Polling thread timed out");
2496 * If the consumer_data_pipe triggered poll go directly to the
2497 * beginning of the loop to update the array. We want to prioritize
2498 * array update over low-priority reads.
2500 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2501 ssize_t pipe_readlen
;
2503 DBG("consumer_data_pipe wake up");
2504 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2505 &new_stream
, sizeof(new_stream
));
2506 if (pipe_readlen
< sizeof(new_stream
)) {
2507 PERROR("Consumer data pipe");
2508 /* Continue so we can at least handle the current stream(s). */
2513 * If the stream is NULL, just ignore it. It's also possible that
2514 * the sessiond poll thread changed the consumer_quit state and is
2515 * waking us up to test it.
2517 if (new_stream
== NULL
) {
2518 validate_endpoint_status_data_stream();
2522 /* Continue to update the local streams and handle prio ones */
2526 /* Handle wakeup pipe. */
2527 if (pollfd
[nb_fd
+ 1].revents
& (POLLIN
| POLLPRI
)) {
2529 ssize_t pipe_readlen
;
2531 pipe_readlen
= lttng_pipe_read(ctx
->consumer_wakeup_pipe
, &dummy
,
2533 if (pipe_readlen
< 0) {
2534 PERROR("Consumer data wakeup pipe");
2536 /* We've been awakened to handle stream(s). */
2537 ctx
->has_wakeup
= 0;
2540 /* Take care of high priority channels first. */
2541 for (i
= 0; i
< nb_fd
; i
++) {
2542 health_code_update();
2544 if (local_stream
[i
] == NULL
) {
2547 if (pollfd
[i
].revents
& POLLPRI
) {
2548 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2550 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2551 /* it's ok to have an unavailable sub-buffer */
2552 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2553 /* Clean the stream and free it. */
2554 consumer_del_stream(local_stream
[i
], data_ht
);
2555 local_stream
[i
] = NULL
;
2556 } else if (len
> 0) {
2557 local_stream
[i
]->data_read
= 1;
2563 * If we read high prio channel in this loop, try again
2564 * for more high prio data.
2570 /* Take care of low priority channels. */
2571 for (i
= 0; i
< nb_fd
; i
++) {
2572 health_code_update();
2574 if (local_stream
[i
] == NULL
) {
2577 if ((pollfd
[i
].revents
& POLLIN
) ||
2578 local_stream
[i
]->hangup_flush_done
||
2579 local_stream
[i
]->has_data
) {
2580 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2581 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2582 /* it's ok to have an unavailable sub-buffer */
2583 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2584 /* Clean the stream and free it. */
2585 consumer_del_stream(local_stream
[i
], data_ht
);
2586 local_stream
[i
] = NULL
;
2587 } else if (len
> 0) {
2588 local_stream
[i
]->data_read
= 1;
2593 /* Handle hangup and errors */
2594 for (i
= 0; i
< nb_fd
; i
++) {
2595 health_code_update();
2597 if (local_stream
[i
] == NULL
) {
2600 if (!local_stream
[i
]->hangup_flush_done
2601 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2602 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2603 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2604 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2606 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2607 /* Attempt read again, for the data we just flushed. */
2608 local_stream
[i
]->data_read
= 1;
2611 * If the poll flag is HUP/ERR/NVAL and we have
2612 * read no data in this pass, we can remove the
2613 * stream from its hash table.
2615 if ((pollfd
[i
].revents
& POLLHUP
)) {
2616 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2617 if (!local_stream
[i
]->data_read
) {
2618 consumer_del_stream(local_stream
[i
], data_ht
);
2619 local_stream
[i
] = NULL
;
2622 } else if (pollfd
[i
].revents
& POLLERR
) {
2623 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2624 if (!local_stream
[i
]->data_read
) {
2625 consumer_del_stream(local_stream
[i
], data_ht
);
2626 local_stream
[i
] = NULL
;
2629 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2630 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2631 if (!local_stream
[i
]->data_read
) {
2632 consumer_del_stream(local_stream
[i
], data_ht
);
2633 local_stream
[i
] = NULL
;
2637 if (local_stream
[i
] != NULL
) {
2638 local_stream
[i
]->data_read
= 0;
2645 DBG("polling thread exiting");
2650 * Close the write side of the pipe so epoll_wait() in
2651 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2652 * read side of the pipe. If we close them both, epoll_wait strangely does
2653 * not return and could create a endless wait period if the pipe is the
2654 * only tracked fd in the poll set. The thread will take care of closing
2657 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2662 ERR("Health error occurred in %s", __func__
);
2664 health_unregister(health_consumerd
);
2666 rcu_unregister_thread();
2671 * Close wake-up end of each stream belonging to the channel. This will
2672 * allow the poll() on the stream read-side to detect when the
2673 * write-side (application) finally closes them.
2676 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2678 struct lttng_ht
*ht
;
2679 struct lttng_consumer_stream
*stream
;
2680 struct lttng_ht_iter iter
;
2682 ht
= consumer_data
.stream_per_chan_id_ht
;
2685 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2686 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2687 ht
->match_fct
, &channel
->key
,
2688 &iter
.iter
, stream
, node_channel_id
.node
) {
2690 * Protect against teardown with mutex.
2692 pthread_mutex_lock(&stream
->lock
);
2693 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2696 switch (consumer_data
.type
) {
2697 case LTTNG_CONSUMER_KERNEL
:
2699 case LTTNG_CONSUMER32_UST
:
2700 case LTTNG_CONSUMER64_UST
:
2701 if (stream
->metadata_flag
) {
2702 /* Safe and protected by the stream lock. */
2703 lttng_ustconsumer_close_metadata(stream
->chan
);
2706 * Note: a mutex is taken internally within
2707 * liblttng-ust-ctl to protect timer wakeup_fd
2708 * use from concurrent close.
2710 lttng_ustconsumer_close_stream_wakeup(stream
);
2714 ERR("Unknown consumer_data type");
2718 pthread_mutex_unlock(&stream
->lock
);
2723 static void destroy_channel_ht(struct lttng_ht
*ht
)
2725 struct lttng_ht_iter iter
;
2726 struct lttng_consumer_channel
*channel
;
2734 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2735 ret
= lttng_ht_del(ht
, &iter
);
2740 lttng_ht_destroy(ht
);
2744 * This thread polls the channel fds to detect when they are being
2745 * closed. It closes all related streams if the channel is detected as
2746 * closed. It is currently only used as a shim layer for UST because the
2747 * consumerd needs to keep the per-stream wakeup end of pipes open for
2750 void *consumer_thread_channel_poll(void *data
)
2752 int ret
, i
, pollfd
, err
= -1;
2753 uint32_t revents
, nb_fd
;
2754 struct lttng_consumer_channel
*chan
= NULL
;
2755 struct lttng_ht_iter iter
;
2756 struct lttng_ht_node_u64
*node
;
2757 struct lttng_poll_event events
;
2758 struct lttng_consumer_local_data
*ctx
= data
;
2759 struct lttng_ht
*channel_ht
;
2761 rcu_register_thread();
2763 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
2765 if (testpoint(consumerd_thread_channel
)) {
2766 goto error_testpoint
;
2769 health_code_update();
2771 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2773 /* ENOMEM at this point. Better to bail out. */
2777 DBG("Thread channel poll started");
2779 /* Size is set to 1 for the consumer_channel pipe */
2780 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2782 ERR("Poll set creation failed");
2786 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2792 DBG("Channel main loop started");
2796 health_code_update();
2797 DBG("Channel poll wait");
2798 health_poll_entry();
2799 ret
= lttng_poll_wait(&events
, -1);
2800 DBG("Channel poll return from wait with %d fd(s)",
2801 LTTNG_POLL_GETNB(&events
));
2803 DBG("Channel event caught in thread");
2805 if (errno
== EINTR
) {
2806 ERR("Poll EINTR caught");
2809 if (LTTNG_POLL_GETNB(&events
) == 0) {
2810 err
= 0; /* All is OK */
2817 /* From here, the event is a channel wait fd */
2818 for (i
= 0; i
< nb_fd
; i
++) {
2819 health_code_update();
2821 revents
= LTTNG_POLL_GETEV(&events
, i
);
2822 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2825 /* No activity for this FD (poll implementation). */
2829 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2830 if (revents
& LPOLLIN
) {
2831 enum consumer_channel_action action
;
2834 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
2837 ERR("Error reading channel pipe");
2839 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2844 case CONSUMER_CHANNEL_ADD
:
2845 DBG("Adding channel %d to poll set",
2848 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
2851 lttng_ht_add_unique_u64(channel_ht
,
2852 &chan
->wait_fd_node
);
2854 /* Add channel to the global poll events list */
2855 lttng_poll_add(&events
, chan
->wait_fd
,
2856 LPOLLERR
| LPOLLHUP
);
2858 case CONSUMER_CHANNEL_DEL
:
2861 * This command should never be called if the channel
2862 * has streams monitored by either the data or metadata
2863 * thread. The consumer only notify this thread with a
2864 * channel del. command if it receives a destroy
2865 * channel command from the session daemon that send it
2866 * if a command prior to the GET_CHANNEL failed.
2870 chan
= consumer_find_channel(key
);
2873 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
2876 lttng_poll_del(&events
, chan
->wait_fd
);
2877 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
2878 ret
= lttng_ht_del(channel_ht
, &iter
);
2881 switch (consumer_data
.type
) {
2882 case LTTNG_CONSUMER_KERNEL
:
2884 case LTTNG_CONSUMER32_UST
:
2885 case LTTNG_CONSUMER64_UST
:
2886 health_code_update();
2887 /* Destroy streams that might have been left in the stream list. */
2888 clean_channel_stream_list(chan
);
2891 ERR("Unknown consumer_data type");
2896 * Release our own refcount. Force channel deletion even if
2897 * streams were not initialized.
2899 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
2900 consumer_del_channel(chan
);
2905 case CONSUMER_CHANNEL_QUIT
:
2907 * Remove the pipe from the poll set and continue the loop
2908 * since their might be data to consume.
2910 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2913 ERR("Unknown action");
2916 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2917 DBG("Channel thread pipe hung up");
2919 * Remove the pipe from the poll set and continue the loop
2920 * since their might be data to consume.
2922 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2925 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2929 /* Handle other stream */
2935 uint64_t tmp_id
= (uint64_t) pollfd
;
2937 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
2939 node
= lttng_ht_iter_get_node_u64(&iter
);
2942 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
2945 /* Check for error event */
2946 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2947 DBG("Channel fd %d is hup|err.", pollfd
);
2949 lttng_poll_del(&events
, chan
->wait_fd
);
2950 ret
= lttng_ht_del(channel_ht
, &iter
);
2954 * This will close the wait fd for each stream associated to
2955 * this channel AND monitored by the data/metadata thread thus
2956 * will be clean by the right thread.
2958 consumer_close_channel_streams(chan
);
2960 /* Release our own refcount */
2961 if (!uatomic_sub_return(&chan
->refcount
, 1)
2962 && !uatomic_read(&chan
->nb_init_stream_left
)) {
2963 consumer_del_channel(chan
);
2966 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2971 /* Release RCU lock for the channel looked up */
2979 lttng_poll_clean(&events
);
2981 destroy_channel_ht(channel_ht
);
2984 DBG("Channel poll thread exiting");
2987 ERR("Health error occurred in %s", __func__
);
2989 health_unregister(health_consumerd
);
2990 rcu_unregister_thread();
2994 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
2995 struct pollfd
*sockpoll
, int client_socket
)
3002 ret
= lttng_consumer_poll_socket(sockpoll
);
3006 DBG("Metadata connection on client_socket");
3008 /* Blocking call, waiting for transmission */
3009 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
3010 if (ctx
->consumer_metadata_socket
< 0) {
3011 WARN("On accept metadata");
3022 * This thread listens on the consumerd socket and receives the file
3023 * descriptors from the session daemon.
3025 void *consumer_thread_sessiond_poll(void *data
)
3027 int sock
= -1, client_socket
, ret
, err
= -1;
3029 * structure to poll for incoming data on communication socket avoids
3030 * making blocking sockets.
3032 struct pollfd consumer_sockpoll
[2];
3033 struct lttng_consumer_local_data
*ctx
= data
;
3035 rcu_register_thread();
3037 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
3039 if (testpoint(consumerd_thread_sessiond
)) {
3040 goto error_testpoint
;
3043 health_code_update();
3045 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
3046 unlink(ctx
->consumer_command_sock_path
);
3047 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
3048 if (client_socket
< 0) {
3049 ERR("Cannot create command socket");
3053 ret
= lttcomm_listen_unix_sock(client_socket
);
3058 DBG("Sending ready command to lttng-sessiond");
3059 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3060 /* return < 0 on error, but == 0 is not fatal */
3062 ERR("Error sending ready command to lttng-sessiond");
3066 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3067 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3068 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3069 consumer_sockpoll
[1].fd
= client_socket
;
3070 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3072 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3080 DBG("Connection on client_socket");
3082 /* Blocking call, waiting for transmission */
3083 sock
= lttcomm_accept_unix_sock(client_socket
);
3090 * Setup metadata socket which is the second socket connection on the
3091 * command unix socket.
3093 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3102 /* This socket is not useful anymore. */
3103 ret
= close(client_socket
);
3105 PERROR("close client_socket");
3109 /* update the polling structure to poll on the established socket */
3110 consumer_sockpoll
[1].fd
= sock
;
3111 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3114 health_code_update();
3116 health_poll_entry();
3117 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3126 DBG("Incoming command on sock");
3127 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3130 * This could simply be a session daemon quitting. Don't output
3133 DBG("Communication interrupted on command socket");
3137 if (consumer_quit
) {
3138 DBG("consumer_thread_receive_fds received quit from signal");
3139 err
= 0; /* All is OK */
3142 DBG("received command on sock");
3148 DBG("Consumer thread sessiond poll exiting");
3151 * Close metadata streams since the producer is the session daemon which
3154 * NOTE: for now, this only applies to the UST tracer.
3156 lttng_consumer_close_all_metadata();
3159 * when all fds have hung up, the polling thread
3165 * Notify the data poll thread to poll back again and test the
3166 * consumer_quit state that we just set so to quit gracefully.
3168 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3170 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3172 notify_health_quit_pipe(health_quit_pipe
);
3174 /* Cleaning up possibly open sockets. */
3178 PERROR("close sock sessiond poll");
3181 if (client_socket
>= 0) {
3182 ret
= close(client_socket
);
3184 PERROR("close client_socket sessiond poll");
3191 ERR("Health error occurred in %s", __func__
);
3193 health_unregister(health_consumerd
);
3195 rcu_unregister_thread();
3199 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3200 struct lttng_consumer_local_data
*ctx
)
3204 pthread_mutex_lock(&stream
->lock
);
3205 if (stream
->metadata_flag
) {
3206 pthread_mutex_lock(&stream
->metadata_rdv_lock
);
3209 switch (consumer_data
.type
) {
3210 case LTTNG_CONSUMER_KERNEL
:
3211 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
3213 case LTTNG_CONSUMER32_UST
:
3214 case LTTNG_CONSUMER64_UST
:
3215 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
3218 ERR("Unknown consumer_data type");
3224 if (stream
->metadata_flag
) {
3225 pthread_cond_broadcast(&stream
->metadata_rdv
);
3226 pthread_mutex_unlock(&stream
->metadata_rdv_lock
);
3228 pthread_mutex_unlock(&stream
->lock
);
3232 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3234 switch (consumer_data
.type
) {
3235 case LTTNG_CONSUMER_KERNEL
:
3236 return lttng_kconsumer_on_recv_stream(stream
);
3237 case LTTNG_CONSUMER32_UST
:
3238 case LTTNG_CONSUMER64_UST
:
3239 return lttng_ustconsumer_on_recv_stream(stream
);
3241 ERR("Unknown consumer_data type");
3248 * Allocate and set consumer data hash tables.
3250 int lttng_consumer_init(void)
3252 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3253 if (!consumer_data
.channel_ht
) {
3257 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3258 if (!consumer_data
.relayd_ht
) {
3262 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3263 if (!consumer_data
.stream_list_ht
) {
3267 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3268 if (!consumer_data
.stream_per_chan_id_ht
) {
3272 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3277 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3289 * Process the ADD_RELAYD command receive by a consumer.
3291 * This will create a relayd socket pair and add it to the relayd hash table.
3292 * The caller MUST acquire a RCU read side lock before calling it.
3294 int consumer_add_relayd_socket(uint64_t net_seq_idx
, int sock_type
,
3295 struct lttng_consumer_local_data
*ctx
, int sock
,
3296 struct pollfd
*consumer_sockpoll
,
3297 struct lttcomm_relayd_sock
*relayd_sock
, uint64_t sessiond_id
,
3298 uint64_t relayd_session_id
)
3300 int fd
= -1, ret
= -1, relayd_created
= 0;
3301 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3302 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3305 assert(relayd_sock
);
3307 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", net_seq_idx
);
3309 /* Get relayd reference if exists. */
3310 relayd
= consumer_find_relayd(net_seq_idx
);
3311 if (relayd
== NULL
) {
3312 assert(sock_type
== LTTNG_STREAM_CONTROL
);
3313 /* Not found. Allocate one. */
3314 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3315 if (relayd
== NULL
) {
3317 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3320 relayd
->sessiond_session_id
= sessiond_id
;
3325 * This code path MUST continue to the consumer send status message to
3326 * we can notify the session daemon and continue our work without
3327 * killing everything.
3331 * relayd key should never be found for control socket.
3333 assert(sock_type
!= LTTNG_STREAM_CONTROL
);
3336 /* First send a status message before receiving the fds. */
3337 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3339 /* Somehow, the session daemon is not responding anymore. */
3340 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3341 goto error_nosignal
;
3344 /* Poll on consumer socket. */
3345 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3347 /* Needing to exit in the middle of a command: error. */
3348 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3350 goto error_nosignal
;
3353 /* Get relayd socket from session daemon */
3354 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3355 if (ret
!= sizeof(fd
)) {
3357 fd
= -1; /* Just in case it gets set with an invalid value. */
3360 * Failing to receive FDs might indicate a major problem such as
3361 * reaching a fd limit during the receive where the kernel returns a
3362 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3363 * don't take any chances and stop everything.
3365 * XXX: Feature request #558 will fix that and avoid this possible
3366 * issue when reaching the fd limit.
3368 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3369 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3373 /* Copy socket information and received FD */
3374 switch (sock_type
) {
3375 case LTTNG_STREAM_CONTROL
:
3376 /* Copy received lttcomm socket */
3377 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3378 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3379 /* Handle create_sock error. */
3381 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3385 * Close the socket created internally by
3386 * lttcomm_create_sock, so we can replace it by the one
3387 * received from sessiond.
3389 if (close(relayd
->control_sock
.sock
.fd
)) {
3393 /* Assign new file descriptor */
3394 relayd
->control_sock
.sock
.fd
= fd
;
3395 fd
= -1; /* For error path */
3396 /* Assign version values. */
3397 relayd
->control_sock
.major
= relayd_sock
->major
;
3398 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3400 relayd
->relayd_session_id
= relayd_session_id
;
3403 case LTTNG_STREAM_DATA
:
3404 /* Copy received lttcomm socket */
3405 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3406 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3407 /* Handle create_sock error. */
3409 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3413 * Close the socket created internally by
3414 * lttcomm_create_sock, so we can replace it by the one
3415 * received from sessiond.
3417 if (close(relayd
->data_sock
.sock
.fd
)) {
3421 /* Assign new file descriptor */
3422 relayd
->data_sock
.sock
.fd
= fd
;
3423 fd
= -1; /* for eventual error paths */
3424 /* Assign version values. */
3425 relayd
->data_sock
.major
= relayd_sock
->major
;
3426 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3429 ERR("Unknown relayd socket type (%d)", sock_type
);
3431 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3435 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3436 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3437 relayd
->net_seq_idx
, fd
);
3439 /* We successfully added the socket. Send status back. */
3440 ret
= consumer_send_status_msg(sock
, ret_code
);
3442 /* Somehow, the session daemon is not responding anymore. */
3443 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3444 goto error_nosignal
;
3448 * Add relayd socket pair to consumer data hashtable. If object already
3449 * exists or on error, the function gracefully returns.
3457 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3458 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3462 /* Close received socket if valid. */
3465 PERROR("close received socket");
3469 if (relayd_created
) {
3477 * Try to lock the stream mutex.
3479 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
3481 static int stream_try_lock(struct lttng_consumer_stream
*stream
)
3488 * Try to lock the stream mutex. On failure, we know that the stream is
3489 * being used else where hence there is data still being extracted.
3491 ret
= pthread_mutex_trylock(&stream
->lock
);
3493 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
3505 * Search for a relayd associated to the session id and return the reference.
3507 * A rcu read side lock MUST be acquire before calling this function and locked
3508 * until the relayd object is no longer necessary.
3510 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3512 struct lttng_ht_iter iter
;
3513 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3515 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3516 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3519 * Check by sessiond id which is unique here where the relayd session
3520 * id might not be when having multiple relayd.
3522 if (relayd
->sessiond_session_id
== id
) {
3523 /* Found the relayd. There can be only one per id. */
3535 * Check if for a given session id there is still data needed to be extract
3538 * Return 1 if data is pending or else 0 meaning ready to be read.
3540 int consumer_data_pending(uint64_t id
)
3543 struct lttng_ht_iter iter
;
3544 struct lttng_ht
*ht
;
3545 struct lttng_consumer_stream
*stream
;
3546 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3547 int (*data_pending
)(struct lttng_consumer_stream
*);
3549 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3552 pthread_mutex_lock(&consumer_data
.lock
);
3554 switch (consumer_data
.type
) {
3555 case LTTNG_CONSUMER_KERNEL
:
3556 data_pending
= lttng_kconsumer_data_pending
;
3558 case LTTNG_CONSUMER32_UST
:
3559 case LTTNG_CONSUMER64_UST
:
3560 data_pending
= lttng_ustconsumer_data_pending
;
3563 ERR("Unknown consumer data type");
3567 /* Ease our life a bit */
3568 ht
= consumer_data
.stream_list_ht
;
3570 relayd
= find_relayd_by_session_id(id
);
3572 /* Send init command for data pending. */
3573 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3574 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3575 relayd
->relayd_session_id
);
3576 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3578 /* Communication error thus the relayd so no data pending. */
3579 goto data_not_pending
;
3583 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3584 ht
->hash_fct(&id
, lttng_ht_seed
),
3586 &iter
.iter
, stream
, node_session_id
.node
) {
3587 /* If this call fails, the stream is being used hence data pending. */
3588 ret
= stream_try_lock(stream
);
3594 * A removed node from the hash table indicates that the stream has
3595 * been deleted thus having a guarantee that the buffers are closed
3596 * on the consumer side. However, data can still be transmitted
3597 * over the network so don't skip the relayd check.
3599 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3601 /* Check the stream if there is data in the buffers. */
3602 ret
= data_pending(stream
);
3604 pthread_mutex_unlock(&stream
->lock
);
3611 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3612 if (stream
->metadata_flag
) {
3613 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3614 stream
->relayd_stream_id
);
3616 ret
= relayd_data_pending(&relayd
->control_sock
,
3617 stream
->relayd_stream_id
,
3618 stream
->next_net_seq_num
- 1);
3620 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3622 pthread_mutex_unlock(&stream
->lock
);
3626 pthread_mutex_unlock(&stream
->lock
);
3630 unsigned int is_data_inflight
= 0;
3632 /* Send init command for data pending. */
3633 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3634 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3635 relayd
->relayd_session_id
, &is_data_inflight
);
3636 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3638 goto data_not_pending
;
3640 if (is_data_inflight
) {
3646 * Finding _no_ node in the hash table and no inflight data means that the
3647 * stream(s) have been removed thus data is guaranteed to be available for
3648 * analysis from the trace files.
3652 /* Data is available to be read by a viewer. */
3653 pthread_mutex_unlock(&consumer_data
.lock
);
3658 /* Data is still being extracted from buffers. */
3659 pthread_mutex_unlock(&consumer_data
.lock
);
3665 * Send a ret code status message to the sessiond daemon.
3667 * Return the sendmsg() return value.
3669 int consumer_send_status_msg(int sock
, int ret_code
)
3671 struct lttcomm_consumer_status_msg msg
;
3673 memset(&msg
, 0, sizeof(msg
));
3674 msg
.ret_code
= ret_code
;
3676 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3680 * Send a channel status message to the sessiond daemon.
3682 * Return the sendmsg() return value.
3684 int consumer_send_status_channel(int sock
,
3685 struct lttng_consumer_channel
*channel
)
3687 struct lttcomm_consumer_status_channel msg
;
3691 memset(&msg
, 0, sizeof(msg
));
3693 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3695 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3696 msg
.key
= channel
->key
;
3697 msg
.stream_count
= channel
->streams
.count
;
3700 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3703 unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos
,
3704 unsigned long produced_pos
, uint64_t nb_packets_per_stream
,
3705 uint64_t max_sb_size
)
3707 unsigned long start_pos
;
3709 if (!nb_packets_per_stream
) {
3710 return consumed_pos
; /* Grab everything */
3712 start_pos
= produced_pos
- offset_align_floor(produced_pos
, max_sb_size
);
3713 start_pos
-= max_sb_size
* nb_packets_per_stream
;
3714 if ((long) (start_pos
- consumed_pos
) < 0) {
3715 return consumed_pos
; /* Grab everything */