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_stream
*stream
,
1427 const struct lttng_buffer_view
*buffer
,
1428 unsigned long padding
)
1431 off_t orig_offset
= stream
->out_fd_offset
;
1432 /* Default is on the disk */
1433 int outfd
= stream
->out_fd
;
1434 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1435 unsigned int relayd_hang_up
= 0;
1436 const size_t subbuf_content_size
= buffer
->size
- padding
;
1439 /* RCU lock for the relayd pointer */
1442 /* Flag that the current stream if set for network streaming. */
1443 if (stream
->relayd_id
!= (uint64_t) -1ULL) {
1444 relayd
= consumer_find_relayd(stream
->relayd_id
);
1445 if (relayd
== NULL
) {
1451 /* Handle stream on the relayd if the output is on the network */
1453 unsigned long netlen
= subbuf_content_size
;
1456 * Lock the control socket for the complete duration of the function
1457 * since from this point on we will use the socket.
1459 if (stream
->metadata_flag
) {
1460 /* Metadata requires the control socket. */
1461 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1462 if (stream
->reset_metadata_flag
) {
1463 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1464 stream
->relayd_stream_id
,
1465 stream
->metadata_version
);
1470 stream
->reset_metadata_flag
= 0;
1472 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1475 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1480 /* Use the returned socket. */
1483 /* Write metadata stream id before payload */
1484 if (stream
->metadata_flag
) {
1485 ret
= write_relayd_metadata_id(outfd
, stream
, relayd
, padding
);
1492 write_len
= subbuf_content_size
;
1494 /* No streaming; we have to write the full padding. */
1495 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1496 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1498 ERR("Reset metadata file");
1501 stream
->reset_metadata_flag
= 0;
1505 * Check if we need to change the tracefile before writing the packet.
1507 if (stream
->chan
->tracefile_size
> 0 &&
1508 (stream
->tracefile_size_current
+ buffer
->size
) >
1509 stream
->chan
->tracefile_size
) {
1510 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1511 stream
->name
, stream
->chan
->tracefile_size
,
1512 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1513 stream
->out_fd
, &(stream
->tracefile_count_current
),
1516 ERR("Rotating output file");
1519 outfd
= stream
->out_fd
;
1521 if (stream
->index_file
) {
1522 lttng_index_file_put(stream
->index_file
);
1523 stream
->index_file
= lttng_index_file_create(stream
->chan
->pathname
,
1524 stream
->name
, stream
->uid
, stream
->gid
,
1525 stream
->chan
->tracefile_size
,
1526 stream
->tracefile_count_current
,
1527 CTF_INDEX_MAJOR
, CTF_INDEX_MINOR
);
1528 if (!stream
->index_file
) {
1533 /* Reset current size because we just perform a rotation. */
1534 stream
->tracefile_size_current
= 0;
1535 stream
->out_fd_offset
= 0;
1538 stream
->tracefile_size_current
+= buffer
->size
;
1539 write_len
= buffer
->size
;
1543 * This call guarantee that len or less is returned. It's impossible to
1544 * receive a ret value that is bigger than len.
1546 ret
= lttng_write(outfd
, buffer
->data
, write_len
);
1547 DBG("Consumer mmap write() ret %zd (len %zu)", ret
, write_len
);
1548 if (ret
< 0 || ((size_t) ret
!= write_len
)) {
1550 * Report error to caller if nothing was written else at least send the
1558 /* Socket operation failed. We consider the relayd dead */
1559 if (errno
== EPIPE
|| errno
== EINVAL
|| errno
== EBADF
) {
1561 * This is possible if the fd is closed on the other side
1562 * (outfd) or any write problem. It can be verbose a bit for a
1563 * normal execution if for instance the relayd is stopped
1564 * abruptly. This can happen so set this to a DBG statement.
1566 DBG("Consumer mmap write detected relayd hang up");
1568 /* Unhandled error, print it and stop function right now. */
1569 PERROR("Error in write mmap (ret %zd != write_len %zu)", ret
,
1574 stream
->output_written
+= ret
;
1576 /* This call is useless on a socket so better save a syscall. */
1578 /* This won't block, but will start writeout asynchronously */
1579 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, write_len
,
1580 SYNC_FILE_RANGE_WRITE
);
1581 stream
->out_fd_offset
+= write_len
;
1582 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1587 * This is a special case that the relayd has closed its socket. Let's
1588 * cleanup the relayd object and all associated streams.
1590 if (relayd
&& relayd_hang_up
) {
1591 ERR("Relayd hangup. Cleaning up relayd %" PRIu64
".", relayd
->id
);
1592 lttng_consumer_cleanup_relayd(relayd
);
1596 /* Unlock only if ctrl socket used */
1597 if (relayd
&& stream
->metadata_flag
) {
1598 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1606 * Splice the data from the ring buffer to the tracefile.
1608 * It must be called with the stream lock held.
1610 * Returns the number of bytes spliced.
1612 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1613 struct lttng_consumer_local_data
*ctx
,
1614 struct lttng_consumer_stream
*stream
, unsigned long len
,
1615 unsigned long padding
)
1617 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1619 off_t orig_offset
= stream
->out_fd_offset
;
1620 int fd
= stream
->wait_fd
;
1621 /* Default is on the disk */
1622 int outfd
= stream
->out_fd
;
1623 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1625 unsigned int relayd_hang_up
= 0;
1627 switch (consumer_data
.type
) {
1628 case LTTNG_CONSUMER_KERNEL
:
1630 case LTTNG_CONSUMER32_UST
:
1631 case LTTNG_CONSUMER64_UST
:
1632 /* Not supported for user space tracing */
1635 ERR("Unknown consumer_data type");
1639 /* RCU lock for the relayd pointer */
1642 /* Flag that the current stream if set for network streaming. */
1643 if (stream
->relayd_id
!= (uint64_t) -1ULL) {
1644 relayd
= consumer_find_relayd(stream
->relayd_id
);
1645 if (relayd
== NULL
) {
1650 splice_pipe
= stream
->splice_pipe
;
1652 /* Write metadata stream id before payload */
1654 unsigned long total_len
= len
;
1656 if (stream
->metadata_flag
) {
1658 * Lock the control socket for the complete duration of the function
1659 * since from this point on we will use the socket.
1661 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1663 if (stream
->reset_metadata_flag
) {
1664 ret
= relayd_reset_metadata(&relayd
->control_sock
,
1665 stream
->relayd_stream_id
,
1666 stream
->metadata_version
);
1671 stream
->reset_metadata_flag
= 0;
1673 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
, relayd
,
1681 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1684 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1690 /* Use the returned socket. */
1693 /* No streaming, we have to set the len with the full padding */
1696 if (stream
->metadata_flag
&& stream
->reset_metadata_flag
) {
1697 ret
= utils_truncate_stream_file(stream
->out_fd
, 0);
1699 ERR("Reset metadata file");
1702 stream
->reset_metadata_flag
= 0;
1705 * Check if we need to change the tracefile before writing the packet.
1707 if (stream
->chan
->tracefile_size
> 0 &&
1708 (stream
->tracefile_size_current
+ len
) >
1709 stream
->chan
->tracefile_size
) {
1710 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1711 stream
->name
, stream
->chan
->tracefile_size
,
1712 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1713 stream
->out_fd
, &(stream
->tracefile_count_current
),
1717 ERR("Rotating output file");
1720 outfd
= stream
->out_fd
;
1722 if (stream
->index_file
) {
1723 lttng_index_file_put(stream
->index_file
);
1724 stream
->index_file
= lttng_index_file_create(stream
->chan
->pathname
,
1725 stream
->name
, stream
->uid
, stream
->gid
,
1726 stream
->chan
->tracefile_size
,
1727 stream
->tracefile_count_current
,
1728 CTF_INDEX_MAJOR
, CTF_INDEX_MINOR
);
1729 if (!stream
->index_file
) {
1734 /* Reset current size because we just perform a rotation. */
1735 stream
->tracefile_size_current
= 0;
1736 stream
->out_fd_offset
= 0;
1739 stream
->tracefile_size_current
+= len
;
1743 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1744 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1745 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1746 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1747 DBG("splice chan to pipe, ret %zd", ret_splice
);
1748 if (ret_splice
< 0) {
1751 PERROR("Error in relay splice");
1755 /* Handle stream on the relayd if the output is on the network */
1756 if (relayd
&& stream
->metadata_flag
) {
1757 size_t metadata_payload_size
=
1758 sizeof(struct lttcomm_relayd_metadata_payload
);
1760 /* Update counter to fit the spliced data */
1761 ret_splice
+= metadata_payload_size
;
1762 len
+= metadata_payload_size
;
1764 * We do this so the return value can match the len passed as
1765 * argument to this function.
1767 written
-= metadata_payload_size
;
1770 /* Splice data out */
1771 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1772 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1773 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
1775 if (ret_splice
< 0) {
1780 } else if (ret_splice
> len
) {
1782 * We don't expect this code path to be executed but you never know
1783 * so this is an extra protection agains a buggy splice().
1786 written
+= ret_splice
;
1787 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice
,
1791 /* All good, update current len and continue. */
1795 /* This call is useless on a socket so better save a syscall. */
1797 /* This won't block, but will start writeout asynchronously */
1798 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1799 SYNC_FILE_RANGE_WRITE
);
1800 stream
->out_fd_offset
+= ret_splice
;
1802 stream
->output_written
+= ret_splice
;
1803 written
+= ret_splice
;
1806 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1812 * This is a special case that the relayd has closed its socket. Let's
1813 * cleanup the relayd object and all associated streams.
1815 if (relayd
&& relayd_hang_up
) {
1816 ERR("Relayd hangup. Cleaning up relayd %" PRIu64
".", relayd
->id
);
1817 lttng_consumer_cleanup_relayd(relayd
);
1818 /* Skip splice error so the consumer does not fail */
1823 /* send the appropriate error description to sessiond */
1826 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1829 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1832 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
1837 if (relayd
&& stream
->metadata_flag
) {
1838 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1846 * Take a snapshot for a specific fd
1848 * Returns 0 on success, < 0 on error
1850 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
1852 switch (consumer_data
.type
) {
1853 case LTTNG_CONSUMER_KERNEL
:
1854 return lttng_kconsumer_take_snapshot(stream
);
1855 case LTTNG_CONSUMER32_UST
:
1856 case LTTNG_CONSUMER64_UST
:
1857 return lttng_ustconsumer_take_snapshot(stream
);
1859 ERR("Unknown consumer_data type");
1866 * Get the produced position
1868 * Returns 0 on success, < 0 on error
1870 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
1873 switch (consumer_data
.type
) {
1874 case LTTNG_CONSUMER_KERNEL
:
1875 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
1876 case LTTNG_CONSUMER32_UST
:
1877 case LTTNG_CONSUMER64_UST
:
1878 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
1880 ERR("Unknown consumer_data type");
1886 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
1887 int sock
, struct pollfd
*consumer_sockpoll
)
1889 switch (consumer_data
.type
) {
1890 case LTTNG_CONSUMER_KERNEL
:
1891 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1892 case LTTNG_CONSUMER32_UST
:
1893 case LTTNG_CONSUMER64_UST
:
1894 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1896 ERR("Unknown consumer_data type");
1902 void lttng_consumer_close_all_metadata(void)
1904 switch (consumer_data
.type
) {
1905 case LTTNG_CONSUMER_KERNEL
:
1907 * The Kernel consumer has a different metadata scheme so we don't
1908 * close anything because the stream will be closed by the session
1912 case LTTNG_CONSUMER32_UST
:
1913 case LTTNG_CONSUMER64_UST
:
1915 * Close all metadata streams. The metadata hash table is passed and
1916 * this call iterates over it by closing all wakeup fd. This is safe
1917 * because at this point we are sure that the metadata producer is
1918 * either dead or blocked.
1920 lttng_ustconsumer_close_all_metadata(metadata_ht
);
1923 ERR("Unknown consumer_data type");
1929 * Clean up a metadata stream and free its memory.
1931 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
1932 struct lttng_ht
*ht
)
1934 struct lttng_consumer_channel
*free_chan
= NULL
;
1938 * This call should NEVER receive regular stream. It must always be
1939 * metadata stream and this is crucial for data structure synchronization.
1941 assert(stream
->metadata_flag
);
1943 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
1945 pthread_mutex_lock(&consumer_data
.lock
);
1946 pthread_mutex_lock(&stream
->chan
->lock
);
1947 pthread_mutex_lock(&stream
->lock
);
1948 if (stream
->chan
->metadata_cache
) {
1949 /* Only applicable to userspace consumers. */
1950 pthread_mutex_lock(&stream
->chan
->metadata_cache
->lock
);
1953 /* Remove any reference to that stream. */
1954 consumer_stream_delete(stream
, ht
);
1956 /* Close down everything including the relayd if one. */
1957 consumer_stream_close(stream
);
1958 /* Destroy tracer buffers of the stream. */
1959 consumer_stream_destroy_buffers(stream
);
1961 /* Atomically decrement channel refcount since other threads can use it. */
1962 if (!uatomic_sub_return(&stream
->chan
->refcount
, 1)
1963 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
1964 /* Go for channel deletion! */
1965 free_chan
= stream
->chan
;
1969 * Nullify the stream reference so it is not used after deletion. The
1970 * channel lock MUST be acquired before being able to check for a NULL
1973 stream
->chan
->metadata_stream
= NULL
;
1975 if (stream
->chan
->metadata_cache
) {
1976 pthread_mutex_unlock(&stream
->chan
->metadata_cache
->lock
);
1978 pthread_mutex_unlock(&stream
->lock
);
1979 pthread_mutex_unlock(&stream
->chan
->lock
);
1980 pthread_mutex_unlock(&consumer_data
.lock
);
1983 consumer_del_channel(free_chan
);
1986 consumer_stream_free(stream
);
1990 * Action done with the metadata stream when adding it to the consumer internal
1991 * data structures to handle it.
1993 int consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
1995 struct lttng_ht
*ht
= metadata_ht
;
1997 struct lttng_ht_iter iter
;
1998 struct lttng_ht_node_u64
*node
;
2003 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2005 pthread_mutex_lock(&consumer_data
.lock
);
2006 pthread_mutex_lock(&stream
->chan
->lock
);
2007 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2008 pthread_mutex_lock(&stream
->lock
);
2011 * From here, refcounts are updated so be _careful_ when returning an error
2018 * Lookup the stream just to make sure it does not exist in our internal
2019 * state. This should NEVER happen.
2021 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2022 node
= lttng_ht_iter_get_node_u64(&iter
);
2026 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2027 * in terms of destroying the associated channel, because the action that
2028 * causes the count to become 0 also causes a stream to be added. The
2029 * channel deletion will thus be triggered by the following removal of this
2032 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2033 /* Increment refcount before decrementing nb_init_stream_left */
2035 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2038 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2040 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
2041 &stream
->node_channel_id
);
2044 * Add stream to the stream_list_ht of the consumer data. No need to steal
2045 * the key since the HT does not use it and we allow to add redundant keys
2048 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2052 pthread_mutex_unlock(&stream
->lock
);
2053 pthread_mutex_unlock(&stream
->chan
->lock
);
2054 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2055 pthread_mutex_unlock(&consumer_data
.lock
);
2060 * Delete data stream that are flagged for deletion (endpoint_status).
2062 static void validate_endpoint_status_data_stream(void)
2064 struct lttng_ht_iter iter
;
2065 struct lttng_consumer_stream
*stream
;
2067 DBG("Consumer delete flagged data stream");
2070 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2071 /* Validate delete flag of the stream */
2072 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2075 /* Delete it right now */
2076 consumer_del_stream(stream
, data_ht
);
2082 * Delete metadata stream that are flagged for deletion (endpoint_status).
2084 static void validate_endpoint_status_metadata_stream(
2085 struct lttng_poll_event
*pollset
)
2087 struct lttng_ht_iter iter
;
2088 struct lttng_consumer_stream
*stream
;
2090 DBG("Consumer delete flagged metadata stream");
2095 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2096 /* Validate delete flag of the stream */
2097 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2101 * Remove from pollset so the metadata thread can continue without
2102 * blocking on a deleted stream.
2104 lttng_poll_del(pollset
, stream
->wait_fd
);
2106 /* Delete it right now */
2107 consumer_del_metadata_stream(stream
, metadata_ht
);
2113 * Thread polls on metadata file descriptor and write them on disk or on the
2116 void *consumer_thread_metadata_poll(void *data
)
2118 int ret
, i
, pollfd
, err
= -1;
2119 uint32_t revents
, nb_fd
;
2120 struct lttng_consumer_stream
*stream
= NULL
;
2121 struct lttng_ht_iter iter
;
2122 struct lttng_ht_node_u64
*node
;
2123 struct lttng_poll_event events
;
2124 struct lttng_consumer_local_data
*ctx
= data
;
2127 rcu_register_thread();
2129 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2131 if (testpoint(consumerd_thread_metadata
)) {
2132 goto error_testpoint
;
2135 health_code_update();
2137 DBG("Thread metadata poll started");
2139 /* Size is set to 1 for the consumer_metadata pipe */
2140 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2142 ERR("Poll set creation failed");
2146 ret
= lttng_poll_add(&events
,
2147 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2153 DBG("Metadata main loop started");
2157 health_code_update();
2158 health_poll_entry();
2159 DBG("Metadata poll wait");
2160 ret
= lttng_poll_wait(&events
, -1);
2161 DBG("Metadata poll return from wait with %d fd(s)",
2162 LTTNG_POLL_GETNB(&events
));
2164 DBG("Metadata event caught in thread");
2166 if (errno
== EINTR
) {
2167 ERR("Poll EINTR caught");
2170 if (LTTNG_POLL_GETNB(&events
) == 0) {
2171 err
= 0; /* All is OK */
2178 /* From here, the event is a metadata wait fd */
2179 for (i
= 0; i
< nb_fd
; i
++) {
2180 health_code_update();
2182 revents
= LTTNG_POLL_GETEV(&events
, i
);
2183 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2186 /* No activity for this FD (poll implementation). */
2190 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2191 if (revents
& LPOLLIN
) {
2194 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2195 &stream
, sizeof(stream
));
2196 if (pipe_len
< sizeof(stream
)) {
2198 PERROR("read metadata stream");
2201 * Remove the pipe from the poll set and continue the loop
2202 * since their might be data to consume.
2204 lttng_poll_del(&events
,
2205 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2206 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2210 /* A NULL stream means that the state has changed. */
2211 if (stream
== NULL
) {
2212 /* Check for deleted streams. */
2213 validate_endpoint_status_metadata_stream(&events
);
2217 DBG("Adding metadata stream %d to poll set",
2220 /* Add metadata stream to the global poll events list */
2221 lttng_poll_add(&events
, stream
->wait_fd
,
2222 LPOLLIN
| LPOLLPRI
| LPOLLHUP
);
2223 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2224 DBG("Metadata thread pipe hung up");
2226 * Remove the pipe from the poll set and continue the loop
2227 * since their might be data to consume.
2229 lttng_poll_del(&events
,
2230 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2231 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2234 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2238 /* Handle other stream */
2244 uint64_t tmp_id
= (uint64_t) pollfd
;
2246 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2248 node
= lttng_ht_iter_get_node_u64(&iter
);
2251 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2254 if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2255 /* Get the data out of the metadata file descriptor */
2256 DBG("Metadata available on fd %d", pollfd
);
2257 assert(stream
->wait_fd
== pollfd
);
2260 health_code_update();
2262 len
= ctx
->on_buffer_ready(stream
, ctx
, false);
2264 * We don't check the return value here since if we get
2265 * a negative len, it means an error occurred thus we
2266 * simply remove it from the poll set and free the
2271 /* It's ok to have an unavailable sub-buffer */
2272 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2273 /* Clean up stream from consumer and free it. */
2274 lttng_poll_del(&events
, stream
->wait_fd
);
2275 consumer_del_metadata_stream(stream
, metadata_ht
);
2277 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2278 DBG("Metadata fd %d is hup|err.", pollfd
);
2279 if (!stream
->hangup_flush_done
2280 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2281 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2282 DBG("Attempting to flush and consume the UST buffers");
2283 lttng_ustconsumer_on_stream_hangup(stream
);
2285 /* We just flushed the stream now read it. */
2287 health_code_update();
2289 len
= ctx
->on_buffer_ready(stream
, ctx
, false);
2291 * We don't check the return value here since if we get
2292 * a negative len, it means an error occurred thus we
2293 * simply remove it from the poll set and free the
2299 lttng_poll_del(&events
, stream
->wait_fd
);
2301 * This call update the channel states, closes file descriptors
2302 * and securely free the stream.
2304 consumer_del_metadata_stream(stream
, metadata_ht
);
2306 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2310 /* Release RCU lock for the stream looked up */
2318 DBG("Metadata poll thread exiting");
2320 lttng_poll_clean(&events
);
2325 ERR("Health error occurred in %s", __func__
);
2327 health_unregister(health_consumerd
);
2328 rcu_unregister_thread();
2333 * This thread polls the fds in the set to consume the data and write
2334 * it to tracefile if necessary.
2336 void *consumer_thread_data_poll(void *data
)
2338 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2339 struct pollfd
*pollfd
= NULL
;
2340 /* local view of the streams */
2341 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2342 /* local view of consumer_data.fds_count */
2344 /* Number of FDs with CONSUMER_ENDPOINT_INACTIVE but still open. */
2345 int nb_inactive_fd
= 0;
2346 struct lttng_consumer_local_data
*ctx
= data
;
2349 rcu_register_thread();
2351 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2353 if (testpoint(consumerd_thread_data
)) {
2354 goto error_testpoint
;
2357 health_code_update();
2359 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
*));
2360 if (local_stream
== NULL
) {
2361 PERROR("local_stream malloc");
2366 health_code_update();
2372 * the fds set has been updated, we need to update our
2373 * local array as well
2375 pthread_mutex_lock(&consumer_data
.lock
);
2376 if (consumer_data
.need_update
) {
2381 local_stream
= NULL
;
2384 * Allocate for all fds +1 for the consumer_data_pipe and +1 for
2387 pollfd
= zmalloc((consumer_data
.stream_count
+ 2) * sizeof(struct pollfd
));
2388 if (pollfd
== NULL
) {
2389 PERROR("pollfd malloc");
2390 pthread_mutex_unlock(&consumer_data
.lock
);
2394 local_stream
= zmalloc((consumer_data
.stream_count
+ 2) *
2395 sizeof(struct lttng_consumer_stream
*));
2396 if (local_stream
== NULL
) {
2397 PERROR("local_stream malloc");
2398 pthread_mutex_unlock(&consumer_data
.lock
);
2401 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2402 data_ht
, &nb_inactive_fd
);
2404 ERR("Error in allocating pollfd or local_outfds");
2405 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2406 pthread_mutex_unlock(&consumer_data
.lock
);
2410 consumer_data
.need_update
= 0;
2412 pthread_mutex_unlock(&consumer_data
.lock
);
2414 /* No FDs and consumer_quit, consumer_cleanup the thread */
2415 if (nb_fd
== 0 && consumer_quit
== 1 && nb_inactive_fd
== 0) {
2416 err
= 0; /* All is OK */
2419 /* poll on the array of fds */
2421 DBG("polling on %d fd", nb_fd
+ 2);
2422 health_poll_entry();
2423 num_rdy
= poll(pollfd
, nb_fd
+ 2, -1);
2425 DBG("poll num_rdy : %d", num_rdy
);
2426 if (num_rdy
== -1) {
2428 * Restart interrupted system call.
2430 if (errno
== EINTR
) {
2433 PERROR("Poll error");
2434 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2436 } else if (num_rdy
== 0) {
2437 DBG("Polling thread timed out");
2442 * If the consumer_data_pipe triggered poll go directly to the
2443 * beginning of the loop to update the array. We want to prioritize
2444 * array update over low-priority reads.
2446 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2447 ssize_t pipe_readlen
;
2449 DBG("consumer_data_pipe wake up");
2450 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2451 &new_stream
, sizeof(new_stream
));
2452 if (pipe_readlen
< sizeof(new_stream
)) {
2453 PERROR("Consumer data pipe");
2454 /* Continue so we can at least handle the current stream(s). */
2459 * If the stream is NULL, just ignore it. It's also possible that
2460 * the sessiond poll thread changed the consumer_quit state and is
2461 * waking us up to test it.
2463 if (new_stream
== NULL
) {
2464 validate_endpoint_status_data_stream();
2468 /* Continue to update the local streams and handle prio ones */
2472 /* Handle wakeup pipe. */
2473 if (pollfd
[nb_fd
+ 1].revents
& (POLLIN
| POLLPRI
)) {
2475 ssize_t pipe_readlen
;
2477 pipe_readlen
= lttng_pipe_read(ctx
->consumer_wakeup_pipe
, &dummy
,
2479 if (pipe_readlen
< 0) {
2480 PERROR("Consumer data wakeup pipe");
2482 /* We've been awakened to handle stream(s). */
2483 ctx
->has_wakeup
= 0;
2486 /* Take care of high priority channels first. */
2487 for (i
= 0; i
< nb_fd
; i
++) {
2488 health_code_update();
2490 if (local_stream
[i
] == NULL
) {
2493 if (pollfd
[i
].revents
& POLLPRI
) {
2494 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2496 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
, false);
2497 /* it's ok to have an unavailable sub-buffer */
2498 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2499 /* Clean the stream and free it. */
2500 consumer_del_stream(local_stream
[i
], data_ht
);
2501 local_stream
[i
] = NULL
;
2502 } else if (len
> 0) {
2503 local_stream
[i
]->data_read
= 1;
2509 * If we read high prio channel in this loop, try again
2510 * for more high prio data.
2516 /* Take care of low priority channels. */
2517 for (i
= 0; i
< nb_fd
; i
++) {
2518 health_code_update();
2520 if (local_stream
[i
] == NULL
) {
2523 if ((pollfd
[i
].revents
& POLLIN
) ||
2524 local_stream
[i
]->hangup_flush_done
||
2525 local_stream
[i
]->has_data
) {
2526 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2527 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
, false);
2528 /* it's ok to have an unavailable sub-buffer */
2529 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2530 /* Clean the stream and free it. */
2531 consumer_del_stream(local_stream
[i
], data_ht
);
2532 local_stream
[i
] = NULL
;
2533 } else if (len
> 0) {
2534 local_stream
[i
]->data_read
= 1;
2539 /* Handle hangup and errors */
2540 for (i
= 0; i
< nb_fd
; i
++) {
2541 health_code_update();
2543 if (local_stream
[i
] == NULL
) {
2546 if (!local_stream
[i
]->hangup_flush_done
2547 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2548 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2549 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2550 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2552 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2553 /* Attempt read again, for the data we just flushed. */
2554 local_stream
[i
]->data_read
= 1;
2557 * If the poll flag is HUP/ERR/NVAL and we have
2558 * read no data in this pass, we can remove the
2559 * stream from its hash table.
2561 if ((pollfd
[i
].revents
& POLLHUP
)) {
2562 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2563 if (!local_stream
[i
]->data_read
) {
2564 consumer_del_stream(local_stream
[i
], data_ht
);
2565 local_stream
[i
] = NULL
;
2568 } else if (pollfd
[i
].revents
& POLLERR
) {
2569 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2570 if (!local_stream
[i
]->data_read
) {
2571 consumer_del_stream(local_stream
[i
], data_ht
);
2572 local_stream
[i
] = NULL
;
2575 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2576 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2577 if (!local_stream
[i
]->data_read
) {
2578 consumer_del_stream(local_stream
[i
], data_ht
);
2579 local_stream
[i
] = NULL
;
2583 if (local_stream
[i
] != NULL
) {
2584 local_stream
[i
]->data_read
= 0;
2591 DBG("polling thread exiting");
2596 * Close the write side of the pipe so epoll_wait() in
2597 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2598 * read side of the pipe. If we close them both, epoll_wait strangely does
2599 * not return and could create a endless wait period if the pipe is the
2600 * only tracked fd in the poll set. The thread will take care of closing
2603 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2608 ERR("Health error occurred in %s", __func__
);
2610 health_unregister(health_consumerd
);
2612 rcu_unregister_thread();
2617 * Close wake-up end of each stream belonging to the channel. This will
2618 * allow the poll() on the stream read-side to detect when the
2619 * write-side (application) finally closes them.
2622 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2624 struct lttng_ht
*ht
;
2625 struct lttng_consumer_stream
*stream
;
2626 struct lttng_ht_iter iter
;
2628 ht
= consumer_data
.stream_per_chan_id_ht
;
2631 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2632 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2633 ht
->match_fct
, &channel
->key
,
2634 &iter
.iter
, stream
, node_channel_id
.node
) {
2636 * Protect against teardown with mutex.
2638 pthread_mutex_lock(&stream
->lock
);
2639 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2642 switch (consumer_data
.type
) {
2643 case LTTNG_CONSUMER_KERNEL
:
2645 case LTTNG_CONSUMER32_UST
:
2646 case LTTNG_CONSUMER64_UST
:
2647 if (stream
->metadata_flag
) {
2648 /* Safe and protected by the stream lock. */
2649 lttng_ustconsumer_close_metadata(stream
->chan
);
2652 * Note: a mutex is taken internally within
2653 * liblttng-ust-ctl to protect timer wakeup_fd
2654 * use from concurrent close.
2656 lttng_ustconsumer_close_stream_wakeup(stream
);
2660 ERR("Unknown consumer_data type");
2664 pthread_mutex_unlock(&stream
->lock
);
2669 static void destroy_channel_ht(struct lttng_ht
*ht
)
2671 struct lttng_ht_iter iter
;
2672 struct lttng_consumer_channel
*channel
;
2680 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2681 ret
= lttng_ht_del(ht
, &iter
);
2686 lttng_ht_destroy(ht
);
2690 * This thread polls the channel fds to detect when they are being
2691 * closed. It closes all related streams if the channel is detected as
2692 * closed. It is currently only used as a shim layer for UST because the
2693 * consumerd needs to keep the per-stream wakeup end of pipes open for
2696 void *consumer_thread_channel_poll(void *data
)
2698 int ret
, i
, pollfd
, err
= -1;
2699 uint32_t revents
, nb_fd
;
2700 struct lttng_consumer_channel
*chan
= NULL
;
2701 struct lttng_ht_iter iter
;
2702 struct lttng_ht_node_u64
*node
;
2703 struct lttng_poll_event events
;
2704 struct lttng_consumer_local_data
*ctx
= data
;
2705 struct lttng_ht
*channel_ht
;
2707 rcu_register_thread();
2709 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
2711 if (testpoint(consumerd_thread_channel
)) {
2712 goto error_testpoint
;
2715 health_code_update();
2717 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2719 /* ENOMEM at this point. Better to bail out. */
2723 DBG("Thread channel poll started");
2725 /* Size is set to 1 for the consumer_channel pipe */
2726 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2728 ERR("Poll set creation failed");
2732 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2738 DBG("Channel main loop started");
2742 health_code_update();
2743 DBG("Channel poll wait");
2744 health_poll_entry();
2745 ret
= lttng_poll_wait(&events
, -1);
2746 DBG("Channel poll return from wait with %d fd(s)",
2747 LTTNG_POLL_GETNB(&events
));
2749 DBG("Channel event caught in thread");
2751 if (errno
== EINTR
) {
2752 ERR("Poll EINTR caught");
2755 if (LTTNG_POLL_GETNB(&events
) == 0) {
2756 err
= 0; /* All is OK */
2763 /* From here, the event is a channel wait fd */
2764 for (i
= 0; i
< nb_fd
; i
++) {
2765 health_code_update();
2767 revents
= LTTNG_POLL_GETEV(&events
, i
);
2768 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2771 /* No activity for this FD (poll implementation). */
2775 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2776 if (revents
& LPOLLIN
) {
2777 enum consumer_channel_action action
;
2780 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
2783 ERR("Error reading channel pipe");
2785 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2790 case CONSUMER_CHANNEL_ADD
:
2791 DBG("Adding channel %d to poll set",
2794 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
2797 lttng_ht_add_unique_u64(channel_ht
,
2798 &chan
->wait_fd_node
);
2800 /* Add channel to the global poll events list */
2801 lttng_poll_add(&events
, chan
->wait_fd
,
2802 LPOLLERR
| LPOLLHUP
);
2804 case CONSUMER_CHANNEL_DEL
:
2807 * This command should never be called if the channel
2808 * has streams monitored by either the data or metadata
2809 * thread. The consumer only notify this thread with a
2810 * channel del. command if it receives a destroy
2811 * channel command from the session daemon that send it
2812 * if a command prior to the GET_CHANNEL failed.
2816 chan
= consumer_find_channel(key
);
2819 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
2822 lttng_poll_del(&events
, chan
->wait_fd
);
2823 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
2824 ret
= lttng_ht_del(channel_ht
, &iter
);
2827 switch (consumer_data
.type
) {
2828 case LTTNG_CONSUMER_KERNEL
:
2830 case LTTNG_CONSUMER32_UST
:
2831 case LTTNG_CONSUMER64_UST
:
2832 health_code_update();
2833 /* Destroy streams that might have been left in the stream list. */
2834 clean_channel_stream_list(chan
);
2837 ERR("Unknown consumer_data type");
2842 * Release our own refcount. Force channel deletion even if
2843 * streams were not initialized.
2845 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
2846 consumer_del_channel(chan
);
2851 case CONSUMER_CHANNEL_QUIT
:
2853 * Remove the pipe from the poll set and continue the loop
2854 * since their might be data to consume.
2856 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2859 ERR("Unknown action");
2862 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2863 DBG("Channel thread pipe hung up");
2865 * Remove the pipe from the poll set and continue the loop
2866 * since their might be data to consume.
2868 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2871 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2875 /* Handle other stream */
2881 uint64_t tmp_id
= (uint64_t) pollfd
;
2883 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
2885 node
= lttng_ht_iter_get_node_u64(&iter
);
2888 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
2891 /* Check for error event */
2892 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2893 DBG("Channel fd %d is hup|err.", pollfd
);
2895 lttng_poll_del(&events
, chan
->wait_fd
);
2896 ret
= lttng_ht_del(channel_ht
, &iter
);
2900 * This will close the wait fd for each stream associated to
2901 * this channel AND monitored by the data/metadata thread thus
2902 * will be clean by the right thread.
2904 consumer_close_channel_streams(chan
);
2906 /* Release our own refcount */
2907 if (!uatomic_sub_return(&chan
->refcount
, 1)
2908 && !uatomic_read(&chan
->nb_init_stream_left
)) {
2909 consumer_del_channel(chan
);
2912 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2917 /* Release RCU lock for the channel looked up */
2925 lttng_poll_clean(&events
);
2927 destroy_channel_ht(channel_ht
);
2930 DBG("Channel poll thread exiting");
2933 ERR("Health error occurred in %s", __func__
);
2935 health_unregister(health_consumerd
);
2936 rcu_unregister_thread();
2940 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
2941 struct pollfd
*sockpoll
, int client_socket
)
2948 ret
= lttng_consumer_poll_socket(sockpoll
);
2952 DBG("Metadata connection on client_socket");
2954 /* Blocking call, waiting for transmission */
2955 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
2956 if (ctx
->consumer_metadata_socket
< 0) {
2957 WARN("On accept metadata");
2968 * This thread listens on the consumerd socket and receives the file
2969 * descriptors from the session daemon.
2971 void *consumer_thread_sessiond_poll(void *data
)
2973 int sock
= -1, client_socket
, ret
, err
= -1;
2975 * structure to poll for incoming data on communication socket avoids
2976 * making blocking sockets.
2978 struct pollfd consumer_sockpoll
[2];
2979 struct lttng_consumer_local_data
*ctx
= data
;
2981 rcu_register_thread();
2983 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
2985 if (testpoint(consumerd_thread_sessiond
)) {
2986 goto error_testpoint
;
2989 health_code_update();
2991 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
2992 unlink(ctx
->consumer_command_sock_path
);
2993 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
2994 if (client_socket
< 0) {
2995 ERR("Cannot create command socket");
2999 ret
= lttcomm_listen_unix_sock(client_socket
);
3004 DBG("Sending ready command to lttng-sessiond");
3005 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3006 /* return < 0 on error, but == 0 is not fatal */
3008 ERR("Error sending ready command to lttng-sessiond");
3012 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3013 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3014 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3015 consumer_sockpoll
[1].fd
= client_socket
;
3016 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3018 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3026 DBG("Connection on client_socket");
3028 /* Blocking call, waiting for transmission */
3029 sock
= lttcomm_accept_unix_sock(client_socket
);
3036 * Setup metadata socket which is the second socket connection on the
3037 * command unix socket.
3039 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3048 /* This socket is not useful anymore. */
3049 ret
= close(client_socket
);
3051 PERROR("close client_socket");
3055 /* update the polling structure to poll on the established socket */
3056 consumer_sockpoll
[1].fd
= sock
;
3057 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3060 health_code_update();
3062 health_poll_entry();
3063 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3072 DBG("Incoming command on sock");
3073 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3076 * This could simply be a session daemon quitting. Don't output
3079 DBG("Communication interrupted on command socket");
3083 if (consumer_quit
) {
3084 DBG("consumer_thread_receive_fds received quit from signal");
3085 err
= 0; /* All is OK */
3088 DBG("received command on sock");
3094 DBG("Consumer thread sessiond poll exiting");
3097 * Close metadata streams since the producer is the session daemon which
3100 * NOTE: for now, this only applies to the UST tracer.
3102 lttng_consumer_close_all_metadata();
3105 * when all fds have hung up, the polling thread
3111 * Notify the data poll thread to poll back again and test the
3112 * consumer_quit state that we just set so to quit gracefully.
3114 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3116 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3118 notify_health_quit_pipe(health_quit_pipe
);
3120 /* Cleaning up possibly open sockets. */
3124 PERROR("close sock sessiond poll");
3127 if (client_socket
>= 0) {
3128 ret
= close(client_socket
);
3130 PERROR("close client_socket sessiond poll");
3137 ERR("Health error occurred in %s", __func__
);
3139 health_unregister(health_consumerd
);
3141 rcu_unregister_thread();
3145 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3146 struct lttng_consumer_local_data
*ctx
,
3147 bool locked_by_caller
)
3149 ssize_t ret
, written_bytes
;
3150 struct stream_subbuffer subbuffer
= {};
3152 if (!locked_by_caller
) {
3153 stream
->read_subbuffer_ops
.lock(stream
);
3156 if (stream
->read_subbuffer_ops
.on_wake_up
) {
3157 ret
= stream
->read_subbuffer_ops
.on_wake_up(stream
);
3163 ret
= stream
->read_subbuffer_ops
.get_next_subbuffer(stream
, &subbuffer
);
3165 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
);
3214 if (stream
->read_subbuffer_ops
.on_sleep
) {
3215 stream
->read_subbuffer_ops
.on_sleep(stream
, ctx
);
3218 ret
= written_bytes
;
3220 if (!locked_by_caller
) {
3221 stream
->read_subbuffer_ops
.unlock(stream
);
3226 (void) stream
->read_subbuffer_ops
.put_next_subbuffer(stream
, &subbuffer
);
3230 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3232 switch (consumer_data
.type
) {
3233 case LTTNG_CONSUMER_KERNEL
:
3234 return lttng_kconsumer_on_recv_stream(stream
);
3235 case LTTNG_CONSUMER32_UST
:
3236 case LTTNG_CONSUMER64_UST
:
3237 return lttng_ustconsumer_on_recv_stream(stream
);
3239 ERR("Unknown consumer_data type");
3246 * Allocate and set consumer data hash tables.
3248 int lttng_consumer_init(void)
3250 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3251 if (!consumer_data
.channel_ht
) {
3255 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3256 if (!consumer_data
.relayd_ht
) {
3260 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3261 if (!consumer_data
.stream_list_ht
) {
3265 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3266 if (!consumer_data
.stream_per_chan_id_ht
) {
3270 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3275 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3287 * Process the ADD_RELAYD command receive by a consumer.
3289 * This will create a relayd socket pair and add it to the relayd hash table.
3290 * The caller MUST acquire a RCU read side lock before calling it.
3292 void consumer_add_relayd_socket(uint64_t relayd_id
, int sock_type
,
3293 struct lttng_consumer_local_data
*ctx
, int sock
,
3294 struct pollfd
*consumer_sockpoll
,
3295 struct lttcomm_relayd_sock
*relayd_sock
, uint64_t sessiond_id
,
3296 uint64_t relayd_session_id
)
3298 int fd
= -1, ret
= -1, relayd_created
= 0;
3299 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3300 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3303 assert(relayd_sock
);
3305 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", relayd_id
);
3307 /* Get relayd reference if exists. */
3308 relayd
= consumer_find_relayd(relayd_id
);
3309 if (relayd
== NULL
) {
3310 assert(sock_type
== LTTNG_STREAM_CONTROL
);
3311 /* Not found. Allocate one. */
3312 relayd
= consumer_allocate_relayd_sock_pair(relayd_id
);
3313 if (relayd
== NULL
) {
3315 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3318 relayd
->sessiond_session_id
= sessiond_id
;
3323 * This code path MUST continue to the consumer send status message to
3324 * we can notify the session daemon and continue our work without
3325 * killing everything.
3329 * relayd key should never be found for control socket.
3331 assert(sock_type
!= LTTNG_STREAM_CONTROL
);
3334 /* First send a status message before receiving the fds. */
3335 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3337 /* Somehow, the session daemon is not responding anymore. */
3338 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3339 goto error_nosignal
;
3342 /* Poll on consumer socket. */
3343 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3345 /* Needing to exit in the middle of a command: error. */
3346 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3348 goto error_nosignal
;
3351 /* Get relayd socket from session daemon */
3352 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3353 if (ret
!= sizeof(fd
)) {
3355 fd
= -1; /* Just in case it gets set with an invalid value. */
3358 * Failing to receive FDs might indicate a major problem such as
3359 * reaching a fd limit during the receive where the kernel returns a
3360 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3361 * don't take any chances and stop everything.
3363 * XXX: Feature request #558 will fix that and avoid this possible
3364 * issue when reaching the fd limit.
3366 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3367 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3371 /* Copy socket information and received FD */
3372 switch (sock_type
) {
3373 case LTTNG_STREAM_CONTROL
:
3374 /* Copy received lttcomm socket */
3375 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3376 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3377 /* Handle create_sock error. */
3379 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3383 * Close the socket created internally by
3384 * lttcomm_create_sock, so we can replace it by the one
3385 * received from sessiond.
3387 if (close(relayd
->control_sock
.sock
.fd
)) {
3391 /* Assign new file descriptor */
3392 relayd
->control_sock
.sock
.fd
= fd
;
3393 fd
= -1; /* For error path */
3394 /* Assign version values. */
3395 relayd
->control_sock
.major
= relayd_sock
->major
;
3396 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3398 relayd
->relayd_session_id
= relayd_session_id
;
3401 case LTTNG_STREAM_DATA
:
3402 /* Copy received lttcomm socket */
3403 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3404 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3405 /* Handle create_sock error. */
3407 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3411 * Close the socket created internally by
3412 * lttcomm_create_sock, so we can replace it by the one
3413 * received from sessiond.
3415 if (close(relayd
->data_sock
.sock
.fd
)) {
3419 /* Assign new file descriptor */
3420 relayd
->data_sock
.sock
.fd
= fd
;
3421 fd
= -1; /* for eventual error paths */
3422 /* Assign version values. */
3423 relayd
->data_sock
.major
= relayd_sock
->major
;
3424 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3427 ERR("Unknown relayd socket type (%d)", sock_type
);
3429 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3433 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3434 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3437 /* We successfully added the socket. Send status back. */
3438 ret
= consumer_send_status_msg(sock
, ret_code
);
3440 /* Somehow, the session daemon is not responding anymore. */
3441 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3442 goto error_nosignal
;
3446 * Add relayd socket pair to consumer data hashtable. If object already
3447 * exists or on error, the function gracefully returns.
3456 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3457 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3461 /* Close received socket if valid. */
3464 PERROR("close received socket");
3468 if (relayd_created
) {
3474 * Search for a relayd associated to the session id and return the reference.
3476 * A rcu read side lock MUST be acquire before calling this function and locked
3477 * until the relayd object is no longer necessary.
3479 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3481 struct lttng_ht_iter iter
;
3482 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3484 /* Iterate over all relayd since they are indexed by relayd_id. */
3485 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3488 * Check by sessiond id which is unique here where the relayd session
3489 * id might not be when having multiple relayd.
3491 if (relayd
->sessiond_session_id
== id
) {
3492 /* Found the relayd. There can be only one per id. */
3504 * Check if for a given session id there is still data needed to be extract
3507 * Return 1 if data is pending or else 0 meaning ready to be read.
3509 int consumer_data_pending(uint64_t id
)
3512 struct lttng_ht_iter iter
;
3513 struct lttng_ht
*ht
;
3514 struct lttng_consumer_stream
*stream
;
3515 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3516 int (*data_pending
)(struct lttng_consumer_stream
*);
3518 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3521 pthread_mutex_lock(&consumer_data
.lock
);
3523 switch (consumer_data
.type
) {
3524 case LTTNG_CONSUMER_KERNEL
:
3525 data_pending
= lttng_kconsumer_data_pending
;
3527 case LTTNG_CONSUMER32_UST
:
3528 case LTTNG_CONSUMER64_UST
:
3529 data_pending
= lttng_ustconsumer_data_pending
;
3532 ERR("Unknown consumer data type");
3536 /* Ease our life a bit */
3537 ht
= consumer_data
.stream_list_ht
;
3539 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3540 ht
->hash_fct(&id
, lttng_ht_seed
),
3542 &iter
.iter
, stream
, node_session_id
.node
) {
3543 pthread_mutex_lock(&stream
->lock
);
3546 * A removed node from the hash table indicates that the stream has
3547 * been deleted thus having a guarantee that the buffers are closed
3548 * on the consumer side. However, data can still be transmitted
3549 * over the network so don't skip the relayd check.
3551 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3553 /* Check the stream if there is data in the buffers. */
3554 ret
= data_pending(stream
);
3556 DBG("Data is pending locally on stream %" PRIu64
, stream
->key
);
3557 pthread_mutex_unlock(&stream
->lock
);
3562 pthread_mutex_unlock(&stream
->lock
);
3565 relayd
= find_relayd_by_session_id(id
);
3567 unsigned int is_data_inflight
= 0;
3569 /* Send init command for data pending. */
3570 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3571 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3572 relayd
->relayd_session_id
);
3574 /* Communication error thus the relayd so no data pending. */
3575 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3576 ERR("Relayd begin data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->id
);
3577 lttng_consumer_cleanup_relayd(relayd
);
3578 goto data_not_pending
;
3581 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3582 ht
->hash_fct(&id
, lttng_ht_seed
),
3584 &iter
.iter
, stream
, node_session_id
.node
) {
3585 if (stream
->metadata_flag
) {
3586 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3587 stream
->relayd_stream_id
);
3589 ret
= relayd_data_pending(&relayd
->control_sock
,
3590 stream
->relayd_stream_id
,
3591 stream
->next_net_seq_num
- 1);
3594 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3598 ERR("Relayd data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->id
);
3599 lttng_consumer_cleanup_relayd(relayd
);
3600 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3601 goto data_not_pending
;
3605 /* Send end command for data pending. */
3606 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3607 relayd
->relayd_session_id
, &is_data_inflight
);
3608 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3610 ERR("Relayd end data pending failed. Cleaning up relayd %" PRIu64
".", relayd
->id
);
3611 lttng_consumer_cleanup_relayd(relayd
);
3612 goto data_not_pending
;
3614 if (is_data_inflight
) {
3615 DBG("Data is in flight on relayd %" PRIu64
, relayd
->id
);
3621 * Finding _no_ node in the hash table and no inflight data means that the
3622 * stream(s) have been removed thus data is guaranteed to be available for
3623 * analysis from the trace files.
3627 /* Data is available to be read by a viewer. */
3628 pthread_mutex_unlock(&consumer_data
.lock
);
3633 /* Data is still being extracted from buffers. */
3634 pthread_mutex_unlock(&consumer_data
.lock
);
3640 * Send a ret code status message to the sessiond daemon.
3642 * Return the sendmsg() return value.
3644 int consumer_send_status_msg(int sock
, int ret_code
)
3646 struct lttcomm_consumer_status_msg msg
;
3648 memset(&msg
, 0, sizeof(msg
));
3649 msg
.ret_code
= ret_code
;
3651 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3655 * Send a channel status message to the sessiond daemon.
3657 * Return the sendmsg() return value.
3659 int consumer_send_status_channel(int sock
,
3660 struct lttng_consumer_channel
*channel
)
3662 struct lttcomm_consumer_status_channel msg
;
3666 memset(&msg
, 0, sizeof(msg
));
3668 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3670 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3671 msg
.key
= channel
->key
;
3672 msg
.stream_count
= channel
->streams
.count
;
3675 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3678 unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos
,
3679 unsigned long produced_pos
, uint64_t nb_packets_per_stream
,
3680 uint64_t max_sb_size
)
3682 unsigned long start_pos
;
3684 if (!nb_packets_per_stream
) {
3685 return consumed_pos
; /* Grab everything */
3687 start_pos
= produced_pos
- offset_align_floor(produced_pos
, max_sb_size
);
3688 start_pos
-= max_sb_size
* nb_packets_per_stream
;
3689 if ((long) (start_pos
- consumed_pos
) < 0) {
3690 return consumed_pos
; /* Grab everything */