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/consumer-timer.h>
46 #include <common/consumer/consumer.h>
47 #include <common/consumer/consumer-stream.h>
48 #include <common/consumer/consumer-testpoint.h>
49 #include <common/align.h>
50 #include <common/consumer/consumer-metadata-cache.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 void lttng_consumer_cleanup_relayd(struct consumer_relayd_sock_pair
*relayd
)
467 DBG("Cleaning up relayd object ID %"PRIu64
, relayd
->net_seq_idx
);
469 /* Save the net sequence index before destroying the object */
470 netidx
= relayd
->net_seq_idx
;
473 * Delete the relayd from the relayd hash table, close the sockets and free
474 * the object in a RCU call.
476 consumer_destroy_relayd(relayd
);
478 /* Set inactive endpoint to all streams */
479 update_endpoint_status_by_netidx(netidx
, CONSUMER_ENDPOINT_INACTIVE
);
482 * With a local data context, notify the threads that the streams' state
483 * have changed. The write() action on the pipe acts as an "implicit"
484 * memory barrier ordering the updates of the end point status from the
485 * read of this status which happens AFTER receiving this notify.
487 notify_thread_lttng_pipe(relayd
->ctx
->consumer_data_pipe
);
488 notify_thread_lttng_pipe(relayd
->ctx
->consumer_metadata_pipe
);
492 * Flag a relayd socket pair for destruction. Destroy it if the refcount
495 * RCU read side lock MUST be aquired before calling this function.
497 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair
*relayd
)
501 /* Set destroy flag for this object */
502 uatomic_set(&relayd
->destroy_flag
, 1);
504 /* Destroy the relayd if refcount is 0 */
505 if (uatomic_read(&relayd
->refcount
) == 0) {
506 consumer_destroy_relayd(relayd
);
511 * Completly destroy stream from every visiable data structure and the given
514 * One this call returns, the stream object is not longer usable nor visible.
516 void consumer_del_stream(struct lttng_consumer_stream
*stream
,
519 consumer_stream_destroy(stream
, ht
);
523 * XXX naming of del vs destroy is all mixed up.
525 void consumer_del_stream_for_data(struct lttng_consumer_stream
*stream
)
527 consumer_stream_destroy(stream
, data_ht
);
530 void consumer_del_stream_for_metadata(struct lttng_consumer_stream
*stream
)
532 consumer_stream_destroy(stream
, metadata_ht
);
535 struct lttng_consumer_stream
*consumer_allocate_stream(uint64_t channel_key
,
537 enum lttng_consumer_stream_state state
,
538 const char *channel_name
,
545 enum consumer_channel_type type
,
546 unsigned int monitor
)
549 struct lttng_consumer_stream
*stream
;
551 stream
= zmalloc(sizeof(*stream
));
552 if (stream
== NULL
) {
553 PERROR("malloc struct lttng_consumer_stream");
560 stream
->key
= stream_key
;
562 stream
->out_fd_offset
= 0;
563 stream
->output_written
= 0;
564 stream
->state
= state
;
567 stream
->net_seq_idx
= relayd_id
;
568 stream
->session_id
= session_id
;
569 stream
->monitor
= monitor
;
570 stream
->endpoint_status
= CONSUMER_ENDPOINT_ACTIVE
;
571 stream
->index_file
= NULL
;
572 stream
->last_sequence_number
= -1ULL;
573 pthread_mutex_init(&stream
->lock
, NULL
);
574 pthread_mutex_init(&stream
->metadata_timer_lock
, NULL
);
576 /* If channel is the metadata, flag this stream as metadata. */
577 if (type
== CONSUMER_CHANNEL_TYPE_METADATA
) {
578 stream
->metadata_flag
= 1;
579 /* Metadata is flat out. */
580 strncpy(stream
->name
, DEFAULT_METADATA_NAME
, sizeof(stream
->name
));
581 /* Live rendez-vous point. */
582 pthread_cond_init(&stream
->metadata_rdv
, NULL
);
583 pthread_mutex_init(&stream
->metadata_rdv_lock
, NULL
);
585 /* Format stream name to <channel_name>_<cpu_number> */
586 ret
= snprintf(stream
->name
, sizeof(stream
->name
), "%s_%d",
589 PERROR("snprintf stream name");
594 /* Key is always the wait_fd for streams. */
595 lttng_ht_node_init_u64(&stream
->node
, stream
->key
);
597 /* Init node per channel id key */
598 lttng_ht_node_init_u64(&stream
->node_channel_id
, channel_key
);
600 /* Init session id node with the stream session id */
601 lttng_ht_node_init_u64(&stream
->node_session_id
, stream
->session_id
);
603 DBG3("Allocated stream %s (key %" PRIu64
", chan_key %" PRIu64
604 " relayd_id %" PRIu64
", session_id %" PRIu64
,
605 stream
->name
, stream
->key
, channel_key
,
606 stream
->net_seq_idx
, stream
->session_id
);
622 * Add a stream to the global list protected by a mutex.
624 int consumer_add_data_stream(struct lttng_consumer_stream
*stream
)
626 struct lttng_ht
*ht
= data_ht
;
632 DBG3("Adding consumer stream %" PRIu64
, stream
->key
);
634 pthread_mutex_lock(&consumer_data
.lock
);
635 pthread_mutex_lock(&stream
->chan
->lock
);
636 pthread_mutex_lock(&stream
->chan
->timer_lock
);
637 pthread_mutex_lock(&stream
->lock
);
640 /* Steal stream identifier to avoid having streams with the same key */
641 steal_stream_key(stream
->key
, ht
);
643 lttng_ht_add_unique_u64(ht
, &stream
->node
);
645 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
646 &stream
->node_channel_id
);
649 * Add stream to the stream_list_ht of the consumer data. No need to steal
650 * the key since the HT does not use it and we allow to add redundant keys
653 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
656 * When nb_init_stream_left reaches 0, we don't need to trigger any action
657 * in terms of destroying the associated channel, because the action that
658 * causes the count to become 0 also causes a stream to be added. The
659 * channel deletion will thus be triggered by the following removal of this
662 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
663 /* Increment refcount before decrementing nb_init_stream_left */
665 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
668 /* Update consumer data once the node is inserted. */
669 consumer_data
.stream_count
++;
670 consumer_data
.need_update
= 1;
673 pthread_mutex_unlock(&stream
->lock
);
674 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
675 pthread_mutex_unlock(&stream
->chan
->lock
);
676 pthread_mutex_unlock(&consumer_data
.lock
);
681 void consumer_del_data_stream(struct lttng_consumer_stream
*stream
)
683 consumer_del_stream(stream
, data_ht
);
687 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
688 * be acquired before calling this.
690 static int add_relayd(struct consumer_relayd_sock_pair
*relayd
)
693 struct lttng_ht_node_u64
*node
;
694 struct lttng_ht_iter iter
;
698 lttng_ht_lookup(consumer_data
.relayd_ht
,
699 &relayd
->net_seq_idx
, &iter
);
700 node
= lttng_ht_iter_get_node_u64(&iter
);
704 lttng_ht_add_unique_u64(consumer_data
.relayd_ht
, &relayd
->node
);
711 * Allocate and return a consumer relayd socket.
713 static struct consumer_relayd_sock_pair
*consumer_allocate_relayd_sock_pair(
714 uint64_t net_seq_idx
)
716 struct consumer_relayd_sock_pair
*obj
= NULL
;
718 /* net sequence index of -1 is a failure */
719 if (net_seq_idx
== (uint64_t) -1ULL) {
723 obj
= zmalloc(sizeof(struct consumer_relayd_sock_pair
));
725 PERROR("zmalloc relayd sock");
729 obj
->net_seq_idx
= net_seq_idx
;
731 obj
->destroy_flag
= 0;
732 obj
->control_sock
.sock
.fd
= -1;
733 obj
->data_sock
.sock
.fd
= -1;
734 lttng_ht_node_init_u64(&obj
->node
, obj
->net_seq_idx
);
735 pthread_mutex_init(&obj
->ctrl_sock_mutex
, NULL
);
742 * Find a relayd socket pair in the global consumer data.
744 * Return the object if found else NULL.
745 * RCU read-side lock must be held across this call and while using the
748 struct consumer_relayd_sock_pair
*consumer_find_relayd(uint64_t key
)
750 struct lttng_ht_iter iter
;
751 struct lttng_ht_node_u64
*node
;
752 struct consumer_relayd_sock_pair
*relayd
= NULL
;
754 /* Negative keys are lookup failures */
755 if (key
== (uint64_t) -1ULL) {
759 lttng_ht_lookup(consumer_data
.relayd_ht
, &key
,
761 node
= lttng_ht_iter_get_node_u64(&iter
);
763 relayd
= caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
771 * Find a relayd and send the stream
773 * Returns 0 on success, < 0 on error
775 int consumer_send_relayd_stream(struct lttng_consumer_stream
*stream
,
779 struct consumer_relayd_sock_pair
*relayd
;
782 assert(stream
->net_seq_idx
!= -1ULL);
785 /* The stream is not metadata. Get relayd reference if exists. */
787 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
788 if (relayd
!= NULL
) {
789 /* Add stream on the relayd */
790 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
791 ret
= relayd_add_stream(&relayd
->control_sock
, stream
->name
,
792 path
, &stream
->relayd_stream_id
,
793 stream
->chan
->tracefile_size
, stream
->chan
->tracefile_count
);
795 ERR("Relayd add stream failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
796 lttng_consumer_cleanup_relayd(relayd
);
797 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
800 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
802 uatomic_inc(&relayd
->refcount
);
803 stream
->sent_to_relayd
= 1;
805 ERR("Stream %" PRIu64
" relayd ID %" PRIu64
" unknown. Can't send it.",
806 stream
->key
, stream
->net_seq_idx
);
811 DBG("Stream %s with key %" PRIu64
" sent to relayd id %" PRIu64
,
812 stream
->name
, stream
->key
, stream
->net_seq_idx
);
820 * Find a relayd and send the streams sent message
822 * Returns 0 on success, < 0 on error
824 int consumer_send_relayd_streams_sent(uint64_t net_seq_idx
)
827 struct consumer_relayd_sock_pair
*relayd
;
829 assert(net_seq_idx
!= -1ULL);
831 /* The stream is not metadata. Get relayd reference if exists. */
833 relayd
= consumer_find_relayd(net_seq_idx
);
834 if (relayd
!= NULL
) {
835 /* Add stream on the relayd */
836 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
837 ret
= relayd_streams_sent(&relayd
->control_sock
);
839 ERR("Relayd streams sent failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
840 lttng_consumer_cleanup_relayd(relayd
);
841 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
844 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
846 ERR("Relayd ID %" PRIu64
" unknown. Can't send streams_sent.",
853 DBG("All streams sent relayd id %" PRIu64
, net_seq_idx
);
861 * Find a relayd and close the stream
863 void close_relayd_stream(struct lttng_consumer_stream
*stream
)
865 struct consumer_relayd_sock_pair
*relayd
;
867 /* The stream is not metadata. Get relayd reference if exists. */
869 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
871 consumer_stream_relayd_close(stream
, relayd
);
877 * Handle stream for relayd transmission if the stream applies for network
878 * streaming where the net sequence index is set.
880 * Return destination file descriptor or negative value on error.
882 static int write_relayd_stream_header(struct lttng_consumer_stream
*stream
,
883 size_t data_size
, unsigned long padding
,
884 struct consumer_relayd_sock_pair
*relayd
)
887 struct lttcomm_relayd_data_hdr data_hdr
;
893 /* Reset data header */
894 memset(&data_hdr
, 0, sizeof(data_hdr
));
896 if (stream
->metadata_flag
) {
897 /* Caller MUST acquire the relayd control socket lock */
898 ret
= relayd_send_metadata(&relayd
->control_sock
, data_size
);
903 /* Metadata are always sent on the control socket. */
904 outfd
= relayd
->control_sock
.sock
.fd
;
906 /* Set header with stream information */
907 data_hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
908 data_hdr
.data_size
= htobe32(data_size
);
909 data_hdr
.padding_size
= htobe32(padding
);
911 * Note that net_seq_num below is assigned with the *current* value of
912 * next_net_seq_num and only after that the next_net_seq_num will be
913 * increment. This is why when issuing a command on the relayd using
914 * this next value, 1 should always be substracted in order to compare
915 * the last seen sequence number on the relayd side to the last sent.
917 data_hdr
.net_seq_num
= htobe64(stream
->next_net_seq_num
);
918 /* Other fields are zeroed previously */
920 ret
= relayd_send_data_hdr(&relayd
->data_sock
, &data_hdr
,
926 ++stream
->next_net_seq_num
;
928 /* Set to go on data socket */
929 outfd
= relayd
->data_sock
.sock
.fd
;
937 * Allocate and return a new lttng_consumer_channel object using the given key
938 * to initialize the hash table node.
940 * On error, return NULL.
942 struct lttng_consumer_channel
*consumer_allocate_channel(uint64_t key
,
944 const char *pathname
,
949 enum lttng_event_output output
,
950 uint64_t tracefile_size
,
951 uint64_t tracefile_count
,
952 uint64_t session_id_per_pid
,
953 unsigned int monitor
,
954 unsigned int live_timer_interval
,
955 const char *root_shm_path
,
956 const char *shm_path
)
958 struct lttng_consumer_channel
*channel
;
960 channel
= zmalloc(sizeof(*channel
));
961 if (channel
== NULL
) {
962 PERROR("malloc struct lttng_consumer_channel");
967 channel
->refcount
= 0;
968 channel
->session_id
= session_id
;
969 channel
->session_id_per_pid
= session_id_per_pid
;
972 channel
->relayd_id
= relayd_id
;
973 channel
->tracefile_size
= tracefile_size
;
974 channel
->tracefile_count
= tracefile_count
;
975 channel
->monitor
= monitor
;
976 channel
->live_timer_interval
= live_timer_interval
;
977 pthread_mutex_init(&channel
->lock
, NULL
);
978 pthread_mutex_init(&channel
->timer_lock
, NULL
);
981 case LTTNG_EVENT_SPLICE
:
982 channel
->output
= CONSUMER_CHANNEL_SPLICE
;
984 case LTTNG_EVENT_MMAP
:
985 channel
->output
= CONSUMER_CHANNEL_MMAP
;
995 * In monitor mode, the streams associated with the channel will be put in
996 * a special list ONLY owned by this channel. So, the refcount is set to 1
997 * here meaning that the channel itself has streams that are referenced.
999 * On a channel deletion, once the channel is no longer visible, the
1000 * refcount is decremented and checked for a zero value to delete it. With
1001 * streams in no monitor mode, it will now be safe to destroy the channel.
1003 if (!channel
->monitor
) {
1004 channel
->refcount
= 1;
1007 strncpy(channel
->pathname
, pathname
, sizeof(channel
->pathname
));
1008 channel
->pathname
[sizeof(channel
->pathname
) - 1] = '\0';
1010 strncpy(channel
->name
, name
, sizeof(channel
->name
));
1011 channel
->name
[sizeof(channel
->name
) - 1] = '\0';
1013 if (root_shm_path
) {
1014 strncpy(channel
->root_shm_path
, root_shm_path
, sizeof(channel
->root_shm_path
));
1015 channel
->root_shm_path
[sizeof(channel
->root_shm_path
) - 1] = '\0';
1018 strncpy(channel
->shm_path
, shm_path
, sizeof(channel
->shm_path
));
1019 channel
->shm_path
[sizeof(channel
->shm_path
) - 1] = '\0';
1022 lttng_ht_node_init_u64(&channel
->node
, channel
->key
);
1024 channel
->wait_fd
= -1;
1026 CDS_INIT_LIST_HEAD(&channel
->streams
.head
);
1028 DBG("Allocated channel (key %" PRIu64
")", channel
->key
);
1035 * Add a channel to the global list protected by a mutex.
1037 * Always return 0 indicating success.
1039 int consumer_add_channel(struct lttng_consumer_channel
*channel
,
1040 struct lttng_consumer_local_data
*ctx
)
1042 pthread_mutex_lock(&consumer_data
.lock
);
1043 pthread_mutex_lock(&channel
->lock
);
1044 pthread_mutex_lock(&channel
->timer_lock
);
1047 * This gives us a guarantee that the channel we are about to add to the
1048 * channel hash table will be unique. See this function comment on the why
1049 * we need to steel the channel key at this stage.
1051 steal_channel_key(channel
->key
);
1054 lttng_ht_add_unique_u64(consumer_data
.channel_ht
, &channel
->node
);
1057 pthread_mutex_unlock(&channel
->timer_lock
);
1058 pthread_mutex_unlock(&channel
->lock
);
1059 pthread_mutex_unlock(&consumer_data
.lock
);
1061 if (channel
->wait_fd
!= -1 && channel
->type
== CONSUMER_CHANNEL_TYPE_DATA
) {
1062 notify_channel_pipe(ctx
, channel
, -1, CONSUMER_CHANNEL_ADD
);
1069 * Allocate the pollfd structure and the local view of the out fds to avoid
1070 * doing a lookup in the linked list and concurrency issues when writing is
1071 * needed. Called with consumer_data.lock held.
1073 * Returns the number of fds in the structures.
1075 static int update_poll_array(struct lttng_consumer_local_data
*ctx
,
1076 struct pollfd
**pollfd
, struct lttng_consumer_stream
**local_stream
,
1077 struct lttng_ht
*ht
, int *nb_inactive_fd
)
1080 struct lttng_ht_iter iter
;
1081 struct lttng_consumer_stream
*stream
;
1086 assert(local_stream
);
1088 DBG("Updating poll fd array");
1089 *nb_inactive_fd
= 0;
1091 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1093 * Only active streams with an active end point can be added to the
1094 * poll set and local stream storage of the thread.
1096 * There is a potential race here for endpoint_status to be updated
1097 * just after the check. However, this is OK since the stream(s) will
1098 * be deleted once the thread is notified that the end point state has
1099 * changed where this function will be called back again.
1101 * We track the number of inactive FDs because they still need to be
1102 * closed by the polling thread after a wakeup on the data_pipe or
1105 if (stream
->state
!= LTTNG_CONSUMER_ACTIVE_STREAM
||
1106 stream
->endpoint_status
== CONSUMER_ENDPOINT_INACTIVE
) {
1107 (*nb_inactive_fd
)++;
1111 * This clobbers way too much the debug output. Uncomment that if you
1112 * need it for debugging purposes.
1114 * DBG("Active FD %d", stream->wait_fd);
1116 (*pollfd
)[i
].fd
= stream
->wait_fd
;
1117 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1118 local_stream
[i
] = stream
;
1124 * Insert the consumer_data_pipe at the end of the array and don't
1125 * increment i so nb_fd is the number of real FD.
1127 (*pollfd
)[i
].fd
= lttng_pipe_get_readfd(ctx
->consumer_data_pipe
);
1128 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1130 (*pollfd
)[i
+ 1].fd
= lttng_pipe_get_readfd(ctx
->consumer_wakeup_pipe
);
1131 (*pollfd
)[i
+ 1].events
= POLLIN
| POLLPRI
;
1136 * Poll on the should_quit pipe and the command socket return -1 on
1137 * error, 1 if should exit, 0 if data is available on the command socket
1139 int lttng_consumer_poll_socket(struct pollfd
*consumer_sockpoll
)
1144 num_rdy
= poll(consumer_sockpoll
, 2, -1);
1145 if (num_rdy
== -1) {
1147 * Restart interrupted system call.
1149 if (errno
== EINTR
) {
1152 PERROR("Poll error");
1155 if (consumer_sockpoll
[0].revents
& (POLLIN
| POLLPRI
)) {
1156 DBG("consumer_should_quit wake up");
1163 * Set the error socket.
1165 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data
*ctx
,
1168 ctx
->consumer_error_socket
= sock
;
1172 * Set the command socket path.
1174 void lttng_consumer_set_command_sock_path(
1175 struct lttng_consumer_local_data
*ctx
, char *sock
)
1177 ctx
->consumer_command_sock_path
= sock
;
1181 * Send return code to the session daemon.
1182 * If the socket is not defined, we return 0, it is not a fatal error
1184 int lttng_consumer_send_error(struct lttng_consumer_local_data
*ctx
, int cmd
)
1186 if (ctx
->consumer_error_socket
> 0) {
1187 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
1188 sizeof(enum lttcomm_sessiond_command
));
1195 * Close all the tracefiles and stream fds and MUST be called when all
1196 * instances are destroyed i.e. when all threads were joined and are ended.
1198 void lttng_consumer_cleanup(void)
1200 struct lttng_ht_iter iter
;
1201 struct lttng_consumer_channel
*channel
;
1205 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
, channel
,
1207 consumer_del_channel(channel
);
1212 lttng_ht_destroy(consumer_data
.channel_ht
);
1214 cleanup_relayd_ht();
1216 lttng_ht_destroy(consumer_data
.stream_per_chan_id_ht
);
1219 * This HT contains streams that are freed by either the metadata thread or
1220 * the data thread so we do *nothing* on the hash table and simply destroy
1223 lttng_ht_destroy(consumer_data
.stream_list_ht
);
1227 * Called from signal handler.
1229 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
1234 ret
= lttng_write(ctx
->consumer_should_quit
[1], "4", 1);
1236 PERROR("write consumer quit");
1239 DBG("Consumer flag that it should quit");
1244 * Flush pending writes to trace output disk file.
1247 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1251 int outfd
= stream
->out_fd
;
1254 * This does a blocking write-and-wait on any page that belongs to the
1255 * subbuffer prior to the one we just wrote.
1256 * Don't care about error values, as these are just hints and ways to
1257 * limit the amount of page cache used.
1259 if (orig_offset
< stream
->max_sb_size
) {
1262 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1263 stream
->max_sb_size
,
1264 SYNC_FILE_RANGE_WAIT_BEFORE
1265 | SYNC_FILE_RANGE_WRITE
1266 | SYNC_FILE_RANGE_WAIT_AFTER
);
1268 * Give hints to the kernel about how we access the file:
1269 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1272 * We need to call fadvise again after the file grows because the
1273 * kernel does not seem to apply fadvise to non-existing parts of the
1276 * Call fadvise _after_ having waited for the page writeback to
1277 * complete because the dirty page writeback semantic is not well
1278 * defined. So it can be expected to lead to lower throughput in
1281 ret
= posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1282 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1283 if (ret
&& ret
!= -ENOSYS
) {
1285 PERROR("posix_fadvise on fd %i", outfd
);
1290 * Initialise the necessary environnement :
1291 * - create a new context
1292 * - create the poll_pipe
1293 * - create the should_quit pipe (for signal handler)
1294 * - create the thread pipe (for splice)
1296 * Takes a function pointer as argument, this function is called when data is
1297 * available on a buffer. This function is responsible to do the
1298 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1299 * buffer configuration and then kernctl_put_next_subbuf at the end.
1301 * Returns a pointer to the new context or NULL on error.
1303 struct lttng_consumer_local_data
*lttng_consumer_create(
1304 enum lttng_consumer_type type
,
1305 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1306 struct lttng_consumer_local_data
*ctx
),
1307 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1308 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1309 int (*update_stream
)(uint64_t stream_key
, uint32_t state
))
1312 struct lttng_consumer_local_data
*ctx
;
1314 assert(consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1315 consumer_data
.type
== type
);
1316 consumer_data
.type
= type
;
1318 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
1320 PERROR("allocating context");
1324 ctx
->consumer_error_socket
= -1;
1325 ctx
->consumer_metadata_socket
= -1;
1326 pthread_mutex_init(&ctx
->metadata_socket_lock
, NULL
);
1327 /* assign the callbacks */
1328 ctx
->on_buffer_ready
= buffer_ready
;
1329 ctx
->on_recv_channel
= recv_channel
;
1330 ctx
->on_recv_stream
= recv_stream
;
1331 ctx
->on_update_stream
= update_stream
;
1333 ctx
->consumer_data_pipe
= lttng_pipe_open(0);
1334 if (!ctx
->consumer_data_pipe
) {
1335 goto error_poll_pipe
;
1338 ctx
->consumer_wakeup_pipe
= lttng_pipe_open(0);
1339 if (!ctx
->consumer_wakeup_pipe
) {
1340 goto error_wakeup_pipe
;
1343 ret
= pipe(ctx
->consumer_should_quit
);
1345 PERROR("Error creating recv pipe");
1346 goto error_quit_pipe
;
1349 ret
= pipe(ctx
->consumer_channel_pipe
);
1351 PERROR("Error creating channel pipe");
1352 goto error_channel_pipe
;
1355 ctx
->consumer_metadata_pipe
= lttng_pipe_open(0);
1356 if (!ctx
->consumer_metadata_pipe
) {
1357 goto error_metadata_pipe
;
1362 error_metadata_pipe
:
1363 utils_close_pipe(ctx
->consumer_channel_pipe
);
1365 utils_close_pipe(ctx
->consumer_should_quit
);
1367 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1369 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1377 * Iterate over all streams of the hashtable and free them properly.
1379 static void destroy_data_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_stream(stream
, ht
);
1398 lttng_ht_destroy(ht
);
1402 * Iterate over all streams of the metadata hashtable and free them
1405 static void destroy_metadata_stream_ht(struct lttng_ht
*ht
)
1407 struct lttng_ht_iter iter
;
1408 struct lttng_consumer_stream
*stream
;
1415 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1417 * Ignore return value since we are currently cleaning up so any error
1420 (void) consumer_del_metadata_stream(stream
, ht
);
1424 lttng_ht_destroy(ht
);
1428 * Close all fds associated with the instance and free the context.
1430 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1434 DBG("Consumer destroying it. Closing everything.");
1440 destroy_data_stream_ht(data_ht
);
1441 destroy_metadata_stream_ht(metadata_ht
);
1443 ret
= close(ctx
->consumer_error_socket
);
1447 ret
= close(ctx
->consumer_metadata_socket
);
1451 utils_close_pipe(ctx
->consumer_channel_pipe
);
1452 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1453 lttng_pipe_destroy(ctx
->consumer_metadata_pipe
);
1454 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1455 utils_close_pipe(ctx
->consumer_should_quit
);
1457 unlink(ctx
->consumer_command_sock_path
);
1462 * Write the metadata stream id on the specified file descriptor.
1464 static int write_relayd_metadata_id(int fd
,
1465 struct lttng_consumer_stream
*stream
,
1466 struct consumer_relayd_sock_pair
*relayd
, unsigned long padding
)
1469 struct lttcomm_relayd_metadata_payload hdr
;
1471 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1472 hdr
.padding_size
= htobe32(padding
);
1473 ret
= lttng_write(fd
, (void *) &hdr
, sizeof(hdr
));
1474 if (ret
< sizeof(hdr
)) {
1476 * This error means that the fd's end is closed so ignore the PERROR
1477 * not to clubber the error output since this can happen in a normal
1480 if (errno
!= EPIPE
) {
1481 PERROR("write metadata stream id");
1483 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1485 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1486 * handle writting the missing part so report that as an error and
1487 * don't lie to the caller.
1492 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1493 stream
->relayd_stream_id
, padding
);
1500 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1501 * core function for writing trace buffers to either the local filesystem or
1504 * It must be called with the stream lock held.
1506 * Careful review MUST be put if any changes occur!
1508 * Returns the number of bytes written
1510 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1511 struct lttng_consumer_local_data
*ctx
,
1512 struct lttng_consumer_stream
*stream
, unsigned long len
,
1513 unsigned long padding
,
1514 struct ctf_packet_index
*index
)
1516 unsigned long mmap_offset
;
1519 off_t orig_offset
= stream
->out_fd_offset
;
1520 /* Default is on the disk */
1521 int outfd
= stream
->out_fd
;
1522 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1523 unsigned int relayd_hang_up
= 0;
1525 /* RCU lock for the relayd pointer */
1528 /* Flag that the current stream if set for network streaming. */
1529 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1530 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1531 if (relayd
== NULL
) {
1537 /* get the offset inside the fd to mmap */
1538 switch (consumer_data
.type
) {
1539 case LTTNG_CONSUMER_KERNEL
:
1540 mmap_base
= stream
->mmap_base
;
1541 ret
= kernctl_get_mmap_read_offset(stream
->wait_fd
, &mmap_offset
);
1543 PERROR("tracer ctl get_mmap_read_offset");
1547 case LTTNG_CONSUMER32_UST
:
1548 case LTTNG_CONSUMER64_UST
:
1549 mmap_base
= lttng_ustctl_get_mmap_base(stream
);
1551 ERR("read mmap get mmap base for stream %s", stream
->name
);
1555 ret
= lttng_ustctl_get_mmap_read_offset(stream
, &mmap_offset
);
1557 PERROR("tracer ctl get_mmap_read_offset");
1563 ERR("Unknown consumer_data type");
1567 /* Handle stream on the relayd if the output is on the network */
1569 unsigned long netlen
= len
;
1572 * Lock the control socket for the complete duration of the function
1573 * since from this point on we will use the socket.
1575 if (stream
->metadata_flag
) {
1576 /* Metadata requires the control socket. */
1577 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1578 if (stream
->reset_metadata_flag
) {
1579 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1580 stream
->relayd_stream_id
,
1581 stream
->metadata_version
);
1586 stream
->reset_metadata_flag
= 0;
1588 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1591 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1596 /* Use the returned socket. */
1599 /* Write metadata stream id before payload */
1600 if (stream
->metadata_flag
) {
1601 ret
= write_relayd_metadata_id(outfd
, stream
, relayd
, padding
);
1608 /* No streaming, we have to set the len with the full padding */
1611 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1612 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1614 ERR("Reset metadata file");
1617 stream
->reset_metadata_flag
= 0;
1621 * Check if we need to change the tracefile before writing the packet.
1623 if (stream
->chan
->tracefile_size
> 0 &&
1624 (stream
->tracefile_size_current
+ len
) >
1625 stream
->chan
->tracefile_size
) {
1626 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1627 stream
->name
, stream
->chan
->tracefile_size
,
1628 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1629 stream
->out_fd
, &(stream
->tracefile_count_current
),
1632 ERR("Rotating output file");
1635 outfd
= stream
->out_fd
;
1637 if (stream
->index_file
) {
1638 lttng_index_file_put(stream
->index_file
);
1639 stream
->index_file
= lttng_index_file_create(stream
->chan
->pathname
,
1640 stream
->name
, stream
->uid
, stream
->gid
,
1641 stream
->chan
->tracefile_size
,
1642 stream
->tracefile_count_current
,
1643 CTF_INDEX_MAJOR
, CTF_INDEX_MINOR
);
1644 if (!stream
->index_file
) {
1649 /* Reset current size because we just perform a rotation. */
1650 stream
->tracefile_size_current
= 0;
1651 stream
->out_fd_offset
= 0;
1654 stream
->tracefile_size_current
+= len
;
1656 index
->offset
= htobe64(stream
->out_fd_offset
);
1661 * This call guarantee that len or less is returned. It's impossible to
1662 * receive a ret value that is bigger than len.
1664 ret
= lttng_write(outfd
, mmap_base
+ mmap_offset
, len
);
1665 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, len
);
1666 if (ret
< 0 || ((size_t) ret
!= len
)) {
1668 * Report error to caller if nothing was written else at least send the
1676 /* Socket operation failed. We consider the relayd dead */
1677 if (errno
== EPIPE
|| errno
== EINVAL
|| errno
== EBADF
) {
1679 * This is possible if the fd is closed on the other side
1680 * (outfd) or any write problem. It can be verbose a bit for a
1681 * normal execution if for instance the relayd is stopped
1682 * abruptly. This can happen so set this to a DBG statement.
1684 DBG("Consumer mmap write detected relayd hang up");
1686 /* Unhandled error, print it and stop function right now. */
1687 PERROR("Error in write mmap (ret %zd != len %lu)", ret
, len
);
1691 stream
->output_written
+= ret
;
1693 /* This call is useless on a socket so better save a syscall. */
1695 /* This won't block, but will start writeout asynchronously */
1696 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, len
,
1697 SYNC_FILE_RANGE_WRITE
);
1698 stream
->out_fd_offset
+= len
;
1699 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1704 * This is a special case that the relayd has closed its socket. Let's
1705 * cleanup the relayd object and all associated streams.
1707 if (relayd
&& relayd_hang_up
) {
1708 ERR("Relayd hangup. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
1709 lttng_consumer_cleanup_relayd(relayd
);
1713 /* Unlock only if ctrl socket used */
1714 if (relayd
&& stream
->metadata_flag
) {
1715 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1723 * Splice the data from the ring buffer to the tracefile.
1725 * It must be called with the stream lock held.
1727 * Returns the number of bytes spliced.
1729 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1730 struct lttng_consumer_local_data
*ctx
,
1731 struct lttng_consumer_stream
*stream
, unsigned long len
,
1732 unsigned long padding
,
1733 struct ctf_packet_index
*index
)
1735 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1737 off_t orig_offset
= stream
->out_fd_offset
;
1738 int fd
= stream
->wait_fd
;
1739 /* Default is on the disk */
1740 int outfd
= stream
->out_fd
;
1741 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1743 unsigned int relayd_hang_up
= 0;
1745 switch (consumer_data
.type
) {
1746 case LTTNG_CONSUMER_KERNEL
:
1748 case LTTNG_CONSUMER32_UST
:
1749 case LTTNG_CONSUMER64_UST
:
1750 /* Not supported for user space tracing */
1753 ERR("Unknown consumer_data type");
1757 /* RCU lock for the relayd pointer */
1760 /* Flag that the current stream if set for network streaming. */
1761 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1762 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1763 if (relayd
== NULL
) {
1768 splice_pipe
= stream
->splice_pipe
;
1770 /* Write metadata stream id before payload */
1772 unsigned long total_len
= len
;
1774 if (stream
->metadata_flag
) {
1776 * Lock the control socket for the complete duration of the function
1777 * since from this point on we will use the socket.
1779 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1781 if (stream
->reset_metadata_flag
) {
1782 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1783 stream
->relayd_stream_id
,
1784 stream
->metadata_version
);
1789 stream
->reset_metadata_flag
= 0;
1791 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
, relayd
,
1799 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1802 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1808 /* Use the returned socket. */
1811 /* No streaming, we have to set the len with the full padding */
1814 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1815 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1817 ERR("Reset metadata file");
1820 stream
->reset_metadata_flag
= 0;
1823 * Check if we need to change the tracefile before writing the packet.
1825 if (stream
->chan
->tracefile_size
> 0 &&
1826 (stream
->tracefile_size_current
+ len
) >
1827 stream
->chan
->tracefile_size
) {
1828 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1829 stream
->name
, stream
->chan
->tracefile_size
,
1830 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1831 stream
->out_fd
, &(stream
->tracefile_count_current
),
1835 ERR("Rotating output file");
1838 outfd
= stream
->out_fd
;
1840 if (stream
->index_file
) {
1841 lttng_index_file_put(stream
->index_file
);
1842 stream
->index_file
= lttng_index_file_create(stream
->chan
->pathname
,
1843 stream
->name
, stream
->uid
, stream
->gid
,
1844 stream
->chan
->tracefile_size
,
1845 stream
->tracefile_count_current
,
1846 CTF_INDEX_MAJOR
, CTF_INDEX_MINOR
);
1847 if (!stream
->index_file
) {
1852 /* Reset current size because we just perform a rotation. */
1853 stream
->tracefile_size_current
= 0;
1854 stream
->out_fd_offset
= 0;
1857 stream
->tracefile_size_current
+= len
;
1858 index
->offset
= htobe64(stream
->out_fd_offset
);
1862 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1863 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1864 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1865 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1866 DBG("splice chan to pipe, ret %zd", ret_splice
);
1867 if (ret_splice
< 0) {
1870 PERROR("Error in relay splice");
1874 /* Handle stream on the relayd if the output is on the network */
1875 if (relayd
&& stream
->metadata_flag
) {
1876 size_t metadata_payload_size
=
1877 sizeof(struct lttcomm_relayd_metadata_payload
);
1879 /* Update counter to fit the spliced data */
1880 ret_splice
+= metadata_payload_size
;
1881 len
+= metadata_payload_size
;
1883 * We do this so the return value can match the len passed as
1884 * argument to this function.
1886 written
-= metadata_payload_size
;
1889 /* Splice data out */
1890 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1891 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1892 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
1894 if (ret_splice
< 0) {
1899 } else if (ret_splice
> len
) {
1901 * We don't expect this code path to be executed but you never know
1902 * so this is an extra protection agains a buggy splice().
1905 written
+= ret_splice
;
1906 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice
,
1910 /* All good, update current len and continue. */
1914 /* This call is useless on a socket so better save a syscall. */
1916 /* This won't block, but will start writeout asynchronously */
1917 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1918 SYNC_FILE_RANGE_WRITE
);
1919 stream
->out_fd_offset
+= ret_splice
;
1921 stream
->output_written
+= ret_splice
;
1922 written
+= ret_splice
;
1925 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1931 * This is a special case that the relayd has closed its socket. Let's
1932 * cleanup the relayd object and all associated streams.
1934 if (relayd
&& relayd_hang_up
) {
1935 ERR("Relayd hangup. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
1936 lttng_consumer_cleanup_relayd(relayd
);
1937 /* Skip splice error so the consumer does not fail */
1942 /* send the appropriate error description to sessiond */
1945 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1948 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1951 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
1956 if (relayd
&& stream
->metadata_flag
) {
1957 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1965 * Take a snapshot for a specific fd
1967 * Returns 0 on success, < 0 on error
1969 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
1971 switch (consumer_data
.type
) {
1972 case LTTNG_CONSUMER_KERNEL
:
1973 return lttng_kconsumer_take_snapshot(stream
);
1974 case LTTNG_CONSUMER32_UST
:
1975 case LTTNG_CONSUMER64_UST
:
1976 return lttng_ustconsumer_take_snapshot(stream
);
1978 ERR("Unknown consumer_data type");
1985 * Get the produced position
1987 * Returns 0 on success, < 0 on error
1989 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
1992 switch (consumer_data
.type
) {
1993 case LTTNG_CONSUMER_KERNEL
:
1994 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
1995 case LTTNG_CONSUMER32_UST
:
1996 case LTTNG_CONSUMER64_UST
:
1997 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
1999 ERR("Unknown consumer_data type");
2005 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
2006 int sock
, struct pollfd
*consumer_sockpoll
)
2008 switch (consumer_data
.type
) {
2009 case LTTNG_CONSUMER_KERNEL
:
2010 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2011 case LTTNG_CONSUMER32_UST
:
2012 case LTTNG_CONSUMER64_UST
:
2013 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2015 ERR("Unknown consumer_data type");
2021 void lttng_consumer_close_all_metadata(void)
2023 switch (consumer_data
.type
) {
2024 case LTTNG_CONSUMER_KERNEL
:
2026 * The Kernel consumer has a different metadata scheme so we don't
2027 * close anything because the stream will be closed by the session
2031 case LTTNG_CONSUMER32_UST
:
2032 case LTTNG_CONSUMER64_UST
:
2034 * Close all metadata streams. The metadata hash table is passed and
2035 * this call iterates over it by closing all wakeup fd. This is safe
2036 * because at this point we are sure that the metadata producer is
2037 * either dead or blocked.
2039 lttng_ustconsumer_close_all_metadata(metadata_ht
);
2042 ERR("Unknown consumer_data type");
2048 * Clean up a metadata stream and free its memory.
2050 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
2051 struct lttng_ht
*ht
)
2053 struct lttng_consumer_channel
*free_chan
= NULL
;
2057 * This call should NEVER receive regular stream. It must always be
2058 * metadata stream and this is crucial for data structure synchronization.
2060 assert(stream
->metadata_flag
);
2062 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
2064 pthread_mutex_lock(&consumer_data
.lock
);
2065 pthread_mutex_lock(&stream
->chan
->lock
);
2066 pthread_mutex_lock(&stream
->lock
);
2067 if (stream
->chan
->metadata_cache
) {
2068 /* Only applicable to userspace consumers. */
2069 pthread_mutex_lock(&stream
->chan
->metadata_cache
->lock
);
2072 /* Remove any reference to that stream. */
2073 consumer_stream_delete(stream
, ht
);
2075 /* Close down everything including the relayd if one. */
2076 consumer_stream_close(stream
);
2077 /* Destroy tracer buffers of the stream. */
2078 consumer_stream_destroy_buffers(stream
);
2080 /* Atomically decrement channel refcount since other threads can use it. */
2081 if (!uatomic_sub_return(&stream
->chan
->refcount
, 1)
2082 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
2083 /* Go for channel deletion! */
2084 free_chan
= stream
->chan
;
2088 * Nullify the stream reference so it is not used after deletion. The
2089 * channel lock MUST be acquired before being able to check for a NULL
2092 stream
->chan
->metadata_stream
= NULL
;
2094 if (stream
->chan
->metadata_cache
) {
2095 pthread_mutex_unlock(&stream
->chan
->metadata_cache
->lock
);
2097 pthread_mutex_unlock(&stream
->lock
);
2098 pthread_mutex_unlock(&stream
->chan
->lock
);
2099 pthread_mutex_unlock(&consumer_data
.lock
);
2102 consumer_del_channel(free_chan
);
2105 consumer_stream_free(stream
);
2109 * Action done with the metadata stream when adding it to the consumer internal
2110 * data structures to handle it.
2112 int consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
2114 struct lttng_ht
*ht
= metadata_ht
;
2116 struct lttng_ht_iter iter
;
2117 struct lttng_ht_node_u64
*node
;
2122 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2124 pthread_mutex_lock(&consumer_data
.lock
);
2125 pthread_mutex_lock(&stream
->chan
->lock
);
2126 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2127 pthread_mutex_lock(&stream
->lock
);
2130 * From here, refcounts are updated so be _careful_ when returning an error
2137 * Lookup the stream just to make sure it does not exist in our internal
2138 * state. This should NEVER happen.
2140 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2141 node
= lttng_ht_iter_get_node_u64(&iter
);
2145 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2146 * in terms of destroying the associated channel, because the action that
2147 * causes the count to become 0 also causes a stream to be added. The
2148 * channel deletion will thus be triggered by the following removal of this
2151 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2152 /* Increment refcount before decrementing nb_init_stream_left */
2154 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2157 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2159 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
2160 &stream
->node_channel_id
);
2163 * Add stream to the stream_list_ht of the consumer data. No need to steal
2164 * the key since the HT does not use it and we allow to add redundant keys
2167 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2171 pthread_mutex_unlock(&stream
->lock
);
2172 pthread_mutex_unlock(&stream
->chan
->lock
);
2173 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2174 pthread_mutex_unlock(&consumer_data
.lock
);
2179 * Delete data stream that are flagged for deletion (endpoint_status).
2181 static void validate_endpoint_status_data_stream(void)
2183 struct lttng_ht_iter iter
;
2184 struct lttng_consumer_stream
*stream
;
2186 DBG("Consumer delete flagged data stream");
2189 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2190 /* Validate delete flag of the stream */
2191 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2194 /* Delete it right now */
2195 consumer_del_stream(stream
, data_ht
);
2201 * Delete metadata stream that are flagged for deletion (endpoint_status).
2203 static void validate_endpoint_status_metadata_stream(
2204 struct lttng_poll_event
*pollset
)
2206 struct lttng_ht_iter iter
;
2207 struct lttng_consumer_stream
*stream
;
2209 DBG("Consumer delete flagged metadata stream");
2214 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2215 /* Validate delete flag of the stream */
2216 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2220 * Remove from pollset so the metadata thread can continue without
2221 * blocking on a deleted stream.
2223 lttng_poll_del(pollset
, stream
->wait_fd
);
2225 /* Delete it right now */
2226 consumer_del_metadata_stream(stream
, metadata_ht
);
2232 * Thread polls on metadata file descriptor and write them on disk or on the
2235 void *consumer_thread_metadata_poll(void *data
)
2237 int ret
, i
, pollfd
, err
= -1;
2238 uint32_t revents
, nb_fd
;
2239 struct lttng_consumer_stream
*stream
= NULL
;
2240 struct lttng_ht_iter iter
;
2241 struct lttng_ht_node_u64
*node
;
2242 struct lttng_poll_event events
;
2243 struct lttng_consumer_local_data
*ctx
= data
;
2246 rcu_register_thread();
2248 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2250 if (testpoint(consumerd_thread_metadata
)) {
2251 goto error_testpoint
;
2254 health_code_update();
2256 DBG("Thread metadata poll started");
2258 /* Size is set to 1 for the consumer_metadata pipe */
2259 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2261 ERR("Poll set creation failed");
2265 ret
= lttng_poll_add(&events
,
2266 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2272 DBG("Metadata main loop started");
2276 health_code_update();
2277 health_poll_entry();
2278 DBG("Metadata poll wait");
2279 ret
= lttng_poll_wait(&events
, -1);
2280 DBG("Metadata poll return from wait with %d fd(s)",
2281 LTTNG_POLL_GETNB(&events
));
2283 DBG("Metadata event caught in thread");
2285 if (errno
== EINTR
) {
2286 ERR("Poll EINTR caught");
2289 if (LTTNG_POLL_GETNB(&events
) == 0) {
2290 err
= 0; /* All is OK */
2297 /* From here, the event is a metadata wait fd */
2298 for (i
= 0; i
< nb_fd
; i
++) {
2299 health_code_update();
2301 revents
= LTTNG_POLL_GETEV(&events
, i
);
2302 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2305 /* No activity for this FD (poll implementation). */
2309 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2310 if (revents
& LPOLLIN
) {
2313 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2314 &stream
, sizeof(stream
));
2315 if (pipe_len
< sizeof(stream
)) {
2317 PERROR("read metadata stream");
2320 * Remove the pipe from the poll set and continue the loop
2321 * since their might be data to consume.
2323 lttng_poll_del(&events
,
2324 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2325 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2329 /* A NULL stream means that the state has changed. */
2330 if (stream
== NULL
) {
2331 /* Check for deleted streams. */
2332 validate_endpoint_status_metadata_stream(&events
);
2336 DBG("Adding metadata stream %d to poll set",
2339 /* Add metadata stream to the global poll events list */
2340 lttng_poll_add(&events
, stream
->wait_fd
,
2341 LPOLLIN
| LPOLLPRI
| LPOLLHUP
);
2342 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2343 DBG("Metadata thread pipe hung up");
2345 * Remove the pipe from the poll set and continue the loop
2346 * since their might be data to consume.
2348 lttng_poll_del(&events
,
2349 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2350 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2353 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2357 /* Handle other stream */
2363 uint64_t tmp_id
= (uint64_t) pollfd
;
2365 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2367 node
= lttng_ht_iter_get_node_u64(&iter
);
2370 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2373 if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2374 /* Get the data out of the metadata file descriptor */
2375 DBG("Metadata available on fd %d", pollfd
);
2376 assert(stream
->wait_fd
== pollfd
);
2379 health_code_update();
2381 len
= ctx
->on_buffer_ready(stream
, ctx
);
2383 * We don't check the return value here since if we get
2384 * a negative len, it means an error occurred thus we
2385 * simply remove it from the poll set and free the
2390 /* It's ok to have an unavailable sub-buffer */
2391 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2392 /* Clean up stream from consumer and free it. */
2393 lttng_poll_del(&events
, stream
->wait_fd
);
2394 consumer_del_metadata_stream(stream
, metadata_ht
);
2396 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2397 DBG("Metadata fd %d is hup|err.", pollfd
);
2398 if (!stream
->hangup_flush_done
2399 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2400 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2401 DBG("Attempting to flush and consume the UST buffers");
2402 lttng_ustconsumer_on_stream_hangup(stream
);
2404 /* We just flushed the stream now read it. */
2406 health_code_update();
2408 len
= ctx
->on_buffer_ready(stream
, ctx
);
2410 * We don't check the return value here since if we get
2411 * a negative len, it means an error occurred thus we
2412 * simply remove it from the poll set and free the
2418 lttng_poll_del(&events
, stream
->wait_fd
);
2420 * This call update the channel states, closes file descriptors
2421 * and securely free the stream.
2423 consumer_del_metadata_stream(stream
, metadata_ht
);
2425 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2429 /* Release RCU lock for the stream looked up */
2437 DBG("Metadata poll thread exiting");
2439 lttng_poll_clean(&events
);
2444 ERR("Health error occurred in %s", __func__
);
2446 health_unregister(health_consumerd
);
2447 rcu_unregister_thread();
2452 * This thread polls the fds in the set to consume the data and write
2453 * it to tracefile if necessary.
2455 void *consumer_thread_data_poll(void *data
)
2457 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2458 struct pollfd
*pollfd
= NULL
;
2459 /* local view of the streams */
2460 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2461 /* local view of consumer_data.fds_count */
2463 /* Number of FDs with CONSUMER_ENDPOINT_INACTIVE but still open. */
2464 int nb_inactive_fd
= 0;
2465 struct lttng_consumer_local_data
*ctx
= data
;
2468 rcu_register_thread();
2470 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2472 if (testpoint(consumerd_thread_data
)) {
2473 goto error_testpoint
;
2476 health_code_update();
2478 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
*));
2479 if (local_stream
== NULL
) {
2480 PERROR("local_stream malloc");
2485 health_code_update();
2491 * the fds set has been updated, we need to update our
2492 * local array as well
2494 pthread_mutex_lock(&consumer_data
.lock
);
2495 if (consumer_data
.need_update
) {
2500 local_stream
= NULL
;
2503 * Allocate for all fds +1 for the consumer_data_pipe and +1 for
2506 pollfd
= zmalloc((consumer_data
.stream_count
+ 2) * sizeof(struct pollfd
));
2507 if (pollfd
== NULL
) {
2508 PERROR("pollfd malloc");
2509 pthread_mutex_unlock(&consumer_data
.lock
);
2513 local_stream
= zmalloc((consumer_data
.stream_count
+ 2) *
2514 sizeof(struct lttng_consumer_stream
*));
2515 if (local_stream
== NULL
) {
2516 PERROR("local_stream malloc");
2517 pthread_mutex_unlock(&consumer_data
.lock
);
2520 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2521 data_ht
, &nb_inactive_fd
);
2523 ERR("Error in allocating pollfd or local_outfds");
2524 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2525 pthread_mutex_unlock(&consumer_data
.lock
);
2529 consumer_data
.need_update
= 0;
2531 pthread_mutex_unlock(&consumer_data
.lock
);
2533 /* No FDs and consumer_quit, consumer_cleanup the thread */
2534 if (nb_fd
== 0 && consumer_quit
== 1 && nb_inactive_fd
== 0) {
2535 err
= 0; /* All is OK */
2538 /* poll on the array of fds */
2540 DBG("polling on %d fd", nb_fd
+ 2);
2541 health_poll_entry();
2542 num_rdy
= poll(pollfd
, nb_fd
+ 2, -1);
2544 DBG("poll num_rdy : %d", num_rdy
);
2545 if (num_rdy
== -1) {
2547 * Restart interrupted system call.
2549 if (errno
== EINTR
) {
2552 PERROR("Poll error");
2553 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2555 } else if (num_rdy
== 0) {
2556 DBG("Polling thread timed out");
2561 * If the consumer_data_pipe triggered poll go directly to the
2562 * beginning of the loop to update the array. We want to prioritize
2563 * array update over low-priority reads.
2565 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2566 ssize_t pipe_readlen
;
2568 DBG("consumer_data_pipe wake up");
2569 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2570 &new_stream
, sizeof(new_stream
));
2571 if (pipe_readlen
< sizeof(new_stream
)) {
2572 PERROR("Consumer data pipe");
2573 /* Continue so we can at least handle the current stream(s). */
2578 * If the stream is NULL, just ignore it. It's also possible that
2579 * the sessiond poll thread changed the consumer_quit state and is
2580 * waking us up to test it.
2582 if (new_stream
== NULL
) {
2583 validate_endpoint_status_data_stream();
2587 /* Continue to update the local streams and handle prio ones */
2591 /* Handle wakeup pipe. */
2592 if (pollfd
[nb_fd
+ 1].revents
& (POLLIN
| POLLPRI
)) {
2594 ssize_t pipe_readlen
;
2596 pipe_readlen
= lttng_pipe_read(ctx
->consumer_wakeup_pipe
, &dummy
,
2598 if (pipe_readlen
< 0) {
2599 PERROR("Consumer data wakeup pipe");
2601 /* We've been awakened to handle stream(s). */
2602 ctx
->has_wakeup
= 0;
2605 /* Take care of high priority channels first. */
2606 for (i
= 0; i
< nb_fd
; i
++) {
2607 health_code_update();
2609 if (local_stream
[i
] == NULL
) {
2612 if (pollfd
[i
].revents
& POLLPRI
) {
2613 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2615 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2616 /* it's ok to have an unavailable sub-buffer */
2617 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2618 /* Clean the stream and free it. */
2619 consumer_del_stream(local_stream
[i
], data_ht
);
2620 local_stream
[i
] = NULL
;
2621 } else if (len
> 0) {
2622 local_stream
[i
]->data_read
= 1;
2628 * If we read high prio channel in this loop, try again
2629 * for more high prio data.
2635 /* Take care of low priority channels. */
2636 for (i
= 0; i
< nb_fd
; i
++) {
2637 health_code_update();
2639 if (local_stream
[i
] == NULL
) {
2642 if ((pollfd
[i
].revents
& POLLIN
) ||
2643 local_stream
[i
]->hangup_flush_done
||
2644 local_stream
[i
]->has_data
) {
2645 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2646 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2647 /* it's ok to have an unavailable sub-buffer */
2648 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2649 /* Clean the stream and free it. */
2650 consumer_del_stream(local_stream
[i
], data_ht
);
2651 local_stream
[i
] = NULL
;
2652 } else if (len
> 0) {
2653 local_stream
[i
]->data_read
= 1;
2658 /* Handle hangup and errors */
2659 for (i
= 0; i
< nb_fd
; i
++) {
2660 health_code_update();
2662 if (local_stream
[i
] == NULL
) {
2665 if (!local_stream
[i
]->hangup_flush_done
2666 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2667 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2668 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2669 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2671 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2672 /* Attempt read again, for the data we just flushed. */
2673 local_stream
[i
]->data_read
= 1;
2676 * If the poll flag is HUP/ERR/NVAL and we have
2677 * read no data in this pass, we can remove the
2678 * stream from its hash table.
2680 if ((pollfd
[i
].revents
& POLLHUP
)) {
2681 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2682 if (!local_stream
[i
]->data_read
) {
2683 consumer_del_stream(local_stream
[i
], data_ht
);
2684 local_stream
[i
] = NULL
;
2687 } else if (pollfd
[i
].revents
& POLLERR
) {
2688 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2689 if (!local_stream
[i
]->data_read
) {
2690 consumer_del_stream(local_stream
[i
], data_ht
);
2691 local_stream
[i
] = NULL
;
2694 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2695 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2696 if (!local_stream
[i
]->data_read
) {
2697 consumer_del_stream(local_stream
[i
], data_ht
);
2698 local_stream
[i
] = NULL
;
2702 if (local_stream
[i
] != NULL
) {
2703 local_stream
[i
]->data_read
= 0;
2710 DBG("polling thread exiting");
2715 * Close the write side of the pipe so epoll_wait() in
2716 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2717 * read side of the pipe. If we close them both, epoll_wait strangely does
2718 * not return and could create a endless wait period if the pipe is the
2719 * only tracked fd in the poll set. The thread will take care of closing
2722 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2727 ERR("Health error occurred in %s", __func__
);
2729 health_unregister(health_consumerd
);
2731 rcu_unregister_thread();
2736 * Close wake-up end of each stream belonging to the channel. This will
2737 * allow the poll() on the stream read-side to detect when the
2738 * write-side (application) finally closes them.
2741 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2743 struct lttng_ht
*ht
;
2744 struct lttng_consumer_stream
*stream
;
2745 struct lttng_ht_iter iter
;
2747 ht
= consumer_data
.stream_per_chan_id_ht
;
2750 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2751 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2752 ht
->match_fct
, &channel
->key
,
2753 &iter
.iter
, stream
, node_channel_id
.node
) {
2755 * Protect against teardown with mutex.
2757 pthread_mutex_lock(&stream
->lock
);
2758 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2761 switch (consumer_data
.type
) {
2762 case LTTNG_CONSUMER_KERNEL
:
2764 case LTTNG_CONSUMER32_UST
:
2765 case LTTNG_CONSUMER64_UST
:
2766 if (stream
->metadata_flag
) {
2767 /* Safe and protected by the stream lock. */
2768 lttng_ustconsumer_close_metadata(stream
->chan
);
2771 * Note: a mutex is taken internally within
2772 * liblttng-ust-ctl to protect timer wakeup_fd
2773 * use from concurrent close.
2775 lttng_ustconsumer_close_stream_wakeup(stream
);
2779 ERR("Unknown consumer_data type");
2783 pthread_mutex_unlock(&stream
->lock
);
2788 static void destroy_channel_ht(struct lttng_ht
*ht
)
2790 struct lttng_ht_iter iter
;
2791 struct lttng_consumer_channel
*channel
;
2799 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2800 ret
= lttng_ht_del(ht
, &iter
);
2805 lttng_ht_destroy(ht
);
2809 * This thread polls the channel fds to detect when they are being
2810 * closed. It closes all related streams if the channel is detected as
2811 * closed. It is currently only used as a shim layer for UST because the
2812 * consumerd needs to keep the per-stream wakeup end of pipes open for
2815 void *consumer_thread_channel_poll(void *data
)
2817 int ret
, i
, pollfd
, err
= -1;
2818 uint32_t revents
, nb_fd
;
2819 struct lttng_consumer_channel
*chan
= NULL
;
2820 struct lttng_ht_iter iter
;
2821 struct lttng_ht_node_u64
*node
;
2822 struct lttng_poll_event events
;
2823 struct lttng_consumer_local_data
*ctx
= data
;
2824 struct lttng_ht
*channel_ht
;
2826 rcu_register_thread();
2828 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
2830 if (testpoint(consumerd_thread_channel
)) {
2831 goto error_testpoint
;
2834 health_code_update();
2836 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2838 /* ENOMEM at this point. Better to bail out. */
2842 DBG("Thread channel poll started");
2844 /* Size is set to 1 for the consumer_channel pipe */
2845 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2847 ERR("Poll set creation failed");
2851 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2857 DBG("Channel main loop started");
2861 health_code_update();
2862 DBG("Channel poll wait");
2863 health_poll_entry();
2864 ret
= lttng_poll_wait(&events
, -1);
2865 DBG("Channel poll return from wait with %d fd(s)",
2866 LTTNG_POLL_GETNB(&events
));
2868 DBG("Channel event caught in thread");
2870 if (errno
== EINTR
) {
2871 ERR("Poll EINTR caught");
2874 if (LTTNG_POLL_GETNB(&events
) == 0) {
2875 err
= 0; /* All is OK */
2882 /* From here, the event is a channel wait fd */
2883 for (i
= 0; i
< nb_fd
; i
++) {
2884 health_code_update();
2886 revents
= LTTNG_POLL_GETEV(&events
, i
);
2887 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2890 /* No activity for this FD (poll implementation). */
2894 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2895 if (revents
& LPOLLIN
) {
2896 enum consumer_channel_action action
;
2899 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
2902 ERR("Error reading channel pipe");
2904 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2909 case CONSUMER_CHANNEL_ADD
:
2910 DBG("Adding channel %d to poll set",
2913 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
2916 lttng_ht_add_unique_u64(channel_ht
,
2917 &chan
->wait_fd_node
);
2919 /* Add channel to the global poll events list */
2920 lttng_poll_add(&events
, chan
->wait_fd
,
2921 LPOLLERR
| LPOLLHUP
);
2923 case CONSUMER_CHANNEL_DEL
:
2926 * This command should never be called if the channel
2927 * has streams monitored by either the data or metadata
2928 * thread. The consumer only notify this thread with a
2929 * channel del. command if it receives a destroy
2930 * channel command from the session daemon that send it
2931 * if a command prior to the GET_CHANNEL failed.
2935 chan
= consumer_find_channel(key
);
2938 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
2941 lttng_poll_del(&events
, chan
->wait_fd
);
2942 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
2943 ret
= lttng_ht_del(channel_ht
, &iter
);
2946 switch (consumer_data
.type
) {
2947 case LTTNG_CONSUMER_KERNEL
:
2949 case LTTNG_CONSUMER32_UST
:
2950 case LTTNG_CONSUMER64_UST
:
2951 health_code_update();
2952 /* Destroy streams that might have been left in the stream list. */
2953 clean_channel_stream_list(chan
);
2956 ERR("Unknown consumer_data type");
2961 * Release our own refcount. Force channel deletion even if
2962 * streams were not initialized.
2964 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
2965 consumer_del_channel(chan
);
2970 case CONSUMER_CHANNEL_QUIT
:
2972 * Remove the pipe from the poll set and continue the loop
2973 * since their might be data to consume.
2975 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2978 ERR("Unknown action");
2981 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2982 DBG("Channel thread pipe hung up");
2984 * Remove the pipe from the poll set and continue the loop
2985 * since their might be data to consume.
2987 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2990 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2994 /* Handle other stream */
3000 uint64_t tmp_id
= (uint64_t) pollfd
;
3002 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
3004 node
= lttng_ht_iter_get_node_u64(&iter
);
3007 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
3010 /* Check for error event */
3011 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
3012 DBG("Channel fd %d is hup|err.", pollfd
);
3014 lttng_poll_del(&events
, chan
->wait_fd
);
3015 ret
= lttng_ht_del(channel_ht
, &iter
);
3019 * This will close the wait fd for each stream associated to
3020 * this channel AND monitored by the data/metadata thread thus
3021 * will be clean by the right thread.
3023 consumer_close_channel_streams(chan
);
3025 /* Release our own refcount */
3026 if (!uatomic_sub_return(&chan
->refcount
, 1)
3027 && !uatomic_read(&chan
->nb_init_stream_left
)) {
3028 consumer_del_channel(chan
);
3031 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3036 /* Release RCU lock for the channel looked up */
3044 lttng_poll_clean(&events
);
3046 destroy_channel_ht(channel_ht
);
3049 DBG("Channel poll thread exiting");
3052 ERR("Health error occurred in %s", __func__
);
3054 health_unregister(health_consumerd
);
3055 rcu_unregister_thread();
3059 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
3060 struct pollfd
*sockpoll
, int client_socket
)
3067 ret
= lttng_consumer_poll_socket(sockpoll
);
3071 DBG("Metadata connection on client_socket");
3073 /* Blocking call, waiting for transmission */
3074 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
3075 if (ctx
->consumer_metadata_socket
< 0) {
3076 WARN("On accept metadata");
3087 * This thread listens on the consumerd socket and receives the file
3088 * descriptors from the session daemon.
3090 void *consumer_thread_sessiond_poll(void *data
)
3092 int sock
= -1, client_socket
, ret
, err
= -1;
3094 * structure to poll for incoming data on communication socket avoids
3095 * making blocking sockets.
3097 struct pollfd consumer_sockpoll
[2];
3098 struct lttng_consumer_local_data
*ctx
= data
;
3100 rcu_register_thread();
3102 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
3104 if (testpoint(consumerd_thread_sessiond
)) {
3105 goto error_testpoint
;
3108 health_code_update();
3110 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
3111 unlink(ctx
->consumer_command_sock_path
);
3112 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
3113 if (client_socket
< 0) {
3114 ERR("Cannot create command socket");
3118 ret
= lttcomm_listen_unix_sock(client_socket
);
3123 DBG("Sending ready command to lttng-sessiond");
3124 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3125 /* return < 0 on error, but == 0 is not fatal */
3127 ERR("Error sending ready command to lttng-sessiond");
3131 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3132 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3133 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3134 consumer_sockpoll
[1].fd
= client_socket
;
3135 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3137 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3145 DBG("Connection on client_socket");
3147 /* Blocking call, waiting for transmission */
3148 sock
= lttcomm_accept_unix_sock(client_socket
);
3155 * Setup metadata socket which is the second socket connection on the
3156 * command unix socket.
3158 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3167 /* This socket is not useful anymore. */
3168 ret
= close(client_socket
);
3170 PERROR("close client_socket");
3174 /* update the polling structure to poll on the established socket */
3175 consumer_sockpoll
[1].fd
= sock
;
3176 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3179 health_code_update();
3181 health_poll_entry();
3182 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3191 DBG("Incoming command on sock");
3192 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3195 * This could simply be a session daemon quitting. Don't output
3198 DBG("Communication interrupted on command socket");
3202 if (consumer_quit
) {
3203 DBG("consumer_thread_receive_fds received quit from signal");
3204 err
= 0; /* All is OK */
3207 DBG("received command on sock");
3213 DBG("Consumer thread sessiond poll exiting");
3216 * Close metadata streams since the producer is the session daemon which
3219 * NOTE: for now, this only applies to the UST tracer.
3221 lttng_consumer_close_all_metadata();
3224 * when all fds have hung up, the polling thread
3230 * Notify the data poll thread to poll back again and test the
3231 * consumer_quit state that we just set so to quit gracefully.
3233 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3235 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3237 notify_health_quit_pipe(health_quit_pipe
);
3239 /* Cleaning up possibly open sockets. */
3243 PERROR("close sock sessiond poll");
3246 if (client_socket
>= 0) {
3247 ret
= close(client_socket
);
3249 PERROR("close client_socket sessiond poll");
3256 ERR("Health error occurred in %s", __func__
);
3258 health_unregister(health_consumerd
);
3260 rcu_unregister_thread();
3264 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3265 struct lttng_consumer_local_data
*ctx
)
3269 pthread_mutex_lock(&stream
->lock
);
3270 if (stream
->metadata_flag
) {
3271 pthread_mutex_lock(&stream
->metadata_rdv_lock
);
3274 switch (consumer_data
.type
) {
3275 case LTTNG_CONSUMER_KERNEL
:
3276 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
3278 case LTTNG_CONSUMER32_UST
:
3279 case LTTNG_CONSUMER64_UST
:
3280 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
3283 ERR("Unknown consumer_data type");
3289 if (stream
->metadata_flag
) {
3290 pthread_cond_broadcast(&stream
->metadata_rdv
);
3291 pthread_mutex_unlock(&stream
->metadata_rdv_lock
);
3293 pthread_mutex_unlock(&stream
->lock
);
3297 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3299 switch (consumer_data
.type
) {
3300 case LTTNG_CONSUMER_KERNEL
:
3301 return lttng_kconsumer_on_recv_stream(stream
);
3302 case LTTNG_CONSUMER32_UST
:
3303 case LTTNG_CONSUMER64_UST
:
3304 return lttng_ustconsumer_on_recv_stream(stream
);
3306 ERR("Unknown consumer_data type");
3313 * Allocate and set consumer data hash tables.
3315 int lttng_consumer_init(void)
3317 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3318 if (!consumer_data
.channel_ht
) {
3322 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3323 if (!consumer_data
.relayd_ht
) {
3327 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3328 if (!consumer_data
.stream_list_ht
) {
3332 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3333 if (!consumer_data
.stream_per_chan_id_ht
) {
3337 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3342 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3354 * Process the ADD_RELAYD command receive by a consumer.
3356 * This will create a relayd socket pair and add it to the relayd hash table.
3357 * The caller MUST acquire a RCU read side lock before calling it.
3359 int consumer_add_relayd_socket(uint64_t net_seq_idx
, int sock_type
,
3360 struct lttng_consumer_local_data
*ctx
, int sock
,
3361 struct pollfd
*consumer_sockpoll
,
3362 struct lttcomm_relayd_sock
*relayd_sock
, uint64_t sessiond_id
,
3363 uint64_t relayd_session_id
)
3365 int fd
= -1, ret
= -1, relayd_created
= 0;
3366 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3367 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3370 assert(relayd_sock
);
3372 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", net_seq_idx
);
3374 /* Get relayd reference if exists. */
3375 relayd
= consumer_find_relayd(net_seq_idx
);
3376 if (relayd
== NULL
) {
3377 assert(sock_type
== LTTNG_STREAM_CONTROL
);
3378 /* Not found. Allocate one. */
3379 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3380 if (relayd
== NULL
) {
3382 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3385 relayd
->sessiond_session_id
= sessiond_id
;
3390 * This code path MUST continue to the consumer send status message to
3391 * we can notify the session daemon and continue our work without
3392 * killing everything.
3396 * relayd key should never be found for control socket.
3398 assert(sock_type
!= LTTNG_STREAM_CONTROL
);
3401 /* First send a status message before receiving the fds. */
3402 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3404 /* Somehow, the session daemon is not responding anymore. */
3405 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3406 goto error_nosignal
;
3409 /* Poll on consumer socket. */
3410 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3412 /* Needing to exit in the middle of a command: error. */
3413 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3415 goto error_nosignal
;
3418 /* Get relayd socket from session daemon */
3419 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3420 if (ret
!= sizeof(fd
)) {
3422 fd
= -1; /* Just in case it gets set with an invalid value. */
3425 * Failing to receive FDs might indicate a major problem such as
3426 * reaching a fd limit during the receive where the kernel returns a
3427 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3428 * don't take any chances and stop everything.
3430 * XXX: Feature request #558 will fix that and avoid this possible
3431 * issue when reaching the fd limit.
3433 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3434 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3438 /* Copy socket information and received FD */
3439 switch (sock_type
) {
3440 case LTTNG_STREAM_CONTROL
:
3441 /* Copy received lttcomm socket */
3442 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3443 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3444 /* Handle create_sock error. */
3446 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3450 * Close the socket created internally by
3451 * lttcomm_create_sock, so we can replace it by the one
3452 * received from sessiond.
3454 if (close(relayd
->control_sock
.sock
.fd
)) {
3458 /* Assign new file descriptor */
3459 relayd
->control_sock
.sock
.fd
= fd
;
3460 fd
= -1; /* For error path */
3461 /* Assign version values. */
3462 relayd
->control_sock
.major
= relayd_sock
->major
;
3463 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3465 relayd
->relayd_session_id
= relayd_session_id
;
3468 case LTTNG_STREAM_DATA
:
3469 /* Copy received lttcomm socket */
3470 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3471 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3472 /* Handle create_sock error. */
3474 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3478 * Close the socket created internally by
3479 * lttcomm_create_sock, so we can replace it by the one
3480 * received from sessiond.
3482 if (close(relayd
->data_sock
.sock
.fd
)) {
3486 /* Assign new file descriptor */
3487 relayd
->data_sock
.sock
.fd
= fd
;
3488 fd
= -1; /* for eventual error paths */
3489 /* Assign version values. */
3490 relayd
->data_sock
.major
= relayd_sock
->major
;
3491 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3494 ERR("Unknown relayd socket type (%d)", sock_type
);
3496 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3500 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3501 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3502 relayd
->net_seq_idx
, fd
);
3504 /* We successfully added the socket. Send status back. */
3505 ret
= consumer_send_status_msg(sock
, ret_code
);
3507 /* Somehow, the session daemon is not responding anymore. */
3508 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3509 goto error_nosignal
;
3513 * Add relayd socket pair to consumer data hashtable. If object already
3514 * exists or on error, the function gracefully returns.
3523 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3524 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3528 /* Close received socket if valid. */
3531 PERROR("close received socket");
3535 if (relayd_created
) {
3543 * Try to lock the stream mutex.
3545 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
3547 static int stream_try_lock(struct lttng_consumer_stream
*stream
)
3554 * Try to lock the stream mutex. On failure, we know that the stream is
3555 * being used else where hence there is data still being extracted.
3557 ret
= pthread_mutex_trylock(&stream
->lock
);
3559 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
3571 * Search for a relayd associated to the session id and return the reference.
3573 * A rcu read side lock MUST be acquire before calling this function and locked
3574 * until the relayd object is no longer necessary.
3576 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3578 struct lttng_ht_iter iter
;
3579 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3581 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3582 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3585 * Check by sessiond id which is unique here where the relayd session
3586 * id might not be when having multiple relayd.
3588 if (relayd
->sessiond_session_id
== id
) {
3589 /* Found the relayd. There can be only one per id. */
3601 * Check if for a given session id there is still data needed to be extract
3604 * Return 1 if data is pending or else 0 meaning ready to be read.
3606 int consumer_data_pending(uint64_t id
)
3609 struct lttng_ht_iter iter
;
3610 struct lttng_ht
*ht
;
3611 struct lttng_consumer_stream
*stream
;
3612 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3613 int (*data_pending
)(struct lttng_consumer_stream
*);
3615 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3618 pthread_mutex_lock(&consumer_data
.lock
);
3620 switch (consumer_data
.type
) {
3621 case LTTNG_CONSUMER_KERNEL
:
3622 data_pending
= lttng_kconsumer_data_pending
;
3624 case LTTNG_CONSUMER32_UST
:
3625 case LTTNG_CONSUMER64_UST
:
3626 data_pending
= lttng_ustconsumer_data_pending
;
3629 ERR("Unknown consumer data type");
3633 /* Ease our life a bit */
3634 ht
= consumer_data
.stream_list_ht
;
3636 relayd
= find_relayd_by_session_id(id
);
3638 /* Send init command for data pending. */
3639 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3640 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3641 relayd
->relayd_session_id
);
3643 /* Communication error thus the relayd so no data pending. */
3644 ERR("Relayd begin data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
3645 lttng_consumer_cleanup_relayd(relayd
);
3646 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3647 goto data_not_pending
;
3649 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3652 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3653 ht
->hash_fct(&id
, lttng_ht_seed
),
3655 &iter
.iter
, stream
, node_session_id
.node
) {
3656 /* If this call fails, the stream is being used hence data pending. */
3657 ret
= stream_try_lock(stream
);
3663 * A removed node from the hash table indicates that the stream has
3664 * been deleted thus having a guarantee that the buffers are closed
3665 * on the consumer side. However, data can still be transmitted
3666 * over the network so don't skip the relayd check.
3668 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3670 /* Check the stream if there is data in the buffers. */
3671 ret
= data_pending(stream
);
3673 pthread_mutex_unlock(&stream
->lock
);
3680 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3681 if (stream
->metadata_flag
) {
3682 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3683 stream
->relayd_stream_id
);
3685 ret
= relayd_data_pending(&relayd
->control_sock
,
3686 stream
->relayd_stream_id
,
3687 stream
->next_net_seq_num
- 1);
3690 ERR("Relayd data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
3691 lttng_consumer_cleanup_relayd(relayd
);
3692 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3693 pthread_mutex_unlock(&stream
->lock
);
3694 goto data_not_pending
;
3696 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3698 pthread_mutex_unlock(&stream
->lock
);
3702 pthread_mutex_unlock(&stream
->lock
);
3706 unsigned int is_data_inflight
= 0;
3708 /* Send init command for data pending. */
3709 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3710 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3711 relayd
->relayd_session_id
, &is_data_inflight
);
3713 ERR("Relayd end data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->net_seq_idx
);
3714 lttng_consumer_cleanup_relayd(relayd
);
3715 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3716 goto data_not_pending
;
3718 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3719 if (is_data_inflight
) {
3725 * Finding _no_ node in the hash table and no inflight data means that the
3726 * stream(s) have been removed thus data is guaranteed to be available for
3727 * analysis from the trace files.
3731 /* Data is available to be read by a viewer. */
3732 pthread_mutex_unlock(&consumer_data
.lock
);
3737 /* Data is still being extracted from buffers. */
3738 pthread_mutex_unlock(&consumer_data
.lock
);
3744 * Send a ret code status message to the sessiond daemon.
3746 * Return the sendmsg() return value.
3748 int consumer_send_status_msg(int sock
, int ret_code
)
3750 struct lttcomm_consumer_status_msg msg
;
3752 memset(&msg
, 0, sizeof(msg
));
3753 msg
.ret_code
= ret_code
;
3755 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3759 * Send a channel status message to the sessiond daemon.
3761 * Return the sendmsg() return value.
3763 int consumer_send_status_channel(int sock
,
3764 struct lttng_consumer_channel
*channel
)
3766 struct lttcomm_consumer_status_channel msg
;
3770 memset(&msg
, 0, sizeof(msg
));
3772 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3774 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3775 msg
.key
= channel
->key
;
3776 msg
.stream_count
= channel
->streams
.count
;
3779 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3782 unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos
,
3783 unsigned long produced_pos
, uint64_t nb_packets_per_stream
,
3784 uint64_t max_sb_size
)
3786 unsigned long start_pos
;
3788 if (!nb_packets_per_stream
) {
3789 return consumed_pos
; /* Grab everything */
3791 start_pos
= produced_pos
- offset_align_floor(produced_pos
, max_sb_size
);
3792 start_pos
-= max_sb_size
* nb_packets_per_stream
;
3793 if ((long) (start_pos
- consumed_pos
) < 0) {
3794 return consumed_pos
; /* Grab everything */