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.
20 #include "common/index/ctf-index.h"
28 #include <sys/socket.h>
29 #include <sys/types.h>
34 #include <bin/lttng-consumerd/health-consumerd.h>
35 #include <common/common.h>
36 #include <common/utils.h>
37 #include <common/compat/poll.h>
38 #include <common/compat/endian.h>
39 #include <common/index/index.h>
40 #include <common/kernel-ctl/kernel-ctl.h>
41 #include <common/sessiond-comm/relayd.h>
42 #include <common/sessiond-comm/sessiond-comm.h>
43 #include <common/kernel-consumer/kernel-consumer.h>
44 #include <common/relayd/relayd.h>
45 #include <common/ust-consumer/ust-consumer.h>
46 #include <common/consumer/consumer-timer.h>
47 #include <common/consumer/consumer.h>
48 #include <common/consumer/consumer-stream.h>
49 #include <common/consumer/consumer-testpoint.h>
50 #include <common/align.h>
51 #include <common/consumer/consumer-metadata-cache.h>
53 struct lttng_consumer_global_data consumer_data
= {
56 .type
= LTTNG_CONSUMER_UNKNOWN
,
59 enum consumer_channel_action
{
62 CONSUMER_CHANNEL_QUIT
,
65 struct consumer_channel_msg
{
66 enum consumer_channel_action action
;
67 struct lttng_consumer_channel
*chan
; /* add */
68 uint64_t key
; /* del */
72 * Flag to inform the polling thread to quit when all fd hung up. Updated by
73 * the consumer_thread_receive_fds when it notices that all fds has hung up.
74 * Also updated by the signal handler (consumer_should_exit()). Read by the
77 volatile int consumer_quit
;
80 * Global hash table containing respectively metadata and data streams. The
81 * stream element in this ht should only be updated by the metadata poll thread
82 * for the metadata and the data poll thread for the data.
84 static struct lttng_ht
*metadata_ht
;
85 static struct lttng_ht
*data_ht
;
88 * Notify a thread lttng pipe to poll back again. This usually means that some
89 * global state has changed so we just send back the thread in a poll wait
92 static void notify_thread_lttng_pipe(struct lttng_pipe
*pipe
)
94 struct lttng_consumer_stream
*null_stream
= NULL
;
98 (void) lttng_pipe_write(pipe
, &null_stream
, sizeof(null_stream
));
101 static void notify_health_quit_pipe(int *pipe
)
105 ret
= lttng_write(pipe
[1], "4", 1);
107 PERROR("write consumer health quit");
111 static void notify_channel_pipe(struct lttng_consumer_local_data
*ctx
,
112 struct lttng_consumer_channel
*chan
,
114 enum consumer_channel_action action
)
116 struct consumer_channel_msg msg
;
119 memset(&msg
, 0, sizeof(msg
));
124 ret
= lttng_write(ctx
->consumer_channel_pipe
[1], &msg
, sizeof(msg
));
125 if (ret
< sizeof(msg
)) {
126 PERROR("notify_channel_pipe write error");
130 void notify_thread_del_channel(struct lttng_consumer_local_data
*ctx
,
133 notify_channel_pipe(ctx
, NULL
, key
, CONSUMER_CHANNEL_DEL
);
136 static int read_channel_pipe(struct lttng_consumer_local_data
*ctx
,
137 struct lttng_consumer_channel
**chan
,
139 enum consumer_channel_action
*action
)
141 struct consumer_channel_msg msg
;
144 ret
= lttng_read(ctx
->consumer_channel_pipe
[0], &msg
, sizeof(msg
));
145 if (ret
< sizeof(msg
)) {
149 *action
= msg
.action
;
157 * Cleanup the stream list of a channel. Those streams are not yet globally
160 static void clean_channel_stream_list(struct lttng_consumer_channel
*channel
)
162 struct lttng_consumer_stream
*stream
, *stmp
;
166 /* Delete streams that might have been left in the stream list. */
167 cds_list_for_each_entry_safe(stream
, stmp
, &channel
->streams
.head
,
169 cds_list_del(&stream
->send_node
);
171 * Once a stream is added to this list, the buffers were created so we
172 * have a guarantee that this call will succeed. Setting the monitor
173 * mode to 0 so we don't lock nor try to delete the stream from the
177 consumer_stream_destroy(stream
, NULL
);
182 * Find a stream. The consumer_data.lock must be locked during this
185 static struct lttng_consumer_stream
*find_stream(uint64_t key
,
188 struct lttng_ht_iter iter
;
189 struct lttng_ht_node_u64
*node
;
190 struct lttng_consumer_stream
*stream
= NULL
;
194 /* -1ULL keys are lookup failures */
195 if (key
== (uint64_t) -1ULL) {
201 lttng_ht_lookup(ht
, &key
, &iter
);
202 node
= lttng_ht_iter_get_node_u64(&iter
);
204 stream
= caa_container_of(node
, struct lttng_consumer_stream
, node
);
212 static void steal_stream_key(uint64_t key
, struct lttng_ht
*ht
)
214 struct lttng_consumer_stream
*stream
;
217 stream
= find_stream(key
, ht
);
219 stream
->key
= (uint64_t) -1ULL;
221 * We don't want the lookup to match, but we still need
222 * to iterate on this stream when iterating over the hash table. Just
223 * change the node key.
225 stream
->node
.key
= (uint64_t) -1ULL;
231 * Return a channel object for the given key.
233 * RCU read side lock MUST be acquired before calling this function and
234 * protects the channel ptr.
236 struct lttng_consumer_channel
*consumer_find_channel(uint64_t key
)
238 struct lttng_ht_iter iter
;
239 struct lttng_ht_node_u64
*node
;
240 struct lttng_consumer_channel
*channel
= NULL
;
242 /* -1ULL keys are lookup failures */
243 if (key
== (uint64_t) -1ULL) {
247 lttng_ht_lookup(consumer_data
.channel_ht
, &key
, &iter
);
248 node
= lttng_ht_iter_get_node_u64(&iter
);
250 channel
= caa_container_of(node
, struct lttng_consumer_channel
, node
);
257 * There is a possibility that the consumer does not have enough time between
258 * the close of the channel on the session daemon and the cleanup in here thus
259 * once we have a channel add with an existing key, we know for sure that this
260 * channel will eventually get cleaned up by all streams being closed.
262 * This function just nullifies the already existing channel key.
264 static void steal_channel_key(uint64_t key
)
266 struct lttng_consumer_channel
*channel
;
269 channel
= consumer_find_channel(key
);
271 channel
->key
= (uint64_t) -1ULL;
273 * We don't want the lookup to match, but we still need to iterate on
274 * this channel when iterating over the hash table. Just change the
277 channel
->node
.key
= (uint64_t) -1ULL;
282 static void free_channel_rcu(struct rcu_head
*head
)
284 struct lttng_ht_node_u64
*node
=
285 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
286 struct lttng_consumer_channel
*channel
=
287 caa_container_of(node
, struct lttng_consumer_channel
, node
);
289 switch (consumer_data
.type
) {
290 case LTTNG_CONSUMER_KERNEL
:
292 case LTTNG_CONSUMER32_UST
:
293 case LTTNG_CONSUMER64_UST
:
294 lttng_ustconsumer_free_channel(channel
);
297 ERR("Unknown consumer_data type");
304 * RCU protected relayd socket pair free.
306 static void free_relayd_rcu(struct rcu_head
*head
)
308 struct lttng_ht_node_u64
*node
=
309 caa_container_of(head
, struct lttng_ht_node_u64
, head
);
310 struct consumer_relayd_sock_pair
*relayd
=
311 caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
314 * Close all sockets. This is done in the call RCU since we don't want the
315 * socket fds to be reassigned thus potentially creating bad state of the
318 * We do not have to lock the control socket mutex here since at this stage
319 * there is no one referencing to this relayd object.
321 (void) relayd_close(&relayd
->control_sock
);
322 (void) relayd_close(&relayd
->data_sock
);
324 pthread_mutex_destroy(&relayd
->ctrl_sock_mutex
);
329 * Destroy and free relayd socket pair object.
331 void consumer_destroy_relayd(struct consumer_relayd_sock_pair
*relayd
)
334 struct lttng_ht_iter iter
;
336 if (relayd
== NULL
) {
340 DBG("Consumer destroy and close relayd socket pair");
342 iter
.iter
.node
= &relayd
->node
.node
;
343 ret
= lttng_ht_del(consumer_data
.relayd_ht
, &iter
);
345 /* We assume the relayd is being or is destroyed */
349 /* RCU free() call */
350 call_rcu(&relayd
->node
.head
, free_relayd_rcu
);
354 * Remove a channel from the global list protected by a mutex. This function is
355 * also responsible for freeing its data structures.
357 void consumer_del_channel(struct lttng_consumer_channel
*channel
)
360 struct lttng_ht_iter iter
;
362 DBG("Consumer delete channel key %" PRIu64
, channel
->key
);
364 pthread_mutex_lock(&consumer_data
.lock
);
365 pthread_mutex_lock(&channel
->lock
);
367 /* Destroy streams that might have been left in the stream list. */
368 clean_channel_stream_list(channel
);
370 if (channel
->live_timer_enabled
== 1) {
371 consumer_timer_live_stop(channel
);
374 switch (consumer_data
.type
) {
375 case LTTNG_CONSUMER_KERNEL
:
377 case LTTNG_CONSUMER32_UST
:
378 case LTTNG_CONSUMER64_UST
:
379 lttng_ustconsumer_del_channel(channel
);
382 ERR("Unknown consumer_data type");
388 iter
.iter
.node
= &channel
->node
.node
;
389 ret
= lttng_ht_del(consumer_data
.channel_ht
, &iter
);
393 call_rcu(&channel
->node
.head
, free_channel_rcu
);
395 pthread_mutex_unlock(&channel
->lock
);
396 pthread_mutex_unlock(&consumer_data
.lock
);
400 * Iterate over the relayd hash table and destroy each element. Finally,
401 * destroy the whole hash table.
403 static void cleanup_relayd_ht(void)
405 struct lttng_ht_iter iter
;
406 struct consumer_relayd_sock_pair
*relayd
;
410 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
412 consumer_destroy_relayd(relayd
);
417 lttng_ht_destroy(consumer_data
.relayd_ht
);
421 * Update the end point status of all streams having the given relayd id.
423 * It's atomically set without having the stream mutex locked which is fine
424 * because we handle the write/read race with a pipe wakeup for each thread.
426 static void update_endpoint_status_by_netidx(uint64_t relayd_id
,
427 enum consumer_endpoint_status status
)
429 struct lttng_ht_iter iter
;
430 struct lttng_consumer_stream
*stream
;
432 DBG("Consumer set delete flag on stream by idx %" PRIu64
, relayd_id
);
436 /* Let's begin with metadata */
437 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
438 if (stream
->relayd_id
== relayd_id
) {
439 uatomic_set(&stream
->endpoint_status
, status
);
440 DBG("Delete flag set to metadata stream %d", stream
->wait_fd
);
444 /* Follow up by the data streams */
445 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
446 if (stream
->relayd_id
== relayd_id
) {
447 uatomic_set(&stream
->endpoint_status
, status
);
448 DBG("Delete flag set to data stream %d", stream
->wait_fd
);
455 * Cleanup a relayd object by flagging every associated streams for deletion,
456 * destroying the object meaning removing it from the relayd hash table,
457 * closing the sockets and freeing the memory in a RCU call.
459 * If a local data context is available, notify the threads that the streams'
460 * state have changed.
462 void lttng_consumer_cleanup_relayd(struct consumer_relayd_sock_pair
*relayd
)
468 DBG("Cleaning up relayd object ID %"PRIu64
, relayd
->id
);
470 /* Save the net sequence index before destroying the object */
474 * Delete the relayd from the relayd hash table, close the sockets and free
475 * the object in a RCU call.
477 consumer_destroy_relayd(relayd
);
479 /* Set inactive endpoint to all streams */
480 update_endpoint_status_by_netidx(netidx
, CONSUMER_ENDPOINT_INACTIVE
);
483 * With a local data context, notify the threads that the streams' state
484 * have changed. The write() action on the pipe acts as an "implicit"
485 * memory barrier ordering the updates of the end point status from the
486 * read of this status which happens AFTER receiving this notify.
488 notify_thread_lttng_pipe(relayd
->ctx
->consumer_data_pipe
);
489 notify_thread_lttng_pipe(relayd
->ctx
->consumer_metadata_pipe
);
493 * Flag a relayd socket pair for destruction. Destroy it if the refcount
496 * RCU read side lock MUST be aquired before calling this function.
498 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair
*relayd
)
502 /* Set destroy flag for this object */
503 uatomic_set(&relayd
->destroy_flag
, 1);
505 /* Destroy the relayd if refcount is 0 */
506 if (uatomic_read(&relayd
->refcount
) == 0) {
507 consumer_destroy_relayd(relayd
);
512 * Completly destroy stream from every visiable data structure and the given
515 * One this call returns, the stream object is not longer usable nor visible.
517 void consumer_del_stream(struct lttng_consumer_stream
*stream
,
520 consumer_stream_destroy(stream
, ht
);
524 * XXX naming of del vs destroy is all mixed up.
526 void consumer_del_stream_for_data(struct lttng_consumer_stream
*stream
)
528 consumer_stream_destroy(stream
, data_ht
);
531 void consumer_del_stream_for_metadata(struct lttng_consumer_stream
*stream
)
533 consumer_stream_destroy(stream
, metadata_ht
);
537 * Add a stream to the global list protected by a mutex.
539 int consumer_add_data_stream(struct lttng_consumer_stream
*stream
)
541 struct lttng_ht
*ht
= data_ht
;
547 DBG3("Adding consumer stream %" PRIu64
, stream
->key
);
549 pthread_mutex_lock(&consumer_data
.lock
);
550 pthread_mutex_lock(&stream
->chan
->lock
);
551 pthread_mutex_lock(&stream
->chan
->timer_lock
);
552 pthread_mutex_lock(&stream
->lock
);
555 /* Steal stream identifier to avoid having streams with the same key */
556 steal_stream_key(stream
->key
, ht
);
558 lttng_ht_add_unique_u64(ht
, &stream
->node
);
560 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
561 &stream
->node_channel_id
);
564 * Add stream to the stream_list_ht of the consumer data. No need to steal
565 * the key since the HT does not use it and we allow to add redundant keys
568 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
571 * When nb_init_stream_left reaches 0, we don't need to trigger any action
572 * in terms of destroying the associated channel, because the action that
573 * causes the count to become 0 also causes a stream to be added. The
574 * channel deletion will thus be triggered by the following removal of this
577 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
578 /* Increment refcount before decrementing nb_init_stream_left */
580 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
583 /* Update consumer data once the node is inserted. */
584 consumer_data
.stream_count
++;
585 consumer_data
.need_update
= 1;
588 pthread_mutex_unlock(&stream
->lock
);
589 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
590 pthread_mutex_unlock(&stream
->chan
->lock
);
591 pthread_mutex_unlock(&consumer_data
.lock
);
596 void consumer_del_data_stream(struct lttng_consumer_stream
*stream
)
598 consumer_del_stream(stream
, data_ht
);
602 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
603 * be acquired before calling this.
605 static int add_relayd(struct consumer_relayd_sock_pair
*relayd
)
608 struct lttng_ht_node_u64
*node
;
609 struct lttng_ht_iter iter
;
613 lttng_ht_lookup(consumer_data
.relayd_ht
,
615 node
= lttng_ht_iter_get_node_u64(&iter
);
619 lttng_ht_add_unique_u64(consumer_data
.relayd_ht
, &relayd
->node
);
626 * Allocate and return a consumer relayd socket.
628 static struct consumer_relayd_sock_pair
*consumer_allocate_relayd_sock_pair(
631 struct consumer_relayd_sock_pair
*obj
= NULL
;
633 /* net sequence index of -1 is a failure */
634 if (relayd_id
== (uint64_t) -1ULL) {
638 obj
= zmalloc(sizeof(struct consumer_relayd_sock_pair
));
640 PERROR("zmalloc relayd sock");
646 obj
->destroy_flag
= 0;
647 obj
->control_sock
.sock
.fd
= -1;
648 obj
->data_sock
.sock
.fd
= -1;
649 lttng_ht_node_init_u64(&obj
->node
, obj
->id
);
650 pthread_mutex_init(&obj
->ctrl_sock_mutex
, NULL
);
657 * Find a relayd socket pair in the global consumer data.
659 * Return the object if found else NULL.
660 * RCU read-side lock must be held across this call and while using the
663 struct consumer_relayd_sock_pair
*consumer_find_relayd(uint64_t key
)
665 struct lttng_ht_iter iter
;
666 struct lttng_ht_node_u64
*node
;
667 struct consumer_relayd_sock_pair
*relayd
= NULL
;
669 /* Negative keys are lookup failures */
670 if (key
== (uint64_t) -1ULL) {
674 lttng_ht_lookup(consumer_data
.relayd_ht
, &key
,
676 node
= lttng_ht_iter_get_node_u64(&iter
);
678 relayd
= caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
686 * Find a relayd and send the stream
688 * Returns 0 on success, < 0 on error
690 int consumer_send_relayd_stream(struct lttng_consumer_stream
*stream
,
694 struct consumer_relayd_sock_pair
*relayd
;
697 assert(stream
->relayd_id
!= -1ULL);
700 /* The stream is not metadata. Get relayd reference if exists. */
702 relayd
= consumer_find_relayd(stream
->relayd_id
);
703 if (relayd
!= NULL
) {
704 /* Add stream on the relayd */
705 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
706 ret
= relayd_add_stream(&relayd
->control_sock
, stream
->name
,
707 path
, &stream
->relayd_stream_id
,
708 stream
->chan
->tracefile_size
, stream
->chan
->tracefile_count
);
709 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
711 ERR("Relayd add stream failed. Cleaning up relayd %" PRIu64
".", relayd
->id
);
712 lttng_consumer_cleanup_relayd(relayd
);
716 uatomic_inc(&relayd
->refcount
);
717 stream
->sent_to_relayd
= 1;
719 ERR("Stream %" PRIu64
" relayd ID %" PRIu64
" unknown. Can't send it.",
720 stream
->key
, stream
->relayd_id
);
725 DBG("Stream %s with key %" PRIu64
" sent to relayd id %" PRIu64
,
726 stream
->name
, stream
->key
, stream
->relayd_id
);
734 * Find a relayd and send the streams sent message
736 * Returns 0 on success, < 0 on error
738 int consumer_send_relayd_streams_sent(uint64_t relayd_id
)
741 struct consumer_relayd_sock_pair
*relayd
;
743 assert(relayd_id
!= -1ULL);
745 /* The stream is not metadata. Get relayd reference if exists. */
747 relayd
= consumer_find_relayd(relayd_id
);
748 if (relayd
!= NULL
) {
749 /* Add stream on the relayd */
750 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
751 ret
= relayd_streams_sent(&relayd
->control_sock
);
752 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
754 ERR("Relayd streams sent failed. Cleaning up relayd %" PRIu64
".", relayd
->id
);
755 lttng_consumer_cleanup_relayd(relayd
);
759 ERR("Relayd ID %" PRIu64
" unknown. Can't send streams_sent.",
766 DBG("All streams sent relayd id %" PRIu64
, relayd_id
);
774 * Find a relayd and close the stream
776 void close_relayd_stream(struct lttng_consumer_stream
*stream
)
778 struct consumer_relayd_sock_pair
*relayd
;
780 /* The stream is not metadata. Get relayd reference if exists. */
782 relayd
= consumer_find_relayd(stream
->relayd_id
);
784 consumer_stream_relayd_close(stream
, relayd
);
790 * Handle stream for relayd transmission if the stream applies for network
791 * streaming where the net sequence index is set.
793 * Return destination file descriptor or negative value on error.
795 static int write_relayd_stream_header(struct lttng_consumer_stream
*stream
,
796 size_t data_size
, unsigned long padding
,
797 struct consumer_relayd_sock_pair
*relayd
)
800 struct lttcomm_relayd_data_hdr data_hdr
;
806 /* Reset data header */
807 memset(&data_hdr
, 0, sizeof(data_hdr
));
809 if (stream
->metadata_flag
) {
810 /* Caller MUST acquire the relayd control socket lock */
811 ret
= relayd_send_metadata(&relayd
->control_sock
, data_size
);
816 /* Metadata are always sent on the control socket. */
817 outfd
= relayd
->control_sock
.sock
.fd
;
819 /* Set header with stream information */
820 data_hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
821 data_hdr
.data_size
= htobe32(data_size
);
822 data_hdr
.padding_size
= htobe32(padding
);
824 * Note that net_seq_num below is assigned with the *current* value of
825 * next_net_seq_num and only after that the next_net_seq_num will be
826 * increment. This is why when issuing a command on the relayd using
827 * this next value, 1 should always be substracted in order to compare
828 * the last seen sequence number on the relayd side to the last sent.
830 data_hdr
.net_seq_num
= htobe64(stream
->next_net_seq_num
);
831 /* Other fields are zeroed previously */
833 ret
= relayd_send_data_hdr(&relayd
->data_sock
, &data_hdr
,
839 ++stream
->next_net_seq_num
;
841 /* Set to go on data socket */
842 outfd
= relayd
->data_sock
.sock
.fd
;
850 * Allocate and return a new lttng_consumer_channel object using the given key
851 * to initialize the hash table node.
853 * On error, return NULL.
855 struct lttng_consumer_channel
*consumer_allocate_channel(uint64_t key
,
857 const char *pathname
,
862 enum lttng_event_output output
,
863 uint64_t tracefile_size
,
864 uint64_t tracefile_count
,
865 uint64_t session_id_per_pid
,
866 unsigned int monitor
,
867 unsigned int live_timer_interval
,
868 bool is_in_live_session
,
869 const char *root_shm_path
,
870 const char *shm_path
)
872 struct lttng_consumer_channel
*channel
;
874 channel
= zmalloc(sizeof(*channel
));
875 if (channel
== NULL
) {
876 PERROR("malloc struct lttng_consumer_channel");
881 channel
->refcount
= 0;
882 channel
->session_id
= session_id
;
883 channel
->session_id_per_pid
= session_id_per_pid
;
886 channel
->relayd_id
= relayd_id
;
887 channel
->tracefile_size
= tracefile_size
;
888 channel
->tracefile_count
= tracefile_count
;
889 channel
->monitor
= monitor
;
890 channel
->live_timer_interval
= live_timer_interval
;
891 channel
->is_live
= is_in_live_session
;
892 pthread_mutex_init(&channel
->lock
, NULL
);
893 pthread_mutex_init(&channel
->timer_lock
, NULL
);
896 case LTTNG_EVENT_SPLICE
:
897 channel
->output
= CONSUMER_CHANNEL_SPLICE
;
899 case LTTNG_EVENT_MMAP
:
900 channel
->output
= CONSUMER_CHANNEL_MMAP
;
910 * In monitor mode, the streams associated with the channel will be put in
911 * a special list ONLY owned by this channel. So, the refcount is set to 1
912 * here meaning that the channel itself has streams that are referenced.
914 * On a channel deletion, once the channel is no longer visible, the
915 * refcount is decremented and checked for a zero value to delete it. With
916 * streams in no monitor mode, it will now be safe to destroy the channel.
918 if (!channel
->monitor
) {
919 channel
->refcount
= 1;
922 strncpy(channel
->pathname
, pathname
, sizeof(channel
->pathname
));
923 channel
->pathname
[sizeof(channel
->pathname
) - 1] = '\0';
925 strncpy(channel
->name
, name
, sizeof(channel
->name
));
926 channel
->name
[sizeof(channel
->name
) - 1] = '\0';
929 strncpy(channel
->root_shm_path
, root_shm_path
, sizeof(channel
->root_shm_path
));
930 channel
->root_shm_path
[sizeof(channel
->root_shm_path
) - 1] = '\0';
933 strncpy(channel
->shm_path
, shm_path
, sizeof(channel
->shm_path
));
934 channel
->shm_path
[sizeof(channel
->shm_path
) - 1] = '\0';
937 lttng_ht_node_init_u64(&channel
->node
, channel
->key
);
939 channel
->wait_fd
= -1;
941 CDS_INIT_LIST_HEAD(&channel
->streams
.head
);
943 DBG("Allocated channel (key %" PRIu64
")", channel
->key
);
950 * Add a channel to the global list protected by a mutex.
952 * Always return 0 indicating success.
954 int consumer_add_channel(struct lttng_consumer_channel
*channel
,
955 struct lttng_consumer_local_data
*ctx
)
957 pthread_mutex_lock(&consumer_data
.lock
);
958 pthread_mutex_lock(&channel
->lock
);
959 pthread_mutex_lock(&channel
->timer_lock
);
962 * This gives us a guarantee that the channel we are about to add to the
963 * channel hash table will be unique. See this function comment on the why
964 * we need to steel the channel key at this stage.
966 steal_channel_key(channel
->key
);
969 lttng_ht_add_unique_u64(consumer_data
.channel_ht
, &channel
->node
);
972 pthread_mutex_unlock(&channel
->timer_lock
);
973 pthread_mutex_unlock(&channel
->lock
);
974 pthread_mutex_unlock(&consumer_data
.lock
);
976 if (channel
->wait_fd
!= -1 && channel
->type
== CONSUMER_CHANNEL_TYPE_DATA
) {
977 notify_channel_pipe(ctx
, channel
, -1, CONSUMER_CHANNEL_ADD
);
984 * Allocate the pollfd structure and the local view of the out fds to avoid
985 * doing a lookup in the linked list and concurrency issues when writing is
986 * needed. Called with consumer_data.lock held.
988 * Returns the number of fds in the structures.
990 static int update_poll_array(struct lttng_consumer_local_data
*ctx
,
991 struct pollfd
**pollfd
, struct lttng_consumer_stream
**local_stream
,
992 struct lttng_ht
*ht
, int *nb_inactive_fd
)
995 struct lttng_ht_iter iter
;
996 struct lttng_consumer_stream
*stream
;
1001 assert(local_stream
);
1003 DBG("Updating poll fd array");
1004 *nb_inactive_fd
= 0;
1006 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1008 * Only active streams with an active end point can be added to the
1009 * poll set and local stream storage of the thread.
1011 * There is a potential race here for endpoint_status to be updated
1012 * just after the check. However, this is OK since the stream(s) will
1013 * be deleted once the thread is notified that the end point state has
1014 * changed where this function will be called back again.
1016 * We track the number of inactive FDs because they still need to be
1017 * closed by the polling thread after a wakeup on the data_pipe or
1020 if (stream
->state
!= LTTNG_CONSUMER_ACTIVE_STREAM
||
1021 stream
->endpoint_status
== CONSUMER_ENDPOINT_INACTIVE
) {
1022 (*nb_inactive_fd
)++;
1026 * This clobbers way too much the debug output. Uncomment that if you
1027 * need it for debugging purposes.
1029 * DBG("Active FD %d", stream->wait_fd);
1031 (*pollfd
)[i
].fd
= stream
->wait_fd
;
1032 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1033 local_stream
[i
] = stream
;
1039 * Insert the consumer_data_pipe at the end of the array and don't
1040 * increment i so nb_fd is the number of real FD.
1042 (*pollfd
)[i
].fd
= lttng_pipe_get_readfd(ctx
->consumer_data_pipe
);
1043 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1045 (*pollfd
)[i
+ 1].fd
= lttng_pipe_get_readfd(ctx
->consumer_wakeup_pipe
);
1046 (*pollfd
)[i
+ 1].events
= POLLIN
| POLLPRI
;
1051 * Poll on the should_quit pipe and the command socket return -1 on
1052 * error, 1 if should exit, 0 if data is available on the command socket
1054 int lttng_consumer_poll_socket(struct pollfd
*consumer_sockpoll
)
1059 num_rdy
= poll(consumer_sockpoll
, 2, -1);
1060 if (num_rdy
== -1) {
1062 * Restart interrupted system call.
1064 if (errno
== EINTR
) {
1067 PERROR("Poll error");
1070 if (consumer_sockpoll
[0].revents
& (POLLIN
| POLLPRI
)) {
1071 DBG("consumer_should_quit wake up");
1078 * Set the error socket.
1080 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data
*ctx
,
1083 ctx
->consumer_error_socket
= sock
;
1087 * Set the command socket path.
1089 void lttng_consumer_set_command_sock_path(
1090 struct lttng_consumer_local_data
*ctx
, char *sock
)
1092 ctx
->consumer_command_sock_path
= sock
;
1096 * Send return code to the session daemon.
1097 * If the socket is not defined, we return 0, it is not a fatal error
1099 int lttng_consumer_send_error(struct lttng_consumer_local_data
*ctx
, int cmd
)
1101 if (ctx
->consumer_error_socket
> 0) {
1102 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
1103 sizeof(enum lttcomm_sessiond_command
));
1110 * Close all the tracefiles and stream fds and MUST be called when all
1111 * instances are destroyed i.e. when all threads were joined and are ended.
1113 void lttng_consumer_cleanup(void)
1115 struct lttng_ht_iter iter
;
1116 struct lttng_consumer_channel
*channel
;
1120 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
, channel
,
1122 consumer_del_channel(channel
);
1127 lttng_ht_destroy(consumer_data
.channel_ht
);
1129 cleanup_relayd_ht();
1131 lttng_ht_destroy(consumer_data
.stream_per_chan_id_ht
);
1134 * This HT contains streams that are freed by either the metadata thread or
1135 * the data thread so we do *nothing* on the hash table and simply destroy
1138 lttng_ht_destroy(consumer_data
.stream_list_ht
);
1142 * Called from signal handler.
1144 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
1149 ret
= lttng_write(ctx
->consumer_should_quit
[1], "4", 1);
1151 PERROR("write consumer quit");
1154 DBG("Consumer flag that it should quit");
1159 * Flush pending writes to trace output disk file.
1162 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1166 int outfd
= stream
->out_fd
;
1169 * This does a blocking write-and-wait on any page that belongs to the
1170 * subbuffer prior to the one we just wrote.
1171 * Don't care about error values, as these are just hints and ways to
1172 * limit the amount of page cache used.
1174 if (orig_offset
< stream
->max_sb_size
) {
1177 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1178 stream
->max_sb_size
,
1179 SYNC_FILE_RANGE_WAIT_BEFORE
1180 | SYNC_FILE_RANGE_WRITE
1181 | SYNC_FILE_RANGE_WAIT_AFTER
);
1183 * Give hints to the kernel about how we access the file:
1184 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1187 * We need to call fadvise again after the file grows because the
1188 * kernel does not seem to apply fadvise to non-existing parts of the
1191 * Call fadvise _after_ having waited for the page writeback to
1192 * complete because the dirty page writeback semantic is not well
1193 * defined. So it can be expected to lead to lower throughput in
1196 ret
= posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1197 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1198 if (ret
&& ret
!= -ENOSYS
) {
1200 PERROR("posix_fadvise on fd %i", outfd
);
1205 * Initialise the necessary environnement :
1206 * - create a new context
1207 * - create the poll_pipe
1208 * - create the should_quit pipe (for signal handler)
1209 * - create the thread pipe (for splice)
1211 * Takes a function pointer as argument, this function is called when data is
1212 * available on a buffer. This function is responsible to do the
1213 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1214 * buffer configuration and then kernctl_put_next_subbuf at the end.
1216 * Returns a pointer to the new context or NULL on error.
1218 struct lttng_consumer_local_data
*lttng_consumer_create(
1219 enum lttng_consumer_type type
,
1220 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1221 struct lttng_consumer_local_data
*ctx
, bool locked_by_caller
),
1222 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1223 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1224 int (*update_stream
)(uint64_t stream_key
, uint32_t state
))
1227 struct lttng_consumer_local_data
*ctx
;
1229 assert(consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1230 consumer_data
.type
== type
);
1231 consumer_data
.type
= type
;
1233 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
1235 PERROR("allocating context");
1239 ctx
->consumer_error_socket
= -1;
1240 ctx
->consumer_metadata_socket
= -1;
1241 pthread_mutex_init(&ctx
->metadata_socket_lock
, NULL
);
1242 /* assign the callbacks */
1243 ctx
->on_buffer_ready
= buffer_ready
;
1244 ctx
->on_recv_channel
= recv_channel
;
1245 ctx
->on_recv_stream
= recv_stream
;
1246 ctx
->on_update_stream
= update_stream
;
1248 ctx
->consumer_data_pipe
= lttng_pipe_open(0);
1249 if (!ctx
->consumer_data_pipe
) {
1250 goto error_poll_pipe
;
1253 ctx
->consumer_wakeup_pipe
= lttng_pipe_open(0);
1254 if (!ctx
->consumer_wakeup_pipe
) {
1255 goto error_wakeup_pipe
;
1258 ret
= pipe(ctx
->consumer_should_quit
);
1260 PERROR("Error creating recv pipe");
1261 goto error_quit_pipe
;
1264 ret
= pipe(ctx
->consumer_channel_pipe
);
1266 PERROR("Error creating channel pipe");
1267 goto error_channel_pipe
;
1270 ctx
->consumer_metadata_pipe
= lttng_pipe_open(0);
1271 if (!ctx
->consumer_metadata_pipe
) {
1272 goto error_metadata_pipe
;
1277 error_metadata_pipe
:
1278 utils_close_pipe(ctx
->consumer_channel_pipe
);
1280 utils_close_pipe(ctx
->consumer_should_quit
);
1282 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1284 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1292 * Iterate over all streams of the hashtable and free them properly.
1294 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1296 struct lttng_ht_iter iter
;
1297 struct lttng_consumer_stream
*stream
;
1304 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1306 * Ignore return value since we are currently cleaning up so any error
1309 (void) consumer_del_stream(stream
, ht
);
1313 lttng_ht_destroy(ht
);
1317 * Iterate over all streams of the metadata hashtable and free them
1320 static void destroy_metadata_stream_ht(struct lttng_ht
*ht
)
1322 struct lttng_ht_iter iter
;
1323 struct lttng_consumer_stream
*stream
;
1330 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1332 * Ignore return value since we are currently cleaning up so any error
1335 (void) consumer_del_metadata_stream(stream
, ht
);
1339 lttng_ht_destroy(ht
);
1343 * Close all fds associated with the instance and free the context.
1345 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1349 DBG("Consumer destroying it. Closing everything.");
1355 destroy_data_stream_ht(data_ht
);
1356 destroy_metadata_stream_ht(metadata_ht
);
1358 ret
= close(ctx
->consumer_error_socket
);
1362 ret
= close(ctx
->consumer_metadata_socket
);
1366 utils_close_pipe(ctx
->consumer_channel_pipe
);
1367 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1368 lttng_pipe_destroy(ctx
->consumer_metadata_pipe
);
1369 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1370 utils_close_pipe(ctx
->consumer_should_quit
);
1372 unlink(ctx
->consumer_command_sock_path
);
1377 * Write the metadata stream id on the specified file descriptor.
1379 static int write_relayd_metadata_id(int fd
,
1380 struct lttng_consumer_stream
*stream
,
1381 struct consumer_relayd_sock_pair
*relayd
, unsigned long padding
)
1384 struct lttcomm_relayd_metadata_payload hdr
;
1386 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1387 hdr
.padding_size
= htobe32(padding
);
1388 ret
= lttng_write(fd
, (void *) &hdr
, sizeof(hdr
));
1389 if (ret
< sizeof(hdr
)) {
1391 * This error means that the fd's end is closed so ignore the PERROR
1392 * not to clubber the error output since this can happen in a normal
1395 if (errno
!= EPIPE
) {
1396 PERROR("write metadata stream id");
1398 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1400 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1401 * handle writting the missing part so report that as an error and
1402 * don't lie to the caller.
1407 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1408 stream
->relayd_stream_id
, padding
);
1415 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1416 * core function for writing trace buffers to either the local filesystem or
1419 * It must be called with the stream lock held.
1421 * Careful review MUST be put if any changes occur!
1423 * Returns the number of bytes written
1425 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1426 struct lttng_consumer_local_data
*ctx
,
1427 struct lttng_consumer_stream
*stream
,
1428 const struct lttng_buffer_view
*buffer
,
1429 unsigned long padding
)
1432 off_t orig_offset
= stream
->out_fd_offset
;
1433 /* Default is on the disk */
1434 int outfd
= stream
->out_fd
;
1435 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1436 unsigned int relayd_hang_up
= 0;
1437 const size_t subbuf_content_size
= buffer
->size
- padding
;
1440 /* RCU lock for the relayd pointer */
1443 /* Flag that the current stream if set for network streaming. */
1444 if (stream
->relayd_id
!= (uint64_t) -1ULL) {
1445 relayd
= consumer_find_relayd(stream
->relayd_id
);
1446 if (relayd
== NULL
) {
1452 /* Handle stream on the relayd if the output is on the network */
1454 unsigned long netlen
= subbuf_content_size
;
1457 * Lock the control socket for the complete duration of the function
1458 * since from this point on we will use the socket.
1460 if (stream
->metadata_flag
) {
1461 /* Metadata requires the control socket. */
1462 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1463 if (stream
->reset_metadata_flag
) {
1464 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1465 stream
->relayd_stream_id
,
1466 stream
->metadata_version
);
1471 stream
->reset_metadata_flag
= 0;
1473 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1476 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1481 /* Use the returned socket. */
1484 /* Write metadata stream id before payload */
1485 if (stream
->metadata_flag
) {
1486 ret
= write_relayd_metadata_id(outfd
, stream
, relayd
, padding
);
1493 write_len
= subbuf_content_size
;
1495 /* No streaming; we have to write the full padding. */
1496 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1497 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1499 ERR("Reset metadata file");
1502 stream
->reset_metadata_flag
= 0;
1506 * Check if we need to change the tracefile before writing the packet.
1508 if (stream
->chan
->tracefile_size
> 0 &&
1509 (stream
->tracefile_size_current
+ buffer
->size
) >
1510 stream
->chan
->tracefile_size
) {
1511 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1512 stream
->name
, stream
->chan
->tracefile_size
,
1513 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1514 stream
->out_fd
, &(stream
->tracefile_count_current
),
1517 ERR("Rotating output file");
1520 outfd
= stream
->out_fd
;
1522 if (stream
->index_file
) {
1523 lttng_index_file_put(stream
->index_file
);
1524 stream
->index_file
= lttng_index_file_create(stream
->chan
->pathname
,
1525 stream
->name
, stream
->uid
, stream
->gid
,
1526 stream
->chan
->tracefile_size
,
1527 stream
->tracefile_count_current
,
1528 CTF_INDEX_MAJOR
, CTF_INDEX_MINOR
);
1529 if (!stream
->index_file
) {
1534 /* Reset current size because we just perform a rotation. */
1535 stream
->tracefile_size_current
= 0;
1536 stream
->out_fd_offset
= 0;
1539 stream
->tracefile_size_current
+= buffer
->size
;
1540 write_len
= buffer
->size
;
1544 * This call guarantee that len or less is returned. It's impossible to
1545 * receive a ret value that is bigger than len.
1547 ret
= lttng_write(outfd
, buffer
->data
, write_len
);
1548 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, write_len
);
1549 if (ret
< 0 || ((size_t) ret
!= write_len
)) {
1551 * Report error to caller if nothing was written else at least send the
1559 /* Socket operation failed. We consider the relayd dead */
1560 if (errno
== EPIPE
|| errno
== EINVAL
|| errno
== EBADF
) {
1562 * This is possible if the fd is closed on the other side
1563 * (outfd) or any write problem. It can be verbose a bit for a
1564 * normal execution if for instance the relayd is stopped
1565 * abruptly. This can happen so set this to a DBG statement.
1567 DBG("Consumer mmap write detected relayd hang up");
1569 /* Unhandled error, print it and stop function right now. */
1570 PERROR("Error in write mmap (ret %zd != write_len %zu)", ret
,
1575 stream
->output_written
+= ret
;
1577 /* This call is useless on a socket so better save a syscall. */
1579 /* This won't block, but will start writeout asynchronously */
1580 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, write_len
,
1581 SYNC_FILE_RANGE_WRITE
);
1582 stream
->out_fd_offset
+= write_len
;
1583 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1588 * This is a special case that the relayd has closed its socket. Let's
1589 * cleanup the relayd object and all associated streams.
1591 if (relayd
&& relayd_hang_up
) {
1592 ERR("Relayd hangup. Cleaning up relayd %" PRIu64
".", relayd
->id
);
1593 lttng_consumer_cleanup_relayd(relayd
);
1597 /* Unlock only if ctrl socket used */
1598 if (relayd
&& stream
->metadata_flag
) {
1599 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1607 * Splice the data from the ring buffer to the tracefile.
1609 * It must be called with the stream lock held.
1611 * Returns the number of bytes spliced.
1613 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1614 struct lttng_consumer_local_data
*ctx
,
1615 struct lttng_consumer_stream
*stream
, unsigned long len
,
1616 unsigned long padding
)
1618 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1620 off_t orig_offset
= stream
->out_fd_offset
;
1621 int fd
= stream
->wait_fd
;
1622 /* Default is on the disk */
1623 int outfd
= stream
->out_fd
;
1624 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1626 unsigned int relayd_hang_up
= 0;
1628 switch (consumer_data
.type
) {
1629 case LTTNG_CONSUMER_KERNEL
:
1631 case LTTNG_CONSUMER32_UST
:
1632 case LTTNG_CONSUMER64_UST
:
1633 /* Not supported for user space tracing */
1636 ERR("Unknown consumer_data type");
1640 /* RCU lock for the relayd pointer */
1643 /* Flag that the current stream if set for network streaming. */
1644 if (stream
->relayd_id
!= (uint64_t) -1ULL) {
1645 relayd
= consumer_find_relayd(stream
->relayd_id
);
1646 if (relayd
== NULL
) {
1651 splice_pipe
= stream
->splice_pipe
;
1653 /* Write metadata stream id before payload */
1655 unsigned long total_len
= len
;
1657 if (stream
->metadata_flag
) {
1659 * Lock the control socket for the complete duration of the function
1660 * since from this point on we will use the socket.
1662 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1664 if (stream
->reset_metadata_flag
) {
1665 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1666 stream
->relayd_stream_id
,
1667 stream
->metadata_version
);
1672 stream
->reset_metadata_flag
= 0;
1674 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
, relayd
,
1682 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1685 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1691 /* Use the returned socket. */
1694 /* No streaming, we have to set the len with the full padding */
1697 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1698 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1700 ERR("Reset metadata file");
1703 stream
->reset_metadata_flag
= 0;
1706 * Check if we need to change the tracefile before writing the packet.
1708 if (stream
->chan
->tracefile_size
> 0 &&
1709 (stream
->tracefile_size_current
+ len
) >
1710 stream
->chan
->tracefile_size
) {
1711 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1712 stream
->name
, stream
->chan
->tracefile_size
,
1713 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1714 stream
->out_fd
, &(stream
->tracefile_count_current
),
1718 ERR("Rotating output file");
1721 outfd
= stream
->out_fd
;
1723 if (stream
->index_file
) {
1724 lttng_index_file_put(stream
->index_file
);
1725 stream
->index_file
= lttng_index_file_create(stream
->chan
->pathname
,
1726 stream
->name
, stream
->uid
, stream
->gid
,
1727 stream
->chan
->tracefile_size
,
1728 stream
->tracefile_count_current
,
1729 CTF_INDEX_MAJOR
, CTF_INDEX_MINOR
);
1730 if (!stream
->index_file
) {
1735 /* Reset current size because we just perform a rotation. */
1736 stream
->tracefile_size_current
= 0;
1737 stream
->out_fd_offset
= 0;
1740 stream
->tracefile_size_current
+= len
;
1744 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1745 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1746 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1747 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1748 DBG("splice chan to pipe, ret %zd", ret_splice
);
1749 if (ret_splice
< 0) {
1752 PERROR("Error in relay splice");
1756 /* Handle stream on the relayd if the output is on the network */
1757 if (relayd
&& stream
->metadata_flag
) {
1758 size_t metadata_payload_size
=
1759 sizeof(struct lttcomm_relayd_metadata_payload
);
1761 /* Update counter to fit the spliced data */
1762 ret_splice
+= metadata_payload_size
;
1763 len
+= metadata_payload_size
;
1765 * We do this so the return value can match the len passed as
1766 * argument to this function.
1768 written
-= metadata_payload_size
;
1771 /* Splice data out */
1772 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1773 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1774 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
1776 if (ret_splice
< 0) {
1781 } else if (ret_splice
> len
) {
1783 * We don't expect this code path to be executed but you never know
1784 * so this is an extra protection agains a buggy splice().
1787 written
+= ret_splice
;
1788 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice
,
1792 /* All good, update current len and continue. */
1796 /* This call is useless on a socket so better save a syscall. */
1798 /* This won't block, but will start writeout asynchronously */
1799 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1800 SYNC_FILE_RANGE_WRITE
);
1801 stream
->out_fd_offset
+= ret_splice
;
1803 stream
->output_written
+= ret_splice
;
1804 written
+= ret_splice
;
1807 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1813 * This is a special case that the relayd has closed its socket. Let's
1814 * cleanup the relayd object and all associated streams.
1816 if (relayd
&& relayd_hang_up
) {
1817 ERR("Relayd hangup. Cleaning up relayd %" PRIu64
".", relayd
->id
);
1818 lttng_consumer_cleanup_relayd(relayd
);
1819 /* Skip splice error so the consumer does not fail */
1824 /* send the appropriate error description to sessiond */
1827 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1830 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1833 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
1838 if (relayd
&& stream
->metadata_flag
) {
1839 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1847 * Take a snapshot for a specific fd
1849 * Returns 0 on success, < 0 on error
1851 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
1853 switch (consumer_data
.type
) {
1854 case LTTNG_CONSUMER_KERNEL
:
1855 return lttng_kconsumer_take_snapshot(stream
);
1856 case LTTNG_CONSUMER32_UST
:
1857 case LTTNG_CONSUMER64_UST
:
1858 return lttng_ustconsumer_take_snapshot(stream
);
1860 ERR("Unknown consumer_data type");
1867 * Get the produced position
1869 * Returns 0 on success, < 0 on error
1871 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
1874 switch (consumer_data
.type
) {
1875 case LTTNG_CONSUMER_KERNEL
:
1876 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
1877 case LTTNG_CONSUMER32_UST
:
1878 case LTTNG_CONSUMER64_UST
:
1879 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
1881 ERR("Unknown consumer_data type");
1887 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
1888 int sock
, struct pollfd
*consumer_sockpoll
)
1890 switch (consumer_data
.type
) {
1891 case LTTNG_CONSUMER_KERNEL
:
1892 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1893 case LTTNG_CONSUMER32_UST
:
1894 case LTTNG_CONSUMER64_UST
:
1895 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1897 ERR("Unknown consumer_data type");
1903 void lttng_consumer_close_all_metadata(void)
1905 switch (consumer_data
.type
) {
1906 case LTTNG_CONSUMER_KERNEL
:
1908 * The Kernel consumer has a different metadata scheme so we don't
1909 * close anything because the stream will be closed by the session
1913 case LTTNG_CONSUMER32_UST
:
1914 case LTTNG_CONSUMER64_UST
:
1916 * Close all metadata streams. The metadata hash table is passed and
1917 * this call iterates over it by closing all wakeup fd. This is safe
1918 * because at this point we are sure that the metadata producer is
1919 * either dead or blocked.
1921 lttng_ustconsumer_close_all_metadata(metadata_ht
);
1924 ERR("Unknown consumer_data type");
1930 * Clean up a metadata stream and free its memory.
1932 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
1933 struct lttng_ht
*ht
)
1935 struct lttng_consumer_channel
*free_chan
= NULL
;
1939 * This call should NEVER receive regular stream. It must always be
1940 * metadata stream and this is crucial for data structure synchronization.
1942 assert(stream
->metadata_flag
);
1944 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
1946 pthread_mutex_lock(&consumer_data
.lock
);
1947 pthread_mutex_lock(&stream
->chan
->lock
);
1948 pthread_mutex_lock(&stream
->lock
);
1949 if (stream
->chan
->metadata_cache
) {
1950 /* Only applicable to userspace consumers. */
1951 pthread_mutex_lock(&stream
->chan
->metadata_cache
->lock
);
1954 /* Remove any reference to that stream. */
1955 consumer_stream_delete(stream
, ht
);
1957 /* Close down everything including the relayd if one. */
1958 consumer_stream_close(stream
);
1959 /* Destroy tracer buffers of the stream. */
1960 consumer_stream_destroy_buffers(stream
);
1962 /* Atomically decrement channel refcount since other threads can use it. */
1963 if (!uatomic_sub_return(&stream
->chan
->refcount
, 1)
1964 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
1965 /* Go for channel deletion! */
1966 free_chan
= stream
->chan
;
1970 * Nullify the stream reference so it is not used after deletion. The
1971 * channel lock MUST be acquired before being able to check for a NULL
1974 stream
->chan
->metadata_stream
= NULL
;
1976 if (stream
->chan
->metadata_cache
) {
1977 pthread_mutex_unlock(&stream
->chan
->metadata_cache
->lock
);
1979 pthread_mutex_unlock(&stream
->lock
);
1980 pthread_mutex_unlock(&stream
->chan
->lock
);
1981 pthread_mutex_unlock(&consumer_data
.lock
);
1984 consumer_del_channel(free_chan
);
1987 consumer_stream_free(stream
);
1991 * Action done with the metadata stream when adding it to the consumer internal
1992 * data structures to handle it.
1994 int consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
1996 struct lttng_ht
*ht
= metadata_ht
;
1998 struct lttng_ht_iter iter
;
1999 struct lttng_ht_node_u64
*node
;
2004 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2006 pthread_mutex_lock(&consumer_data
.lock
);
2007 pthread_mutex_lock(&stream
->chan
->lock
);
2008 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2009 pthread_mutex_lock(&stream
->lock
);
2012 * From here, refcounts are updated so be _careful_ when returning an error
2019 * Lookup the stream just to make sure it does not exist in our internal
2020 * state. This should NEVER happen.
2022 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2023 node
= lttng_ht_iter_get_node_u64(&iter
);
2027 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2028 * in terms of destroying the associated channel, because the action that
2029 * causes the count to become 0 also causes a stream to be added. The
2030 * channel deletion will thus be triggered by the following removal of this
2033 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2034 /* Increment refcount before decrementing nb_init_stream_left */
2036 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2039 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2041 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
2042 &stream
->node_channel_id
);
2045 * Add stream to the stream_list_ht of the consumer data. No need to steal
2046 * the key since the HT does not use it and we allow to add redundant keys
2049 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2053 pthread_mutex_unlock(&stream
->lock
);
2054 pthread_mutex_unlock(&stream
->chan
->lock
);
2055 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2056 pthread_mutex_unlock(&consumer_data
.lock
);
2061 * Delete data stream that are flagged for deletion (endpoint_status).
2063 static void validate_endpoint_status_data_stream(void)
2065 struct lttng_ht_iter iter
;
2066 struct lttng_consumer_stream
*stream
;
2068 DBG("Consumer delete flagged data stream");
2071 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2072 /* Validate delete flag of the stream */
2073 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2076 /* Delete it right now */
2077 consumer_del_stream(stream
, data_ht
);
2083 * Delete metadata stream that are flagged for deletion (endpoint_status).
2085 static void validate_endpoint_status_metadata_stream(
2086 struct lttng_poll_event
*pollset
)
2088 struct lttng_ht_iter iter
;
2089 struct lttng_consumer_stream
*stream
;
2091 DBG("Consumer delete flagged metadata stream");
2096 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2097 /* Validate delete flag of the stream */
2098 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2102 * Remove from pollset so the metadata thread can continue without
2103 * blocking on a deleted stream.
2105 lttng_poll_del(pollset
, stream
->wait_fd
);
2107 /* Delete it right now */
2108 consumer_del_metadata_stream(stream
, metadata_ht
);
2114 * Thread polls on metadata file descriptor and write them on disk or on the
2117 void *consumer_thread_metadata_poll(void *data
)
2119 int ret
, i
, pollfd
, err
= -1;
2120 uint32_t revents
, nb_fd
;
2121 struct lttng_consumer_stream
*stream
= NULL
;
2122 struct lttng_ht_iter iter
;
2123 struct lttng_ht_node_u64
*node
;
2124 struct lttng_poll_event events
;
2125 struct lttng_consumer_local_data
*ctx
= data
;
2128 rcu_register_thread();
2130 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2132 if (testpoint(consumerd_thread_metadata
)) {
2133 goto error_testpoint
;
2136 health_code_update();
2138 DBG("Thread metadata poll started");
2140 /* Size is set to 1 for the consumer_metadata pipe */
2141 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2143 ERR("Poll set creation failed");
2147 ret
= lttng_poll_add(&events
,
2148 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2154 DBG("Metadata main loop started");
2158 health_code_update();
2159 health_poll_entry();
2160 DBG("Metadata poll wait");
2161 ret
= lttng_poll_wait(&events
, -1);
2162 DBG("Metadata poll return from wait with %d fd(s)",
2163 LTTNG_POLL_GETNB(&events
));
2165 DBG("Metadata event caught in thread");
2167 if (errno
== EINTR
) {
2168 ERR("Poll EINTR caught");
2171 if (LTTNG_POLL_GETNB(&events
) == 0) {
2172 err
= 0; /* All is OK */
2179 /* From here, the event is a metadata wait fd */
2180 for (i
= 0; i
< nb_fd
; i
++) {
2181 health_code_update();
2183 revents
= LTTNG_POLL_GETEV(&events
, i
);
2184 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2187 /* No activity for this FD (poll implementation). */
2191 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2192 if (revents
& LPOLLIN
) {
2195 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2196 &stream
, sizeof(stream
));
2197 if (pipe_len
< sizeof(stream
)) {
2199 PERROR("read metadata stream");
2202 * Remove the pipe from the poll set and continue the loop
2203 * since their might be data to consume.
2205 lttng_poll_del(&events
,
2206 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2207 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2211 /* A NULL stream means that the state has changed. */
2212 if (stream
== NULL
) {
2213 /* Check for deleted streams. */
2214 validate_endpoint_status_metadata_stream(&events
);
2218 DBG("Adding metadata stream %d to poll set",
2221 /* Add metadata stream to the global poll events list */
2222 lttng_poll_add(&events
, stream
->wait_fd
,
2223 LPOLLIN
| LPOLLPRI
| LPOLLHUP
);
2224 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2225 DBG("Metadata thread pipe hung up");
2227 * Remove the pipe from the poll set and continue the loop
2228 * since their might be data to consume.
2230 lttng_poll_del(&events
,
2231 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2232 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2235 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2239 /* Handle other stream */
2245 uint64_t tmp_id
= (uint64_t) pollfd
;
2247 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2249 node
= lttng_ht_iter_get_node_u64(&iter
);
2252 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2255 if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2256 /* Get the data out of the metadata file descriptor */
2257 DBG("Metadata available on fd %d", pollfd
);
2258 assert(stream
->wait_fd
== pollfd
);
2261 health_code_update();
2263 len
= ctx
->on_buffer_ready(stream
, ctx
, false);
2265 * We don't check the return value here since if we get
2266 * a negative len, it means an error occurred thus we
2267 * simply remove it from the poll set and free the
2272 /* It's ok to have an unavailable sub-buffer */
2273 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2274 /* Clean up stream from consumer and free it. */
2275 lttng_poll_del(&events
, stream
->wait_fd
);
2276 consumer_del_metadata_stream(stream
, metadata_ht
);
2278 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2279 DBG("Metadata fd %d is hup|err.", pollfd
);
2280 if (!stream
->hangup_flush_done
2281 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2282 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2283 DBG("Attempting to flush and consume the UST buffers");
2284 lttng_ustconsumer_on_stream_hangup(stream
);
2286 /* We just flushed the stream now read it. */
2288 health_code_update();
2290 len
= ctx
->on_buffer_ready(stream
, ctx
, false);
2292 * We don't check the return value here since if we get
2293 * a negative len, it means an error occurred thus we
2294 * simply remove it from the poll set and free the
2300 lttng_poll_del(&events
, stream
->wait_fd
);
2302 * This call update the channel states, closes file descriptors
2303 * and securely free the stream.
2305 consumer_del_metadata_stream(stream
, metadata_ht
);
2307 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2311 /* Release RCU lock for the stream looked up */
2319 DBG("Metadata poll thread exiting");
2321 lttng_poll_clean(&events
);
2326 ERR("Health error occurred in %s", __func__
);
2328 health_unregister(health_consumerd
);
2329 rcu_unregister_thread();
2334 * This thread polls the fds in the set to consume the data and write
2335 * it to tracefile if necessary.
2337 void *consumer_thread_data_poll(void *data
)
2339 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2340 struct pollfd
*pollfd
= NULL
;
2341 /* local view of the streams */
2342 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2343 /* local view of consumer_data.fds_count */
2345 /* Number of FDs with CONSUMER_ENDPOINT_INACTIVE but still open. */
2346 int nb_inactive_fd
= 0;
2347 struct lttng_consumer_local_data
*ctx
= data
;
2350 rcu_register_thread();
2352 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2354 if (testpoint(consumerd_thread_data
)) {
2355 goto error_testpoint
;
2358 health_code_update();
2360 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
*));
2361 if (local_stream
== NULL
) {
2362 PERROR("local_stream malloc");
2367 health_code_update();
2373 * the fds set has been updated, we need to update our
2374 * local array as well
2376 pthread_mutex_lock(&consumer_data
.lock
);
2377 if (consumer_data
.need_update
) {
2382 local_stream
= NULL
;
2385 * Allocate for all fds +1 for the consumer_data_pipe and +1 for
2388 pollfd
= zmalloc((consumer_data
.stream_count
+ 2) * sizeof(struct pollfd
));
2389 if (pollfd
== NULL
) {
2390 PERROR("pollfd malloc");
2391 pthread_mutex_unlock(&consumer_data
.lock
);
2395 local_stream
= zmalloc((consumer_data
.stream_count
+ 2) *
2396 sizeof(struct lttng_consumer_stream
*));
2397 if (local_stream
== NULL
) {
2398 PERROR("local_stream malloc");
2399 pthread_mutex_unlock(&consumer_data
.lock
);
2402 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2403 data_ht
, &nb_inactive_fd
);
2405 ERR("Error in allocating pollfd or local_outfds");
2406 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2407 pthread_mutex_unlock(&consumer_data
.lock
);
2411 consumer_data
.need_update
= 0;
2413 pthread_mutex_unlock(&consumer_data
.lock
);
2415 /* No FDs and consumer_quit, consumer_cleanup the thread */
2416 if (nb_fd
== 0 && consumer_quit
== 1 && nb_inactive_fd
== 0) {
2417 err
= 0; /* All is OK */
2420 /* poll on the array of fds */
2422 DBG("polling on %d fd", nb_fd
+ 2);
2423 health_poll_entry();
2424 num_rdy
= poll(pollfd
, nb_fd
+ 2, -1);
2426 DBG("poll num_rdy : %d", num_rdy
);
2427 if (num_rdy
== -1) {
2429 * Restart interrupted system call.
2431 if (errno
== EINTR
) {
2434 PERROR("Poll error");
2435 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2437 } else if (num_rdy
== 0) {
2438 DBG("Polling thread timed out");
2443 * If the consumer_data_pipe triggered poll go directly to the
2444 * beginning of the loop to update the array. We want to prioritize
2445 * array update over low-priority reads.
2447 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2448 ssize_t pipe_readlen
;
2450 DBG("consumer_data_pipe wake up");
2451 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2452 &new_stream
, sizeof(new_stream
));
2453 if (pipe_readlen
< sizeof(new_stream
)) {
2454 PERROR("Consumer data pipe");
2455 /* Continue so we can at least handle the current stream(s). */
2460 * If the stream is NULL, just ignore it. It's also possible that
2461 * the sessiond poll thread changed the consumer_quit state and is
2462 * waking us up to test it.
2464 if (new_stream
== NULL
) {
2465 validate_endpoint_status_data_stream();
2469 /* Continue to update the local streams and handle prio ones */
2473 /* Handle wakeup pipe. */
2474 if (pollfd
[nb_fd
+ 1].revents
& (POLLIN
| POLLPRI
)) {
2476 ssize_t pipe_readlen
;
2478 pipe_readlen
= lttng_pipe_read(ctx
->consumer_wakeup_pipe
, &dummy
,
2480 if (pipe_readlen
< 0) {
2481 PERROR("Consumer data wakeup pipe");
2483 /* We've been awakened to handle stream(s). */
2484 ctx
->has_wakeup
= 0;
2487 /* Take care of high priority channels first. */
2488 for (i
= 0; i
< nb_fd
; i
++) {
2489 health_code_update();
2491 if (local_stream
[i
] == NULL
) {
2494 if (pollfd
[i
].revents
& POLLPRI
) {
2495 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2497 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
, false);
2498 /* it's ok to have an unavailable sub-buffer */
2499 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2500 /* Clean the stream and free it. */
2501 consumer_del_stream(local_stream
[i
], data_ht
);
2502 local_stream
[i
] = NULL
;
2503 } else if (len
> 0) {
2504 local_stream
[i
]->data_read
= 1;
2510 * If we read high prio channel in this loop, try again
2511 * for more high prio data.
2517 /* Take care of low priority channels. */
2518 for (i
= 0; i
< nb_fd
; i
++) {
2519 health_code_update();
2521 if (local_stream
[i
] == NULL
) {
2524 if ((pollfd
[i
].revents
& POLLIN
) ||
2525 local_stream
[i
]->hangup_flush_done
||
2526 local_stream
[i
]->has_data
) {
2527 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2528 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
, false);
2529 /* it's ok to have an unavailable sub-buffer */
2530 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2531 /* Clean the stream and free it. */
2532 consumer_del_stream(local_stream
[i
], data_ht
);
2533 local_stream
[i
] = NULL
;
2534 } else if (len
> 0) {
2535 local_stream
[i
]->data_read
= 1;
2540 /* Handle hangup and errors */
2541 for (i
= 0; i
< nb_fd
; i
++) {
2542 health_code_update();
2544 if (local_stream
[i
] == NULL
) {
2547 if (!local_stream
[i
]->hangup_flush_done
2548 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2549 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2550 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2551 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2553 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2554 /* Attempt read again, for the data we just flushed. */
2555 local_stream
[i
]->data_read
= 1;
2558 * If the poll flag is HUP/ERR/NVAL and we have
2559 * read no data in this pass, we can remove the
2560 * stream from its hash table.
2562 if ((pollfd
[i
].revents
& POLLHUP
)) {
2563 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2564 if (!local_stream
[i
]->data_read
) {
2565 consumer_del_stream(local_stream
[i
], data_ht
);
2566 local_stream
[i
] = NULL
;
2569 } else if (pollfd
[i
].revents
& POLLERR
) {
2570 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2571 if (!local_stream
[i
]->data_read
) {
2572 consumer_del_stream(local_stream
[i
], data_ht
);
2573 local_stream
[i
] = NULL
;
2576 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2577 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2578 if (!local_stream
[i
]->data_read
) {
2579 consumer_del_stream(local_stream
[i
], data_ht
);
2580 local_stream
[i
] = NULL
;
2584 if (local_stream
[i
] != NULL
) {
2585 local_stream
[i
]->data_read
= 0;
2592 DBG("polling thread exiting");
2597 * Close the write side of the pipe so epoll_wait() in
2598 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2599 * read side of the pipe. If we close them both, epoll_wait strangely does
2600 * not return and could create a endless wait period if the pipe is the
2601 * only tracked fd in the poll set. The thread will take care of closing
2604 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2609 ERR("Health error occurred in %s", __func__
);
2611 health_unregister(health_consumerd
);
2613 rcu_unregister_thread();
2618 * Close wake-up end of each stream belonging to the channel. This will
2619 * allow the poll() on the stream read-side to detect when the
2620 * write-side (application) finally closes them.
2623 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2625 struct lttng_ht
*ht
;
2626 struct lttng_consumer_stream
*stream
;
2627 struct lttng_ht_iter iter
;
2629 ht
= consumer_data
.stream_per_chan_id_ht
;
2632 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2633 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2634 ht
->match_fct
, &channel
->key
,
2635 &iter
.iter
, stream
, node_channel_id
.node
) {
2637 * Protect against teardown with mutex.
2639 pthread_mutex_lock(&stream
->lock
);
2640 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2643 switch (consumer_data
.type
) {
2644 case LTTNG_CONSUMER_KERNEL
:
2646 case LTTNG_CONSUMER32_UST
:
2647 case LTTNG_CONSUMER64_UST
:
2648 if (stream
->metadata_flag
) {
2649 /* Safe and protected by the stream lock. */
2650 lttng_ustconsumer_close_metadata(stream
->chan
);
2653 * Note: a mutex is taken internally within
2654 * liblttng-ust-ctl to protect timer wakeup_fd
2655 * use from concurrent close.
2657 lttng_ustconsumer_close_stream_wakeup(stream
);
2661 ERR("Unknown consumer_data type");
2665 pthread_mutex_unlock(&stream
->lock
);
2670 static void destroy_channel_ht(struct lttng_ht
*ht
)
2672 struct lttng_ht_iter iter
;
2673 struct lttng_consumer_channel
*channel
;
2681 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2682 ret
= lttng_ht_del(ht
, &iter
);
2687 lttng_ht_destroy(ht
);
2691 * This thread polls the channel fds to detect when they are being
2692 * closed. It closes all related streams if the channel is detected as
2693 * closed. It is currently only used as a shim layer for UST because the
2694 * consumerd needs to keep the per-stream wakeup end of pipes open for
2697 void *consumer_thread_channel_poll(void *data
)
2699 int ret
, i
, pollfd
, err
= -1;
2700 uint32_t revents
, nb_fd
;
2701 struct lttng_consumer_channel
*chan
= NULL
;
2702 struct lttng_ht_iter iter
;
2703 struct lttng_ht_node_u64
*node
;
2704 struct lttng_poll_event events
;
2705 struct lttng_consumer_local_data
*ctx
= data
;
2706 struct lttng_ht
*channel_ht
;
2708 rcu_register_thread();
2710 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
2712 if (testpoint(consumerd_thread_channel
)) {
2713 goto error_testpoint
;
2716 health_code_update();
2718 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2720 /* ENOMEM at this point. Better to bail out. */
2724 DBG("Thread channel poll started");
2726 /* Size is set to 1 for the consumer_channel pipe */
2727 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2729 ERR("Poll set creation failed");
2733 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2739 DBG("Channel main loop started");
2743 health_code_update();
2744 DBG("Channel poll wait");
2745 health_poll_entry();
2746 ret
= lttng_poll_wait(&events
, -1);
2747 DBG("Channel poll return from wait with %d fd(s)",
2748 LTTNG_POLL_GETNB(&events
));
2750 DBG("Channel event caught in thread");
2752 if (errno
== EINTR
) {
2753 ERR("Poll EINTR caught");
2756 if (LTTNG_POLL_GETNB(&events
) == 0) {
2757 err
= 0; /* All is OK */
2764 /* From here, the event is a channel wait fd */
2765 for (i
= 0; i
< nb_fd
; i
++) {
2766 health_code_update();
2768 revents
= LTTNG_POLL_GETEV(&events
, i
);
2769 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2772 /* No activity for this FD (poll implementation). */
2776 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2777 if (revents
& LPOLLIN
) {
2778 enum consumer_channel_action action
;
2781 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
2784 ERR("Error reading channel pipe");
2786 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2791 case CONSUMER_CHANNEL_ADD
:
2792 DBG("Adding channel %d to poll set",
2795 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
2798 lttng_ht_add_unique_u64(channel_ht
,
2799 &chan
->wait_fd_node
);
2801 /* Add channel to the global poll events list */
2802 lttng_poll_add(&events
, chan
->wait_fd
,
2803 LPOLLERR
| LPOLLHUP
);
2805 case CONSUMER_CHANNEL_DEL
:
2808 * This command should never be called if the channel
2809 * has streams monitored by either the data or metadata
2810 * thread. The consumer only notify this thread with a
2811 * channel del. command if it receives a destroy
2812 * channel command from the session daemon that send it
2813 * if a command prior to the GET_CHANNEL failed.
2817 chan
= consumer_find_channel(key
);
2820 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
2823 lttng_poll_del(&events
, chan
->wait_fd
);
2824 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
2825 ret
= lttng_ht_del(channel_ht
, &iter
);
2828 switch (consumer_data
.type
) {
2829 case LTTNG_CONSUMER_KERNEL
:
2831 case LTTNG_CONSUMER32_UST
:
2832 case LTTNG_CONSUMER64_UST
:
2833 health_code_update();
2834 /* Destroy streams that might have been left in the stream list. */
2835 clean_channel_stream_list(chan
);
2838 ERR("Unknown consumer_data type");
2843 * Release our own refcount. Force channel deletion even if
2844 * streams were not initialized.
2846 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
2847 consumer_del_channel(chan
);
2852 case CONSUMER_CHANNEL_QUIT
:
2854 * Remove the pipe from the poll set and continue the loop
2855 * since their might be data to consume.
2857 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2860 ERR("Unknown action");
2863 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2864 DBG("Channel thread pipe hung up");
2866 * Remove the pipe from the poll set and continue the loop
2867 * since their might be data to consume.
2869 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2872 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2876 /* Handle other stream */
2882 uint64_t tmp_id
= (uint64_t) pollfd
;
2884 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
2886 node
= lttng_ht_iter_get_node_u64(&iter
);
2889 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
2892 /* Check for error event */
2893 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2894 DBG("Channel fd %d is hup|err.", pollfd
);
2896 lttng_poll_del(&events
, chan
->wait_fd
);
2897 ret
= lttng_ht_del(channel_ht
, &iter
);
2901 * This will close the wait fd for each stream associated to
2902 * this channel AND monitored by the data/metadata thread thus
2903 * will be clean by the right thread.
2905 consumer_close_channel_streams(chan
);
2907 /* Release our own refcount */
2908 if (!uatomic_sub_return(&chan
->refcount
, 1)
2909 && !uatomic_read(&chan
->nb_init_stream_left
)) {
2910 consumer_del_channel(chan
);
2913 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2918 /* Release RCU lock for the channel looked up */
2926 lttng_poll_clean(&events
);
2928 destroy_channel_ht(channel_ht
);
2931 DBG("Channel poll thread exiting");
2934 ERR("Health error occurred in %s", __func__
);
2936 health_unregister(health_consumerd
);
2937 rcu_unregister_thread();
2941 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
2942 struct pollfd
*sockpoll
, int client_socket
)
2949 ret
= lttng_consumer_poll_socket(sockpoll
);
2953 DBG("Metadata connection on client_socket");
2955 /* Blocking call, waiting for transmission */
2956 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
2957 if (ctx
->consumer_metadata_socket
< 0) {
2958 WARN("On accept metadata");
2969 * This thread listens on the consumerd socket and receives the file
2970 * descriptors from the session daemon.
2972 void *consumer_thread_sessiond_poll(void *data
)
2974 int sock
= -1, client_socket
, ret
, err
= -1;
2976 * structure to poll for incoming data on communication socket avoids
2977 * making blocking sockets.
2979 struct pollfd consumer_sockpoll
[2];
2980 struct lttng_consumer_local_data
*ctx
= data
;
2982 rcu_register_thread();
2984 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
2986 if (testpoint(consumerd_thread_sessiond
)) {
2987 goto error_testpoint
;
2990 health_code_update();
2992 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
2993 unlink(ctx
->consumer_command_sock_path
);
2994 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
2995 if (client_socket
< 0) {
2996 ERR("Cannot create command socket");
3000 ret
= lttcomm_listen_unix_sock(client_socket
);
3005 DBG("Sending ready command to lttng-sessiond");
3006 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3007 /* return < 0 on error, but == 0 is not fatal */
3009 ERR("Error sending ready command to lttng-sessiond");
3013 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3014 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3015 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3016 consumer_sockpoll
[1].fd
= client_socket
;
3017 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3019 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3027 DBG("Connection on client_socket");
3029 /* Blocking call, waiting for transmission */
3030 sock
= lttcomm_accept_unix_sock(client_socket
);
3037 * Setup metadata socket which is the second socket connection on the
3038 * command unix socket.
3040 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3049 /* This socket is not useful anymore. */
3050 ret
= close(client_socket
);
3052 PERROR("close client_socket");
3056 /* update the polling structure to poll on the established socket */
3057 consumer_sockpoll
[1].fd
= sock
;
3058 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3061 health_code_update();
3063 health_poll_entry();
3064 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3073 DBG("Incoming command on sock");
3074 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3077 * This could simply be a session daemon quitting. Don't output
3080 DBG("Communication interrupted on command socket");
3084 if (consumer_quit
) {
3085 DBG("consumer_thread_receive_fds received quit from signal");
3086 err
= 0; /* All is OK */
3089 DBG("received command on sock");
3095 DBG("Consumer thread sessiond poll exiting");
3098 * Close metadata streams since the producer is the session daemon which
3101 * NOTE: for now, this only applies to the UST tracer.
3103 lttng_consumer_close_all_metadata();
3106 * when all fds have hung up, the polling thread
3112 * Notify the data poll thread to poll back again and test the
3113 * consumer_quit state that we just set so to quit gracefully.
3115 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3117 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3119 notify_health_quit_pipe(health_quit_pipe
);
3121 /* Cleaning up possibly open sockets. */
3125 PERROR("close sock sessiond poll");
3128 if (client_socket
>= 0) {
3129 ret
= close(client_socket
);
3131 PERROR("close client_socket sessiond poll");
3138 ERR("Health error occurred in %s", __func__
);
3140 health_unregister(health_consumerd
);
3142 rcu_unregister_thread();
3146 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3147 struct lttng_consumer_local_data
*ctx
,
3148 bool locked_by_caller
)
3150 ssize_t ret
, written_bytes
;
3151 struct stream_subbuffer subbuffer
= {};
3153 if (!locked_by_caller
) {
3154 stream
->read_subbuffer_ops
.lock(stream
);
3157 if (stream
->read_subbuffer_ops
.on_wake_up
) {
3158 ret
= stream
->read_subbuffer_ops
.on_wake_up(stream
);
3164 ret
= stream
->read_subbuffer_ops
.get_next_subbuffer(stream
, &subbuffer
);
3166 if (ret
== -ENODATA
) {
3173 ret
= stream
->read_subbuffer_ops
.pre_consume_subbuffer(
3174 stream
, &subbuffer
);
3176 goto error_put_subbuf
;
3179 written_bytes
= stream
->read_subbuffer_ops
.consume_subbuffer(
3180 ctx
, stream
, &subbuffer
);
3182 * Should write subbuf_size amount of data when network streaming or
3183 * the full padded size when we are not streaming.
3185 if ((written_bytes
!= subbuffer
.info
.data
.subbuf_size
&&
3186 stream
->relayd_id
!= (uint64_t) -1ULL) ||
3187 (written_bytes
!= subbuffer
.info
.data
.padded_subbuf_size
&&
3188 stream
->relayd_id
==
3189 (uint64_t) -1ULL)) {
3191 * Display the error but continue processing to try to
3192 * release the subbuffer. This is a DBG statement
3193 * since this can happen without being a critical
3196 DBG("Failed to write to tracefile (written_bytes: %zd != padded subbuffer size: %lu, subbuffer size: %lu)",
3197 written_bytes
, subbuffer
.info
.data
.subbuf_size
,
3198 subbuffer
.info
.data
.padded_subbuf_size
);
3201 ret
= stream
->read_subbuffer_ops
.put_next_subbuffer(stream
, &subbuffer
);
3206 if (stream
->read_subbuffer_ops
.post_consume
) {
3207 ret
= stream
->read_subbuffer_ops
.post_consume(stream
, &subbuffer
, ctx
);
3213 if (stream
->read_subbuffer_ops
.on_sleep
) {
3214 stream
->read_subbuffer_ops
.on_sleep(stream
, ctx
);
3217 ret
= written_bytes
;
3219 if (!locked_by_caller
) {
3220 stream
->read_subbuffer_ops
.unlock(stream
);
3225 (void) stream
->read_subbuffer_ops
.put_next_subbuffer(stream
, &subbuffer
);
3229 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3231 switch (consumer_data
.type
) {
3232 case LTTNG_CONSUMER_KERNEL
:
3233 return lttng_kconsumer_on_recv_stream(stream
);
3234 case LTTNG_CONSUMER32_UST
:
3235 case LTTNG_CONSUMER64_UST
:
3236 return lttng_ustconsumer_on_recv_stream(stream
);
3238 ERR("Unknown consumer_data type");
3245 * Allocate and set consumer data hash tables.
3247 int lttng_consumer_init(void)
3249 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3250 if (!consumer_data
.channel_ht
) {
3254 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3255 if (!consumer_data
.relayd_ht
) {
3259 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3260 if (!consumer_data
.stream_list_ht
) {
3264 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3265 if (!consumer_data
.stream_per_chan_id_ht
) {
3269 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3274 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3286 * Process the ADD_RELAYD command receive by a consumer.
3288 * This will create a relayd socket pair and add it to the relayd hash table.
3289 * The caller MUST acquire a RCU read side lock before calling it.
3291 void consumer_add_relayd_socket(uint64_t relayd_id
, int sock_type
,
3292 struct lttng_consumer_local_data
*ctx
, int sock
,
3293 struct pollfd
*consumer_sockpoll
,
3294 struct lttcomm_relayd_sock
*relayd_sock
, uint64_t sessiond_id
,
3295 uint64_t relayd_session_id
)
3297 int fd
= -1, ret
= -1, relayd_created
= 0;
3298 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3299 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3302 assert(relayd_sock
);
3304 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", relayd_id
);
3306 /* Get relayd reference if exists. */
3307 relayd
= consumer_find_relayd(relayd_id
);
3308 if (relayd
== NULL
) {
3309 assert(sock_type
== LTTNG_STREAM_CONTROL
);
3310 /* Not found. Allocate one. */
3311 relayd
= consumer_allocate_relayd_sock_pair(relayd_id
);
3312 if (relayd
== NULL
) {
3314 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3317 relayd
->sessiond_session_id
= sessiond_id
;
3322 * This code path MUST continue to the consumer send status message to
3323 * we can notify the session daemon and continue our work without
3324 * killing everything.
3328 * relayd key should never be found for control socket.
3330 assert(sock_type
!= LTTNG_STREAM_CONTROL
);
3333 /* First send a status message before receiving the fds. */
3334 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3336 /* Somehow, the session daemon is not responding anymore. */
3337 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3338 goto error_nosignal
;
3341 /* Poll on consumer socket. */
3342 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3344 /* Needing to exit in the middle of a command: error. */
3345 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3347 goto error_nosignal
;
3350 /* Get relayd socket from session daemon */
3351 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3352 if (ret
!= sizeof(fd
)) {
3354 fd
= -1; /* Just in case it gets set with an invalid value. */
3357 * Failing to receive FDs might indicate a major problem such as
3358 * reaching a fd limit during the receive where the kernel returns a
3359 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3360 * don't take any chances and stop everything.
3362 * XXX: Feature request #558 will fix that and avoid this possible
3363 * issue when reaching the fd limit.
3365 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3366 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3370 /* Copy socket information and received FD */
3371 switch (sock_type
) {
3372 case LTTNG_STREAM_CONTROL
:
3373 /* Copy received lttcomm socket */
3374 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3375 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3376 /* Handle create_sock error. */
3378 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3382 * Close the socket created internally by
3383 * lttcomm_create_sock, so we can replace it by the one
3384 * received from sessiond.
3386 if (close(relayd
->control_sock
.sock
.fd
)) {
3390 /* Assign new file descriptor */
3391 relayd
->control_sock
.sock
.fd
= fd
;
3392 fd
= -1; /* For error path */
3393 /* Assign version values. */
3394 relayd
->control_sock
.major
= relayd_sock
->major
;
3395 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3397 relayd
->relayd_session_id
= relayd_session_id
;
3400 case LTTNG_STREAM_DATA
:
3401 /* Copy received lttcomm socket */
3402 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3403 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3404 /* Handle create_sock error. */
3406 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3410 * Close the socket created internally by
3411 * lttcomm_create_sock, so we can replace it by the one
3412 * received from sessiond.
3414 if (close(relayd
->data_sock
.sock
.fd
)) {
3418 /* Assign new file descriptor */
3419 relayd
->data_sock
.sock
.fd
= fd
;
3420 fd
= -1; /* for eventual error paths */
3421 /* Assign version values. */
3422 relayd
->data_sock
.major
= relayd_sock
->major
;
3423 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3426 ERR("Unknown relayd socket type (%d)", sock_type
);
3428 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3432 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3433 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3436 /* We successfully added the socket. Send status back. */
3437 ret
= consumer_send_status_msg(sock
, ret_code
);
3439 /* Somehow, the session daemon is not responding anymore. */
3440 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3441 goto error_nosignal
;
3445 * Add relayd socket pair to consumer data hashtable. If object already
3446 * exists or on error, the function gracefully returns.
3455 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3456 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3460 /* Close received socket if valid. */
3463 PERROR("close received socket");
3467 if (relayd_created
) {
3473 * Search for a relayd associated to the session id and return the reference.
3475 * A rcu read side lock MUST be acquire before calling this function and locked
3476 * until the relayd object is no longer necessary.
3478 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3480 struct lttng_ht_iter iter
;
3481 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3483 /* Iterate over all relayd since they are indexed by relayd_id. */
3484 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3487 * Check by sessiond id which is unique here where the relayd session
3488 * id might not be when having multiple relayd.
3490 if (relayd
->sessiond_session_id
== id
) {
3491 /* Found the relayd. There can be only one per id. */
3503 * Check if for a given session id there is still data needed to be extract
3506 * Return 1 if data is pending or else 0 meaning ready to be read.
3508 int consumer_data_pending(uint64_t id
)
3511 struct lttng_ht_iter iter
;
3512 struct lttng_ht
*ht
;
3513 struct lttng_consumer_stream
*stream
;
3514 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3515 int (*data_pending
)(struct lttng_consumer_stream
*);
3517 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3520 pthread_mutex_lock(&consumer_data
.lock
);
3522 switch (consumer_data
.type
) {
3523 case LTTNG_CONSUMER_KERNEL
:
3524 data_pending
= lttng_kconsumer_data_pending
;
3526 case LTTNG_CONSUMER32_UST
:
3527 case LTTNG_CONSUMER64_UST
:
3528 data_pending
= lttng_ustconsumer_data_pending
;
3531 ERR("Unknown consumer data type");
3535 /* Ease our life a bit */
3536 ht
= consumer_data
.stream_list_ht
;
3538 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3539 ht
->hash_fct(&id
, lttng_ht_seed
),
3541 &iter
.iter
, stream
, node_session_id
.node
) {
3542 pthread_mutex_lock(&stream
->lock
);
3545 * A removed node from the hash table indicates that the stream has
3546 * been deleted thus having a guarantee that the buffers are closed
3547 * on the consumer side. However, data can still be transmitted
3548 * over the network so don't skip the relayd check.
3550 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3552 /* Check the stream if there is data in the buffers. */
3553 ret
= data_pending(stream
);
3555 DBG("Data is pending locally on stream %" PRIu64
, stream
->key
);
3556 pthread_mutex_unlock(&stream
->lock
);
3561 pthread_mutex_unlock(&stream
->lock
);
3564 relayd
= find_relayd_by_session_id(id
);
3566 unsigned int is_data_inflight
= 0;
3568 /* Send init command for data pending. */
3569 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3570 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3571 relayd
->relayd_session_id
);
3573 /* Communication error thus the relayd so no data pending. */
3574 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3575 ERR("Relayd begin data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->id
);
3576 lttng_consumer_cleanup_relayd(relayd
);
3577 goto data_not_pending
;
3580 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3581 ht
->hash_fct(&id
, lttng_ht_seed
),
3583 &iter
.iter
, stream
, node_session_id
.node
) {
3584 if (stream
->metadata_flag
) {
3585 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3586 stream
->relayd_stream_id
);
3588 ret
= relayd_data_pending(&relayd
->control_sock
,
3589 stream
->relayd_stream_id
,
3590 stream
->next_net_seq_num
- 1);
3593 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3597 ERR("Relayd data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->id
);
3598 lttng_consumer_cleanup_relayd(relayd
);
3599 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3600 goto data_not_pending
;
3604 /* Send end command for data pending. */
3605 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3606 relayd
->relayd_session_id
, &is_data_inflight
);
3607 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3609 ERR("Relayd end data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->id
);
3610 lttng_consumer_cleanup_relayd(relayd
);
3611 goto data_not_pending
;
3613 if (is_data_inflight
) {
3614 DBG("Data is in flight on relayd %" PRIu64
, relayd
->id
);
3620 * Finding _no_ node in the hash table and no inflight data means that the
3621 * stream(s) have been removed thus data is guaranteed to be available for
3622 * analysis from the trace files.
3626 /* Data is available to be read by a viewer. */
3627 pthread_mutex_unlock(&consumer_data
.lock
);
3632 /* Data is still being extracted from buffers. */
3633 pthread_mutex_unlock(&consumer_data
.lock
);
3639 * Send a ret code status message to the sessiond daemon.
3641 * Return the sendmsg() return value.
3643 int consumer_send_status_msg(int sock
, int ret_code
)
3645 struct lttcomm_consumer_status_msg msg
;
3647 memset(&msg
, 0, sizeof(msg
));
3648 msg
.ret_code
= ret_code
;
3650 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3654 * Send a channel status message to the sessiond daemon.
3656 * Return the sendmsg() return value.
3658 int consumer_send_status_channel(int sock
,
3659 struct lttng_consumer_channel
*channel
)
3661 struct lttcomm_consumer_status_channel msg
;
3665 memset(&msg
, 0, sizeof(msg
));
3667 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3669 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3670 msg
.key
= channel
->key
;
3671 msg
.stream_count
= channel
->streams
.count
;
3674 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3677 unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos
,
3678 unsigned long produced_pos
, uint64_t nb_packets_per_stream
,
3679 uint64_t max_sb_size
)
3681 unsigned long start_pos
;
3683 if (!nb_packets_per_stream
) {
3684 return consumed_pos
; /* Grab everything */
3686 start_pos
= produced_pos
- offset_align_floor(produced_pos
, max_sb_size
);
3687 start_pos
-= max_sb_size
* nb_packets_per_stream
;
3688 if ((long) (start_pos
- consumed_pos
) < 0) {
3689 return consumed_pos
; /* Grab everything */