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
= close(stream
->index_fd
);
1603 PERROR("Closing index");
1606 stream
->index_fd
= -1;
1607 ret
= index_create_file(stream
->chan
->pathname
,
1608 stream
->name
, stream
->uid
, stream
->gid
,
1609 stream
->chan
->tracefile_size
,
1610 stream
->tracefile_count_current
);
1614 stream
->index_fd
= ret
;
1617 /* Reset current size because we just perform a rotation. */
1618 stream
->tracefile_size_current
= 0;
1619 stream
->out_fd_offset
= 0;
1622 stream
->tracefile_size_current
+= len
;
1624 index
->offset
= htobe64(stream
->out_fd_offset
);
1629 * This call guarantee that len or less is returned. It's impossible to
1630 * receive a ret value that is bigger than len.
1632 ret
= lttng_write(outfd
, mmap_base
+ mmap_offset
, len
);
1633 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, len
);
1634 if (ret
< 0 || ((size_t) ret
!= len
)) {
1636 * Report error to caller if nothing was written else at least send the
1644 /* Socket operation failed. We consider the relayd dead */
1645 if (errno
== EPIPE
|| errno
== EINVAL
|| errno
== EBADF
) {
1647 * This is possible if the fd is closed on the other side
1648 * (outfd) or any write problem. It can be verbose a bit for a
1649 * normal execution if for instance the relayd is stopped
1650 * abruptly. This can happen so set this to a DBG statement.
1652 DBG("Consumer mmap write detected relayd hang up");
1654 /* Unhandled error, print it and stop function right now. */
1655 PERROR("Error in write mmap (ret %zd != len %lu)", ret
, len
);
1659 stream
->output_written
+= ret
;
1661 /* This call is useless on a socket so better save a syscall. */
1663 /* This won't block, but will start writeout asynchronously */
1664 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, len
,
1665 SYNC_FILE_RANGE_WRITE
);
1666 stream
->out_fd_offset
+= len
;
1668 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1672 * This is a special case that the relayd has closed its socket. Let's
1673 * cleanup the relayd object and all associated streams.
1675 if (relayd
&& relayd_hang_up
) {
1676 cleanup_relayd(relayd
, ctx
);
1680 /* Unlock only if ctrl socket used */
1681 if (relayd
&& stream
->metadata_flag
) {
1682 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1690 * Splice the data from the ring buffer to the tracefile.
1692 * It must be called with the stream lock held.
1694 * Returns the number of bytes spliced.
1696 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1697 struct lttng_consumer_local_data
*ctx
,
1698 struct lttng_consumer_stream
*stream
, unsigned long len
,
1699 unsigned long padding
,
1700 struct ctf_packet_index
*index
)
1702 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1704 off_t orig_offset
= stream
->out_fd_offset
;
1705 int fd
= stream
->wait_fd
;
1706 /* Default is on the disk */
1707 int outfd
= stream
->out_fd
;
1708 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1710 unsigned int relayd_hang_up
= 0;
1712 switch (consumer_data
.type
) {
1713 case LTTNG_CONSUMER_KERNEL
:
1715 case LTTNG_CONSUMER32_UST
:
1716 case LTTNG_CONSUMER64_UST
:
1717 /* Not supported for user space tracing */
1720 ERR("Unknown consumer_data type");
1724 /* RCU lock for the relayd pointer */
1727 /* Flag that the current stream if set for network streaming. */
1728 if (stream
->net_seq_idx
!= (uint64_t) -1ULL) {
1729 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1730 if (relayd
== NULL
) {
1735 splice_pipe
= stream
->splice_pipe
;
1737 /* Write metadata stream id before payload */
1739 unsigned long total_len
= len
;
1741 if (stream
->metadata_flag
) {
1743 * Lock the control socket for the complete duration of the function
1744 * since from this point on we will use the socket.
1746 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1748 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
, relayd
,
1756 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1759 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1765 /* Use the returned socket. */
1768 /* No streaming, we have to set the len with the full padding */
1772 * Check if we need to change the tracefile before writing the packet.
1774 if (stream
->chan
->tracefile_size
> 0 &&
1775 (stream
->tracefile_size_current
+ len
) >
1776 stream
->chan
->tracefile_size
) {
1777 ret
= utils_rotate_stream_file(stream
->chan
->pathname
,
1778 stream
->name
, stream
->chan
->tracefile_size
,
1779 stream
->chan
->tracefile_count
, stream
->uid
, stream
->gid
,
1780 stream
->out_fd
, &(stream
->tracefile_count_current
),
1784 ERR("Rotating output file");
1787 outfd
= stream
->out_fd
;
1789 if (stream
->index_fd
>= 0) {
1790 ret
= close(stream
->index_fd
);
1792 PERROR("Closing index");
1795 stream
->index_fd
= -1;
1796 ret
= index_create_file(stream
->chan
->pathname
,
1797 stream
->name
, stream
->uid
, stream
->gid
,
1798 stream
->chan
->tracefile_size
,
1799 stream
->tracefile_count_current
);
1804 stream
->index_fd
= ret
;
1807 /* Reset current size because we just perform a rotation. */
1808 stream
->tracefile_size_current
= 0;
1809 stream
->out_fd_offset
= 0;
1812 stream
->tracefile_size_current
+= len
;
1813 index
->offset
= htobe64(stream
->out_fd_offset
);
1817 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1818 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1819 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1820 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1821 DBG("splice chan to pipe, ret %zd", ret_splice
);
1822 if (ret_splice
< 0) {
1825 PERROR("Error in relay splice");
1829 /* Handle stream on the relayd if the output is on the network */
1830 if (relayd
&& stream
->metadata_flag
) {
1831 size_t metadata_payload_size
=
1832 sizeof(struct lttcomm_relayd_metadata_payload
);
1834 /* Update counter to fit the spliced data */
1835 ret_splice
+= metadata_payload_size
;
1836 len
+= metadata_payload_size
;
1838 * We do this so the return value can match the len passed as
1839 * argument to this function.
1841 written
-= metadata_payload_size
;
1844 /* Splice data out */
1845 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1846 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1847 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
1849 if (ret_splice
< 0) {
1854 } else if (ret_splice
> len
) {
1856 * We don't expect this code path to be executed but you never know
1857 * so this is an extra protection agains a buggy splice().
1860 written
+= ret_splice
;
1861 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice
,
1865 /* All good, update current len and continue. */
1869 /* This call is useless on a socket so better save a syscall. */
1871 /* This won't block, but will start writeout asynchronously */
1872 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1873 SYNC_FILE_RANGE_WRITE
);
1874 stream
->out_fd_offset
+= ret_splice
;
1876 stream
->output_written
+= ret_splice
;
1877 written
+= ret_splice
;
1879 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1884 * This is a special case that the relayd has closed its socket. Let's
1885 * cleanup the relayd object and all associated streams.
1887 if (relayd
&& relayd_hang_up
) {
1888 cleanup_relayd(relayd
, ctx
);
1889 /* Skip splice error so the consumer does not fail */
1894 /* send the appropriate error description to sessiond */
1897 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1900 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1903 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
1908 if (relayd
&& stream
->metadata_flag
) {
1909 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1917 * Take a snapshot for a specific fd
1919 * Returns 0 on success, < 0 on error
1921 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
1923 switch (consumer_data
.type
) {
1924 case LTTNG_CONSUMER_KERNEL
:
1925 return lttng_kconsumer_take_snapshot(stream
);
1926 case LTTNG_CONSUMER32_UST
:
1927 case LTTNG_CONSUMER64_UST
:
1928 return lttng_ustconsumer_take_snapshot(stream
);
1930 ERR("Unknown consumer_data type");
1937 * Get the produced position
1939 * Returns 0 on success, < 0 on error
1941 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
1944 switch (consumer_data
.type
) {
1945 case LTTNG_CONSUMER_KERNEL
:
1946 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
1947 case LTTNG_CONSUMER32_UST
:
1948 case LTTNG_CONSUMER64_UST
:
1949 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
1951 ERR("Unknown consumer_data type");
1957 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
1958 int sock
, struct pollfd
*consumer_sockpoll
)
1960 switch (consumer_data
.type
) {
1961 case LTTNG_CONSUMER_KERNEL
:
1962 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1963 case LTTNG_CONSUMER32_UST
:
1964 case LTTNG_CONSUMER64_UST
:
1965 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1967 ERR("Unknown consumer_data type");
1973 void lttng_consumer_close_all_metadata(void)
1975 switch (consumer_data
.type
) {
1976 case LTTNG_CONSUMER_KERNEL
:
1978 * The Kernel consumer has a different metadata scheme so we don't
1979 * close anything because the stream will be closed by the session
1983 case LTTNG_CONSUMER32_UST
:
1984 case LTTNG_CONSUMER64_UST
:
1986 * Close all metadata streams. The metadata hash table is passed and
1987 * this call iterates over it by closing all wakeup fd. This is safe
1988 * because at this point we are sure that the metadata producer is
1989 * either dead or blocked.
1991 lttng_ustconsumer_close_all_metadata(metadata_ht
);
1994 ERR("Unknown consumer_data type");
2000 * Clean up a metadata stream and free its memory.
2002 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
2003 struct lttng_ht
*ht
)
2005 struct lttng_consumer_channel
*free_chan
= NULL
;
2009 * This call should NEVER receive regular stream. It must always be
2010 * metadata stream and this is crucial for data structure synchronization.
2012 assert(stream
->metadata_flag
);
2014 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
2016 pthread_mutex_lock(&consumer_data
.lock
);
2017 pthread_mutex_lock(&stream
->chan
->lock
);
2018 pthread_mutex_lock(&stream
->lock
);
2020 /* Remove any reference to that stream. */
2021 consumer_stream_delete(stream
, ht
);
2023 /* Close down everything including the relayd if one. */
2024 consumer_stream_close(stream
);
2025 /* Destroy tracer buffers of the stream. */
2026 consumer_stream_destroy_buffers(stream
);
2028 /* Atomically decrement channel refcount since other threads can use it. */
2029 if (!uatomic_sub_return(&stream
->chan
->refcount
, 1)
2030 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
2031 /* Go for channel deletion! */
2032 free_chan
= stream
->chan
;
2036 * Nullify the stream reference so it is not used after deletion. The
2037 * channel lock MUST be acquired before being able to check for a NULL
2040 stream
->chan
->metadata_stream
= NULL
;
2042 pthread_mutex_unlock(&stream
->lock
);
2043 pthread_mutex_unlock(&stream
->chan
->lock
);
2044 pthread_mutex_unlock(&consumer_data
.lock
);
2047 consumer_del_channel(free_chan
);
2050 consumer_stream_free(stream
);
2054 * Action done with the metadata stream when adding it to the consumer internal
2055 * data structures to handle it.
2057 int consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
2059 struct lttng_ht
*ht
= metadata_ht
;
2061 struct lttng_ht_iter iter
;
2062 struct lttng_ht_node_u64
*node
;
2067 DBG3("Adding metadata stream %" PRIu64
" to hash table", stream
->key
);
2069 pthread_mutex_lock(&consumer_data
.lock
);
2070 pthread_mutex_lock(&stream
->chan
->lock
);
2071 pthread_mutex_lock(&stream
->chan
->timer_lock
);
2072 pthread_mutex_lock(&stream
->lock
);
2075 * From here, refcounts are updated so be _careful_ when returning an error
2082 * Lookup the stream just to make sure it does not exist in our internal
2083 * state. This should NEVER happen.
2085 lttng_ht_lookup(ht
, &stream
->key
, &iter
);
2086 node
= lttng_ht_iter_get_node_u64(&iter
);
2090 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2091 * in terms of destroying the associated channel, because the action that
2092 * causes the count to become 0 also causes a stream to be added. The
2093 * channel deletion will thus be triggered by the following removal of this
2096 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
2097 /* Increment refcount before decrementing nb_init_stream_left */
2099 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
2102 lttng_ht_add_unique_u64(ht
, &stream
->node
);
2104 lttng_ht_add_unique_u64(consumer_data
.stream_per_chan_id_ht
,
2105 &stream
->node_channel_id
);
2108 * Add stream to the stream_list_ht of the consumer data. No need to steal
2109 * the key since the HT does not use it and we allow to add redundant keys
2112 lttng_ht_add_u64(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
2116 pthread_mutex_unlock(&stream
->lock
);
2117 pthread_mutex_unlock(&stream
->chan
->lock
);
2118 pthread_mutex_unlock(&stream
->chan
->timer_lock
);
2119 pthread_mutex_unlock(&consumer_data
.lock
);
2124 * Delete data stream that are flagged for deletion (endpoint_status).
2126 static void validate_endpoint_status_data_stream(void)
2128 struct lttng_ht_iter iter
;
2129 struct lttng_consumer_stream
*stream
;
2131 DBG("Consumer delete flagged data stream");
2134 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2135 /* Validate delete flag of the stream */
2136 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2139 /* Delete it right now */
2140 consumer_del_stream(stream
, data_ht
);
2146 * Delete metadata stream that are flagged for deletion (endpoint_status).
2148 static void validate_endpoint_status_metadata_stream(
2149 struct lttng_poll_event
*pollset
)
2151 struct lttng_ht_iter iter
;
2152 struct lttng_consumer_stream
*stream
;
2154 DBG("Consumer delete flagged metadata stream");
2159 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
2160 /* Validate delete flag of the stream */
2161 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
2165 * Remove from pollset so the metadata thread can continue without
2166 * blocking on a deleted stream.
2168 lttng_poll_del(pollset
, stream
->wait_fd
);
2170 /* Delete it right now */
2171 consumer_del_metadata_stream(stream
, metadata_ht
);
2177 * Thread polls on metadata file descriptor and write them on disk or on the
2180 void *consumer_thread_metadata_poll(void *data
)
2182 int ret
, i
, pollfd
, err
= -1;
2183 uint32_t revents
, nb_fd
;
2184 struct lttng_consumer_stream
*stream
= NULL
;
2185 struct lttng_ht_iter iter
;
2186 struct lttng_ht_node_u64
*node
;
2187 struct lttng_poll_event events
;
2188 struct lttng_consumer_local_data
*ctx
= data
;
2191 rcu_register_thread();
2193 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_METADATA
);
2195 if (testpoint(consumerd_thread_metadata
)) {
2196 goto error_testpoint
;
2199 health_code_update();
2201 DBG("Thread metadata poll started");
2203 /* Size is set to 1 for the consumer_metadata pipe */
2204 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2206 ERR("Poll set creation failed");
2210 ret
= lttng_poll_add(&events
,
2211 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
), LPOLLIN
);
2217 DBG("Metadata main loop started");
2221 health_code_update();
2222 health_poll_entry();
2223 DBG("Metadata poll wait");
2224 ret
= lttng_poll_wait(&events
, -1);
2225 DBG("Metadata poll return from wait with %d fd(s)",
2226 LTTNG_POLL_GETNB(&events
));
2228 DBG("Metadata event caught in thread");
2230 if (errno
== EINTR
) {
2231 ERR("Poll EINTR caught");
2234 if (LTTNG_POLL_GETNB(&events
) == 0) {
2235 err
= 0; /* All is OK */
2242 /* From here, the event is a metadata wait fd */
2243 for (i
= 0; i
< nb_fd
; i
++) {
2244 health_code_update();
2246 revents
= LTTNG_POLL_GETEV(&events
, i
);
2247 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2250 /* No activity for this FD (poll implementation). */
2254 if (pollfd
== lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
)) {
2255 if (revents
& LPOLLIN
) {
2258 pipe_len
= lttng_pipe_read(ctx
->consumer_metadata_pipe
,
2259 &stream
, sizeof(stream
));
2260 if (pipe_len
< sizeof(stream
)) {
2262 PERROR("read metadata stream");
2265 * Remove the pipe from the poll set and continue the loop
2266 * since their might be data to consume.
2268 lttng_poll_del(&events
,
2269 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2270 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2274 /* A NULL stream means that the state has changed. */
2275 if (stream
== NULL
) {
2276 /* Check for deleted streams. */
2277 validate_endpoint_status_metadata_stream(&events
);
2281 DBG("Adding metadata stream %d to poll set",
2284 /* Add metadata stream to the global poll events list */
2285 lttng_poll_add(&events
, stream
->wait_fd
,
2286 LPOLLIN
| LPOLLPRI
| LPOLLHUP
);
2287 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2288 DBG("Metadata thread pipe hung up");
2290 * Remove the pipe from the poll set and continue the loop
2291 * since their might be data to consume.
2293 lttng_poll_del(&events
,
2294 lttng_pipe_get_readfd(ctx
->consumer_metadata_pipe
));
2295 lttng_pipe_read_close(ctx
->consumer_metadata_pipe
);
2298 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2302 /* Handle other stream */
2308 uint64_t tmp_id
= (uint64_t) pollfd
;
2310 lttng_ht_lookup(metadata_ht
, &tmp_id
, &iter
);
2312 node
= lttng_ht_iter_get_node_u64(&iter
);
2315 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2318 if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2319 /* Get the data out of the metadata file descriptor */
2320 DBG("Metadata available on fd %d", pollfd
);
2321 assert(stream
->wait_fd
== pollfd
);
2324 health_code_update();
2326 len
= ctx
->on_buffer_ready(stream
, ctx
);
2328 * We don't check the return value here since if we get
2329 * a negative len, it means an error occured thus we
2330 * simply remove it from the poll set and free the
2335 /* It's ok to have an unavailable sub-buffer */
2336 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2337 /* Clean up stream from consumer and free it. */
2338 lttng_poll_del(&events
, stream
->wait_fd
);
2339 consumer_del_metadata_stream(stream
, metadata_ht
);
2341 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2342 DBG("Metadata fd %d is hup|err.", pollfd
);
2343 if (!stream
->hangup_flush_done
2344 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2345 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2346 DBG("Attempting to flush and consume the UST buffers");
2347 lttng_ustconsumer_on_stream_hangup(stream
);
2349 /* We just flushed the stream now read it. */
2351 health_code_update();
2353 len
= ctx
->on_buffer_ready(stream
, ctx
);
2355 * We don't check the return value here since if we get
2356 * a negative len, it means an error occured thus we
2357 * simply remove it from the poll set and free the
2363 lttng_poll_del(&events
, stream
->wait_fd
);
2365 * This call update the channel states, closes file descriptors
2366 * and securely free the stream.
2368 consumer_del_metadata_stream(stream
, metadata_ht
);
2370 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2374 /* Release RCU lock for the stream looked up */
2382 DBG("Metadata poll thread exiting");
2384 lttng_poll_clean(&events
);
2389 ERR("Health error occurred in %s", __func__
);
2391 health_unregister(health_consumerd
);
2392 rcu_unregister_thread();
2397 * This thread polls the fds in the set to consume the data and write
2398 * it to tracefile if necessary.
2400 void *consumer_thread_data_poll(void *data
)
2402 int num_rdy
, num_hup
, high_prio
, ret
, i
, err
= -1;
2403 struct pollfd
*pollfd
= NULL
;
2404 /* local view of the streams */
2405 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2406 /* local view of consumer_data.fds_count */
2408 struct lttng_consumer_local_data
*ctx
= data
;
2411 rcu_register_thread();
2413 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_DATA
);
2415 if (testpoint(consumerd_thread_data
)) {
2416 goto error_testpoint
;
2419 health_code_update();
2421 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
*));
2422 if (local_stream
== NULL
) {
2423 PERROR("local_stream malloc");
2428 health_code_update();
2434 * the fds set has been updated, we need to update our
2435 * local array as well
2437 pthread_mutex_lock(&consumer_data
.lock
);
2438 if (consumer_data
.need_update
) {
2443 local_stream
= NULL
;
2446 * Allocate for all fds +1 for the consumer_data_pipe and +1 for
2449 pollfd
= zmalloc((consumer_data
.stream_count
+ 2) * sizeof(struct pollfd
));
2450 if (pollfd
== NULL
) {
2451 PERROR("pollfd malloc");
2452 pthread_mutex_unlock(&consumer_data
.lock
);
2456 local_stream
= zmalloc((consumer_data
.stream_count
+ 2) *
2457 sizeof(struct lttng_consumer_stream
*));
2458 if (local_stream
== NULL
) {
2459 PERROR("local_stream malloc");
2460 pthread_mutex_unlock(&consumer_data
.lock
);
2463 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2466 ERR("Error in allocating pollfd or local_outfds");
2467 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2468 pthread_mutex_unlock(&consumer_data
.lock
);
2472 consumer_data
.need_update
= 0;
2474 pthread_mutex_unlock(&consumer_data
.lock
);
2476 /* No FDs and consumer_quit, consumer_cleanup the thread */
2477 if (nb_fd
== 0 && consumer_quit
== 1) {
2478 err
= 0; /* All is OK */
2481 /* poll on the array of fds */
2483 DBG("polling on %d fd", nb_fd
+ 2);
2484 health_poll_entry();
2485 num_rdy
= poll(pollfd
, nb_fd
+ 2, -1);
2487 DBG("poll num_rdy : %d", num_rdy
);
2488 if (num_rdy
== -1) {
2490 * Restart interrupted system call.
2492 if (errno
== EINTR
) {
2495 PERROR("Poll error");
2496 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2498 } else if (num_rdy
== 0) {
2499 DBG("Polling thread timed out");
2504 * If the consumer_data_pipe triggered poll go directly to the
2505 * beginning of the loop to update the array. We want to prioritize
2506 * array update over low-priority reads.
2508 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2509 ssize_t pipe_readlen
;
2511 DBG("consumer_data_pipe wake up");
2512 pipe_readlen
= lttng_pipe_read(ctx
->consumer_data_pipe
,
2513 &new_stream
, sizeof(new_stream
));
2514 if (pipe_readlen
< sizeof(new_stream
)) {
2515 PERROR("Consumer data pipe");
2516 /* Continue so we can at least handle the current stream(s). */
2521 * If the stream is NULL, just ignore it. It's also possible that
2522 * the sessiond poll thread changed the consumer_quit state and is
2523 * waking us up to test it.
2525 if (new_stream
== NULL
) {
2526 validate_endpoint_status_data_stream();
2530 /* Continue to update the local streams and handle prio ones */
2534 /* Handle wakeup pipe. */
2535 if (pollfd
[nb_fd
+ 1].revents
& (POLLIN
| POLLPRI
)) {
2537 ssize_t pipe_readlen
;
2539 pipe_readlen
= lttng_pipe_read(ctx
->consumer_wakeup_pipe
, &dummy
,
2541 if (pipe_readlen
< 0) {
2542 PERROR("Consumer data wakeup pipe");
2544 /* We've been awakened to handle stream(s). */
2545 ctx
->has_wakeup
= 0;
2548 /* Take care of high priority channels first. */
2549 for (i
= 0; i
< nb_fd
; i
++) {
2550 health_code_update();
2552 if (local_stream
[i
] == NULL
) {
2555 if (pollfd
[i
].revents
& POLLPRI
) {
2556 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2558 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2559 /* it's ok to have an unavailable sub-buffer */
2560 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2561 /* Clean the stream and free it. */
2562 consumer_del_stream(local_stream
[i
], data_ht
);
2563 local_stream
[i
] = NULL
;
2564 } else if (len
> 0) {
2565 local_stream
[i
]->data_read
= 1;
2571 * If we read high prio channel in this loop, try again
2572 * for more high prio data.
2578 /* Take care of low priority channels. */
2579 for (i
= 0; i
< nb_fd
; i
++) {
2580 health_code_update();
2582 if (local_stream
[i
] == NULL
) {
2585 if ((pollfd
[i
].revents
& POLLIN
) ||
2586 local_stream
[i
]->hangup_flush_done
||
2587 local_stream
[i
]->has_data
) {
2588 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2589 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2590 /* it's ok to have an unavailable sub-buffer */
2591 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2592 /* Clean the stream and free it. */
2593 consumer_del_stream(local_stream
[i
], data_ht
);
2594 local_stream
[i
] = NULL
;
2595 } else if (len
> 0) {
2596 local_stream
[i
]->data_read
= 1;
2601 /* Handle hangup and errors */
2602 for (i
= 0; i
< nb_fd
; i
++) {
2603 health_code_update();
2605 if (local_stream
[i
] == NULL
) {
2608 if (!local_stream
[i
]->hangup_flush_done
2609 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2610 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2611 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2612 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2614 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2615 /* Attempt read again, for the data we just flushed. */
2616 local_stream
[i
]->data_read
= 1;
2619 * If the poll flag is HUP/ERR/NVAL and we have
2620 * read no data in this pass, we can remove the
2621 * stream from its hash table.
2623 if ((pollfd
[i
].revents
& POLLHUP
)) {
2624 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2625 if (!local_stream
[i
]->data_read
) {
2626 consumer_del_stream(local_stream
[i
], data_ht
);
2627 local_stream
[i
] = NULL
;
2630 } else if (pollfd
[i
].revents
& POLLERR
) {
2631 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2632 if (!local_stream
[i
]->data_read
) {
2633 consumer_del_stream(local_stream
[i
], data_ht
);
2634 local_stream
[i
] = NULL
;
2637 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2638 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2639 if (!local_stream
[i
]->data_read
) {
2640 consumer_del_stream(local_stream
[i
], data_ht
);
2641 local_stream
[i
] = NULL
;
2645 if (local_stream
[i
] != NULL
) {
2646 local_stream
[i
]->data_read
= 0;
2653 DBG("polling thread exiting");
2658 * Close the write side of the pipe so epoll_wait() in
2659 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2660 * read side of the pipe. If we close them both, epoll_wait strangely does
2661 * not return and could create a endless wait period if the pipe is the
2662 * only tracked fd in the poll set. The thread will take care of closing
2665 (void) lttng_pipe_write_close(ctx
->consumer_metadata_pipe
);
2670 ERR("Health error occurred in %s", __func__
);
2672 health_unregister(health_consumerd
);
2674 rcu_unregister_thread();
2679 * Close wake-up end of each stream belonging to the channel. This will
2680 * allow the poll() on the stream read-side to detect when the
2681 * write-side (application) finally closes them.
2684 void consumer_close_channel_streams(struct lttng_consumer_channel
*channel
)
2686 struct lttng_ht
*ht
;
2687 struct lttng_consumer_stream
*stream
;
2688 struct lttng_ht_iter iter
;
2690 ht
= consumer_data
.stream_per_chan_id_ht
;
2693 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2694 ht
->hash_fct(&channel
->key
, lttng_ht_seed
),
2695 ht
->match_fct
, &channel
->key
,
2696 &iter
.iter
, stream
, node_channel_id
.node
) {
2698 * Protect against teardown with mutex.
2700 pthread_mutex_lock(&stream
->lock
);
2701 if (cds_lfht_is_node_deleted(&stream
->node
.node
)) {
2704 switch (consumer_data
.type
) {
2705 case LTTNG_CONSUMER_KERNEL
:
2707 case LTTNG_CONSUMER32_UST
:
2708 case LTTNG_CONSUMER64_UST
:
2709 if (stream
->metadata_flag
) {
2710 /* Safe and protected by the stream lock. */
2711 lttng_ustconsumer_close_metadata(stream
->chan
);
2714 * Note: a mutex is taken internally within
2715 * liblttng-ust-ctl to protect timer wakeup_fd
2716 * use from concurrent close.
2718 lttng_ustconsumer_close_stream_wakeup(stream
);
2722 ERR("Unknown consumer_data type");
2726 pthread_mutex_unlock(&stream
->lock
);
2731 static void destroy_channel_ht(struct lttng_ht
*ht
)
2733 struct lttng_ht_iter iter
;
2734 struct lttng_consumer_channel
*channel
;
2742 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, channel
, wait_fd_node
.node
) {
2743 ret
= lttng_ht_del(ht
, &iter
);
2748 lttng_ht_destroy(ht
);
2752 * This thread polls the channel fds to detect when they are being
2753 * closed. It closes all related streams if the channel is detected as
2754 * closed. It is currently only used as a shim layer for UST because the
2755 * consumerd needs to keep the per-stream wakeup end of pipes open for
2758 void *consumer_thread_channel_poll(void *data
)
2760 int ret
, i
, pollfd
, err
= -1;
2761 uint32_t revents
, nb_fd
;
2762 struct lttng_consumer_channel
*chan
= NULL
;
2763 struct lttng_ht_iter iter
;
2764 struct lttng_ht_node_u64
*node
;
2765 struct lttng_poll_event events
;
2766 struct lttng_consumer_local_data
*ctx
= data
;
2767 struct lttng_ht
*channel_ht
;
2769 rcu_register_thread();
2771 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_CHANNEL
);
2773 if (testpoint(consumerd_thread_channel
)) {
2774 goto error_testpoint
;
2777 health_code_update();
2779 channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
2781 /* ENOMEM at this point. Better to bail out. */
2785 DBG("Thread channel poll started");
2787 /* Size is set to 1 for the consumer_channel pipe */
2788 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
2790 ERR("Poll set creation failed");
2794 ret
= lttng_poll_add(&events
, ctx
->consumer_channel_pipe
[0], LPOLLIN
);
2800 DBG("Channel main loop started");
2804 health_code_update();
2805 DBG("Channel poll wait");
2806 health_poll_entry();
2807 ret
= lttng_poll_wait(&events
, -1);
2808 DBG("Channel poll return from wait with %d fd(s)",
2809 LTTNG_POLL_GETNB(&events
));
2811 DBG("Channel event caught in thread");
2813 if (errno
== EINTR
) {
2814 ERR("Poll EINTR caught");
2817 if (LTTNG_POLL_GETNB(&events
) == 0) {
2818 err
= 0; /* All is OK */
2825 /* From here, the event is a channel wait fd */
2826 for (i
= 0; i
< nb_fd
; i
++) {
2827 health_code_update();
2829 revents
= LTTNG_POLL_GETEV(&events
, i
);
2830 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2833 /* No activity for this FD (poll implementation). */
2837 if (pollfd
== ctx
->consumer_channel_pipe
[0]) {
2838 if (revents
& LPOLLIN
) {
2839 enum consumer_channel_action action
;
2842 ret
= read_channel_pipe(ctx
, &chan
, &key
, &action
);
2845 ERR("Error reading channel pipe");
2847 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2852 case CONSUMER_CHANNEL_ADD
:
2853 DBG("Adding channel %d to poll set",
2856 lttng_ht_node_init_u64(&chan
->wait_fd_node
,
2859 lttng_ht_add_unique_u64(channel_ht
,
2860 &chan
->wait_fd_node
);
2862 /* Add channel to the global poll events list */
2863 lttng_poll_add(&events
, chan
->wait_fd
,
2864 LPOLLERR
| LPOLLHUP
);
2866 case CONSUMER_CHANNEL_DEL
:
2869 * This command should never be called if the channel
2870 * has streams monitored by either the data or metadata
2871 * thread. The consumer only notify this thread with a
2872 * channel del. command if it receives a destroy
2873 * channel command from the session daemon that send it
2874 * if a command prior to the GET_CHANNEL failed.
2878 chan
= consumer_find_channel(key
);
2881 ERR("UST consumer get channel key %" PRIu64
" not found for del channel", key
);
2884 lttng_poll_del(&events
, chan
->wait_fd
);
2885 iter
.iter
.node
= &chan
->wait_fd_node
.node
;
2886 ret
= lttng_ht_del(channel_ht
, &iter
);
2889 switch (consumer_data
.type
) {
2890 case LTTNG_CONSUMER_KERNEL
:
2892 case LTTNG_CONSUMER32_UST
:
2893 case LTTNG_CONSUMER64_UST
:
2894 health_code_update();
2895 /* Destroy streams that might have been left in the stream list. */
2896 clean_channel_stream_list(chan
);
2899 ERR("Unknown consumer_data type");
2904 * Release our own refcount. Force channel deletion even if
2905 * streams were not initialized.
2907 if (!uatomic_sub_return(&chan
->refcount
, 1)) {
2908 consumer_del_channel(chan
);
2913 case CONSUMER_CHANNEL_QUIT
:
2915 * Remove the pipe from the poll set and continue the loop
2916 * since their might be data to consume.
2918 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2921 ERR("Unknown action");
2924 } else if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2925 DBG("Channel thread pipe hung up");
2927 * Remove the pipe from the poll set and continue the loop
2928 * since their might be data to consume.
2930 lttng_poll_del(&events
, ctx
->consumer_channel_pipe
[0]);
2933 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2937 /* Handle other stream */
2943 uint64_t tmp_id
= (uint64_t) pollfd
;
2945 lttng_ht_lookup(channel_ht
, &tmp_id
, &iter
);
2947 node
= lttng_ht_iter_get_node_u64(&iter
);
2950 chan
= caa_container_of(node
, struct lttng_consumer_channel
,
2953 /* Check for error event */
2954 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2955 DBG("Channel fd %d is hup|err.", pollfd
);
2957 lttng_poll_del(&events
, chan
->wait_fd
);
2958 ret
= lttng_ht_del(channel_ht
, &iter
);
2962 * This will close the wait fd for each stream associated to
2963 * this channel AND monitored by the data/metadata thread thus
2964 * will be clean by the right thread.
2966 consumer_close_channel_streams(chan
);
2968 /* Release our own refcount */
2969 if (!uatomic_sub_return(&chan
->refcount
, 1)
2970 && !uatomic_read(&chan
->nb_init_stream_left
)) {
2971 consumer_del_channel(chan
);
2974 ERR("Unexpected poll events %u for sock %d", revents
, pollfd
);
2979 /* Release RCU lock for the channel looked up */
2987 lttng_poll_clean(&events
);
2989 destroy_channel_ht(channel_ht
);
2992 DBG("Channel poll thread exiting");
2995 ERR("Health error occurred in %s", __func__
);
2997 health_unregister(health_consumerd
);
2998 rcu_unregister_thread();
3002 static int set_metadata_socket(struct lttng_consumer_local_data
*ctx
,
3003 struct pollfd
*sockpoll
, int client_socket
)
3010 ret
= lttng_consumer_poll_socket(sockpoll
);
3014 DBG("Metadata connection on client_socket");
3016 /* Blocking call, waiting for transmission */
3017 ctx
->consumer_metadata_socket
= lttcomm_accept_unix_sock(client_socket
);
3018 if (ctx
->consumer_metadata_socket
< 0) {
3019 WARN("On accept metadata");
3030 * This thread listens on the consumerd socket and receives the file
3031 * descriptors from the session daemon.
3033 void *consumer_thread_sessiond_poll(void *data
)
3035 int sock
= -1, client_socket
, ret
, err
= -1;
3037 * structure to poll for incoming data on communication socket avoids
3038 * making blocking sockets.
3040 struct pollfd consumer_sockpoll
[2];
3041 struct lttng_consumer_local_data
*ctx
= data
;
3043 rcu_register_thread();
3045 health_register(health_consumerd
, HEALTH_CONSUMERD_TYPE_SESSIOND
);
3047 if (testpoint(consumerd_thread_sessiond
)) {
3048 goto error_testpoint
;
3051 health_code_update();
3053 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
3054 unlink(ctx
->consumer_command_sock_path
);
3055 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
3056 if (client_socket
< 0) {
3057 ERR("Cannot create command socket");
3061 ret
= lttcomm_listen_unix_sock(client_socket
);
3066 DBG("Sending ready command to lttng-sessiond");
3067 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
3068 /* return < 0 on error, but == 0 is not fatal */
3070 ERR("Error sending ready command to lttng-sessiond");
3074 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3075 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
3076 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
3077 consumer_sockpoll
[1].fd
= client_socket
;
3078 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3080 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3088 DBG("Connection on client_socket");
3090 /* Blocking call, waiting for transmission */
3091 sock
= lttcomm_accept_unix_sock(client_socket
);
3098 * Setup metadata socket which is the second socket connection on the
3099 * command unix socket.
3101 ret
= set_metadata_socket(ctx
, consumer_sockpoll
, client_socket
);
3110 /* This socket is not useful anymore. */
3111 ret
= close(client_socket
);
3113 PERROR("close client_socket");
3117 /* update the polling structure to poll on the established socket */
3118 consumer_sockpoll
[1].fd
= sock
;
3119 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
3122 health_code_update();
3124 health_poll_entry();
3125 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3134 DBG("Incoming command on sock");
3135 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
3138 * This could simply be a session daemon quitting. Don't output
3141 DBG("Communication interrupted on command socket");
3145 if (consumer_quit
) {
3146 DBG("consumer_thread_receive_fds received quit from signal");
3147 err
= 0; /* All is OK */
3150 DBG("received command on sock");
3156 DBG("Consumer thread sessiond poll exiting");
3159 * Close metadata streams since the producer is the session daemon which
3162 * NOTE: for now, this only applies to the UST tracer.
3164 lttng_consumer_close_all_metadata();
3167 * when all fds have hung up, the polling thread
3173 * Notify the data poll thread to poll back again and test the
3174 * consumer_quit state that we just set so to quit gracefully.
3176 notify_thread_lttng_pipe(ctx
->consumer_data_pipe
);
3178 notify_channel_pipe(ctx
, NULL
, -1, CONSUMER_CHANNEL_QUIT
);
3180 notify_health_quit_pipe(health_quit_pipe
);
3182 /* Cleaning up possibly open sockets. */
3186 PERROR("close sock sessiond poll");
3189 if (client_socket
>= 0) {
3190 ret
= close(client_socket
);
3192 PERROR("close client_socket sessiond poll");
3199 ERR("Health error occurred in %s", __func__
);
3201 health_unregister(health_consumerd
);
3203 rcu_unregister_thread();
3207 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
3208 struct lttng_consumer_local_data
*ctx
)
3212 pthread_mutex_lock(&stream
->lock
);
3213 if (stream
->metadata_flag
) {
3214 pthread_mutex_lock(&stream
->metadata_rdv_lock
);
3217 switch (consumer_data
.type
) {
3218 case LTTNG_CONSUMER_KERNEL
:
3219 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
3221 case LTTNG_CONSUMER32_UST
:
3222 case LTTNG_CONSUMER64_UST
:
3223 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
3226 ERR("Unknown consumer_data type");
3232 if (stream
->metadata_flag
) {
3233 pthread_cond_broadcast(&stream
->metadata_rdv
);
3234 pthread_mutex_unlock(&stream
->metadata_rdv_lock
);
3236 pthread_mutex_unlock(&stream
->lock
);
3240 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
3242 switch (consumer_data
.type
) {
3243 case LTTNG_CONSUMER_KERNEL
:
3244 return lttng_kconsumer_on_recv_stream(stream
);
3245 case LTTNG_CONSUMER32_UST
:
3246 case LTTNG_CONSUMER64_UST
:
3247 return lttng_ustconsumer_on_recv_stream(stream
);
3249 ERR("Unknown consumer_data type");
3256 * Allocate and set consumer data hash tables.
3258 int lttng_consumer_init(void)
3260 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3261 if (!consumer_data
.channel_ht
) {
3265 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3266 if (!consumer_data
.relayd_ht
) {
3270 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3271 if (!consumer_data
.stream_list_ht
) {
3275 consumer_data
.stream_per_chan_id_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3276 if (!consumer_data
.stream_per_chan_id_ht
) {
3280 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3285 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_U64
);
3297 * Process the ADD_RELAYD command receive by a consumer.
3299 * This will create a relayd socket pair and add it to the relayd hash table.
3300 * The caller MUST acquire a RCU read side lock before calling it.
3302 int consumer_add_relayd_socket(uint64_t net_seq_idx
, int sock_type
,
3303 struct lttng_consumer_local_data
*ctx
, int sock
,
3304 struct pollfd
*consumer_sockpoll
,
3305 struct lttcomm_relayd_sock
*relayd_sock
, uint64_t sessiond_id
,
3306 uint64_t relayd_session_id
)
3308 int fd
= -1, ret
= -1, relayd_created
= 0;
3309 enum lttcomm_return_code ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3310 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3313 assert(relayd_sock
);
3315 DBG("Consumer adding relayd socket (idx: %" PRIu64
")", net_seq_idx
);
3317 /* Get relayd reference if exists. */
3318 relayd
= consumer_find_relayd(net_seq_idx
);
3319 if (relayd
== NULL
) {
3320 assert(sock_type
== LTTNG_STREAM_CONTROL
);
3321 /* Not found. Allocate one. */
3322 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
3323 if (relayd
== NULL
) {
3325 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3328 relayd
->sessiond_session_id
= sessiond_id
;
3333 * This code path MUST continue to the consumer send status message to
3334 * we can notify the session daemon and continue our work without
3335 * killing everything.
3339 * relayd key should never be found for control socket.
3341 assert(sock_type
!= LTTNG_STREAM_CONTROL
);
3344 /* First send a status message before receiving the fds. */
3345 ret
= consumer_send_status_msg(sock
, LTTCOMM_CONSUMERD_SUCCESS
);
3347 /* Somehow, the session daemon is not responding anymore. */
3348 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3349 goto error_nosignal
;
3352 /* Poll on consumer socket. */
3353 ret
= lttng_consumer_poll_socket(consumer_sockpoll
);
3355 /* Needing to exit in the middle of a command: error. */
3356 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
3358 goto error_nosignal
;
3361 /* Get relayd socket from session daemon */
3362 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
3363 if (ret
!= sizeof(fd
)) {
3365 fd
= -1; /* Just in case it gets set with an invalid value. */
3368 * Failing to receive FDs might indicate a major problem such as
3369 * reaching a fd limit during the receive where the kernel returns a
3370 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3371 * don't take any chances and stop everything.
3373 * XXX: Feature request #558 will fix that and avoid this possible
3374 * issue when reaching the fd limit.
3376 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
3377 ret_code
= LTTCOMM_CONSUMERD_ERROR_RECV_FD
;
3381 /* Copy socket information and received FD */
3382 switch (sock_type
) {
3383 case LTTNG_STREAM_CONTROL
:
3384 /* Copy received lttcomm socket */
3385 lttcomm_copy_sock(&relayd
->control_sock
.sock
, &relayd_sock
->sock
);
3386 ret
= lttcomm_create_sock(&relayd
->control_sock
.sock
);
3387 /* Handle create_sock error. */
3389 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3393 * Close the socket created internally by
3394 * lttcomm_create_sock, so we can replace it by the one
3395 * received from sessiond.
3397 if (close(relayd
->control_sock
.sock
.fd
)) {
3401 /* Assign new file descriptor */
3402 relayd
->control_sock
.sock
.fd
= fd
;
3403 fd
= -1; /* For error path */
3404 /* Assign version values. */
3405 relayd
->control_sock
.major
= relayd_sock
->major
;
3406 relayd
->control_sock
.minor
= relayd_sock
->minor
;
3408 relayd
->relayd_session_id
= relayd_session_id
;
3411 case LTTNG_STREAM_DATA
:
3412 /* Copy received lttcomm socket */
3413 lttcomm_copy_sock(&relayd
->data_sock
.sock
, &relayd_sock
->sock
);
3414 ret
= lttcomm_create_sock(&relayd
->data_sock
.sock
);
3415 /* Handle create_sock error. */
3417 ret_code
= LTTCOMM_CONSUMERD_ENOMEM
;
3421 * Close the socket created internally by
3422 * lttcomm_create_sock, so we can replace it by the one
3423 * received from sessiond.
3425 if (close(relayd
->data_sock
.sock
.fd
)) {
3429 /* Assign new file descriptor */
3430 relayd
->data_sock
.sock
.fd
= fd
;
3431 fd
= -1; /* for eventual error paths */
3432 /* Assign version values. */
3433 relayd
->data_sock
.major
= relayd_sock
->major
;
3434 relayd
->data_sock
.minor
= relayd_sock
->minor
;
3437 ERR("Unknown relayd socket type (%d)", sock_type
);
3439 ret_code
= LTTCOMM_CONSUMERD_FATAL
;
3443 DBG("Consumer %s socket created successfully with net idx %" PRIu64
" (fd: %d)",
3444 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
3445 relayd
->net_seq_idx
, fd
);
3447 /* We successfully added the socket. Send status back. */
3448 ret
= consumer_send_status_msg(sock
, ret_code
);
3450 /* Somehow, the session daemon is not responding anymore. */
3451 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3452 goto error_nosignal
;
3456 * Add relayd socket pair to consumer data hashtable. If object already
3457 * exists or on error, the function gracefully returns.
3465 if (consumer_send_status_msg(sock
, ret_code
) < 0) {
3466 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_FATAL
);
3470 /* Close received socket if valid. */
3473 PERROR("close received socket");
3477 if (relayd_created
) {
3485 * Try to lock the stream mutex.
3487 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
3489 static int stream_try_lock(struct lttng_consumer_stream
*stream
)
3496 * Try to lock the stream mutex. On failure, we know that the stream is
3497 * being used else where hence there is data still being extracted.
3499 ret
= pthread_mutex_trylock(&stream
->lock
);
3501 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
3513 * Search for a relayd associated to the session id and return the reference.
3515 * A rcu read side lock MUST be acquire before calling this function and locked
3516 * until the relayd object is no longer necessary.
3518 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
3520 struct lttng_ht_iter iter
;
3521 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3523 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3524 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
3527 * Check by sessiond id which is unique here where the relayd session
3528 * id might not be when having multiple relayd.
3530 if (relayd
->sessiond_session_id
== id
) {
3531 /* Found the relayd. There can be only one per id. */
3543 * Check if for a given session id there is still data needed to be extract
3546 * Return 1 if data is pending or else 0 meaning ready to be read.
3548 int consumer_data_pending(uint64_t id
)
3551 struct lttng_ht_iter iter
;
3552 struct lttng_ht
*ht
;
3553 struct lttng_consumer_stream
*stream
;
3554 struct consumer_relayd_sock_pair
*relayd
= NULL
;
3555 int (*data_pending
)(struct lttng_consumer_stream
*);
3557 DBG("Consumer data pending command on session id %" PRIu64
, id
);
3560 pthread_mutex_lock(&consumer_data
.lock
);
3562 switch (consumer_data
.type
) {
3563 case LTTNG_CONSUMER_KERNEL
:
3564 data_pending
= lttng_kconsumer_data_pending
;
3566 case LTTNG_CONSUMER32_UST
:
3567 case LTTNG_CONSUMER64_UST
:
3568 data_pending
= lttng_ustconsumer_data_pending
;
3571 ERR("Unknown consumer data type");
3575 /* Ease our life a bit */
3576 ht
= consumer_data
.stream_list_ht
;
3578 relayd
= find_relayd_by_session_id(id
);
3580 /* Send init command for data pending. */
3581 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3582 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
3583 relayd
->relayd_session_id
);
3584 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3586 /* Communication error thus the relayd so no data pending. */
3587 goto data_not_pending
;
3591 cds_lfht_for_each_entry_duplicate(ht
->ht
,
3592 ht
->hash_fct(&id
, lttng_ht_seed
),
3594 &iter
.iter
, stream
, node_session_id
.node
) {
3595 /* If this call fails, the stream is being used hence data pending. */
3596 ret
= stream_try_lock(stream
);
3602 * A removed node from the hash table indicates that the stream has
3603 * been deleted thus having a guarantee that the buffers are closed
3604 * on the consumer side. However, data can still be transmitted
3605 * over the network so don't skip the relayd check.
3607 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
3609 /* Check the stream if there is data in the buffers. */
3610 ret
= data_pending(stream
);
3612 pthread_mutex_unlock(&stream
->lock
);
3619 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3620 if (stream
->metadata_flag
) {
3621 ret
= relayd_quiescent_control(&relayd
->control_sock
,
3622 stream
->relayd_stream_id
);
3624 ret
= relayd_data_pending(&relayd
->control_sock
,
3625 stream
->relayd_stream_id
,
3626 stream
->next_net_seq_num
- 1);
3628 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3630 pthread_mutex_unlock(&stream
->lock
);
3634 pthread_mutex_unlock(&stream
->lock
);
3638 unsigned int is_data_inflight
= 0;
3640 /* Send init command for data pending. */
3641 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
3642 ret
= relayd_end_data_pending(&relayd
->control_sock
,
3643 relayd
->relayd_session_id
, &is_data_inflight
);
3644 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
3646 goto data_not_pending
;
3648 if (is_data_inflight
) {
3654 * Finding _no_ node in the hash table and no inflight data means that the
3655 * stream(s) have been removed thus data is guaranteed to be available for
3656 * analysis from the trace files.
3660 /* Data is available to be read by a viewer. */
3661 pthread_mutex_unlock(&consumer_data
.lock
);
3666 /* Data is still being extracted from buffers. */
3667 pthread_mutex_unlock(&consumer_data
.lock
);
3673 * Send a ret code status message to the sessiond daemon.
3675 * Return the sendmsg() return value.
3677 int consumer_send_status_msg(int sock
, int ret_code
)
3679 struct lttcomm_consumer_status_msg msg
;
3681 memset(&msg
, 0, sizeof(msg
));
3682 msg
.ret_code
= ret_code
;
3684 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3688 * Send a channel status message to the sessiond daemon.
3690 * Return the sendmsg() return value.
3692 int consumer_send_status_channel(int sock
,
3693 struct lttng_consumer_channel
*channel
)
3695 struct lttcomm_consumer_status_channel msg
;
3699 memset(&msg
, 0, sizeof(msg
));
3701 msg
.ret_code
= LTTCOMM_CONSUMERD_CHANNEL_FAIL
;
3703 msg
.ret_code
= LTTCOMM_CONSUMERD_SUCCESS
;
3704 msg
.key
= channel
->key
;
3705 msg
.stream_count
= channel
->streams
.count
;
3708 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
3711 unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos
,
3712 unsigned long produced_pos
, uint64_t nb_packets_per_stream
,
3713 uint64_t max_sb_size
)
3715 unsigned long start_pos
;
3717 if (!nb_packets_per_stream
) {
3718 return consumed_pos
; /* Grab everything */
3720 start_pos
= produced_pos
- offset_align_floor(produced_pos
, max_sb_size
);
3721 start_pos
-= max_sb_size
* nb_packets_per_stream
;
3722 if ((long) (start_pos
- consumed_pos
) < 0) {
3723 return consumed_pos
; /* Grab everything */