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 */
70 /* Flag used to temporarily pause data consumption from testpoints. */
71 int data_consumption_paused
;
74 * Flag to inform the polling thread to quit when all fd hung up. Updated by
75 * the consumer_thread_receive_fds when it notices that all fds has hung up.
76 * Also updated by the signal handler (consumer_should_exit()). Read by the
82 * Global hash table containing respectively metadata and data streams. The
83 * stream element in this ht should only be updated by the metadata poll thread
84 * for the metadata and the data poll thread for the data.
86 static struct lttng_ht
*metadata_ht
;
87 static struct lttng_ht
*data_ht
;
90 * Notify a thread lttng pipe to poll back again. This usually means that some
91 * global state has changed so we just send back the thread in a poll wait
94 static void notify_thread_lttng_pipe(struct lttng_pipe
*pipe
)
96 struct lttng_consumer_stream
*null_stream
= NULL
;
100 (void) lttng_pipe_write(pipe
, &null_stream
, sizeof(null_stream
));
103 static void notify_health_quit_pipe(int *pipe
)
107 ret
= lttng_write(pipe
[1], "4", 1);
109 PERROR("write consumer health quit");
113 static void notify_channel_pipe(struct lttng_consumer_local_data
*ctx
,
114 struct lttng_consumer_channel
*chan
,
116 enum consumer_channel_action action
)
118 struct consumer_channel_msg msg
;
121 memset(&msg
, 0, sizeof(msg
));
126 ret
= lttng_write(ctx
->consumer_channel_pipe
[1], &msg
, sizeof(msg
));
127 if (ret
< sizeof(msg
)) {
128 PERROR("notify_channel_pipe write error");
132 void notify_thread_del_channel(struct lttng_consumer_local_data
*ctx
,
135 notify_channel_pipe(ctx
, NULL
, key
, CONSUMER_CHANNEL_DEL
);
138 static int read_channel_pipe(struct lttng_consumer_local_data
*ctx
,
139 struct lttng_consumer_channel
**chan
,
141 enum consumer_channel_action
*action
)
143 struct consumer_channel_msg msg
;
146 ret
= lttng_read(ctx
->consumer_channel_pipe
[0], &msg
, sizeof(msg
));
147 if (ret
< sizeof(msg
)) {
151 *action
= msg
.action
;
159 * Cleanup the stream list of a channel. Those streams are not yet globally
162 static void clean_channel_stream_list(struct lttng_consumer_channel
*channel
)
164 struct lttng_consumer_stream
*stream
, *stmp
;
168 /* Delete streams that might have been left in the stream list. */
169 cds_list_for_each_entry_safe(stream
, stmp
, &channel
->streams
.head
,
171 cds_list_del(&stream
->send_node
);
173 * Once a stream is added to this list, the buffers were created so we
174 * have a guarantee that this call will succeed. Setting the monitor
175 * mode to 0 so we don't lock nor try to delete the stream from the
179 consumer_stream_destroy(stream
, NULL
);
184 * Find a stream. The consumer_data.lock must be locked during this
187 static struct lttng_consumer_stream
*find_stream(uint64_t key
,
190 struct lttng_ht_iter iter
;
191 struct lttng_ht_node_u64
*node
;
192 struct lttng_consumer_stream
*stream
= NULL
;
196 /* -1ULL keys are lookup failures */
197 if (key
== (uint64_t) -1ULL) {
203 lttng_ht_lookup(ht
, &key
, &iter
);
204 node
= lttng_ht_iter_get_node_u64(&iter
);
206 stream
= caa_container_of(node
, struct lttng_consumer_stream
, node
);
214 static void steal_stream_key(uint64_t key
, struct lttng_ht
*ht
)
216 struct lttng_consumer_stream
*stream
;
219 stream
= find_stream(key
, ht
);
221 stream
->key
= (uint64_t) -1ULL;
223 * We don't want the lookup to match, but we still need
224 * to iterate on this stream when iterating over the hash table. Just
225 * change the node key.
227 stream
->node
.key
= (uint64_t) -1ULL;
233 * Return a channel object for the given key.
235 * RCU read side lock MUST be acquired before calling this function and
236 * protects the channel ptr.
238 struct lttng_consumer_channel
*consumer_find_channel(uint64_t key
)
240 struct lttng_ht_iter iter
;
241 struct lttng_ht_node_u64
*node
;
242 struct lttng_consumer_channel
*channel
= NULL
;
244 /* -1ULL keys are lookup failures */
245 if (key
== (uint64_t) -1ULL) {
249 lttng_ht_lookup(consumer_data
.channel_ht
, &key
, &iter
);
250 node
= lttng_ht_iter_get_node_u64(&iter
);
252 channel
= caa_container_of(node
, struct lttng_consumer_channel
, node
);
259 * There is a possibility that the consumer does not have enough time between
260 * the close of the channel on the session daemon and the cleanup in here thus
261 * once we have a channel add with an existing key, we know for sure that this
262 * channel will eventually get cleaned up by all streams being closed.
264 * This function just nullifies the already existing channel key.
266 static void steal_channel_key(uint64_t key
)
268 struct lttng_consumer_channel
*channel
;
271 channel
= consumer_find_channel(key
);
273 channel
->key
= (uint64_t) -1ULL;
275 * We don't want the lookup to match, but we still need to iterate on
276 * this channel when iterating over the hash table. Just change the
279 channel
->node
.key
= (uint64_t) -1ULL;
284 static void free_channel_rcu(struct rcu_head
*head
)
286 struct lttng_ht_node_u64
*node
=
287 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
288 struct lttng_consumer_channel
*channel
=
289 caa_container_of(node
, struct lttng_consumer_channel
, node
);
291 switch (consumer_data
.type
) {
292 case LTTNG_CONSUMER_KERNEL
:
294 case LTTNG_CONSUMER32_UST
:
295 case LTTNG_CONSUMER64_UST
:
296 lttng_ustconsumer_free_channel(channel
);
299 ERR("Unknown consumer_data type");
306 * RCU protected relayd socket pair free.
308 static void free_relayd_rcu(struct rcu_head
*head
)
310 struct lttng_ht_node_u64
*node
=
311 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
312 struct consumer_relayd_sock_pair
*relayd
=
313 caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
316 * Close all sockets. This is done in the call RCU since we don't want the
317 * socket fds to be reassigned thus potentially creating bad state of the
320 * We do not have to lock the control socket mutex here since at this stage
321 * there is no one referencing to this relayd object.
323 (void) relayd_close(&relayd
->control_sock
);
324 (void) relayd_close(&relayd
->data_sock
);
330 * Destroy and free relayd socket pair object.
332 void consumer_destroy_relayd(struct consumer_relayd_sock_pair
*relayd
)
335 struct lttng_ht_iter iter
;
337 if (relayd
== NULL
) {
341 DBG("Consumer destroy and close relayd socket pair");
343 iter
.iter
.node
= &relayd
->node
.node
;
344 ret
= lttng_ht_del(consumer_data
.relayd_ht
, &iter
);
346 /* We assume the relayd is being or is destroyed */
350 /* RCU free() call */
351 call_rcu(&relayd
->node
.head
, free_relayd_rcu
);
355 * Remove a channel from the global list protected by a mutex. This function is
356 * also responsible for freeing its data structures.
358 void consumer_del_channel(struct lttng_consumer_channel
*channel
)
361 struct lttng_ht_iter iter
;
363 DBG("Consumer delete channel key %" PRIu64
, channel
->key
);
365 pthread_mutex_lock(&consumer_data
.lock
);
366 pthread_mutex_lock(&channel
->lock
);
368 /* Destroy streams that might have been left in the stream list. */
369 clean_channel_stream_list(channel
);
371 if (channel
->live_timer_enabled
== 1) {
372 consumer_timer_live_stop(channel
);
374 if (channel
->monitor_timer_enabled
== 1) {
375 consumer_timer_monitor_stop(channel
);
378 switch (consumer_data
.type
) {
379 case LTTNG_CONSUMER_KERNEL
:
381 case LTTNG_CONSUMER32_UST
:
382 case LTTNG_CONSUMER64_UST
:
383 lttng_ustconsumer_del_channel(channel
);
386 ERR("Unknown consumer_data type");
392 iter
.iter
.node
= &channel
->node
.node
;
393 ret
= lttng_ht_del(consumer_data
.channel_ht
, &iter
);
397 call_rcu(&channel
->node
.head
, free_channel_rcu
);
399 pthread_mutex_unlock(&channel
->lock
);
400 pthread_mutex_unlock(&consumer_data
.lock
);
404 * Iterate over the relayd hash table and destroy each element. Finally,
405 * destroy the whole hash table.
407 static void cleanup_relayd_ht(void)
409 struct lttng_ht_iter iter
;
410 struct consumer_relayd_sock_pair
*relayd
;
414 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
416 consumer_destroy_relayd(relayd
);
421 lttng_ht_destroy(consumer_data
.relayd_ht
);
425 * Update the end point status of all streams having the given network sequence
426 * index (relayd index).
428 * It's atomically set without having the stream mutex locked which is fine
429 * because we handle the write/read race with a pipe wakeup for each thread.
431 static void update_endpoint_status_by_netidx(uint64_t net_seq_idx
,
432 enum consumer_endpoint_status status
)
434 struct lttng_ht_iter iter
;
435 struct lttng_consumer_stream
*stream
;
437 DBG("Consumer set delete flag on stream by idx %" PRIu64
, net_seq_idx
);
441 /* Let's begin with metadata */
442 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
443 if (stream
->net_seq_idx
== net_seq_idx
) {
444 uatomic_set(&stream
->endpoint_status
, status
);
445 DBG("Delete flag set to metadata stream %d", stream
->wait_fd
);
449 /* Follow up by the data streams */
450 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
451 if (stream
->net_seq_idx
== net_seq_idx
) {
452 uatomic_set(&stream
->endpoint_status
, status
);
453 DBG("Delete flag set to data stream %d", stream
->wait_fd
);
460 * Cleanup a relayd object by flagging every associated streams for deletion,
461 * destroying the object meaning removing it from the relayd hash table,
462 * closing the sockets and freeing the memory in a RCU call.
464 * If a local data context is available, notify the threads that the streams'
465 * state have changed.
467 static void cleanup_relayd(struct consumer_relayd_sock_pair
*relayd
,
468 struct lttng_consumer_local_data
*ctx
)
474 DBG("Cleaning up relayd sockets");
476 /* Save the net sequence index before destroying the object */
477 netidx
= relayd
->net_seq_idx
;
480 * Delete the relayd from the relayd hash table, close the sockets and free
481 * the object in a RCU call.
483 consumer_destroy_relayd(relayd
);
485 /* Set inactive endpoint to all streams */
486 update_endpoint_status_by_netidx(netidx
, CONSUMER_ENDPOINT_INACTIVE
);
489 * With a local data context, notify the threads that the streams' state
490 * have changed. The write() action on the pipe acts as an "implicit"
491 * memory barrier ordering the updates of the end point status from the
492 * read of this status which happens AFTER receiving this notify.
495 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
496 notify_thread_lttng_pipe(ctx
->consumer_metadata_pipe
);
501 * Flag a relayd socket pair for destruction. Destroy it if the refcount
504 * RCU read side lock MUST be aquired before calling this function.
506 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair
*relayd
)
510 /* Set destroy flag for this object */
511 uatomic_set(&relayd
->destroy_flag
, 1);
513 /* Destroy the relayd if refcount is 0 */
514 if (uatomic_read(&relayd
->refcount
) == 0) {
515 consumer_destroy_relayd(relayd
);
520 * Completly destroy stream from every visiable data structure and the given
523 * One this call returns, the stream object is not longer usable nor visible.
525 void consumer_del_stream(struct lttng_consumer_stream
*stream
,
528 consumer_stream_destroy(stream
, ht
);
532 * XXX naming of del vs destroy is all mixed up.
534 void consumer_del_stream_for_data(struct lttng_consumer_stream
*stream
)
536 consumer_stream_destroy(stream
, data_ht
);
539 void consumer_del_stream_for_metadata(struct lttng_consumer_stream
*stream
)
541 consumer_stream_destroy(stream
, metadata_ht
);
544 struct lttng_consumer_stream
*consumer_allocate_stream(uint64_t channel_key
,
546 enum lttng_consumer_stream_state state
,
547 const char *channel_name
,
554 enum consumer_channel_type type
,
555 unsigned int monitor
)
558 struct lttng_consumer_stream
*stream
;
560 stream
= zmalloc(sizeof(*stream
));
561 if (stream
== NULL
) {
562 PERROR("malloc struct lttng_consumer_stream");
569 stream
->key
= stream_key
;
571 stream
->out_fd_offset
= 0;
572 stream
->output_written
= 0;
573 stream
->state
= state
;
576 stream
->net_seq_idx
= relayd_id
;
577 stream
->session_id
= session_id
;
578 stream
->monitor
= monitor
;
579 stream
->endpoint_status
= CONSUMER_ENDPOINT_ACTIVE
;
580 stream
->index_file
= NULL
;
581 stream
->last_sequence_number
= -1ULL;
582 pthread_mutex_init(&stream
->lock
, NULL
);
583 pthread_mutex_init(&stream
->metadata_timer_lock
, NULL
);
585 /* If channel is the metadata, flag this stream as metadata. */
586 if (type
== CONSUMER_CHANNEL_TYPE_METADATA
) {
587 stream
->metadata_flag
= 1;
588 /* Metadata is flat out. */
589 strncpy(stream
->name
, DEFAULT_METADATA_NAME
, sizeof(stream
->name
));
590 /* Live rendez-vous point. */
591 pthread_cond_init(&stream
->metadata_rdv
, NULL
);
592 pthread_mutex_init(&stream
->metadata_rdv_lock
, NULL
);
594 /* Format stream name to <channel_name>_<cpu_number> */
595 ret
= snprintf(stream
->name
, sizeof(stream
->name
), "%s_%d",
598 PERROR("snprintf stream name");
603 /* Key is always the wait_fd for streams. */
604 lttng_ht_node_init_u64(&stream
->node
, stream
->key
);
606 /* Init node per channel id key */
607 lttng_ht_node_init_u64(&stream
->node_channel_id
, channel_key
);
609 /* Init session id node with the stream session id */
610 lttng_ht_node_init_u64(&stream
->node_session_id
, stream
->session_id
);
612 DBG3("Allocated stream %s (key %" PRIu64
", chan_key %" PRIu64
613 " relayd_id %" PRIu64
", session_id %" PRIu64
,
614 stream
->name
, stream
->key
, channel_key
,
615 stream
->net_seq_idx
, stream
->session_id
);
631 * Add a stream to the global list protected by a mutex.
633 void consumer_add_data_stream(struct lttng_consumer_stream
*stream
)
635 struct lttng_ht
*ht
= data_ht
;
640 DBG3("Adding consumer stream %" PRIu64
, stream
->key
);
642 pthread_mutex_lock(&consumer_data
.lock
);
643 pthread_mutex_lock(&stream
->chan
->lock
);
644 pthread_mutex_lock(&stream
->chan
->timer_lock
);
645 pthread_mutex_lock(&stream
->lock
);
648 /* Steal stream identifier to avoid having streams with the same key */
649 steal_stream_key(stream
->key
, ht
);
651 lttng_ht_add_unique_u64(ht
, &stream
->node
);
653 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
654 &stream
->node_channel_id
);
657 * Add stream to the stream_list_ht of the consumer data. No need to steal
658 * the key since the HT does not use it and we allow to add redundant keys
661 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
664 * When nb_init_stream_left reaches 0, we don't need to trigger any action
665 * in terms of destroying the associated channel, because the action that
666 * causes the count to become 0 also causes a stream to be added. The
667 * channel deletion will thus be triggered by the following removal of this
670 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
671 /* Increment refcount before decrementing nb_init_stream_left */
673 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
676 /* Update consumer data once the node is inserted. */
677 consumer_data
.stream_count
++;
678 consumer_data
.need_update
= 1;
681 pthread_mutex_unlock(&stream
->lock
);
682 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
683 pthread_mutex_unlock(&stream
->chan
->lock
);
684 pthread_mutex_unlock(&consumer_data
.lock
);
687 void consumer_del_data_stream(struct lttng_consumer_stream
*stream
)
689 consumer_del_stream(stream
, data_ht
);
693 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
694 * be acquired before calling this.
696 static int add_relayd(struct consumer_relayd_sock_pair
*relayd
)
699 struct lttng_ht_node_u64
*node
;
700 struct lttng_ht_iter iter
;
704 lttng_ht_lookup(consumer_data
.relayd_ht
,
705 &relayd
->net_seq_idx
, &iter
);
706 node
= lttng_ht_iter_get_node_u64(&iter
);
710 lttng_ht_add_unique_u64(consumer_data
.relayd_ht
, &relayd
->node
);
717 * Allocate and return a consumer relayd socket.
719 static struct consumer_relayd_sock_pair
*consumer_allocate_relayd_sock_pair(
720 uint64_t net_seq_idx
)
722 struct consumer_relayd_sock_pair
*obj
= NULL
;
724 /* net sequence index of -1 is a failure */
725 if (net_seq_idx
== (uint64_t) -1ULL) {
729 obj
= zmalloc(sizeof(struct consumer_relayd_sock_pair
));
731 PERROR("zmalloc relayd sock");
735 obj
->net_seq_idx
= net_seq_idx
;
737 obj
->destroy_flag
= 0;
738 obj
->control_sock
.sock
.fd
= -1;
739 obj
->data_sock
.sock
.fd
= -1;
740 lttng_ht_node_init_u64(&obj
->node
, obj
->net_seq_idx
);
741 pthread_mutex_init(&obj
->ctrl_sock_mutex
, NULL
);
748 * Find a relayd socket pair in the global consumer data.
750 * Return the object if found else NULL.
751 * RCU read-side lock must be held across this call and while using the
754 struct consumer_relayd_sock_pair
*consumer_find_relayd(uint64_t key
)
756 struct lttng_ht_iter iter
;
757 struct lttng_ht_node_u64
*node
;
758 struct consumer_relayd_sock_pair
*relayd
= NULL
;
760 /* Negative keys are lookup failures */
761 if (key
== (uint64_t) -1ULL) {
765 lttng_ht_lookup(consumer_data
.relayd_ht
, &key
,
767 node
= lttng_ht_iter_get_node_u64(&iter
);
769 relayd
= caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
777 * Find a relayd and send the stream
779 * Returns 0 on success, < 0 on error
781 int consumer_send_relayd_stream(struct lttng_consumer_stream
*stream
,
785 struct consumer_relayd_sock_pair
*relayd
;
788 assert(stream
->net_seq_idx
!= -1ULL);
791 /* The stream is not metadata. Get relayd reference if exists. */
793 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
794 if (relayd
!= NULL
) {
795 /* Add stream on the relayd */
796 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
797 ret
= relayd_add_stream(&relayd
->control_sock
, stream
->name
,
798 path
, &stream
->relayd_stream_id
,
799 stream
->chan
->tracefile_size
, stream
->chan
->tracefile_count
);
800 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
805 uatomic_inc(&relayd
->refcount
);
806 stream
->sent_to_relayd
= 1;
808 ERR("Stream %" PRIu64
" relayd ID %" PRIu64
" unknown. Can't send it.",
809 stream
->key
, stream
->net_seq_idx
);
814 DBG("Stream %s with key %" PRIu64
" sent to relayd id %" PRIu64
,
815 stream
->name
, stream
->key
, stream
->net_seq_idx
);
823 * Find a relayd and send the streams sent message
825 * Returns 0 on success, < 0 on error
827 int consumer_send_relayd_streams_sent(uint64_t net_seq_idx
)
830 struct consumer_relayd_sock_pair
*relayd
;
832 assert(net_seq_idx
!= -1ULL);
834 /* The stream is not metadata. Get relayd reference if exists. */
836 relayd
= consumer_find_relayd(net_seq_idx
);
837 if (relayd
!= NULL
) {
838 /* Add stream on the relayd */
839 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
840 ret
= relayd_streams_sent(&relayd
->control_sock
);
841 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
)
1233 CMM_STORE_SHARED(consumer_quit
, 1);
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
;
1360 ctx
->channel_monitor_pipe
= -1;
1364 error_metadata_pipe
:
1365 utils_close_pipe(ctx
->consumer_channel_pipe
);
1367 utils_close_pipe(ctx
->consumer_should_quit
);
1369 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1371 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1379 * Iterate over all streams of the hashtable and free them properly.
1381 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1383 struct lttng_ht_iter iter
;
1384 struct lttng_consumer_stream
*stream
;
1391 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1393 * Ignore return value since we are currently cleaning up so any error
1396 (void) consumer_del_stream(stream
, ht
);
1400 lttng_ht_destroy(ht
);
1404 * Iterate over all streams of the metadata hashtable and free them
1407 static void destroy_metadata_stream_ht(struct lttng_ht
*ht
)
1409 struct lttng_ht_iter iter
;
1410 struct lttng_consumer_stream
*stream
;
1417 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1419 * Ignore return value since we are currently cleaning up so any error
1422 (void) consumer_del_metadata_stream(stream
, ht
);
1426 lttng_ht_destroy(ht
);
1430 * Close all fds associated with the instance and free the context.
1432 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1436 DBG("Consumer destroying it. Closing everything.");
1442 destroy_data_stream_ht(data_ht
);
1443 destroy_metadata_stream_ht(metadata_ht
);
1445 ret
= close(ctx
->consumer_error_socket
);
1449 ret
= close(ctx
->consumer_metadata_socket
);
1453 utils_close_pipe(ctx
->consumer_channel_pipe
);
1454 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1455 lttng_pipe_destroy(ctx
->consumer_metadata_pipe
);
1456 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1457 utils_close_pipe(ctx
->consumer_should_quit
);
1459 unlink(ctx
->consumer_command_sock_path
);
1464 * Write the metadata stream id on the specified file descriptor.
1466 static int write_relayd_metadata_id(int fd
,
1467 struct lttng_consumer_stream
*stream
,
1468 struct consumer_relayd_sock_pair
*relayd
, unsigned long padding
)
1471 struct lttcomm_relayd_metadata_payload hdr
;
1473 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1474 hdr
.padding_size
= htobe32(padding
);
1475 ret
= lttng_write(fd
, (void *) &hdr
, sizeof(hdr
));
1476 if (ret
< sizeof(hdr
)) {
1478 * This error means that the fd's end is closed so ignore the PERROR
1479 * not to clubber the error output since this can happen in a normal
1482 if (errno
!= EPIPE
) {
1483 PERROR("write metadata stream id");
1485 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1487 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1488 * handle writting the missing part so report that as an error and
1489 * don't lie to the caller.
1494 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1495 stream
->relayd_stream_id
, padding
);
1502 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1503 * core function for writing trace buffers to either the local filesystem or
1506 * It must be called with the stream lock held.
1508 * Careful review MUST be put if any changes occur!
1510 * Returns the number of bytes written
1512 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1513 struct lttng_consumer_local_data
*ctx
,
1514 struct lttng_consumer_stream
*stream
, unsigned long len
,
1515 unsigned long padding
,
1516 struct ctf_packet_index
*index
)
1518 unsigned long mmap_offset
;
1521 off_t orig_offset
= stream
->out_fd_offset
;
1522 /* Default is on the disk */
1523 int outfd
= stream
->out_fd
;
1524 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1525 unsigned int relayd_hang_up
= 0;
1527 /* RCU lock for the relayd pointer */
1530 /* Flag that the current stream if set for network streaming. */
1531 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1532 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1533 if (relayd
== NULL
) {
1539 /* get the offset inside the fd to mmap */
1540 switch (consumer_data
.type
) {
1541 case LTTNG_CONSUMER_KERNEL
:
1542 mmap_base
= stream
->mmap_base
;
1543 ret
= kernctl_get_mmap_read_offset(stream
->wait_fd
, &mmap_offset
);
1545 PERROR("tracer ctl get_mmap_read_offset");
1549 case LTTNG_CONSUMER32_UST
:
1550 case LTTNG_CONSUMER64_UST
:
1551 mmap_base
= lttng_ustctl_get_mmap_base(stream
);
1553 ERR("read mmap get mmap base for stream %s", stream
->name
);
1557 ret
= lttng_ustctl_get_mmap_read_offset(stream
, &mmap_offset
);
1559 PERROR("tracer ctl get_mmap_read_offset");
1565 ERR("Unknown consumer_data type");
1569 /* Handle stream on the relayd if the output is on the network */
1571 unsigned long netlen
= len
;
1574 * Lock the control socket for the complete duration of the function
1575 * since from this point on we will use the socket.
1577 if (stream
->metadata_flag
) {
1578 /* Metadata requires the control socket. */
1579 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1580 if (stream
->reset_metadata_flag
) {
1581 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1582 stream
->relayd_stream_id
,
1583 stream
->metadata_version
);
1588 stream
->reset_metadata_flag
= 0;
1590 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1593 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1598 /* Use the returned socket. */
1601 /* Write metadata stream id before payload */
1602 if (stream
->metadata_flag
) {
1603 ret
= write_relayd_metadata_id(outfd
, stream
, relayd
, padding
);
1610 /* No streaming, we have to set the len with the full padding */
1613 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1614 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1616 ERR("Reset metadata file");
1619 stream
->reset_metadata_flag
= 0;
1623 * Check if we need to change the tracefile before writing the packet.
1625 if (stream
->chan
->tracefile_size
> 0 &&
1626 (stream
->tracefile_size_current
+ len
) >
1627 stream
->chan
->tracefile_size
) {
1628 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1629 stream
->name
, stream
->chan
->tracefile_size
,
1630 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1631 stream
->out_fd
, &(stream
->tracefile_count_current
),
1634 ERR("Rotating output file");
1637 outfd
= stream
->out_fd
;
1639 if (stream
->index_file
) {
1640 lttng_index_file_put(stream
->index_file
);
1641 stream
->index_file
= lttng_index_file_create(stream
->chan
->pathname
,
1642 stream
->name
, stream
->uid
, stream
->gid
,
1643 stream
->chan
->tracefile_size
,
1644 stream
->tracefile_count_current
,
1645 CTF_INDEX_MAJOR
, CTF_INDEX_MINOR
);
1646 if (!stream
->index_file
) {
1651 /* Reset current size because we just perform a rotation. */
1652 stream
->tracefile_size_current
= 0;
1653 stream
->out_fd_offset
= 0;
1656 stream
->tracefile_size_current
+= len
;
1658 index
->offset
= htobe64(stream
->out_fd_offset
);
1663 * This call guarantee that len or less is returned. It's impossible to
1664 * receive a ret value that is bigger than len.
1666 ret
= lttng_write(outfd
, mmap_base
+ mmap_offset
, len
);
1667 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, len
);
1668 if (ret
< 0 || ((size_t) ret
!= len
)) {
1670 * Report error to caller if nothing was written else at least send the
1678 /* Socket operation failed. We consider the relayd dead */
1679 if (errno
== EPIPE
|| errno
== EINVAL
|| errno
== EBADF
) {
1681 * This is possible if the fd is closed on the other side
1682 * (outfd) or any write problem. It can be verbose a bit for a
1683 * normal execution if for instance the relayd is stopped
1684 * abruptly. This can happen so set this to a DBG statement.
1686 DBG("Consumer mmap write detected relayd hang up");
1688 /* Unhandled error, print it and stop function right now. */
1689 PERROR("Error in write mmap (ret %zd != len %lu)", ret
, len
);
1693 stream
->output_written
+= ret
;
1695 /* This call is useless on a socket so better save a syscall. */
1697 /* This won't block, but will start writeout asynchronously */
1698 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, len
,
1699 SYNC_FILE_RANGE_WRITE
);
1700 stream
->out_fd_offset
+= len
;
1701 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1706 * This is a special case that the relayd has closed its socket. Let's
1707 * cleanup the relayd object and all associated streams.
1709 if (relayd
&& relayd_hang_up
) {
1710 cleanup_relayd(relayd
, ctx
);
1714 /* Unlock only if ctrl socket used */
1715 if (relayd
&& stream
->metadata_flag
) {
1716 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1724 * Splice the data from the ring buffer to the tracefile.
1726 * It must be called with the stream lock held.
1728 * Returns the number of bytes spliced.
1730 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1731 struct lttng_consumer_local_data
*ctx
,
1732 struct lttng_consumer_stream
*stream
, unsigned long len
,
1733 unsigned long padding
,
1734 struct ctf_packet_index
*index
)
1736 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1738 off_t orig_offset
= stream
->out_fd_offset
;
1739 int fd
= stream
->wait_fd
;
1740 /* Default is on the disk */
1741 int outfd
= stream
->out_fd
;
1742 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1744 unsigned int relayd_hang_up
= 0;
1746 switch (consumer_data
.type
) {
1747 case LTTNG_CONSUMER_KERNEL
:
1749 case LTTNG_CONSUMER32_UST
:
1750 case LTTNG_CONSUMER64_UST
:
1751 /* Not supported for user space tracing */
1754 ERR("Unknown consumer_data type");
1758 /* RCU lock for the relayd pointer */
1761 /* Flag that the current stream if set for network streaming. */
1762 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1763 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1764 if (relayd
== NULL
) {
1769 splice_pipe
= stream
->splice_pipe
;
1771 /* Write metadata stream id before payload */
1773 unsigned long total_len
= len
;
1775 if (stream
->metadata_flag
) {
1777 * Lock the control socket for the complete duration of the function
1778 * since from this point on we will use the socket.
1780 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1782 if (stream
->reset_metadata_flag
) {
1783 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1784 stream
->relayd_stream_id
,
1785 stream
->metadata_version
);
1790 stream
->reset_metadata_flag
= 0;
1792 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
, relayd
,
1800 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1803 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1809 /* Use the returned socket. */
1812 /* No streaming, we have to set the len with the full padding */
1815 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1816 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1818 ERR("Reset metadata file");
1821 stream
->reset_metadata_flag
= 0;
1824 * Check if we need to change the tracefile before writing the packet.
1826 if (stream
->chan
->tracefile_size
> 0 &&
1827 (stream
->tracefile_size_current
+ len
) >
1828 stream
->chan
->tracefile_size
) {
1829 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1830 stream
->name
, stream
->chan
->tracefile_size
,
1831 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1832 stream
->out_fd
, &(stream
->tracefile_count_current
),
1836 ERR("Rotating output file");
1839 outfd
= stream
->out_fd
;
1841 if (stream
->index_file
) {
1842 lttng_index_file_put(stream
->index_file
);
1843 stream
->index_file
= lttng_index_file_create(stream
->chan
->pathname
,
1844 stream
->name
, stream
->uid
, stream
->gid
,
1845 stream
->chan
->tracefile_size
,
1846 stream
->tracefile_count_current
,
1847 CTF_INDEX_MAJOR
, CTF_INDEX_MINOR
);
1848 if (!stream
->index_file
) {
1853 /* Reset current size because we just perform a rotation. */
1854 stream
->tracefile_size_current
= 0;
1855 stream
->out_fd_offset
= 0;
1858 stream
->tracefile_size_current
+= len
;
1859 index
->offset
= htobe64(stream
->out_fd_offset
);
1863 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1864 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1865 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1866 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1867 DBG("splice chan to pipe, ret %zd", ret_splice
);
1868 if (ret_splice
< 0) {
1871 PERROR("Error in relay splice");
1875 /* Handle stream on the relayd if the output is on the network */
1876 if (relayd
&& stream
->metadata_flag
) {
1877 size_t metadata_payload_size
=
1878 sizeof(struct lttcomm_relayd_metadata_payload
);
1880 /* Update counter to fit the spliced data */
1881 ret_splice
+= metadata_payload_size
;
1882 len
+= metadata_payload_size
;
1884 * We do this so the return value can match the len passed as
1885 * argument to this function.
1887 written
-= metadata_payload_size
;
1890 /* Splice data out */
1891 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1892 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1893 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
1895 if (ret_splice
< 0) {
1900 } else if (ret_splice
> len
) {
1902 * We don't expect this code path to be executed but you never know
1903 * so this is an extra protection agains a buggy splice().
1906 written
+= ret_splice
;
1907 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice
,
1911 /* All good, update current len and continue. */
1915 /* This call is useless on a socket so better save a syscall. */
1917 /* This won't block, but will start writeout asynchronously */
1918 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1919 SYNC_FILE_RANGE_WRITE
);
1920 stream
->out_fd_offset
+= ret_splice
;
1922 stream
->output_written
+= ret_splice
;
1923 written
+= ret_splice
;
1926 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1932 * This is a special case that the relayd has closed its socket. Let's
1933 * cleanup the relayd object and all associated streams.
1935 if (relayd
&& relayd_hang_up
) {
1936 cleanup_relayd(relayd
, ctx
);
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 void consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
2114 struct lttng_ht
*ht
= metadata_ht
;
2115 struct lttng_ht_iter iter
;
2116 struct lttng_ht_node_u64
*node
;
2121 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2123 pthread_mutex_lock(&consumer_data
.lock
);
2124 pthread_mutex_lock(&stream
->chan
->lock
);
2125 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2126 pthread_mutex_lock(&stream
->lock
);
2129 * From here, refcounts are updated so be _careful_ when returning an error
2136 * Lookup the stream just to make sure it does not exist in our internal
2137 * state. This should NEVER happen.
2139 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2140 node
= lttng_ht_iter_get_node_u64(&iter
);
2144 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2145 * in terms of destroying the associated channel, because the action that
2146 * causes the count to become 0 also causes a stream to be added. The
2147 * channel deletion will thus be triggered by the following removal of this
2150 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2151 /* Increment refcount before decrementing nb_init_stream_left */
2153 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2156 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2158 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
2159 &stream
->node_channel_id
);
2162 * Add stream to the stream_list_ht of the consumer data. No need to steal
2163 * the key since the HT does not use it and we allow to add redundant keys
2166 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2170 pthread_mutex_unlock(&stream
->lock
);
2171 pthread_mutex_unlock(&stream
->chan
->lock
);
2172 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2173 pthread_mutex_unlock(&consumer_data
.lock
);
2177 * Delete data stream that are flagged for deletion (endpoint_status).
2179 static void validate_endpoint_status_data_stream(void)
2181 struct lttng_ht_iter iter
;
2182 struct lttng_consumer_stream
*stream
;
2184 DBG("Consumer delete flagged data stream");
2187 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2188 /* Validate delete flag of the stream */
2189 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2192 /* Delete it right now */
2193 consumer_del_stream(stream
, data_ht
);
2199 * Delete metadata stream that are flagged for deletion (endpoint_status).
2201 static void validate_endpoint_status_metadata_stream(
2202 struct lttng_poll_event
*pollset
)
2204 struct lttng_ht_iter iter
;
2205 struct lttng_consumer_stream
*stream
;
2207 DBG("Consumer delete flagged metadata stream");
2212 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2213 /* Validate delete flag of the stream */
2214 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2218 * Remove from pollset so the metadata thread can continue without
2219 * blocking on a deleted stream.
2221 lttng_poll_del(pollset
, stream
->wait_fd
);
2223 /* Delete it right now */
2224 consumer_del_metadata_stream(stream
, metadata_ht
);
2230 * Thread polls on metadata file descriptor and write them on disk or on the
2233 void *consumer_thread_metadata_poll(void *data
)
2235 int ret
, i
, pollfd
, err
= -1;
2236 uint32_t revents
, nb_fd
;
2237 struct lttng_consumer_stream
*stream
= NULL
;
2238 struct lttng_ht_iter iter
;
2239 struct lttng_ht_node_u64
*node
;
2240 struct lttng_poll_event events
;
2241 struct lttng_consumer_local_data
*ctx
= data
;
2244 rcu_register_thread();
2246 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2248 if (testpoint(consumerd_thread_metadata
)) {
2249 goto error_testpoint
;
2252 health_code_update();
2254 DBG("Thread metadata poll started");
2256 /* Size is set to 1 for the consumer_metadata pipe */
2257 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2259 ERR("Poll set creation failed");
2263 ret
= lttng_poll_add(&events
,
2264 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2270 DBG("Metadata main loop started");
2274 health_code_update();
2275 health_poll_entry();
2276 DBG("Metadata poll wait");
2277 ret
= lttng_poll_wait(&events
, -1);
2278 DBG("Metadata poll return from wait with %d fd(s)",
2279 LTTNG_POLL_GETNB(&events
));
2281 DBG("Metadata event caught in thread");
2283 if (errno
== EINTR
) {
2284 ERR("Poll EINTR caught");
2287 if (LTTNG_POLL_GETNB(&events
) == 0) {
2288 err
= 0; /* All is OK */
2295 /* From here, the event is a metadata wait fd */
2296 for (i
= 0; i
< nb_fd
; i
++) {
2297 health_code_update();
2299 revents
= LTTNG_POLL_GETEV(&events
, i
);
2300 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2303 /* No activity for this FD (poll implementation). */
2307 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2308 if (revents
& LPOLLIN
) {
2311 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2312 &stream
, sizeof(stream
));
2313 if (pipe_len
< sizeof(stream
)) {
2315 PERROR("read metadata stream");
2318 * Remove the pipe from the poll set and continue the loop
2319 * since their might be data to consume.
2321 lttng_poll_del(&events
,
2322 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2323 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2327 /* A NULL stream means that the state has changed. */
2328 if (stream
== NULL
) {
2329 /* Check for deleted streams. */
2330 validate_endpoint_status_metadata_stream(&events
);
2334 DBG("Adding metadata stream %d to poll set",
2337 /* Add metadata stream to the global poll events list */
2338 lttng_poll_add(&events
, stream
->wait_fd
,
2339 LPOLLIN
| LPOLLPRI
| LPOLLHUP
);
2340 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2341 DBG("Metadata thread pipe hung up");
2343 * Remove the pipe from the poll set and continue the loop
2344 * since their might be data to consume.
2346 lttng_poll_del(&events
,
2347 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2348 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2351 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2355 /* Handle other stream */
2361 uint64_t tmp_id
= (uint64_t) pollfd
;
2363 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2365 node
= lttng_ht_iter_get_node_u64(&iter
);
2368 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2371 if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2372 /* Get the data out of the metadata file descriptor */
2373 DBG("Metadata available on fd %d", pollfd
);
2374 assert(stream
->wait_fd
== pollfd
);
2377 health_code_update();
2379 len
= ctx
->on_buffer_ready(stream
, ctx
);
2381 * We don't check the return value here since if we get
2382 * a negative len, it means an error occurred thus we
2383 * simply remove it from the poll set and free the
2388 /* It's ok to have an unavailable sub-buffer */
2389 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2390 /* Clean up stream from consumer and free it. */
2391 lttng_poll_del(&events
, stream
->wait_fd
);
2392 consumer_del_metadata_stream(stream
, metadata_ht
);
2394 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2395 DBG("Metadata fd %d is hup|err.", pollfd
);
2396 if (!stream
->hangup_flush_done
2397 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2398 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2399 DBG("Attempting to flush and consume the UST buffers");
2400 lttng_ustconsumer_on_stream_hangup(stream
);
2402 /* We just flushed the stream now read it. */
2404 health_code_update();
2406 len
= ctx
->on_buffer_ready(stream
, ctx
);
2408 * We don't check the return value here since if we get
2409 * a negative len, it means an error occurred thus we
2410 * simply remove it from the poll set and free the
2416 lttng_poll_del(&events
, stream
->wait_fd
);
2418 * This call update the channel states, closes file descriptors
2419 * and securely free the stream.
2421 consumer_del_metadata_stream(stream
, metadata_ht
);
2423 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2427 /* Release RCU lock for the stream looked up */
2435 DBG("Metadata poll thread exiting");
2437 lttng_poll_clean(&events
);
2442 ERR("Health error occurred in %s", __func__
);
2444 health_unregister(health_consumerd
);
2445 rcu_unregister_thread();
2450 * This thread polls the fds in the set to consume the data and write
2451 * it to tracefile if necessary.
2453 void *consumer_thread_data_poll(void *data
)
2455 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2456 struct pollfd
*pollfd
= NULL
;
2457 /* local view of the streams */
2458 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2459 /* local view of consumer_data.fds_count */
2460 int nb_fd
= 0, nb_pipes_fd
;
2461 /* Number of FDs with CONSUMER_ENDPOINT_INACTIVE but still open. */
2462 int nb_inactive_fd
= 0;
2463 struct lttng_consumer_local_data
*ctx
= data
;
2466 rcu_register_thread();
2468 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2470 if (testpoint(consumerd_thread_data
)) {
2471 goto error_testpoint
;
2474 health_code_update();
2476 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
*));
2477 if (local_stream
== NULL
) {
2478 PERROR("local_stream malloc");
2483 health_code_update();
2489 * the fds set has been updated, we need to update our
2490 * local array as well
2492 pthread_mutex_lock(&consumer_data
.lock
);
2493 if (consumer_data
.need_update
) {
2498 local_stream
= NULL
;
2501 * Allocate for all fds + 2:
2502 * +1 for the consumer_data_pipe
2503 * +1 for wake up pipe
2506 pollfd
= zmalloc((consumer_data
.stream_count
+ nb_pipes_fd
) * 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
+ nb_pipes_fd
) *
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 && nb_inactive_fd
== 0 &&
2535 CMM_LOAD_SHARED(consumer_quit
) == 1) {
2536 err
= 0; /* All is OK */
2539 /* poll on the array of fds */
2541 DBG("polling on %d fd", nb_fd
+ nb_pipes_fd
);
2542 if (testpoint(consumerd_thread_data_poll
)) {
2545 health_poll_entry();
2546 num_rdy
= poll(pollfd
, nb_fd
+ nb_pipes_fd
, -1);
2548 DBG("poll num_rdy : %d", num_rdy
);
2549 if (num_rdy
== -1) {
2551 * Restart interrupted system call.
2553 if (errno
== EINTR
) {
2556 PERROR("Poll error");
2557 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2559 } else if (num_rdy
== 0) {
2560 DBG("Polling thread timed out");
2564 if (caa_unlikely(data_consumption_paused
)) {
2565 DBG("Data consumption paused, sleeping...");
2571 * If the consumer_data_pipe triggered poll go directly to the
2572 * beginning of the loop to update the array. We want to prioritize
2573 * array update over low-priority reads.
2575 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2576 ssize_t pipe_readlen
;
2578 DBG("consumer_data_pipe wake up");
2579 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2580 &new_stream
, sizeof(new_stream
));
2581 if (pipe_readlen
< sizeof(new_stream
)) {
2582 PERROR("Consumer data pipe");
2583 /* Continue so we can at least handle the current stream(s). */
2588 * If the stream is NULL, just ignore it. It's also possible that
2589 * the sessiond poll thread changed the consumer_quit state and is
2590 * waking us up to test it.
2592 if (new_stream
== NULL
) {
2593 validate_endpoint_status_data_stream();
2597 /* Continue to update the local streams and handle prio ones */
2601 /* Handle wakeup pipe. */
2602 if (pollfd
[nb_fd
+ 1].revents
& (POLLIN
| POLLPRI
)) {
2604 ssize_t pipe_readlen
;
2606 pipe_readlen
= lttng_pipe_read(ctx
->consumer_wakeup_pipe
, &dummy
,
2608 if (pipe_readlen
< 0) {
2609 PERROR("Consumer data wakeup pipe");
2611 /* We've been awakened to handle stream(s). */
2612 ctx
->has_wakeup
= 0;
2615 /* Take care of high priority channels first. */
2616 for (i
= 0; i
< nb_fd
; i
++) {
2617 health_code_update();
2619 if (local_stream
[i
] == NULL
) {
2622 if (pollfd
[i
].revents
& POLLPRI
) {
2623 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2625 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2626 /* it's ok to have an unavailable sub-buffer */
2627 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2628 /* Clean the stream and free it. */
2629 consumer_del_stream(local_stream
[i
], data_ht
);
2630 local_stream
[i
] = NULL
;
2631 } else if (len
> 0) {
2632 local_stream
[i
]->data_read
= 1;
2638 * If we read high prio channel in this loop, try again
2639 * for more high prio data.
2645 /* Take care of low priority channels. */
2646 for (i
= 0; i
< nb_fd
; i
++) {
2647 health_code_update();
2649 if (local_stream
[i
] == NULL
) {
2652 if ((pollfd
[i
].revents
& POLLIN
) ||
2653 local_stream
[i
]->hangup_flush_done
||
2654 local_stream
[i
]->has_data
) {
2655 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2656 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2657 /* it's ok to have an unavailable sub-buffer */
2658 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2659 /* Clean the stream and free it. */
2660 consumer_del_stream(local_stream
[i
], data_ht
);
2661 local_stream
[i
] = NULL
;
2662 } else if (len
> 0) {
2663 local_stream
[i
]->data_read
= 1;
2668 /* Handle hangup and errors */
2669 for (i
= 0; i
< nb_fd
; i
++) {
2670 health_code_update();
2672 if (local_stream
[i
] == NULL
) {
2675 if (!local_stream
[i
]->hangup_flush_done
2676 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2677 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2678 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2679 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2681 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2682 /* Attempt read again, for the data we just flushed. */
2683 local_stream
[i
]->data_read
= 1;
2686 * If the poll flag is HUP/ERR/NVAL and we have
2687 * read no data in this pass, we can remove the
2688 * stream from its hash table.
2690 if ((pollfd
[i
].revents
& POLLHUP
)) {
2691 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2692 if (!local_stream
[i
]->data_read
) {
2693 consumer_del_stream(local_stream
[i
], data_ht
);
2694 local_stream
[i
] = NULL
;
2697 } else if (pollfd
[i
].revents
& POLLERR
) {
2698 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2699 if (!local_stream
[i
]->data_read
) {
2700 consumer_del_stream(local_stream
[i
], data_ht
);
2701 local_stream
[i
] = NULL
;
2704 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2705 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2706 if (!local_stream
[i
]->data_read
) {
2707 consumer_del_stream(local_stream
[i
], data_ht
);
2708 local_stream
[i
] = NULL
;
2712 if (local_stream
[i
] != NULL
) {
2713 local_stream
[i
]->data_read
= 0;
2720 DBG("polling thread exiting");
2725 * Close the write side of the pipe so epoll_wait() in
2726 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2727 * read side of the pipe. If we close them both, epoll_wait strangely does
2728 * not return and could create a endless wait period if the pipe is the
2729 * only tracked fd in the poll set. The thread will take care of closing
2732 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2737 ERR("Health error occurred in %s", __func__
);
2739 health_unregister(health_consumerd
);
2741 rcu_unregister_thread();
2746 * Close wake-up end of each stream belonging to the channel. This will
2747 * allow the poll() on the stream read-side to detect when the
2748 * write-side (application) finally closes them.
2751 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2753 struct lttng_ht
*ht
;
2754 struct lttng_consumer_stream
*stream
;
2755 struct lttng_ht_iter iter
;
2757 ht
= consumer_data
.stream_per_chan_id_ht
;
2760 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2761 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2762 ht
->match_fct
, &channel
->key
,
2763 &iter
.iter
, stream
, node_channel_id
.node
) {
2765 * Protect against teardown with mutex.
2767 pthread_mutex_lock(&stream
->lock
);
2768 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2771 switch (consumer_data
.type
) {
2772 case LTTNG_CONSUMER_KERNEL
:
2774 case LTTNG_CONSUMER32_UST
:
2775 case LTTNG_CONSUMER64_UST
:
2776 if (stream
->metadata_flag
) {
2777 /* Safe and protected by the stream lock. */
2778 lttng_ustconsumer_close_metadata(stream
->chan
);
2781 * Note: a mutex is taken internally within
2782 * liblttng-ust-ctl to protect timer wakeup_fd
2783 * use from concurrent close.
2785 lttng_ustconsumer_close_stream_wakeup(stream
);
2789 ERR("Unknown consumer_data type");
2793 pthread_mutex_unlock(&stream
->lock
);
2798 static void destroy_channel_ht(struct lttng_ht
*ht
)
2800 struct lttng_ht_iter iter
;
2801 struct lttng_consumer_channel
*channel
;
2809 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2810 ret
= lttng_ht_del(ht
, &iter
);
2815 lttng_ht_destroy(ht
);
2819 * This thread polls the channel fds to detect when they are being
2820 * closed. It closes all related streams if the channel is detected as
2821 * closed. It is currently only used as a shim layer for UST because the
2822 * consumerd needs to keep the per-stream wakeup end of pipes open for
2825 void *consumer_thread_channel_poll(void *data
)
2827 int ret
, i
, pollfd
, err
= -1;
2828 uint32_t revents
, nb_fd
;
2829 struct lttng_consumer_channel
*chan
= NULL
;
2830 struct lttng_ht_iter iter
;
2831 struct lttng_ht_node_u64
*node
;
2832 struct lttng_poll_event events
;
2833 struct lttng_consumer_local_data
*ctx
= data
;
2834 struct lttng_ht
*channel_ht
;
2836 rcu_register_thread();
2838 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
2840 if (testpoint(consumerd_thread_channel
)) {
2841 goto error_testpoint
;
2844 health_code_update();
2846 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2848 /* ENOMEM at this point. Better to bail out. */
2852 DBG("Thread channel poll started");
2854 /* Size is set to 1 for the consumer_channel pipe */
2855 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2857 ERR("Poll set creation failed");
2861 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2867 DBG("Channel main loop started");
2871 health_code_update();
2872 DBG("Channel poll wait");
2873 health_poll_entry();
2874 ret
= lttng_poll_wait(&events
, -1);
2875 DBG("Channel poll return from wait with %d fd(s)",
2876 LTTNG_POLL_GETNB(&events
));
2878 DBG("Channel event caught in thread");
2880 if (errno
== EINTR
) {
2881 ERR("Poll EINTR caught");
2884 if (LTTNG_POLL_GETNB(&events
) == 0) {
2885 err
= 0; /* All is OK */
2892 /* From here, the event is a channel wait fd */
2893 for (i
= 0; i
< nb_fd
; i
++) {
2894 health_code_update();
2896 revents
= LTTNG_POLL_GETEV(&events
, i
);
2897 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2900 /* No activity for this FD (poll implementation). */
2904 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2905 if (revents
& LPOLLIN
) {
2906 enum consumer_channel_action action
;
2909 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
2912 ERR("Error reading channel pipe");
2914 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2919 case CONSUMER_CHANNEL_ADD
:
2920 DBG("Adding channel %d to poll set",
2923 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
2926 lttng_ht_add_unique_u64(channel_ht
,
2927 &chan
->wait_fd_node
);
2929 /* Add channel to the global poll events list */
2930 lttng_poll_add(&events
, chan
->wait_fd
,
2931 LPOLLERR
| LPOLLHUP
);
2933 case CONSUMER_CHANNEL_DEL
:
2936 * This command should never be called if the channel
2937 * has streams monitored by either the data or metadata
2938 * thread. The consumer only notify this thread with a
2939 * channel del. command if it receives a destroy
2940 * channel command from the session daemon that send it
2941 * if a command prior to the GET_CHANNEL failed.
2945 chan
= consumer_find_channel(key
);
2948 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
2951 lttng_poll_del(&events
, chan
->wait_fd
);
2952 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
2953 ret
= lttng_ht_del(channel_ht
, &iter
);
2956 switch (consumer_data
.type
) {
2957 case LTTNG_CONSUMER_KERNEL
:
2959 case LTTNG_CONSUMER32_UST
:
2960 case LTTNG_CONSUMER64_UST
:
2961 health_code_update();
2962 /* Destroy streams that might have been left in the stream list. */
2963 clean_channel_stream_list(chan
);
2966 ERR("Unknown consumer_data type");
2971 * Release our own refcount. Force channel deletion even if
2972 * streams were not initialized.
2974 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
2975 consumer_del_channel(chan
);
2980 case CONSUMER_CHANNEL_QUIT
:
2982 * Remove the pipe from the poll set and continue the loop
2983 * since their might be data to consume.
2985 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2988 ERR("Unknown action");
2991 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2992 DBG("Channel thread pipe hung up");
2994 * Remove the pipe from the poll set and continue the loop
2995 * since their might be data to consume.
2997 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
3000 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3004 /* Handle other stream */
3010 uint64_t tmp_id
= (uint64_t) pollfd
;
3012 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
3014 node
= lttng_ht_iter_get_node_u64(&iter
);
3017 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
3020 /* Check for error event */
3021 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
3022 DBG("Channel fd %d is hup|err.", pollfd
);
3024 lttng_poll_del(&events
, chan
->wait_fd
);
3025 ret
= lttng_ht_del(channel_ht
, &iter
);
3029 * This will close the wait fd for each stream associated to
3030 * this channel AND monitored by the data/metadata thread thus
3031 * will be clean by the right thread.
3033 consumer_close_channel_streams(chan
);
3035 /* Release our own refcount */
3036 if (!uatomic_sub_return(&chan
->refcount
, 1)
3037 && !uatomic_read(&chan
->nb_init_stream_left
)) {
3038 consumer_del_channel(chan
);
3041 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
3046 /* Release RCU lock for the channel looked up */
3054 lttng_poll_clean(&events
);
3056 destroy_channel_ht(channel_ht
);
3059 DBG("Channel poll thread exiting");
3062 ERR("Health error occurred in %s", __func__
);
3064 health_unregister(health_consumerd
);
3065 rcu_unregister_thread();
3069 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
3070 struct pollfd
*sockpoll
, int client_socket
)
3077 ret
= lttng_consumer_poll_socket(sockpoll
);
3081 DBG("Metadata connection on client_socket");
3083 /* Blocking call, waiting for transmission */
3084 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
3085 if (ctx
->consumer_metadata_socket
< 0) {
3086 WARN("On accept metadata");
3097 * This thread listens on the consumerd socket and receives the file
3098 * descriptors from the session daemon.
3100 void *consumer_thread_sessiond_poll(void *data
)
3102 int sock
= -1, client_socket
, ret
, err
= -1;
3104 * structure to poll for incoming data on communication socket avoids
3105 * making blocking sockets.
3107 struct pollfd consumer_sockpoll
[2];
3108 struct lttng_consumer_local_data
*ctx
= data
;
3110 rcu_register_thread();
3112 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
3114 if (testpoint(consumerd_thread_sessiond
)) {
3115 goto error_testpoint
;
3118 health_code_update();
3120 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
3121 unlink(ctx
->consumer_command_sock_path
);
3122 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
3123 if (client_socket
< 0) {
3124 ERR("Cannot create command socket");
3128 ret
= lttcomm_listen_unix_sock(client_socket
);
3133 DBG("Sending ready command to lttng-sessiond");
3134 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3135 /* return < 0 on error, but == 0 is not fatal */
3137 ERR("Error sending ready command to lttng-sessiond");
3141 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3142 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3143 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3144 consumer_sockpoll
[1].fd
= client_socket
;
3145 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3147 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3155 DBG("Connection on client_socket");
3157 /* Blocking call, waiting for transmission */
3158 sock
= lttcomm_accept_unix_sock(client_socket
);
3165 * Setup metadata socket which is the second socket connection on the
3166 * command unix socket.
3168 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3177 /* This socket is not useful anymore. */
3178 ret
= close(client_socket
);
3180 PERROR("close client_socket");
3184 /* update the polling structure to poll on the established socket */
3185 consumer_sockpoll
[1].fd
= sock
;
3186 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3189 health_code_update();
3191 health_poll_entry();
3192 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3201 DBG("Incoming command on sock");
3202 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3205 * This could simply be a session daemon quitting. Don't output
3208 DBG("Communication interrupted on command socket");
3212 if (CMM_LOAD_SHARED(consumer_quit
)) {
3213 DBG("consumer_thread_receive_fds received quit from signal");
3214 err
= 0; /* All is OK */
3217 DBG("received command on sock");
3223 DBG("Consumer thread sessiond poll exiting");
3226 * Close metadata streams since the producer is the session daemon which
3229 * NOTE: for now, this only applies to the UST tracer.
3231 lttng_consumer_close_all_metadata();
3234 * when all fds have hung up, the polling thread
3237 CMM_STORE_SHARED(consumer_quit
, 1);
3240 * Notify the data poll thread to poll back again and test the
3241 * consumer_quit state that we just set so to quit gracefully.
3243 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3245 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3247 notify_health_quit_pipe(health_quit_pipe
);
3249 /* Cleaning up possibly open sockets. */
3253 PERROR("close sock sessiond poll");
3256 if (client_socket
>= 0) {
3257 ret
= close(client_socket
);
3259 PERROR("close client_socket sessiond poll");
3266 ERR("Health error occurred in %s", __func__
);
3268 health_unregister(health_consumerd
);
3270 rcu_unregister_thread();
3274 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3275 struct lttng_consumer_local_data
*ctx
)
3279 pthread_mutex_lock(&stream
->lock
);
3280 if (stream
->metadata_flag
) {
3281 pthread_mutex_lock(&stream
->metadata_rdv_lock
);
3284 switch (consumer_data
.type
) {
3285 case LTTNG_CONSUMER_KERNEL
:
3286 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
3288 case LTTNG_CONSUMER32_UST
:
3289 case LTTNG_CONSUMER64_UST
:
3290 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
3293 ERR("Unknown consumer_data type");
3299 if (stream
->metadata_flag
) {
3300 pthread_cond_broadcast(&stream
->metadata_rdv
);
3301 pthread_mutex_unlock(&stream
->metadata_rdv_lock
);
3303 pthread_mutex_unlock(&stream
->lock
);
3307 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3309 switch (consumer_data
.type
) {
3310 case LTTNG_CONSUMER_KERNEL
:
3311 return lttng_kconsumer_on_recv_stream(stream
);
3312 case LTTNG_CONSUMER32_UST
:
3313 case LTTNG_CONSUMER64_UST
:
3314 return lttng_ustconsumer_on_recv_stream(stream
);
3316 ERR("Unknown consumer_data type");
3323 * Allocate and set consumer data hash tables.
3325 int lttng_consumer_init(void)
3327 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3328 if (!consumer_data
.channel_ht
) {
3332 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3333 if (!consumer_data
.relayd_ht
) {
3337 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3338 if (!consumer_data
.stream_list_ht
) {
3342 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3343 if (!consumer_data
.stream_per_chan_id_ht
) {
3347 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3352 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3364 * Process the ADD_RELAYD command receive by a consumer.
3366 * This will create a relayd socket pair and add it to the relayd hash table.
3367 * The caller MUST acquire a RCU read side lock before calling it.
3369 void consumer_add_relayd_socket(uint64_t net_seq_idx
, int sock_type
,
3370 struct lttng_consumer_local_data
*ctx
, int sock
,
3371 struct pollfd
*consumer_sockpoll
,
3372 struct lttcomm_relayd_sock
*relayd_sock
, uint64_t sessiond_id
,
3373 uint64_t relayd_session_id
)
3375 int fd
= -1, ret
= -1, relayd_created
= 0;
3376 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3377 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3380 assert(relayd_sock
);
3382 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", net_seq_idx
);
3384 /* Get relayd reference if exists. */
3385 relayd
= consumer_find_relayd(net_seq_idx
);
3386 if (relayd
== NULL
) {
3387 assert(sock_type
== LTTNG_STREAM_CONTROL
);
3388 /* Not found. Allocate one. */
3389 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3390 if (relayd
== NULL
) {
3391 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3394 relayd
->sessiond_session_id
= sessiond_id
;
3399 * This code path MUST continue to the consumer send status message to
3400 * we can notify the session daemon and continue our work without
3401 * killing everything.
3405 * relayd key should never be found for control socket.
3407 assert(sock_type
!= LTTNG_STREAM_CONTROL
);
3410 /* First send a status message before receiving the fds. */
3411 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3413 /* Somehow, the session daemon is not responding anymore. */
3414 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3415 goto error_nosignal
;
3418 /* Poll on consumer socket. */
3419 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3421 /* Needing to exit in the middle of a command: error. */
3422 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3423 goto error_nosignal
;
3426 /* Get relayd socket from session daemon */
3427 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3428 if (ret
!= sizeof(fd
)) {
3429 fd
= -1; /* Just in case it gets set with an invalid value. */
3432 * Failing to receive FDs might indicate a major problem such as
3433 * reaching a fd limit during the receive where the kernel returns a
3434 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3435 * don't take any chances and stop everything.
3437 * XXX: Feature request #558 will fix that and avoid this possible
3438 * issue when reaching the fd limit.
3440 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3441 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3445 /* Copy socket information and received FD */
3446 switch (sock_type
) {
3447 case LTTNG_STREAM_CONTROL
:
3448 /* Copy received lttcomm socket */
3449 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3450 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3451 /* Handle create_sock error. */
3453 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3457 * Close the socket created internally by
3458 * lttcomm_create_sock, so we can replace it by the one
3459 * received from sessiond.
3461 if (close(relayd
->control_sock
.sock
.fd
)) {
3465 /* Assign new file descriptor */
3466 relayd
->control_sock
.sock
.fd
= fd
;
3467 fd
= -1; /* For error path */
3468 /* Assign version values. */
3469 relayd
->control_sock
.major
= relayd_sock
->major
;
3470 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3472 relayd
->relayd_session_id
= relayd_session_id
;
3475 case LTTNG_STREAM_DATA
:
3476 /* Copy received lttcomm socket */
3477 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3478 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3479 /* Handle create_sock error. */
3481 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3485 * Close the socket created internally by
3486 * lttcomm_create_sock, so we can replace it by the one
3487 * received from sessiond.
3489 if (close(relayd
->data_sock
.sock
.fd
)) {
3493 /* Assign new file descriptor */
3494 relayd
->data_sock
.sock
.fd
= fd
;
3495 fd
= -1; /* for eventual error paths */
3496 /* Assign version values. */
3497 relayd
->data_sock
.major
= relayd_sock
->major
;
3498 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3501 ERR("Unknown relayd socket type (%d)", sock_type
);
3502 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3506 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3507 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3508 relayd
->net_seq_idx
, fd
);
3510 /* We successfully added the socket. Send status back. */
3511 ret
= consumer_send_status_msg(sock
, ret_code
);
3513 /* Somehow, the session daemon is not responding anymore. */
3514 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3515 goto error_nosignal
;
3519 * Add relayd socket pair to consumer data hashtable. If object already
3520 * exists or on error, the function gracefully returns.
3528 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3529 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3533 /* Close received socket if valid. */
3536 PERROR("close received socket");
3540 if (relayd_created
) {
3546 * Try to lock the stream mutex.
3548 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
3550 static int stream_try_lock(struct lttng_consumer_stream
*stream
)
3557 * Try to lock the stream mutex. On failure, we know that the stream is
3558 * being used else where hence there is data still being extracted.
3560 ret
= pthread_mutex_trylock(&stream
->lock
);
3562 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
3574 * Search for a relayd associated to the session id and return the reference.
3576 * A rcu read side lock MUST be acquire before calling this function and locked
3577 * until the relayd object is no longer necessary.
3579 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3581 struct lttng_ht_iter iter
;
3582 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3584 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3585 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3588 * Check by sessiond id which is unique here where the relayd session
3589 * id might not be when having multiple relayd.
3591 if (relayd
->sessiond_session_id
== id
) {
3592 /* Found the relayd. There can be only one per id. */
3604 * Check if for a given session id there is still data needed to be extract
3607 * Return 1 if data is pending or else 0 meaning ready to be read.
3609 int consumer_data_pending(uint64_t id
)
3612 struct lttng_ht_iter iter
;
3613 struct lttng_ht
*ht
;
3614 struct lttng_consumer_stream
*stream
;
3615 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3616 int (*data_pending
)(struct lttng_consumer_stream
*);
3618 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3621 pthread_mutex_lock(&consumer_data
.lock
);
3623 switch (consumer_data
.type
) {
3624 case LTTNG_CONSUMER_KERNEL
:
3625 data_pending
= lttng_kconsumer_data_pending
;
3627 case LTTNG_CONSUMER32_UST
:
3628 case LTTNG_CONSUMER64_UST
:
3629 data_pending
= lttng_ustconsumer_data_pending
;
3632 ERR("Unknown consumer data type");
3636 /* Ease our life a bit */
3637 ht
= consumer_data
.stream_list_ht
;
3639 relayd
= find_relayd_by_session_id(id
);
3641 /* Send init command for data pending. */
3642 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3643 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3644 relayd
->relayd_session_id
);
3645 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3647 /* Communication error thus the relayd so no data pending. */
3648 goto data_not_pending
;
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);
3689 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3691 pthread_mutex_unlock(&stream
->lock
);
3695 pthread_mutex_unlock(&stream
->lock
);
3699 unsigned int is_data_inflight
= 0;
3701 /* Send init command for data pending. */
3702 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3703 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3704 relayd
->relayd_session_id
, &is_data_inflight
);
3705 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3707 goto data_not_pending
;
3709 if (is_data_inflight
) {
3715 * Finding _no_ node in the hash table and no inflight data means that the
3716 * stream(s) have been removed thus data is guaranteed to be available for
3717 * analysis from the trace files.
3721 /* Data is available to be read by a viewer. */
3722 pthread_mutex_unlock(&consumer_data
.lock
);
3727 /* Data is still being extracted from buffers. */
3728 pthread_mutex_unlock(&consumer_data
.lock
);
3734 * Send a ret code status message to the sessiond daemon.
3736 * Return the sendmsg() return value.
3738 int consumer_send_status_msg(int sock
, int ret_code
)
3740 struct lttcomm_consumer_status_msg msg
;
3742 memset(&msg
, 0, sizeof(msg
));
3743 msg
.ret_code
= ret_code
;
3745 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3749 * Send a channel status message to the sessiond daemon.
3751 * Return the sendmsg() return value.
3753 int consumer_send_status_channel(int sock
,
3754 struct lttng_consumer_channel
*channel
)
3756 struct lttcomm_consumer_status_channel msg
;
3760 memset(&msg
, 0, sizeof(msg
));
3762 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3764 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3765 msg
.key
= channel
->key
;
3766 msg
.stream_count
= channel
->streams
.count
;
3769 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3772 unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos
,
3773 unsigned long produced_pos
, uint64_t nb_packets_per_stream
,
3774 uint64_t max_sb_size
)
3776 unsigned long start_pos
;
3778 if (!nb_packets_per_stream
) {
3779 return consumed_pos
; /* Grab everything */
3781 start_pos
= produced_pos
- offset_align_floor(produced_pos
, max_sb_size
);
3782 start_pos
-= max_sb_size
* nb_packets_per_stream
;
3783 if ((long) (start_pos
- consumed_pos
) < 0) {
3784 return consumed_pos
; /* Grab everything */
3790 int rotate_rename_local(const char *old_path
, const char *new_path
,
3791 uid_t uid
, gid_t gid
)
3798 ret
= utils_mkdir_recursive(new_path
, S_IRWXU
| S_IRWXG
, uid
, gid
);
3800 ERR("Create directory on rotate");
3804 ret
= rename(old_path
, new_path
);
3805 if (ret
< 0 && errno
!= ENOENT
) {
3806 PERROR("Rename completed rotation chunk");
3816 int rotate_rename_relay(const char *old_path
, const char *new_path
,
3820 struct consumer_relayd_sock_pair
*relayd
;
3822 relayd
= consumer_find_relayd(relayd_id
);
3824 ERR("Failed to find relayd while running rotate_rename_relay command");
3829 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3830 ret
= relayd_rotate_rename(&relayd
->control_sock
, old_path
, new_path
);
3831 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3836 int lttng_consumer_rotate_rename(const char *old_path
, const char *new_path
,
3837 uid_t uid
, gid_t gid
, uint64_t relayd_id
)
3839 if (relayd_id
!= -1ULL) {
3840 return rotate_rename_relay(old_path
, new_path
, relayd_id
);
3842 return rotate_rename_local(old_path
, new_path
, uid
, gid
);
3847 int mkdir_local(const char *path
, uid_t uid
, gid_t gid
)
3851 ret
= utils_mkdir_recursive(path
, S_IRWXU
| S_IRWXG
, uid
, gid
);
3853 /* utils_mkdir_recursive logs an error. */
3863 int mkdir_relay(const char *path
, uint64_t relayd_id
)
3866 struct consumer_relayd_sock_pair
*relayd
;
3868 relayd
= consumer_find_relayd(relayd_id
);
3870 ERR("Failed to find relayd");
3875 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3876 ret
= relayd_mkdir(&relayd
->control_sock
, path
);
3877 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3884 int lttng_consumer_mkdir(const char *path
, uid_t uid
, gid_t gid
,
3887 if (relayd_id
!= -1ULL) {
3888 return mkdir_relay(path
, relayd_id
);
3890 return mkdir_local(path
, uid
, gid
);