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.
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-timer.h>
49 #include "consumer-stream.h"
50 #include "consumer-testpoint.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
);
328 * Destroy and free relayd socket pair object.
330 void consumer_destroy_relayd(struct consumer_relayd_sock_pair
*relayd
)
333 struct lttng_ht_iter iter
;
335 if (relayd
== NULL
) {
339 DBG("Consumer destroy and close relayd socket pair");
341 iter
.iter
.node
= &relayd
->node
.node
;
342 ret
= lttng_ht_del(consumer_data
.relayd_ht
, &iter
);
344 /* We assume the relayd is being or is destroyed */
348 /* RCU free() call */
349 call_rcu(&relayd
->node
.head
, free_relayd_rcu
);
353 * Remove a channel from the global list protected by a mutex. This function is
354 * also responsible for freeing its data structures.
356 void consumer_del_channel(struct lttng_consumer_channel
*channel
)
359 struct lttng_ht_iter iter
;
361 DBG("Consumer delete channel key %" PRIu64
, channel
->key
);
363 pthread_mutex_lock(&consumer_data
.lock
);
364 pthread_mutex_lock(&channel
->lock
);
366 /* Destroy streams that might have been left in the stream list. */
367 clean_channel_stream_list(channel
);
369 if (channel
->live_timer_enabled
== 1) {
370 consumer_timer_live_stop(channel
);
373 switch (consumer_data
.type
) {
374 case LTTNG_CONSUMER_KERNEL
:
376 case LTTNG_CONSUMER32_UST
:
377 case LTTNG_CONSUMER64_UST
:
378 lttng_ustconsumer_del_channel(channel
);
381 ERR("Unknown consumer_data type");
387 iter
.iter
.node
= &channel
->node
.node
;
388 ret
= lttng_ht_del(consumer_data
.channel_ht
, &iter
);
392 call_rcu(&channel
->node
.head
, free_channel_rcu
);
394 pthread_mutex_unlock(&channel
->lock
);
395 pthread_mutex_unlock(&consumer_data
.lock
);
399 * Iterate over the relayd hash table and destroy each element. Finally,
400 * destroy the whole hash table.
402 static void cleanup_relayd_ht(void)
404 struct lttng_ht_iter iter
;
405 struct consumer_relayd_sock_pair
*relayd
;
409 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
411 consumer_destroy_relayd(relayd
);
416 lttng_ht_destroy(consumer_data
.relayd_ht
);
420 * Update the end point status of all streams having the given network sequence
421 * index (relayd index).
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 net_seq_idx
,
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
, net_seq_idx
);
436 /* Let's begin with metadata */
437 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
438 if (stream
->net_seq_idx
== net_seq_idx
) {
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
->net_seq_idx
== net_seq_idx
) {
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 static void cleanup_relayd(struct consumer_relayd_sock_pair
*relayd
,
463 struct lttng_consumer_local_data
*ctx
)
469 DBG("Cleaning up relayd sockets");
471 /* Save the net sequence index before destroying the object */
472 netidx
= relayd
->net_seq_idx
;
475 * Delete the relayd from the relayd hash table, close the sockets and free
476 * the object in a RCU call.
478 consumer_destroy_relayd(relayd
);
480 /* Set inactive endpoint to all streams */
481 update_endpoint_status_by_netidx(netidx
, CONSUMER_ENDPOINT_INACTIVE
);
484 * With a local data context, notify the threads that the streams' state
485 * have changed. The write() action on the pipe acts as an "implicit"
486 * memory barrier ordering the updates of the end point status from the
487 * read of this status which happens AFTER receiving this notify.
490 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
491 notify_thread_lttng_pipe(ctx
->consumer_metadata_pipe
);
496 * Flag a relayd socket pair for destruction. Destroy it if the refcount
499 * RCU read side lock MUST be aquired before calling this function.
501 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair
*relayd
)
505 /* Set destroy flag for this object */
506 uatomic_set(&relayd
->destroy_flag
, 1);
508 /* Destroy the relayd if refcount is 0 */
509 if (uatomic_read(&relayd
->refcount
) == 0) {
510 consumer_destroy_relayd(relayd
);
515 * Completly destroy stream from every visiable data structure and the given
518 * One this call returns, the stream object is not longer usable nor visible.
520 void consumer_del_stream(struct lttng_consumer_stream
*stream
,
523 consumer_stream_destroy(stream
, ht
);
527 * XXX naming of del vs destroy is all mixed up.
529 void consumer_del_stream_for_data(struct lttng_consumer_stream
*stream
)
531 consumer_stream_destroy(stream
, data_ht
);
534 void consumer_del_stream_for_metadata(struct lttng_consumer_stream
*stream
)
536 consumer_stream_destroy(stream
, metadata_ht
);
539 struct lttng_consumer_stream
*consumer_allocate_stream(uint64_t channel_key
,
541 enum lttng_consumer_stream_state state
,
542 const char *channel_name
,
549 enum consumer_channel_type type
,
550 unsigned int monitor
)
553 struct lttng_consumer_stream
*stream
;
555 stream
= zmalloc(sizeof(*stream
));
556 if (stream
== NULL
) {
557 PERROR("malloc struct lttng_consumer_stream");
564 stream
->key
= stream_key
;
566 stream
->out_fd_offset
= 0;
567 stream
->output_written
= 0;
568 stream
->state
= state
;
571 stream
->net_seq_idx
= relayd_id
;
572 stream
->session_id
= session_id
;
573 stream
->monitor
= monitor
;
574 stream
->endpoint_status
= CONSUMER_ENDPOINT_ACTIVE
;
575 stream
->index_fd
= -1;
576 pthread_mutex_init(&stream
->lock
, NULL
);
577 pthread_mutex_init(&stream
->metadata_timer_lock
, NULL
);
579 /* If channel is the metadata, flag this stream as metadata. */
580 if (type
== CONSUMER_CHANNEL_TYPE_METADATA
) {
581 stream
->metadata_flag
= 1;
582 /* Metadata is flat out. */
583 strncpy(stream
->name
, DEFAULT_METADATA_NAME
, sizeof(stream
->name
));
584 /* Live rendez-vous point. */
585 pthread_cond_init(&stream
->metadata_rdv
, NULL
);
586 pthread_mutex_init(&stream
->metadata_rdv_lock
, NULL
);
588 /* Format stream name to <channel_name>_<cpu_number> */
589 ret
= snprintf(stream
->name
, sizeof(stream
->name
), "%s_%d",
592 PERROR("snprintf stream name");
597 /* Key is always the wait_fd for streams. */
598 lttng_ht_node_init_u64(&stream
->node
, stream
->key
);
600 /* Init node per channel id key */
601 lttng_ht_node_init_u64(&stream
->node_channel_id
, channel_key
);
603 /* Init session id node with the stream session id */
604 lttng_ht_node_init_u64(&stream
->node_session_id
, stream
->session_id
);
606 DBG3("Allocated stream %s (key %" PRIu64
", chan_key %" PRIu64
607 " relayd_id %" PRIu64
", session_id %" PRIu64
,
608 stream
->name
, stream
->key
, channel_key
,
609 stream
->net_seq_idx
, stream
->session_id
);
625 * Add a stream to the global list protected by a mutex.
627 int consumer_add_data_stream(struct lttng_consumer_stream
*stream
)
629 struct lttng_ht
*ht
= data_ht
;
635 DBG3("Adding consumer stream %" PRIu64
, stream
->key
);
637 pthread_mutex_lock(&consumer_data
.lock
);
638 pthread_mutex_lock(&stream
->chan
->lock
);
639 pthread_mutex_lock(&stream
->chan
->timer_lock
);
640 pthread_mutex_lock(&stream
->lock
);
643 /* Steal stream identifier to avoid having streams with the same key */
644 steal_stream_key(stream
->key
, ht
);
646 lttng_ht_add_unique_u64(ht
, &stream
->node
);
648 lttng_ht_add_u64(consumer_data
.stream_per_chan_id_ht
,
649 &stream
->node_channel_id
);
652 * Add stream to the stream_list_ht of the consumer data. No need to steal
653 * the key since the HT does not use it and we allow to add redundant keys
656 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
659 * When nb_init_stream_left reaches 0, we don't need to trigger any action
660 * in terms of destroying the associated channel, because the action that
661 * causes the count to become 0 also causes a stream to be added. The
662 * channel deletion will thus be triggered by the following removal of this
665 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
666 /* Increment refcount before decrementing nb_init_stream_left */
668 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
671 /* Update consumer data once the node is inserted. */
672 consumer_data
.stream_count
++;
673 consumer_data
.need_update
= 1;
676 pthread_mutex_unlock(&stream
->lock
);
677 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
678 pthread_mutex_unlock(&stream
->chan
->lock
);
679 pthread_mutex_unlock(&consumer_data
.lock
);
684 void consumer_del_data_stream(struct lttng_consumer_stream
*stream
)
686 consumer_del_stream(stream
, data_ht
);
690 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
691 * be acquired before calling this.
693 static int add_relayd(struct consumer_relayd_sock_pair
*relayd
)
696 struct lttng_ht_node_u64
*node
;
697 struct lttng_ht_iter iter
;
701 lttng_ht_lookup(consumer_data
.relayd_ht
,
702 &relayd
->net_seq_idx
, &iter
);
703 node
= lttng_ht_iter_get_node_u64(&iter
);
707 lttng_ht_add_unique_u64(consumer_data
.relayd_ht
, &relayd
->node
);
714 * Allocate and return a consumer relayd socket.
716 struct consumer_relayd_sock_pair
*consumer_allocate_relayd_sock_pair(
717 uint64_t net_seq_idx
)
719 struct consumer_relayd_sock_pair
*obj
= NULL
;
721 /* net sequence index of -1 is a failure */
722 if (net_seq_idx
== (uint64_t) -1ULL) {
726 obj
= zmalloc(sizeof(struct consumer_relayd_sock_pair
));
728 PERROR("zmalloc relayd sock");
732 obj
->net_seq_idx
= net_seq_idx
;
734 obj
->destroy_flag
= 0;
735 obj
->control_sock
.sock
.fd
= -1;
736 obj
->data_sock
.sock
.fd
= -1;
737 lttng_ht_node_init_u64(&obj
->node
, obj
->net_seq_idx
);
738 pthread_mutex_init(&obj
->ctrl_sock_mutex
, NULL
);
745 * Find a relayd socket pair in the global consumer data.
747 * Return the object if found else NULL.
748 * RCU read-side lock must be held across this call and while using the
751 struct consumer_relayd_sock_pair
*consumer_find_relayd(uint64_t key
)
753 struct lttng_ht_iter iter
;
754 struct lttng_ht_node_u64
*node
;
755 struct consumer_relayd_sock_pair
*relayd
= NULL
;
757 /* Negative keys are lookup failures */
758 if (key
== (uint64_t) -1ULL) {
762 lttng_ht_lookup(consumer_data
.relayd_ht
, &key
,
764 node
= lttng_ht_iter_get_node_u64(&iter
);
766 relayd
= caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
774 * Find a relayd and send the stream
776 * Returns 0 on success, < 0 on error
778 int consumer_send_relayd_stream(struct lttng_consumer_stream
*stream
,
782 struct consumer_relayd_sock_pair
*relayd
;
785 assert(stream
->net_seq_idx
!= -1ULL);
788 /* The stream is not metadata. Get relayd reference if exists. */
790 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
791 if (relayd
!= NULL
) {
792 /* Add stream on the relayd */
793 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
794 ret
= relayd_add_stream(&relayd
->control_sock
, stream
->name
,
795 path
, &stream
->relayd_stream_id
,
796 stream
->chan
->tracefile_size
, stream
->chan
->tracefile_count
);
797 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
802 uatomic_inc(&relayd
->refcount
);
803 stream
->sent_to_relayd
= 1;
805 ERR("Stream %" PRIu64
" relayd ID %" PRIu64
" unknown. Can't send it.",
806 stream
->key
, stream
->net_seq_idx
);
811 DBG("Stream %s with key %" PRIu64
" sent to relayd id %" PRIu64
,
812 stream
->name
, stream
->key
, stream
->net_seq_idx
);
820 * Find a relayd and send the streams sent message
822 * Returns 0 on success, < 0 on error
824 int consumer_send_relayd_streams_sent(uint64_t net_seq_idx
)
827 struct consumer_relayd_sock_pair
*relayd
;
829 assert(net_seq_idx
!= -1ULL);
831 /* The stream is not metadata. Get relayd reference if exists. */
833 relayd
= consumer_find_relayd(net_seq_idx
);
834 if (relayd
!= NULL
) {
835 /* Add stream on the relayd */
836 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
837 ret
= relayd_streams_sent(&relayd
->control_sock
);
838 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
843 ERR("Relayd ID %" PRIu64
" unknown. Can't send streams_sent.",
850 DBG("All streams sent relayd id %" PRIu64
, net_seq_idx
);
858 * Find a relayd and close the stream
860 void close_relayd_stream(struct lttng_consumer_stream
*stream
)
862 struct consumer_relayd_sock_pair
*relayd
;
864 /* The stream is not metadata. Get relayd reference if exists. */
866 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
868 consumer_stream_relayd_close(stream
, relayd
);
874 * Handle stream for relayd transmission if the stream applies for network
875 * streaming where the net sequence index is set.
877 * Return destination file descriptor or negative value on error.
879 static int write_relayd_stream_header(struct lttng_consumer_stream
*stream
,
880 size_t data_size
, unsigned long padding
,
881 struct consumer_relayd_sock_pair
*relayd
)
884 struct lttcomm_relayd_data_hdr data_hdr
;
890 /* Reset data header */
891 memset(&data_hdr
, 0, sizeof(data_hdr
));
893 if (stream
->metadata_flag
) {
894 /* Caller MUST acquire the relayd control socket lock */
895 ret
= relayd_send_metadata(&relayd
->control_sock
, data_size
);
900 /* Metadata are always sent on the control socket. */
901 outfd
= relayd
->control_sock
.sock
.fd
;
903 /* Set header with stream information */
904 data_hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
905 data_hdr
.data_size
= htobe32(data_size
);
906 data_hdr
.padding_size
= htobe32(padding
);
908 * Note that net_seq_num below is assigned with the *current* value of
909 * next_net_seq_num and only after that the next_net_seq_num will be
910 * increment. This is why when issuing a command on the relayd using
911 * this next value, 1 should always be substracted in order to compare
912 * the last seen sequence number on the relayd side to the last sent.
914 data_hdr
.net_seq_num
= htobe64(stream
->next_net_seq_num
);
915 /* Other fields are zeroed previously */
917 ret
= relayd_send_data_hdr(&relayd
->data_sock
, &data_hdr
,
923 ++stream
->next_net_seq_num
;
925 /* Set to go on data socket */
926 outfd
= relayd
->data_sock
.sock
.fd
;
934 * Allocate and return a new lttng_consumer_channel object using the given key
935 * to initialize the hash table node.
937 * On error, return NULL.
939 struct lttng_consumer_channel
*consumer_allocate_channel(uint64_t key
,
941 const char *pathname
,
946 enum lttng_event_output output
,
947 uint64_t tracefile_size
,
948 uint64_t tracefile_count
,
949 uint64_t session_id_per_pid
,
950 unsigned int monitor
,
951 unsigned int live_timer_interval
,
952 const char *root_shm_path
,
953 const char *shm_path
)
955 struct lttng_consumer_channel
*channel
;
957 channel
= zmalloc(sizeof(*channel
));
958 if (channel
== NULL
) {
959 PERROR("malloc struct lttng_consumer_channel");
964 channel
->refcount
= 0;
965 channel
->session_id
= session_id
;
966 channel
->session_id_per_pid
= session_id_per_pid
;
969 channel
->relayd_id
= relayd_id
;
970 channel
->tracefile_size
= tracefile_size
;
971 channel
->tracefile_count
= tracefile_count
;
972 channel
->monitor
= monitor
;
973 channel
->live_timer_interval
= live_timer_interval
;
974 pthread_mutex_init(&channel
->lock
, NULL
);
975 pthread_mutex_init(&channel
->timer_lock
, NULL
);
978 case LTTNG_EVENT_SPLICE
:
979 channel
->output
= CONSUMER_CHANNEL_SPLICE
;
981 case LTTNG_EVENT_MMAP
:
982 channel
->output
= CONSUMER_CHANNEL_MMAP
;
992 * In monitor mode, the streams associated with the channel will be put in
993 * a special list ONLY owned by this channel. So, the refcount is set to 1
994 * here meaning that the channel itself has streams that are referenced.
996 * On a channel deletion, once the channel is no longer visible, the
997 * refcount is decremented and checked for a zero value to delete it. With
998 * streams in no monitor mode, it will now be safe to destroy the channel.
1000 if (!channel
->monitor
) {
1001 channel
->refcount
= 1;
1004 strncpy(channel
->pathname
, pathname
, sizeof(channel
->pathname
));
1005 channel
->pathname
[sizeof(channel
->pathname
) - 1] = '\0';
1007 strncpy(channel
->name
, name
, sizeof(channel
->name
));
1008 channel
->name
[sizeof(channel
->name
) - 1] = '\0';
1010 if (root_shm_path
) {
1011 strncpy(channel
->root_shm_path
, root_shm_path
, sizeof(channel
->root_shm_path
));
1012 channel
->root_shm_path
[sizeof(channel
->root_shm_path
) - 1] = '\0';
1015 strncpy(channel
->shm_path
, shm_path
, sizeof(channel
->shm_path
));
1016 channel
->shm_path
[sizeof(channel
->shm_path
) - 1] = '\0';
1019 lttng_ht_node_init_u64(&channel
->node
, channel
->key
);
1021 channel
->wait_fd
= -1;
1023 CDS_INIT_LIST_HEAD(&channel
->streams
.head
);
1025 DBG("Allocated channel (key %" PRIu64
")", channel
->key
)
1032 * Add a channel to the global list protected by a mutex.
1034 * Always return 0 indicating success.
1036 int consumer_add_channel(struct lttng_consumer_channel
*channel
,
1037 struct lttng_consumer_local_data
*ctx
)
1039 pthread_mutex_lock(&consumer_data
.lock
);
1040 pthread_mutex_lock(&channel
->lock
);
1041 pthread_mutex_lock(&channel
->timer_lock
);
1044 * This gives us a guarantee that the channel we are about to add to the
1045 * channel hash table will be unique. See this function comment on the why
1046 * we need to steel the channel key at this stage.
1048 steal_channel_key(channel
->key
);
1051 lttng_ht_add_unique_u64(consumer_data
.channel_ht
, &channel
->node
);
1054 pthread_mutex_unlock(&channel
->timer_lock
);
1055 pthread_mutex_unlock(&channel
->lock
);
1056 pthread_mutex_unlock(&consumer_data
.lock
);
1058 if (channel
->wait_fd
!= -1 && channel
->type
== CONSUMER_CHANNEL_TYPE_DATA
) {
1059 notify_channel_pipe(ctx
, channel
, -1, CONSUMER_CHANNEL_ADD
);
1066 * Allocate the pollfd structure and the local view of the out fds to avoid
1067 * doing a lookup in the linked list and concurrency issues when writing is
1068 * needed. Called with consumer_data.lock held.
1070 * Returns the number of fds in the structures.
1072 static int update_poll_array(struct lttng_consumer_local_data
*ctx
,
1073 struct pollfd
**pollfd
, struct lttng_consumer_stream
**local_stream
,
1074 struct lttng_ht
*ht
)
1077 struct lttng_ht_iter iter
;
1078 struct lttng_consumer_stream
*stream
;
1083 assert(local_stream
);
1085 DBG("Updating poll fd array");
1087 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1089 * Only active streams with an active end point can be added to the
1090 * poll set and local stream storage of the thread.
1092 * There is a potential race here for endpoint_status to be updated
1093 * just after the check. However, this is OK since the stream(s) will
1094 * be deleted once the thread is notified that the end point state has
1095 * changed where this function will be called back again.
1097 if (stream
->state
!= LTTNG_CONSUMER_ACTIVE_STREAM
||
1098 stream
->endpoint_status
== CONSUMER_ENDPOINT_INACTIVE
) {
1102 * This clobbers way too much the debug output. Uncomment that if you
1103 * need it for debugging purposes.
1105 * DBG("Active FD %d", stream->wait_fd);
1107 (*pollfd
)[i
].fd
= stream
->wait_fd
;
1108 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1109 local_stream
[i
] = stream
;
1115 * Insert the consumer_data_pipe at the end of the array and don't
1116 * increment i so nb_fd is the number of real FD.
1118 (*pollfd
)[i
].fd
= lttng_pipe_get_readfd(ctx
->consumer_data_pipe
);
1119 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
1121 (*pollfd
)[i
+ 1].fd
= lttng_pipe_get_readfd(ctx
->consumer_wakeup_pipe
);
1122 (*pollfd
)[i
+ 1].events
= POLLIN
| POLLPRI
;
1127 * Poll on the should_quit pipe and the command socket return -1 on
1128 * error, 1 if should exit, 0 if data is available on the command socket
1130 int lttng_consumer_poll_socket(struct pollfd
*consumer_sockpoll
)
1135 num_rdy
= poll(consumer_sockpoll
, 2, -1);
1136 if (num_rdy
== -1) {
1138 * Restart interrupted system call.
1140 if (errno
== EINTR
) {
1143 PERROR("Poll error");
1146 if (consumer_sockpoll
[0].revents
& (POLLIN
| POLLPRI
)) {
1147 DBG("consumer_should_quit wake up");
1154 * Set the error socket.
1156 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data
*ctx
,
1159 ctx
->consumer_error_socket
= sock
;
1163 * Set the command socket path.
1165 void lttng_consumer_set_command_sock_path(
1166 struct lttng_consumer_local_data
*ctx
, char *sock
)
1168 ctx
->consumer_command_sock_path
= sock
;
1172 * Send return code to the session daemon.
1173 * If the socket is not defined, we return 0, it is not a fatal error
1175 int lttng_consumer_send_error(struct lttng_consumer_local_data
*ctx
, int cmd
)
1177 if (ctx
->consumer_error_socket
> 0) {
1178 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
1179 sizeof(enum lttcomm_sessiond_command
));
1186 * Close all the tracefiles and stream fds and MUST be called when all
1187 * instances are destroyed i.e. when all threads were joined and are ended.
1189 void lttng_consumer_cleanup(void)
1191 struct lttng_ht_iter iter
;
1192 struct lttng_consumer_channel
*channel
;
1196 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
, channel
,
1198 consumer_del_channel(channel
);
1203 lttng_ht_destroy(consumer_data
.channel_ht
);
1205 cleanup_relayd_ht();
1207 lttng_ht_destroy(consumer_data
.stream_per_chan_id_ht
);
1210 * This HT contains streams that are freed by either the metadata thread or
1211 * the data thread so we do *nothing* on the hash table and simply destroy
1214 lttng_ht_destroy(consumer_data
.stream_list_ht
);
1218 * Called from signal handler.
1220 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
1225 ret
= lttng_write(ctx
->consumer_should_quit
[1], "4", 1);
1227 PERROR("write consumer quit");
1230 DBG("Consumer flag that it should quit");
1233 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1236 int outfd
= stream
->out_fd
;
1239 * This does a blocking write-and-wait on any page that belongs to the
1240 * subbuffer prior to the one we just wrote.
1241 * Don't care about error values, as these are just hints and ways to
1242 * limit the amount of page cache used.
1244 if (orig_offset
< stream
->max_sb_size
) {
1247 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1248 stream
->max_sb_size
,
1249 SYNC_FILE_RANGE_WAIT_BEFORE
1250 | SYNC_FILE_RANGE_WRITE
1251 | SYNC_FILE_RANGE_WAIT_AFTER
);
1253 * Give hints to the kernel about how we access the file:
1254 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1257 * We need to call fadvise again after the file grows because the
1258 * kernel does not seem to apply fadvise to non-existing parts of the
1261 * Call fadvise _after_ having waited for the page writeback to
1262 * complete because the dirty page writeback semantic is not well
1263 * defined. So it can be expected to lead to lower throughput in
1266 posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1267 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1271 * Initialise the necessary environnement :
1272 * - create a new context
1273 * - create the poll_pipe
1274 * - create the should_quit pipe (for signal handler)
1275 * - create the thread pipe (for splice)
1277 * Takes a function pointer as argument, this function is called when data is
1278 * available on a buffer. This function is responsible to do the
1279 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1280 * buffer configuration and then kernctl_put_next_subbuf at the end.
1282 * Returns a pointer to the new context or NULL on error.
1284 struct lttng_consumer_local_data
*lttng_consumer_create(
1285 enum lttng_consumer_type type
,
1286 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1287 struct lttng_consumer_local_data
*ctx
),
1288 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1289 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1290 int (*update_stream
)(uint64_t stream_key
, uint32_t state
))
1293 struct lttng_consumer_local_data
*ctx
;
1295 assert(consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1296 consumer_data
.type
== type
);
1297 consumer_data
.type
= type
;
1299 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
1301 PERROR("allocating context");
1305 ctx
->consumer_error_socket
= -1;
1306 ctx
->consumer_metadata_socket
= -1;
1307 pthread_mutex_init(&ctx
->metadata_socket_lock
, NULL
);
1308 /* assign the callbacks */
1309 ctx
->on_buffer_ready
= buffer_ready
;
1310 ctx
->on_recv_channel
= recv_channel
;
1311 ctx
->on_recv_stream
= recv_stream
;
1312 ctx
->on_update_stream
= update_stream
;
1314 ctx
->consumer_data_pipe
= lttng_pipe_open(0);
1315 if (!ctx
->consumer_data_pipe
) {
1316 goto error_poll_pipe
;
1319 ctx
->consumer_wakeup_pipe
= lttng_pipe_open(0);
1320 if (!ctx
->consumer_wakeup_pipe
) {
1321 goto error_wakeup_pipe
;
1324 ret
= pipe(ctx
->consumer_should_quit
);
1326 PERROR("Error creating recv pipe");
1327 goto error_quit_pipe
;
1330 ret
= pipe(ctx
->consumer_channel_pipe
);
1332 PERROR("Error creating channel pipe");
1333 goto error_channel_pipe
;
1336 ctx
->consumer_metadata_pipe
= lttng_pipe_open(0);
1337 if (!ctx
->consumer_metadata_pipe
) {
1338 goto error_metadata_pipe
;
1343 error_metadata_pipe
:
1344 utils_close_pipe(ctx
->consumer_channel_pipe
);
1346 utils_close_pipe(ctx
->consumer_should_quit
);
1348 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1350 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1358 * Iterate over all streams of the hashtable and free them properly.
1360 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1362 struct lttng_ht_iter iter
;
1363 struct lttng_consumer_stream
*stream
;
1370 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1372 * Ignore return value since we are currently cleaning up so any error
1375 (void) consumer_del_stream(stream
, ht
);
1379 lttng_ht_destroy(ht
);
1383 * Iterate over all streams of the metadata hashtable and free them
1386 static void destroy_metadata_stream_ht(struct lttng_ht
*ht
)
1388 struct lttng_ht_iter iter
;
1389 struct lttng_consumer_stream
*stream
;
1396 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1398 * Ignore return value since we are currently cleaning up so any error
1401 (void) consumer_del_metadata_stream(stream
, ht
);
1405 lttng_ht_destroy(ht
);
1409 * Close all fds associated with the instance and free the context.
1411 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1415 DBG("Consumer destroying it. Closing everything.");
1421 destroy_data_stream_ht(data_ht
);
1422 destroy_metadata_stream_ht(metadata_ht
);
1424 ret
= close(ctx
->consumer_error_socket
);
1428 ret
= close(ctx
->consumer_metadata_socket
);
1432 utils_close_pipe(ctx
->consumer_channel_pipe
);
1433 lttng_pipe_destroy(ctx
->consumer_data_pipe
);
1434 lttng_pipe_destroy(ctx
->consumer_metadata_pipe
);
1435 lttng_pipe_destroy(ctx
->consumer_wakeup_pipe
);
1436 utils_close_pipe(ctx
->consumer_should_quit
);
1438 unlink(ctx
->consumer_command_sock_path
);
1443 * Write the metadata stream id on the specified file descriptor.
1445 static int write_relayd_metadata_id(int fd
,
1446 struct lttng_consumer_stream
*stream
,
1447 struct consumer_relayd_sock_pair
*relayd
, unsigned long padding
)
1450 struct lttcomm_relayd_metadata_payload hdr
;
1452 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1453 hdr
.padding_size
= htobe32(padding
);
1454 ret
= lttng_write(fd
, (void *) &hdr
, sizeof(hdr
));
1455 if (ret
< sizeof(hdr
)) {
1457 * This error means that the fd's end is closed so ignore the PERROR
1458 * not to clubber the error output since this can happen in a normal
1461 if (errno
!= EPIPE
) {
1462 PERROR("write metadata stream id");
1464 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1466 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1467 * handle writting the missing part so report that as an error and
1468 * don't lie to the caller.
1473 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1474 stream
->relayd_stream_id
, padding
);
1481 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1482 * core function for writing trace buffers to either the local filesystem or
1485 * It must be called with the stream lock held.
1487 * Careful review MUST be put if any changes occur!
1489 * Returns the number of bytes written
1491 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1492 struct lttng_consumer_local_data
*ctx
,
1493 struct lttng_consumer_stream
*stream
, unsigned long len
,
1494 unsigned long padding
,
1495 struct ctf_packet_index
*index
)
1497 unsigned long mmap_offset
;
1500 off_t orig_offset
= stream
->out_fd_offset
;
1501 /* Default is on the disk */
1502 int outfd
= stream
->out_fd
;
1503 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1504 unsigned int relayd_hang_up
= 0;
1506 /* RCU lock for the relayd pointer */
1509 /* Flag that the current stream if set for network streaming. */
1510 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1511 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1512 if (relayd
== NULL
) {
1518 /* get the offset inside the fd to mmap */
1519 switch (consumer_data
.type
) {
1520 case LTTNG_CONSUMER_KERNEL
:
1521 mmap_base
= stream
->mmap_base
;
1522 ret
= kernctl_get_mmap_read_offset(stream
->wait_fd
, &mmap_offset
);
1525 PERROR("tracer ctl get_mmap_read_offset");
1529 case LTTNG_CONSUMER32_UST
:
1530 case LTTNG_CONSUMER64_UST
:
1531 mmap_base
= lttng_ustctl_get_mmap_base(stream
);
1533 ERR("read mmap get mmap base for stream %s", stream
->name
);
1537 ret
= lttng_ustctl_get_mmap_read_offset(stream
, &mmap_offset
);
1539 PERROR("tracer ctl get_mmap_read_offset");
1545 ERR("Unknown consumer_data type");
1549 /* Handle stream on the relayd if the output is on the network */
1551 unsigned long netlen
= len
;
1554 * Lock the control socket for the complete duration of the function
1555 * since from this point on we will use the socket.
1557 if (stream
->metadata_flag
) {
1558 /* Metadata requires the control socket. */
1559 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1560 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1563 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1568 /* Use the returned socket. */
1571 /* Write metadata stream id before payload */
1572 if (stream
->metadata_flag
) {
1573 ret
= write_relayd_metadata_id(outfd
, stream
, relayd
, padding
);
1580 /* No streaming, we have to set the len with the full padding */
1584 * Check if we need to change the tracefile before writing the packet.
1586 if (stream
->chan
->tracefile_size
> 0 &&
1587 (stream
->tracefile_size_current
+ len
) >
1588 stream
->chan
->tracefile_size
) {
1589 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1590 stream
->name
, stream
->chan
->tracefile_size
,
1591 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1592 stream
->out_fd
, &(stream
->tracefile_count_current
),
1595 ERR("Rotating output file");
1598 outfd
= stream
->out_fd
;
1600 if (stream
->index_fd
>= 0) {
1601 ret
= index_create_file(stream
->chan
->pathname
,
1602 stream
->name
, stream
->uid
, stream
->gid
,
1603 stream
->chan
->tracefile_size
,
1604 stream
->tracefile_count_current
);
1608 stream
->index_fd
= ret
;
1611 /* Reset current size because we just perform a rotation. */
1612 stream
->tracefile_size_current
= 0;
1613 stream
->out_fd_offset
= 0;
1616 stream
->tracefile_size_current
+= len
;
1618 index
->offset
= htobe64(stream
->out_fd_offset
);
1623 * This call guarantee that len or less is returned. It's impossible to
1624 * receive a ret value that is bigger than len.
1626 ret
= lttng_write(outfd
, mmap_base
+ mmap_offset
, len
);
1627 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, len
);
1628 if (ret
< 0 || ((size_t) ret
!= len
)) {
1630 * Report error to caller if nothing was written else at least send the
1638 /* Socket operation failed. We consider the relayd dead */
1639 if (errno
== EPIPE
|| errno
== EINVAL
|| errno
== EBADF
) {
1641 * This is possible if the fd is closed on the other side
1642 * (outfd) or any write problem. It can be verbose a bit for a
1643 * normal execution if for instance the relayd is stopped
1644 * abruptly. This can happen so set this to a DBG statement.
1646 DBG("Consumer mmap write detected relayd hang up");
1648 /* Unhandled error, print it and stop function right now. */
1649 PERROR("Error in write mmap (ret %zd != len %lu)", ret
, len
);
1653 stream
->output_written
+= ret
;
1655 /* This call is useless on a socket so better save a syscall. */
1657 /* This won't block, but will start writeout asynchronously */
1658 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, len
,
1659 SYNC_FILE_RANGE_WRITE
);
1660 stream
->out_fd_offset
+= len
;
1662 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1666 * This is a special case that the relayd has closed its socket. Let's
1667 * cleanup the relayd object and all associated streams.
1669 if (relayd
&& relayd_hang_up
) {
1670 cleanup_relayd(relayd
, ctx
);
1674 /* Unlock only if ctrl socket used */
1675 if (relayd
&& stream
->metadata_flag
) {
1676 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1684 * Splice the data from the ring buffer to the tracefile.
1686 * It must be called with the stream lock held.
1688 * Returns the number of bytes spliced.
1690 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1691 struct lttng_consumer_local_data
*ctx
,
1692 struct lttng_consumer_stream
*stream
, unsigned long len
,
1693 unsigned long padding
,
1694 struct ctf_packet_index
*index
)
1696 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1698 off_t orig_offset
= stream
->out_fd_offset
;
1699 int fd
= stream
->wait_fd
;
1700 /* Default is on the disk */
1701 int outfd
= stream
->out_fd
;
1702 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1704 unsigned int relayd_hang_up
= 0;
1706 switch (consumer_data
.type
) {
1707 case LTTNG_CONSUMER_KERNEL
:
1709 case LTTNG_CONSUMER32_UST
:
1710 case LTTNG_CONSUMER64_UST
:
1711 /* Not supported for user space tracing */
1714 ERR("Unknown consumer_data type");
1718 /* RCU lock for the relayd pointer */
1721 /* Flag that the current stream if set for network streaming. */
1722 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1723 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1724 if (relayd
== NULL
) {
1729 splice_pipe
= stream
->splice_pipe
;
1731 /* Write metadata stream id before payload */
1733 unsigned long total_len
= len
;
1735 if (stream
->metadata_flag
) {
1737 * Lock the control socket for the complete duration of the function
1738 * since from this point on we will use the socket.
1740 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1742 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
, relayd
,
1750 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1753 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1759 /* Use the returned socket. */
1762 /* No streaming, we have to set the len with the full padding */
1766 * Check if we need to change the tracefile before writing the packet.
1768 if (stream
->chan
->tracefile_size
> 0 &&
1769 (stream
->tracefile_size_current
+ len
) >
1770 stream
->chan
->tracefile_size
) {
1771 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1772 stream
->name
, stream
->chan
->tracefile_size
,
1773 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1774 stream
->out_fd
, &(stream
->tracefile_count_current
),
1778 ERR("Rotating output file");
1781 outfd
= stream
->out_fd
;
1783 if (stream
->index_fd
>= 0) {
1784 ret
= close(stream
->index_fd
);
1786 PERROR("Closing index");
1789 stream
->index_fd
= -1;
1790 ret
= index_create_file(stream
->chan
->pathname
,
1791 stream
->name
, stream
->uid
, stream
->gid
,
1792 stream
->chan
->tracefile_size
,
1793 stream
->tracefile_count_current
);
1798 stream
->index_fd
= ret
;
1801 /* Reset current size because we just perform a rotation. */
1802 stream
->tracefile_size_current
= 0;
1803 stream
->out_fd_offset
= 0;
1806 stream
->tracefile_size_current
+= len
;
1807 index
->offset
= htobe64(stream
->out_fd_offset
);
1811 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1812 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1813 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1814 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1815 DBG("splice chan to pipe, ret %zd", ret_splice
);
1816 if (ret_splice
< 0) {
1819 PERROR("Error in relay splice");
1823 /* Handle stream on the relayd if the output is on the network */
1824 if (relayd
&& stream
->metadata_flag
) {
1825 size_t metadata_payload_size
=
1826 sizeof(struct lttcomm_relayd_metadata_payload
);
1828 /* Update counter to fit the spliced data */
1829 ret_splice
+= metadata_payload_size
;
1830 len
+= metadata_payload_size
;
1832 * We do this so the return value can match the len passed as
1833 * argument to this function.
1835 written
-= metadata_payload_size
;
1838 /* Splice data out */
1839 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1840 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1841 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
1843 if (ret_splice
< 0) {
1848 } else if (ret_splice
> len
) {
1850 * We don't expect this code path to be executed but you never know
1851 * so this is an extra protection agains a buggy splice().
1854 written
+= ret_splice
;
1855 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice
,
1859 /* All good, update current len and continue. */
1863 /* This call is useless on a socket so better save a syscall. */
1865 /* This won't block, but will start writeout asynchronously */
1866 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1867 SYNC_FILE_RANGE_WRITE
);
1868 stream
->out_fd_offset
+= ret_splice
;
1870 stream
->output_written
+= ret_splice
;
1871 written
+= ret_splice
;
1873 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1878 * This is a special case that the relayd has closed its socket. Let's
1879 * cleanup the relayd object and all associated streams.
1881 if (relayd
&& relayd_hang_up
) {
1882 cleanup_relayd(relayd
, ctx
);
1883 /* Skip splice error so the consumer does not fail */
1888 /* send the appropriate error description to sessiond */
1891 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1894 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1897 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
1902 if (relayd
&& stream
->metadata_flag
) {
1903 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1911 * Take a snapshot for a specific fd
1913 * Returns 0 on success, < 0 on error
1915 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
1917 switch (consumer_data
.type
) {
1918 case LTTNG_CONSUMER_KERNEL
:
1919 return lttng_kconsumer_take_snapshot(stream
);
1920 case LTTNG_CONSUMER32_UST
:
1921 case LTTNG_CONSUMER64_UST
:
1922 return lttng_ustconsumer_take_snapshot(stream
);
1924 ERR("Unknown consumer_data type");
1931 * Get the produced position
1933 * Returns 0 on success, < 0 on error
1935 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
1938 switch (consumer_data
.type
) {
1939 case LTTNG_CONSUMER_KERNEL
:
1940 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
1941 case LTTNG_CONSUMER32_UST
:
1942 case LTTNG_CONSUMER64_UST
:
1943 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
1945 ERR("Unknown consumer_data type");
1951 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
1952 int sock
, struct pollfd
*consumer_sockpoll
)
1954 switch (consumer_data
.type
) {
1955 case LTTNG_CONSUMER_KERNEL
:
1956 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1957 case LTTNG_CONSUMER32_UST
:
1958 case LTTNG_CONSUMER64_UST
:
1959 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1961 ERR("Unknown consumer_data type");
1967 void lttng_consumer_close_all_metadata(void)
1969 switch (consumer_data
.type
) {
1970 case LTTNG_CONSUMER_KERNEL
:
1972 * The Kernel consumer has a different metadata scheme so we don't
1973 * close anything because the stream will be closed by the session
1977 case LTTNG_CONSUMER32_UST
:
1978 case LTTNG_CONSUMER64_UST
:
1980 * Close all metadata streams. The metadata hash table is passed and
1981 * this call iterates over it by closing all wakeup fd. This is safe
1982 * because at this point we are sure that the metadata producer is
1983 * either dead or blocked.
1985 lttng_ustconsumer_close_all_metadata(metadata_ht
);
1988 ERR("Unknown consumer_data type");
1994 * Clean up a metadata stream and free its memory.
1996 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
1997 struct lttng_ht
*ht
)
1999 struct lttng_consumer_channel
*free_chan
= NULL
;
2003 * This call should NEVER receive regular stream. It must always be
2004 * metadata stream and this is crucial for data structure synchronization.
2006 assert(stream
->metadata_flag
);
2008 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
2010 pthread_mutex_lock(&consumer_data
.lock
);
2011 pthread_mutex_lock(&stream
->chan
->lock
);
2012 pthread_mutex_lock(&stream
->lock
);
2014 /* Remove any reference to that stream. */
2015 consumer_stream_delete(stream
, ht
);
2017 /* Close down everything including the relayd if one. */
2018 consumer_stream_close(stream
);
2019 /* Destroy tracer buffers of the stream. */
2020 consumer_stream_destroy_buffers(stream
);
2022 /* Atomically decrement channel refcount since other threads can use it. */
2023 if (!uatomic_sub_return(&stream
->chan
->refcount
, 1)
2024 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
2025 /* Go for channel deletion! */
2026 free_chan
= stream
->chan
;
2030 * Nullify the stream reference so it is not used after deletion. The
2031 * channel lock MUST be acquired before being able to check for a NULL
2034 stream
->chan
->metadata_stream
= NULL
;
2036 pthread_mutex_unlock(&stream
->lock
);
2037 pthread_mutex_unlock(&stream
->chan
->lock
);
2038 pthread_mutex_unlock(&consumer_data
.lock
);
2041 consumer_del_channel(free_chan
);
2044 consumer_stream_free(stream
);
2048 * Action done with the metadata stream when adding it to the consumer internal
2049 * data structures to handle it.
2051 int consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
2053 struct lttng_ht
*ht
= metadata_ht
;
2055 struct lttng_ht_iter iter
;
2056 struct lttng_ht_node_u64
*node
;
2061 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2063 pthread_mutex_lock(&consumer_data
.lock
);
2064 pthread_mutex_lock(&stream
->chan
->lock
);
2065 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2066 pthread_mutex_lock(&stream
->lock
);
2069 * From here, refcounts are updated so be _careful_ when returning an error
2076 * Lookup the stream just to make sure it does not exist in our internal
2077 * state. This should NEVER happen.
2079 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2080 node
= lttng_ht_iter_get_node_u64(&iter
);
2084 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2085 * in terms of destroying the associated channel, because the action that
2086 * causes the count to become 0 also causes a stream to be added. The
2087 * channel deletion will thus be triggered by the following removal of this
2090 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2091 /* Increment refcount before decrementing nb_init_stream_left */
2093 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2096 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2098 lttng_ht_add_unique_u64(consumer_data
.stream_per_chan_id_ht
,
2099 &stream
->node_channel_id
);
2102 * Add stream to the stream_list_ht of the consumer data. No need to steal
2103 * the key since the HT does not use it and we allow to add redundant keys
2106 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2110 pthread_mutex_unlock(&stream
->lock
);
2111 pthread_mutex_unlock(&stream
->chan
->lock
);
2112 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2113 pthread_mutex_unlock(&consumer_data
.lock
);
2118 * Delete data stream that are flagged for deletion (endpoint_status).
2120 static void validate_endpoint_status_data_stream(void)
2122 struct lttng_ht_iter iter
;
2123 struct lttng_consumer_stream
*stream
;
2125 DBG("Consumer delete flagged data stream");
2128 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2129 /* Validate delete flag of the stream */
2130 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2133 /* Delete it right now */
2134 consumer_del_stream(stream
, data_ht
);
2140 * Delete metadata stream that are flagged for deletion (endpoint_status).
2142 static void validate_endpoint_status_metadata_stream(
2143 struct lttng_poll_event
*pollset
)
2145 struct lttng_ht_iter iter
;
2146 struct lttng_consumer_stream
*stream
;
2148 DBG("Consumer delete flagged metadata stream");
2153 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2154 /* Validate delete flag of the stream */
2155 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2159 * Remove from pollset so the metadata thread can continue without
2160 * blocking on a deleted stream.
2162 lttng_poll_del(pollset
, stream
->wait_fd
);
2164 /* Delete it right now */
2165 consumer_del_metadata_stream(stream
, metadata_ht
);
2171 * Thread polls on metadata file descriptor and write them on disk or on the
2174 void *consumer_thread_metadata_poll(void *data
)
2176 int ret
, i
, pollfd
, err
= -1;
2177 uint32_t revents
, nb_fd
;
2178 struct lttng_consumer_stream
*stream
= NULL
;
2179 struct lttng_ht_iter iter
;
2180 struct lttng_ht_node_u64
*node
;
2181 struct lttng_poll_event events
;
2182 struct lttng_consumer_local_data
*ctx
= data
;
2185 rcu_register_thread();
2187 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2189 if (testpoint(consumerd_thread_metadata
)) {
2190 goto error_testpoint
;
2193 health_code_update();
2195 DBG("Thread metadata poll started");
2197 /* Size is set to 1 for the consumer_metadata pipe */
2198 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2200 ERR("Poll set creation failed");
2204 ret
= lttng_poll_add(&events
,
2205 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2211 DBG("Metadata main loop started");
2215 health_code_update();
2216 health_poll_entry();
2217 DBG("Metadata poll wait");
2218 ret
= lttng_poll_wait(&events
, -1);
2219 DBG("Metadata poll return from wait with %d fd(s)",
2220 LTTNG_POLL_GETNB(&events
));
2222 DBG("Metadata event catched in thread");
2224 if (errno
== EINTR
) {
2225 ERR("Poll EINTR catched");
2228 if (LTTNG_POLL_GETNB(&events
) == 0) {
2229 err
= 0; /* All is OK */
2236 /* From here, the event is a metadata wait fd */
2237 for (i
= 0; i
< nb_fd
; i
++) {
2238 health_code_update();
2240 revents
= LTTNG_POLL_GETEV(&events
, i
);
2241 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2244 /* No activity for this FD (poll implementation). */
2248 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2249 if (revents
& LPOLLIN
) {
2252 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2253 &stream
, sizeof(stream
));
2254 if (pipe_len
< sizeof(stream
)) {
2256 PERROR("read metadata stream");
2259 * Remove the pipe from the poll set and continue the loop
2260 * since their might be data to consume.
2262 lttng_poll_del(&events
,
2263 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2264 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2268 /* A NULL stream means that the state has changed. */
2269 if (stream
== NULL
) {
2270 /* Check for deleted streams. */
2271 validate_endpoint_status_metadata_stream(&events
);
2275 DBG("Adding metadata stream %d to poll set",
2278 /* Add metadata stream to the global poll events list */
2279 lttng_poll_add(&events
, stream
->wait_fd
,
2280 LPOLLIN
| LPOLLPRI
| LPOLLHUP
);
2281 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2282 DBG("Metadata thread pipe hung up");
2284 * Remove the pipe from the poll set and continue the loop
2285 * since their might be data to consume.
2287 lttng_poll_del(&events
,
2288 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2289 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2292 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2296 /* Handle other stream */
2302 uint64_t tmp_id
= (uint64_t) pollfd
;
2304 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2306 node
= lttng_ht_iter_get_node_u64(&iter
);
2309 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2312 if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2313 /* Get the data out of the metadata file descriptor */
2314 DBG("Metadata available on fd %d", pollfd
);
2315 assert(stream
->wait_fd
== pollfd
);
2318 health_code_update();
2320 len
= ctx
->on_buffer_ready(stream
, ctx
);
2322 * We don't check the return value here since if we get
2323 * a negative len, it means an error occured thus we
2324 * simply remove it from the poll set and free the
2329 /* It's ok to have an unavailable sub-buffer */
2330 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2331 /* Clean up stream from consumer and free it. */
2332 lttng_poll_del(&events
, stream
->wait_fd
);
2333 consumer_del_metadata_stream(stream
, metadata_ht
);
2335 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2336 DBG("Metadata fd %d is hup|err.", pollfd
);
2337 if (!stream
->hangup_flush_done
2338 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2339 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2340 DBG("Attempting to flush and consume the UST buffers");
2341 lttng_ustconsumer_on_stream_hangup(stream
);
2343 /* We just flushed the stream now read it. */
2345 health_code_update();
2347 len
= ctx
->on_buffer_ready(stream
, ctx
);
2349 * We don't check the return value here since if we get
2350 * a negative len, it means an error occured thus we
2351 * simply remove it from the poll set and free the
2357 lttng_poll_del(&events
, stream
->wait_fd
);
2359 * This call update the channel states, closes file descriptors
2360 * and securely free the stream.
2362 consumer_del_metadata_stream(stream
, metadata_ht
);
2364 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2368 /* Release RCU lock for the stream looked up */
2376 DBG("Metadata poll thread exiting");
2378 lttng_poll_clean(&events
);
2383 ERR("Health error occurred in %s", __func__
);
2385 health_unregister(health_consumerd
);
2386 rcu_unregister_thread();
2391 * This thread polls the fds in the set to consume the data and write
2392 * it to tracefile if necessary.
2394 void *consumer_thread_data_poll(void *data
)
2396 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2397 struct pollfd
*pollfd
= NULL
;
2398 /* local view of the streams */
2399 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2400 /* local view of consumer_data.fds_count */
2402 struct lttng_consumer_local_data
*ctx
= data
;
2405 rcu_register_thread();
2407 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2409 if (testpoint(consumerd_thread_data
)) {
2410 goto error_testpoint
;
2413 health_code_update();
2415 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
*));
2416 if (local_stream
== NULL
) {
2417 PERROR("local_stream malloc");
2422 health_code_update();
2428 * the fds set has been updated, we need to update our
2429 * local array as well
2431 pthread_mutex_lock(&consumer_data
.lock
);
2432 if (consumer_data
.need_update
) {
2437 local_stream
= NULL
;
2440 * Allocate for all fds +1 for the consumer_data_pipe and +1 for
2443 pollfd
= zmalloc((consumer_data
.stream_count
+ 2) * sizeof(struct pollfd
));
2444 if (pollfd
== NULL
) {
2445 PERROR("pollfd malloc");
2446 pthread_mutex_unlock(&consumer_data
.lock
);
2450 local_stream
= zmalloc((consumer_data
.stream_count
+ 2) *
2451 sizeof(struct lttng_consumer_stream
*));
2452 if (local_stream
== NULL
) {
2453 PERROR("local_stream malloc");
2454 pthread_mutex_unlock(&consumer_data
.lock
);
2457 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2460 ERR("Error in allocating pollfd or local_outfds");
2461 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2462 pthread_mutex_unlock(&consumer_data
.lock
);
2466 consumer_data
.need_update
= 0;
2468 pthread_mutex_unlock(&consumer_data
.lock
);
2470 /* No FDs and consumer_quit, consumer_cleanup the thread */
2471 if (nb_fd
== 0 && consumer_quit
== 1) {
2472 err
= 0; /* All is OK */
2475 /* poll on the array of fds */
2477 DBG("polling on %d fd", nb_fd
+ 2);
2478 health_poll_entry();
2479 num_rdy
= poll(pollfd
, nb_fd
+ 2, -1);
2481 DBG("poll num_rdy : %d", num_rdy
);
2482 if (num_rdy
== -1) {
2484 * Restart interrupted system call.
2486 if (errno
== EINTR
) {
2489 PERROR("Poll error");
2490 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2492 } else if (num_rdy
== 0) {
2493 DBG("Polling thread timed out");
2498 * If the consumer_data_pipe triggered poll go directly to the
2499 * beginning of the loop to update the array. We want to prioritize
2500 * array update over low-priority reads.
2502 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2503 ssize_t pipe_readlen
;
2505 DBG("consumer_data_pipe wake up");
2506 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2507 &new_stream
, sizeof(new_stream
));
2508 if (pipe_readlen
< sizeof(new_stream
)) {
2509 PERROR("Consumer data pipe");
2510 /* Continue so we can at least handle the current stream(s). */
2515 * If the stream is NULL, just ignore it. It's also possible that
2516 * the sessiond poll thread changed the consumer_quit state and is
2517 * waking us up to test it.
2519 if (new_stream
== NULL
) {
2520 validate_endpoint_status_data_stream();
2524 /* Continue to update the local streams and handle prio ones */
2528 /* Handle wakeup pipe. */
2529 if (pollfd
[nb_fd
+ 1].revents
& (POLLIN
| POLLPRI
)) {
2531 ssize_t pipe_readlen
;
2533 pipe_readlen
= lttng_pipe_read(ctx
->consumer_wakeup_pipe
, &dummy
,
2535 if (pipe_readlen
< 0) {
2536 PERROR("Consumer data wakeup pipe");
2538 /* We've been awakened to handle stream(s). */
2539 ctx
->has_wakeup
= 0;
2542 /* Take care of high priority channels first. */
2543 for (i
= 0; i
< nb_fd
; i
++) {
2544 health_code_update();
2546 if (local_stream
[i
] == NULL
) {
2549 if (pollfd
[i
].revents
& POLLPRI
) {
2550 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2552 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2553 /* it's ok to have an unavailable sub-buffer */
2554 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2555 /* Clean the stream and free it. */
2556 consumer_del_stream(local_stream
[i
], data_ht
);
2557 local_stream
[i
] = NULL
;
2558 } else if (len
> 0) {
2559 local_stream
[i
]->data_read
= 1;
2565 * If we read high prio channel in this loop, try again
2566 * for more high prio data.
2572 /* Take care of low priority channels. */
2573 for (i
= 0; i
< nb_fd
; i
++) {
2574 health_code_update();
2576 if (local_stream
[i
] == NULL
) {
2579 if ((pollfd
[i
].revents
& POLLIN
) ||
2580 local_stream
[i
]->hangup_flush_done
||
2581 local_stream
[i
]->has_data
) {
2582 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2583 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2584 /* it's ok to have an unavailable sub-buffer */
2585 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2586 /* Clean the stream and free it. */
2587 consumer_del_stream(local_stream
[i
], data_ht
);
2588 local_stream
[i
] = NULL
;
2589 } else if (len
> 0) {
2590 local_stream
[i
]->data_read
= 1;
2595 /* Handle hangup and errors */
2596 for (i
= 0; i
< nb_fd
; i
++) {
2597 health_code_update();
2599 if (local_stream
[i
] == NULL
) {
2602 if (!local_stream
[i
]->hangup_flush_done
2603 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2604 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2605 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2606 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2608 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2609 /* Attempt read again, for the data we just flushed. */
2610 local_stream
[i
]->data_read
= 1;
2613 * If the poll flag is HUP/ERR/NVAL and we have
2614 * read no data in this pass, we can remove the
2615 * stream from its hash table.
2617 if ((pollfd
[i
].revents
& POLLHUP
)) {
2618 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2619 if (!local_stream
[i
]->data_read
) {
2620 consumer_del_stream(local_stream
[i
], data_ht
);
2621 local_stream
[i
] = NULL
;
2624 } else if (pollfd
[i
].revents
& POLLERR
) {
2625 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2626 if (!local_stream
[i
]->data_read
) {
2627 consumer_del_stream(local_stream
[i
], data_ht
);
2628 local_stream
[i
] = NULL
;
2631 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2632 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2633 if (!local_stream
[i
]->data_read
) {
2634 consumer_del_stream(local_stream
[i
], data_ht
);
2635 local_stream
[i
] = NULL
;
2639 if (local_stream
[i
] != NULL
) {
2640 local_stream
[i
]->data_read
= 0;
2647 DBG("polling thread exiting");
2652 * Close the write side of the pipe so epoll_wait() in
2653 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2654 * read side of the pipe. If we close them both, epoll_wait strangely does
2655 * not return and could create a endless wait period if the pipe is the
2656 * only tracked fd in the poll set. The thread will take care of closing
2659 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2664 ERR("Health error occurred in %s", __func__
);
2666 health_unregister(health_consumerd
);
2668 rcu_unregister_thread();
2673 * Close wake-up end of each stream belonging to the channel. This will
2674 * allow the poll() on the stream read-side to detect when the
2675 * write-side (application) finally closes them.
2678 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2680 struct lttng_ht
*ht
;
2681 struct lttng_consumer_stream
*stream
;
2682 struct lttng_ht_iter iter
;
2684 ht
= consumer_data
.stream_per_chan_id_ht
;
2687 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2688 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2689 ht
->match_fct
, &channel
->key
,
2690 &iter
.iter
, stream
, node_channel_id
.node
) {
2692 * Protect against teardown with mutex.
2694 pthread_mutex_lock(&stream
->lock
);
2695 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2698 switch (consumer_data
.type
) {
2699 case LTTNG_CONSUMER_KERNEL
:
2701 case LTTNG_CONSUMER32_UST
:
2702 case LTTNG_CONSUMER64_UST
:
2703 if (stream
->metadata_flag
) {
2704 /* Safe and protected by the stream lock. */
2705 lttng_ustconsumer_close_metadata(stream
->chan
);
2708 * Note: a mutex is taken internally within
2709 * liblttng-ust-ctl to protect timer wakeup_fd
2710 * use from concurrent close.
2712 lttng_ustconsumer_close_stream_wakeup(stream
);
2716 ERR("Unknown consumer_data type");
2720 pthread_mutex_unlock(&stream
->lock
);
2725 static void destroy_channel_ht(struct lttng_ht
*ht
)
2727 struct lttng_ht_iter iter
;
2728 struct lttng_consumer_channel
*channel
;
2736 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2737 ret
= lttng_ht_del(ht
, &iter
);
2742 lttng_ht_destroy(ht
);
2746 * This thread polls the channel fds to detect when they are being
2747 * closed. It closes all related streams if the channel is detected as
2748 * closed. It is currently only used as a shim layer for UST because the
2749 * consumerd needs to keep the per-stream wakeup end of pipes open for
2752 void *consumer_thread_channel_poll(void *data
)
2754 int ret
, i
, pollfd
, err
= -1;
2755 uint32_t revents
, nb_fd
;
2756 struct lttng_consumer_channel
*chan
= NULL
;
2757 struct lttng_ht_iter iter
;
2758 struct lttng_ht_node_u64
*node
;
2759 struct lttng_poll_event events
;
2760 struct lttng_consumer_local_data
*ctx
= data
;
2761 struct lttng_ht
*channel_ht
;
2763 rcu_register_thread();
2765 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
2767 if (testpoint(consumerd_thread_channel
)) {
2768 goto error_testpoint
;
2771 health_code_update();
2773 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2775 /* ENOMEM at this point. Better to bail out. */
2779 DBG("Thread channel poll started");
2781 /* Size is set to 1 for the consumer_channel pipe */
2782 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2784 ERR("Poll set creation failed");
2788 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2794 DBG("Channel main loop started");
2798 health_code_update();
2799 DBG("Channel poll wait");
2800 health_poll_entry();
2801 ret
= lttng_poll_wait(&events
, -1);
2802 DBG("Channel poll return from wait with %d fd(s)",
2803 LTTNG_POLL_GETNB(&events
));
2805 DBG("Channel event catched in thread");
2807 if (errno
== EINTR
) {
2808 ERR("Poll EINTR catched");
2811 if (LTTNG_POLL_GETNB(&events
) == 0) {
2812 err
= 0; /* All is OK */
2819 /* From here, the event is a channel wait fd */
2820 for (i
= 0; i
< nb_fd
; i
++) {
2821 health_code_update();
2823 revents
= LTTNG_POLL_GETEV(&events
, i
);
2824 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2827 /* No activity for this FD (poll implementation). */
2831 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2832 if (revents
& LPOLLIN
) {
2833 enum consumer_channel_action action
;
2836 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
2839 ERR("Error reading channel pipe");
2841 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2846 case CONSUMER_CHANNEL_ADD
:
2847 DBG("Adding channel %d to poll set",
2850 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
2853 lttng_ht_add_unique_u64(channel_ht
,
2854 &chan
->wait_fd_node
);
2856 /* Add channel to the global poll events list */
2857 lttng_poll_add(&events
, chan
->wait_fd
,
2858 LPOLLERR
| LPOLLHUP
);
2860 case CONSUMER_CHANNEL_DEL
:
2863 * This command should never be called if the channel
2864 * has streams monitored by either the data or metadata
2865 * thread. The consumer only notify this thread with a
2866 * channel del. command if it receives a destroy
2867 * channel command from the session daemon that send it
2868 * if a command prior to the GET_CHANNEL failed.
2872 chan
= consumer_find_channel(key
);
2875 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
2878 lttng_poll_del(&events
, chan
->wait_fd
);
2879 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
2880 ret
= lttng_ht_del(channel_ht
, &iter
);
2883 switch (consumer_data
.type
) {
2884 case LTTNG_CONSUMER_KERNEL
:
2886 case LTTNG_CONSUMER32_UST
:
2887 case LTTNG_CONSUMER64_UST
:
2888 health_code_update();
2889 /* Destroy streams that might have been left in the stream list. */
2890 clean_channel_stream_list(chan
);
2893 ERR("Unknown consumer_data type");
2898 * Release our own refcount. Force channel deletion even if
2899 * streams were not initialized.
2901 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
2902 consumer_del_channel(chan
);
2907 case CONSUMER_CHANNEL_QUIT
:
2909 * Remove the pipe from the poll set and continue the loop
2910 * since their might be data to consume.
2912 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2915 ERR("Unknown action");
2918 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2919 DBG("Channel thread pipe hung up");
2921 * Remove the pipe from the poll set and continue the loop
2922 * since their might be data to consume.
2924 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2927 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2931 /* Handle other stream */
2937 uint64_t tmp_id
= (uint64_t) pollfd
;
2939 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
2941 node
= lttng_ht_iter_get_node_u64(&iter
);
2944 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
2947 /* Check for error event */
2948 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2949 DBG("Channel fd %d is hup|err.", pollfd
);
2951 lttng_poll_del(&events
, chan
->wait_fd
);
2952 ret
= lttng_ht_del(channel_ht
, &iter
);
2956 * This will close the wait fd for each stream associated to
2957 * this channel AND monitored by the data/metadata thread thus
2958 * will be clean by the right thread.
2960 consumer_close_channel_streams(chan
);
2962 /* Release our own refcount */
2963 if (!uatomic_sub_return(&chan
->refcount
, 1)
2964 && !uatomic_read(&chan
->nb_init_stream_left
)) {
2965 consumer_del_channel(chan
);
2968 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2973 /* Release RCU lock for the channel looked up */
2981 lttng_poll_clean(&events
);
2983 destroy_channel_ht(channel_ht
);
2986 DBG("Channel poll thread exiting");
2989 ERR("Health error occurred in %s", __func__
);
2991 health_unregister(health_consumerd
);
2992 rcu_unregister_thread();
2996 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
2997 struct pollfd
*sockpoll
, int client_socket
)
3004 ret
= lttng_consumer_poll_socket(sockpoll
);
3008 DBG("Metadata connection on client_socket");
3010 /* Blocking call, waiting for transmission */
3011 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
3012 if (ctx
->consumer_metadata_socket
< 0) {
3013 WARN("On accept metadata");
3024 * This thread listens on the consumerd socket and receives the file
3025 * descriptors from the session daemon.
3027 void *consumer_thread_sessiond_poll(void *data
)
3029 int sock
= -1, client_socket
, ret
, err
= -1;
3031 * structure to poll for incoming data on communication socket avoids
3032 * making blocking sockets.
3034 struct pollfd consumer_sockpoll
[2];
3035 struct lttng_consumer_local_data
*ctx
= data
;
3037 rcu_register_thread();
3039 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
3041 if (testpoint(consumerd_thread_sessiond
)) {
3042 goto error_testpoint
;
3045 health_code_update();
3047 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
3048 unlink(ctx
->consumer_command_sock_path
);
3049 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
3050 if (client_socket
< 0) {
3051 ERR("Cannot create command socket");
3055 ret
= lttcomm_listen_unix_sock(client_socket
);
3060 DBG("Sending ready command to lttng-sessiond");
3061 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3062 /* return < 0 on error, but == 0 is not fatal */
3064 ERR("Error sending ready command to lttng-sessiond");
3068 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3069 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3070 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3071 consumer_sockpoll
[1].fd
= client_socket
;
3072 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3074 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3082 DBG("Connection on client_socket");
3084 /* Blocking call, waiting for transmission */
3085 sock
= lttcomm_accept_unix_sock(client_socket
);
3092 * Setup metadata socket which is the second socket connection on the
3093 * command unix socket.
3095 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3104 /* This socket is not useful anymore. */
3105 ret
= close(client_socket
);
3107 PERROR("close client_socket");
3111 /* update the polling structure to poll on the established socket */
3112 consumer_sockpoll
[1].fd
= sock
;
3113 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3116 health_code_update();
3118 health_poll_entry();
3119 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3128 DBG("Incoming command on sock");
3129 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3132 * This could simply be a session daemon quitting. Don't output
3135 DBG("Communication interrupted on command socket");
3139 if (consumer_quit
) {
3140 DBG("consumer_thread_receive_fds received quit from signal");
3141 err
= 0; /* All is OK */
3144 DBG("received command on sock");
3150 DBG("Consumer thread sessiond poll exiting");
3153 * Close metadata streams since the producer is the session daemon which
3156 * NOTE: for now, this only applies to the UST tracer.
3158 lttng_consumer_close_all_metadata();
3161 * when all fds have hung up, the polling thread
3167 * Notify the data poll thread to poll back again and test the
3168 * consumer_quit state that we just set so to quit gracefully.
3170 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3172 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3174 notify_health_quit_pipe(health_quit_pipe
);
3176 /* Cleaning up possibly open sockets. */
3180 PERROR("close sock sessiond poll");
3183 if (client_socket
>= 0) {
3184 ret
= close(client_socket
);
3186 PERROR("close client_socket sessiond poll");
3193 ERR("Health error occurred in %s", __func__
);
3195 health_unregister(health_consumerd
);
3197 rcu_unregister_thread();
3201 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3202 struct lttng_consumer_local_data
*ctx
)
3206 pthread_mutex_lock(&stream
->lock
);
3207 if (stream
->metadata_flag
) {
3208 pthread_mutex_lock(&stream
->metadata_rdv_lock
);
3211 switch (consumer_data
.type
) {
3212 case LTTNG_CONSUMER_KERNEL
:
3213 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
3215 case LTTNG_CONSUMER32_UST
:
3216 case LTTNG_CONSUMER64_UST
:
3217 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
3220 ERR("Unknown consumer_data type");
3226 if (stream
->metadata_flag
) {
3227 pthread_cond_broadcast(&stream
->metadata_rdv
);
3228 pthread_mutex_unlock(&stream
->metadata_rdv_lock
);
3230 pthread_mutex_unlock(&stream
->lock
);
3234 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3236 switch (consumer_data
.type
) {
3237 case LTTNG_CONSUMER_KERNEL
:
3238 return lttng_kconsumer_on_recv_stream(stream
);
3239 case LTTNG_CONSUMER32_UST
:
3240 case LTTNG_CONSUMER64_UST
:
3241 return lttng_ustconsumer_on_recv_stream(stream
);
3243 ERR("Unknown consumer_data type");
3250 * Allocate and set consumer data hash tables.
3252 int lttng_consumer_init(void)
3254 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3255 if (!consumer_data
.channel_ht
) {
3259 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3260 if (!consumer_data
.relayd_ht
) {
3264 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3265 if (!consumer_data
.stream_list_ht
) {
3269 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3270 if (!consumer_data
.stream_per_chan_id_ht
) {
3274 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3279 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3291 * Process the ADD_RELAYD command receive by a consumer.
3293 * This will create a relayd socket pair and add it to the relayd hash table.
3294 * The caller MUST acquire a RCU read side lock before calling it.
3296 int consumer_add_relayd_socket(uint64_t net_seq_idx
, int sock_type
,
3297 struct lttng_consumer_local_data
*ctx
, int sock
,
3298 struct pollfd
*consumer_sockpoll
,
3299 struct lttcomm_relayd_sock
*relayd_sock
, uint64_t sessiond_id
,
3300 uint64_t relayd_session_id
)
3302 int fd
= -1, ret
= -1, relayd_created
= 0;
3303 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3304 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3307 assert(relayd_sock
);
3309 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", net_seq_idx
);
3311 /* Get relayd reference if exists. */
3312 relayd
= consumer_find_relayd(net_seq_idx
);
3313 if (relayd
== NULL
) {
3314 assert(sock_type
== LTTNG_STREAM_CONTROL
);
3315 /* Not found. Allocate one. */
3316 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3317 if (relayd
== NULL
) {
3319 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3322 relayd
->sessiond_session_id
= sessiond_id
;
3327 * This code path MUST continue to the consumer send status message to
3328 * we can notify the session daemon and continue our work without
3329 * killing everything.
3333 * relayd key should never be found for control socket.
3335 assert(sock_type
!= LTTNG_STREAM_CONTROL
);
3338 /* First send a status message before receiving the fds. */
3339 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3341 /* Somehow, the session daemon is not responding anymore. */
3342 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3343 goto error_nosignal
;
3346 /* Poll on consumer socket. */
3347 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3349 /* Needing to exit in the middle of a command: error. */
3350 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3352 goto error_nosignal
;
3355 /* Get relayd socket from session daemon */
3356 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3357 if (ret
!= sizeof(fd
)) {
3359 fd
= -1; /* Just in case it gets set with an invalid value. */
3362 * Failing to receive FDs might indicate a major problem such as
3363 * reaching a fd limit during the receive where the kernel returns a
3364 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3365 * don't take any chances and stop everything.
3367 * XXX: Feature request #558 will fix that and avoid this possible
3368 * issue when reaching the fd limit.
3370 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3371 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3375 /* Copy socket information and received FD */
3376 switch (sock_type
) {
3377 case LTTNG_STREAM_CONTROL
:
3378 /* Copy received lttcomm socket */
3379 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3380 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3381 /* Handle create_sock error. */
3383 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3387 * Close the socket created internally by
3388 * lttcomm_create_sock, so we can replace it by the one
3389 * received from sessiond.
3391 if (close(relayd
->control_sock
.sock
.fd
)) {
3395 /* Assign new file descriptor */
3396 relayd
->control_sock
.sock
.fd
= fd
;
3397 fd
= -1; /* For error path */
3398 /* Assign version values. */
3399 relayd
->control_sock
.major
= relayd_sock
->major
;
3400 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3402 relayd
->relayd_session_id
= relayd_session_id
;
3405 case LTTNG_STREAM_DATA
:
3406 /* Copy received lttcomm socket */
3407 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3408 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3409 /* Handle create_sock error. */
3411 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3415 * Close the socket created internally by
3416 * lttcomm_create_sock, so we can replace it by the one
3417 * received from sessiond.
3419 if (close(relayd
->data_sock
.sock
.fd
)) {
3423 /* Assign new file descriptor */
3424 relayd
->data_sock
.sock
.fd
= fd
;
3425 fd
= -1; /* for eventual error paths */
3426 /* Assign version values. */
3427 relayd
->data_sock
.major
= relayd_sock
->major
;
3428 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3431 ERR("Unknown relayd socket type (%d)", sock_type
);
3433 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3437 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3438 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3439 relayd
->net_seq_idx
, fd
);
3441 /* We successfully added the socket. Send status back. */
3442 ret
= consumer_send_status_msg(sock
, ret_code
);
3444 /* Somehow, the session daemon is not responding anymore. */
3445 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3446 goto error_nosignal
;
3450 * Add relayd socket pair to consumer data hashtable. If object already
3451 * exists or on error, the function gracefully returns.
3459 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3460 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3464 /* Close received socket if valid. */
3467 PERROR("close received socket");
3471 if (relayd_created
) {
3479 * Try to lock the stream mutex.
3481 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
3483 static int stream_try_lock(struct lttng_consumer_stream
*stream
)
3490 * Try to lock the stream mutex. On failure, we know that the stream is
3491 * being used else where hence there is data still being extracted.
3493 ret
= pthread_mutex_trylock(&stream
->lock
);
3495 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
3507 * Search for a relayd associated to the session id and return the reference.
3509 * A rcu read side lock MUST be acquire before calling this function and locked
3510 * until the relayd object is no longer necessary.
3512 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3514 struct lttng_ht_iter iter
;
3515 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3517 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3518 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3521 * Check by sessiond id which is unique here where the relayd session
3522 * id might not be when having multiple relayd.
3524 if (relayd
->sessiond_session_id
== id
) {
3525 /* Found the relayd. There can be only one per id. */
3537 * Check if for a given session id there is still data needed to be extract
3540 * Return 1 if data is pending or else 0 meaning ready to be read.
3542 int consumer_data_pending(uint64_t id
)
3545 struct lttng_ht_iter iter
;
3546 struct lttng_ht
*ht
;
3547 struct lttng_consumer_stream
*stream
;
3548 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3549 int (*data_pending
)(struct lttng_consumer_stream
*);
3551 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3554 pthread_mutex_lock(&consumer_data
.lock
);
3556 switch (consumer_data
.type
) {
3557 case LTTNG_CONSUMER_KERNEL
:
3558 data_pending
= lttng_kconsumer_data_pending
;
3560 case LTTNG_CONSUMER32_UST
:
3561 case LTTNG_CONSUMER64_UST
:
3562 data_pending
= lttng_ustconsumer_data_pending
;
3565 ERR("Unknown consumer data type");
3569 /* Ease our life a bit */
3570 ht
= consumer_data
.stream_list_ht
;
3572 relayd
= find_relayd_by_session_id(id
);
3574 /* Send init command for data pending. */
3575 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3576 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3577 relayd
->relayd_session_id
);
3578 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3580 /* Communication error thus the relayd so no data pending. */
3581 goto data_not_pending
;
3585 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3586 ht
->hash_fct(&id
, lttng_ht_seed
),
3588 &iter
.iter
, stream
, node_session_id
.node
) {
3589 /* If this call fails, the stream is being used hence data pending. */
3590 ret
= stream_try_lock(stream
);
3596 * A removed node from the hash table indicates that the stream has
3597 * been deleted thus having a guarantee that the buffers are closed
3598 * on the consumer side. However, data can still be transmitted
3599 * over the network so don't skip the relayd check.
3601 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3603 /* Check the stream if there is data in the buffers. */
3604 ret
= data_pending(stream
);
3606 pthread_mutex_unlock(&stream
->lock
);
3613 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3614 if (stream
->metadata_flag
) {
3615 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3616 stream
->relayd_stream_id
);
3618 ret
= relayd_data_pending(&relayd
->control_sock
,
3619 stream
->relayd_stream_id
,
3620 stream
->next_net_seq_num
- 1);
3622 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3624 pthread_mutex_unlock(&stream
->lock
);
3628 pthread_mutex_unlock(&stream
->lock
);
3632 unsigned int is_data_inflight
= 0;
3634 /* Send init command for data pending. */
3635 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3636 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3637 relayd
->relayd_session_id
, &is_data_inflight
);
3638 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3640 goto data_not_pending
;
3642 if (is_data_inflight
) {
3648 * Finding _no_ node in the hash table and no inflight data means that the
3649 * stream(s) have been removed thus data is guaranteed to be available for
3650 * analysis from the trace files.
3654 /* Data is available to be read by a viewer. */
3655 pthread_mutex_unlock(&consumer_data
.lock
);
3660 /* Data is still being extracted from buffers. */
3661 pthread_mutex_unlock(&consumer_data
.lock
);
3667 * Send a ret code status message to the sessiond daemon.
3669 * Return the sendmsg() return value.
3671 int consumer_send_status_msg(int sock
, int ret_code
)
3673 struct lttcomm_consumer_status_msg msg
;
3675 memset(&msg
, 0, sizeof(msg
));
3676 msg
.ret_code
= ret_code
;
3678 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3682 * Send a channel status message to the sessiond daemon.
3684 * Return the sendmsg() return value.
3686 int consumer_send_status_channel(int sock
,
3687 struct lttng_consumer_channel
*channel
)
3689 struct lttcomm_consumer_status_channel msg
;
3693 memset(&msg
, 0, sizeof(msg
));
3695 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3697 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3698 msg
.key
= channel
->key
;
3699 msg
.stream_count
= channel
->streams
.count
;
3702 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3705 unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos
,
3706 unsigned long produced_pos
, uint64_t nb_packets_per_stream
,
3707 uint64_t max_sb_size
)
3709 unsigned long start_pos
;
3711 if (!nb_packets_per_stream
) {
3712 return consumed_pos
; /* Grab everything */
3714 start_pos
= produced_pos
- offset_align_floor(produced_pos
, max_sb_size
);
3715 start_pos
-= max_sb_size
* nb_packets_per_stream
;
3716 if ((long) (start_pos
- consumed_pos
) < 0) {
3717 return consumed_pos
; /* Grab everything */