2 * Copyright (C) 2011 - Julien Desfossez <julien.desfossez@polymtl.ca>
3 * Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
4 * 2012 - David Goulet <dgoulet@efficios.com>
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License, version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
15 * You should have received a copy of the GNU General Public License along
16 * with this program; if not, write to the Free Software Foundation, Inc.,
17 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
27 #include <sys/socket.h>
28 #include <sys/types.h>
32 #include <common/common.h>
33 #include <common/utils.h>
34 #include <common/compat/poll.h>
35 #include <common/kernel-ctl/kernel-ctl.h>
36 #include <common/sessiond-comm/relayd.h>
37 #include <common/sessiond-comm/sessiond-comm.h>
38 #include <common/kernel-consumer/kernel-consumer.h>
39 #include <common/relayd/relayd.h>
40 #include <common/ust-consumer/ust-consumer.h>
44 struct lttng_consumer_global_data consumer_data
= {
47 .type
= LTTNG_CONSUMER_UNKNOWN
,
51 * Flag to inform the polling thread to quit when all fd hung up. Updated by
52 * the consumer_thread_receive_fds when it notices that all fds has hung up.
53 * Also updated by the signal handler (consumer_should_exit()). Read by the
56 volatile int consumer_quit
;
59 * Global hash table containing respectively metadata and data streams. The
60 * stream element in this ht should only be updated by the metadata poll thread
61 * for the metadata and the data poll thread for the data.
63 static struct lttng_ht
*metadata_ht
;
64 static struct lttng_ht
*data_ht
;
67 * Notify a thread pipe to poll back again. This usually means that some global
68 * state has changed so we just send back the thread in a poll wait call.
70 static void notify_thread_pipe(int wpipe
)
75 struct lttng_consumer_stream
*null_stream
= NULL
;
77 ret
= write(wpipe
, &null_stream
, sizeof(null_stream
));
78 } while (ret
< 0 && errno
== EINTR
);
82 * Find a stream. The consumer_data.lock must be locked during this
85 static struct lttng_consumer_stream
*find_stream(int key
,
88 struct lttng_ht_iter iter
;
89 struct lttng_ht_node_ulong
*node
;
90 struct lttng_consumer_stream
*stream
= NULL
;
94 /* Negative keys are lookup failures */
101 lttng_ht_lookup(ht
, (void *)((unsigned long) key
), &iter
);
102 node
= lttng_ht_iter_get_node_ulong(&iter
);
104 stream
= caa_container_of(node
, struct lttng_consumer_stream
, node
);
112 static void steal_stream_key(int key
, struct lttng_ht
*ht
)
114 struct lttng_consumer_stream
*stream
;
117 stream
= find_stream(key
, ht
);
121 * We don't want the lookup to match, but we still need
122 * to iterate on this stream when iterating over the hash table. Just
123 * change the node key.
125 stream
->node
.key
= -1;
131 * Return a channel object for the given key.
133 * RCU read side lock MUST be acquired before calling this function and
134 * protects the channel ptr.
136 struct lttng_consumer_channel
*consumer_find_channel(unsigned long key
)
138 struct lttng_ht_iter iter
;
139 struct lttng_ht_node_ulong
*node
;
140 struct lttng_consumer_channel
*channel
= NULL
;
142 /* Negative keys are lookup failures */
147 lttng_ht_lookup(consumer_data
.channel_ht
, (void *) key
, &iter
);
148 node
= lttng_ht_iter_get_node_ulong(&iter
);
150 channel
= caa_container_of(node
, struct lttng_consumer_channel
, node
);
156 static void free_stream_rcu(struct rcu_head
*head
)
158 struct lttng_ht_node_ulong
*node
=
159 caa_container_of(head
, struct lttng_ht_node_ulong
, head
);
160 struct lttng_consumer_stream
*stream
=
161 caa_container_of(node
, struct lttng_consumer_stream
, node
);
166 static void free_channel_rcu(struct rcu_head
*head
)
168 struct lttng_ht_node_ulong
*node
=
169 caa_container_of(head
, struct lttng_ht_node_ulong
, head
);
170 struct lttng_consumer_channel
*channel
=
171 caa_container_of(node
, struct lttng_consumer_channel
, node
);
177 * RCU protected relayd socket pair free.
179 static void free_relayd_rcu(struct rcu_head
*head
)
181 struct lttng_ht_node_ulong
*node
=
182 caa_container_of(head
, struct lttng_ht_node_ulong
, head
);
183 struct consumer_relayd_sock_pair
*relayd
=
184 caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
187 * Close all sockets. This is done in the call RCU since we don't want the
188 * socket fds to be reassigned thus potentially creating bad state of the
191 * We do not have to lock the control socket mutex here since at this stage
192 * there is no one referencing to this relayd object.
194 (void) relayd_close(&relayd
->control_sock
);
195 (void) relayd_close(&relayd
->data_sock
);
201 * Destroy and free relayd socket pair object.
203 * This function MUST be called with the consumer_data lock acquired.
205 static void destroy_relayd(struct consumer_relayd_sock_pair
*relayd
)
208 struct lttng_ht_iter iter
;
210 if (relayd
== NULL
) {
214 DBG("Consumer destroy and close relayd socket pair");
216 iter
.iter
.node
= &relayd
->node
.node
;
217 ret
= lttng_ht_del(consumer_data
.relayd_ht
, &iter
);
219 /* We assume the relayd is being or is destroyed */
223 /* RCU free() call */
224 call_rcu(&relayd
->node
.head
, free_relayd_rcu
);
228 * Remove a channel from the global list protected by a mutex. This function is
229 * also responsible for freeing its data structures.
231 void consumer_del_channel(struct lttng_consumer_channel
*channel
)
234 struct lttng_ht_iter iter
;
236 DBG("Consumer delete channel key %d", channel
->key
);
238 pthread_mutex_lock(&consumer_data
.lock
);
240 switch (consumer_data
.type
) {
241 case LTTNG_CONSUMER_KERNEL
:
243 case LTTNG_CONSUMER32_UST
:
244 case LTTNG_CONSUMER64_UST
:
245 lttng_ustconsumer_del_channel(channel
);
248 ERR("Unknown consumer_data type");
254 iter
.iter
.node
= &channel
->node
.node
;
255 ret
= lttng_ht_del(consumer_data
.channel_ht
, &iter
);
259 call_rcu(&channel
->node
.head
, free_channel_rcu
);
261 pthread_mutex_unlock(&consumer_data
.lock
);
265 * Iterate over the relayd hash table and destroy each element. Finally,
266 * destroy the whole hash table.
268 static void cleanup_relayd_ht(void)
270 struct lttng_ht_iter iter
;
271 struct consumer_relayd_sock_pair
*relayd
;
275 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
277 destroy_relayd(relayd
);
280 lttng_ht_destroy(consumer_data
.relayd_ht
);
286 * Update the end point status of all streams having the given network sequence
287 * index (relayd index).
289 * It's atomically set without having the stream mutex locked which is fine
290 * because we handle the write/read race with a pipe wakeup for each thread.
292 static void update_endpoint_status_by_netidx(int net_seq_idx
,
293 enum consumer_endpoint_status status
)
295 struct lttng_ht_iter iter
;
296 struct lttng_consumer_stream
*stream
;
298 DBG("Consumer set delete flag on stream by idx %d", net_seq_idx
);
302 /* Let's begin with metadata */
303 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
304 if (stream
->net_seq_idx
== net_seq_idx
) {
305 uatomic_set(&stream
->endpoint_status
, status
);
306 DBG("Delete flag set to metadata stream %d", stream
->wait_fd
);
310 /* Follow up by the data streams */
311 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
312 if (stream
->net_seq_idx
== net_seq_idx
) {
313 uatomic_set(&stream
->endpoint_status
, status
);
314 DBG("Delete flag set to data stream %d", stream
->wait_fd
);
321 * Cleanup a relayd object by flagging every associated streams for deletion,
322 * destroying the object meaning removing it from the relayd hash table,
323 * closing the sockets and freeing the memory in a RCU call.
325 * If a local data context is available, notify the threads that the streams'
326 * state have changed.
328 static void cleanup_relayd(struct consumer_relayd_sock_pair
*relayd
,
329 struct lttng_consumer_local_data
*ctx
)
335 DBG("Cleaning up relayd sockets");
337 /* Save the net sequence index before destroying the object */
338 netidx
= relayd
->net_seq_idx
;
341 * Delete the relayd from the relayd hash table, close the sockets and free
342 * the object in a RCU call.
344 destroy_relayd(relayd
);
346 /* Set inactive endpoint to all streams */
347 update_endpoint_status_by_netidx(netidx
, CONSUMER_ENDPOINT_INACTIVE
);
350 * With a local data context, notify the threads that the streams' state
351 * have changed. The write() action on the pipe acts as an "implicit"
352 * memory barrier ordering the updates of the end point status from the
353 * read of this status which happens AFTER receiving this notify.
356 notify_thread_pipe(ctx
->consumer_data_pipe
[1]);
357 notify_thread_pipe(ctx
->consumer_metadata_pipe
[1]);
362 * Flag a relayd socket pair for destruction. Destroy it if the refcount
365 * RCU read side lock MUST be aquired before calling this function.
367 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair
*relayd
)
371 /* Set destroy flag for this object */
372 uatomic_set(&relayd
->destroy_flag
, 1);
374 /* Destroy the relayd if refcount is 0 */
375 if (uatomic_read(&relayd
->refcount
) == 0) {
376 destroy_relayd(relayd
);
381 * Remove a stream from the global list protected by a mutex. This
382 * function is also responsible for freeing its data structures.
384 void consumer_del_stream(struct lttng_consumer_stream
*stream
,
388 struct lttng_ht_iter iter
;
389 struct lttng_consumer_channel
*free_chan
= NULL
;
390 struct consumer_relayd_sock_pair
*relayd
;
394 DBG("Consumer del stream %d", stream
->wait_fd
);
397 /* Means the stream was allocated but not successfully added */
398 goto free_stream_rcu
;
401 pthread_mutex_lock(&consumer_data
.lock
);
402 pthread_mutex_lock(&stream
->lock
);
404 switch (consumer_data
.type
) {
405 case LTTNG_CONSUMER_KERNEL
:
406 if (stream
->mmap_base
!= NULL
) {
407 ret
= munmap(stream
->mmap_base
, stream
->mmap_len
);
413 case LTTNG_CONSUMER32_UST
:
414 case LTTNG_CONSUMER64_UST
:
415 lttng_ustconsumer_del_stream(stream
);
418 ERR("Unknown consumer_data type");
424 iter
.iter
.node
= &stream
->node
.node
;
425 ret
= lttng_ht_del(ht
, &iter
);
428 /* Remove node session id from the consumer_data stream ht */
429 iter
.iter
.node
= &stream
->node_session_id
.node
;
430 ret
= lttng_ht_del(consumer_data
.stream_list_ht
, &iter
);
434 assert(consumer_data
.stream_count
> 0);
435 consumer_data
.stream_count
--;
437 if (stream
->out_fd
>= 0) {
438 ret
= close(stream
->out_fd
);
444 /* Check and cleanup relayd */
446 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
447 if (relayd
!= NULL
) {
448 uatomic_dec(&relayd
->refcount
);
449 assert(uatomic_read(&relayd
->refcount
) >= 0);
451 /* Closing streams requires to lock the control socket. */
452 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
453 ret
= relayd_send_close_stream(&relayd
->control_sock
,
454 stream
->relayd_stream_id
,
455 stream
->next_net_seq_num
- 1);
456 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
458 DBG("Unable to close stream on the relayd. Continuing");
460 * Continue here. There is nothing we can do for the relayd.
461 * Chances are that the relayd has closed the socket so we just
462 * continue cleaning up.
466 /* Both conditions are met, we destroy the relayd. */
467 if (uatomic_read(&relayd
->refcount
) == 0 &&
468 uatomic_read(&relayd
->destroy_flag
)) {
469 destroy_relayd(relayd
);
474 uatomic_dec(&stream
->chan
->refcount
);
475 if (!uatomic_read(&stream
->chan
->refcount
)
476 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
477 free_chan
= stream
->chan
;
481 consumer_data
.need_update
= 1;
482 pthread_mutex_unlock(&stream
->lock
);
483 pthread_mutex_unlock(&consumer_data
.lock
);
486 consumer_del_channel(free_chan
);
490 call_rcu(&stream
->node
.head
, free_stream_rcu
);
493 struct lttng_consumer_stream
*consumer_allocate_stream(int channel_key
,
495 enum lttng_consumer_stream_state state
,
496 const char *channel_name
,
503 enum consumer_channel_type type
)
506 struct lttng_consumer_stream
*stream
;
508 stream
= zmalloc(sizeof(*stream
));
509 if (stream
== NULL
) {
510 PERROR("malloc struct lttng_consumer_stream");
517 stream
->key
= stream_key
;
519 stream
->out_fd_offset
= 0;
520 stream
->state
= state
;
523 stream
->net_seq_idx
= relayd_id
;
524 stream
->session_id
= session_id
;
525 pthread_mutex_init(&stream
->lock
, NULL
);
527 /* If channel is the metadata, flag this stream as metadata. */
528 if (type
== CONSUMER_CHANNEL_TYPE_METADATA
) {
529 stream
->metadata_flag
= 1;
530 /* Metadata is flat out. */
531 strncpy(stream
->name
, DEFAULT_METADATA_NAME
, sizeof(stream
->name
));
533 /* Format stream name to <channel_name>_<cpu_number> */
534 ret
= snprintf(stream
->name
, sizeof(stream
->name
), "%s_%d",
537 PERROR("snprintf stream name");
542 /* Key is always the wait_fd for streams. */
543 lttng_ht_node_init_ulong(&stream
->node
, stream
->key
);
545 /* Init session id node with the stream session id */
546 lttng_ht_node_init_ulong(&stream
->node_session_id
, stream
->session_id
);
548 DBG3("Allocated stream %s (key %d, relayd_id %d, session_id %" PRIu64
,
549 stream
->name
, stream
->key
, stream
->net_seq_idx
, stream
->session_id
);
565 * Add a stream to the global list protected by a mutex.
567 static int add_stream(struct lttng_consumer_stream
*stream
,
571 struct consumer_relayd_sock_pair
*relayd
;
576 DBG3("Adding consumer stream %d", stream
->key
);
578 pthread_mutex_lock(&consumer_data
.lock
);
579 pthread_mutex_lock(&stream
->lock
);
582 /* Steal stream identifier to avoid having streams with the same key */
583 steal_stream_key(stream
->key
, ht
);
585 lttng_ht_add_unique_ulong(ht
, &stream
->node
);
588 * Add stream to the stream_list_ht of the consumer data. No need to steal
589 * the key since the HT does not use it and we allow to add redundant keys
592 lttng_ht_add_ulong(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
594 /* Check and cleanup relayd */
595 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
596 if (relayd
!= NULL
) {
597 uatomic_inc(&relayd
->refcount
);
600 /* Update channel refcount once added without error(s). */
601 uatomic_inc(&stream
->chan
->refcount
);
604 * When nb_init_stream_left reaches 0, we don't need to trigger any action
605 * in terms of destroying the associated channel, because the action that
606 * causes the count to become 0 also causes a stream to be added. The
607 * channel deletion will thus be triggered by the following removal of this
610 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
611 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
614 /* Update consumer data once the node is inserted. */
615 consumer_data
.stream_count
++;
616 consumer_data
.need_update
= 1;
619 pthread_mutex_unlock(&stream
->lock
);
620 pthread_mutex_unlock(&consumer_data
.lock
);
626 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
627 * be acquired before calling this.
629 static int add_relayd(struct consumer_relayd_sock_pair
*relayd
)
632 struct lttng_ht_node_ulong
*node
;
633 struct lttng_ht_iter iter
;
637 lttng_ht_lookup(consumer_data
.relayd_ht
,
638 (void *)((unsigned long) relayd
->net_seq_idx
), &iter
);
639 node
= lttng_ht_iter_get_node_ulong(&iter
);
643 lttng_ht_add_unique_ulong(consumer_data
.relayd_ht
, &relayd
->node
);
650 * Allocate and return a consumer relayd socket.
652 struct consumer_relayd_sock_pair
*consumer_allocate_relayd_sock_pair(
655 struct consumer_relayd_sock_pair
*obj
= NULL
;
657 /* Negative net sequence index is a failure */
658 if (net_seq_idx
< 0) {
662 obj
= zmalloc(sizeof(struct consumer_relayd_sock_pair
));
664 PERROR("zmalloc relayd sock");
668 obj
->net_seq_idx
= net_seq_idx
;
670 obj
->destroy_flag
= 0;
671 lttng_ht_node_init_ulong(&obj
->node
, obj
->net_seq_idx
);
672 pthread_mutex_init(&obj
->ctrl_sock_mutex
, NULL
);
679 * Find a relayd socket pair in the global consumer data.
681 * Return the object if found else NULL.
682 * RCU read-side lock must be held across this call and while using the
685 struct consumer_relayd_sock_pair
*consumer_find_relayd(int key
)
687 struct lttng_ht_iter iter
;
688 struct lttng_ht_node_ulong
*node
;
689 struct consumer_relayd_sock_pair
*relayd
= NULL
;
691 /* Negative keys are lookup failures */
696 lttng_ht_lookup(consumer_data
.relayd_ht
, (void *)((unsigned long) key
),
698 node
= lttng_ht_iter_get_node_ulong(&iter
);
700 relayd
= caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
708 * Handle stream for relayd transmission if the stream applies for network
709 * streaming where the net sequence index is set.
711 * Return destination file descriptor or negative value on error.
713 static int write_relayd_stream_header(struct lttng_consumer_stream
*stream
,
714 size_t data_size
, unsigned long padding
,
715 struct consumer_relayd_sock_pair
*relayd
)
718 struct lttcomm_relayd_data_hdr data_hdr
;
724 /* Reset data header */
725 memset(&data_hdr
, 0, sizeof(data_hdr
));
727 if (stream
->metadata_flag
) {
728 /* Caller MUST acquire the relayd control socket lock */
729 ret
= relayd_send_metadata(&relayd
->control_sock
, data_size
);
734 /* Metadata are always sent on the control socket. */
735 outfd
= relayd
->control_sock
.fd
;
737 /* Set header with stream information */
738 data_hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
739 data_hdr
.data_size
= htobe32(data_size
);
740 data_hdr
.padding_size
= htobe32(padding
);
742 * Note that net_seq_num below is assigned with the *current* value of
743 * next_net_seq_num and only after that the next_net_seq_num will be
744 * increment. This is why when issuing a command on the relayd using
745 * this next value, 1 should always be substracted in order to compare
746 * the last seen sequence number on the relayd side to the last sent.
748 data_hdr
.net_seq_num
= htobe64(stream
->next_net_seq_num
);
749 /* Other fields are zeroed previously */
751 ret
= relayd_send_data_hdr(&relayd
->data_sock
, &data_hdr
,
757 ++stream
->next_net_seq_num
;
759 /* Set to go on data socket */
760 outfd
= relayd
->data_sock
.fd
;
768 * Allocate and return a new lttng_consumer_channel object using the given key
769 * to initialize the hash table node.
771 * On error, return NULL.
773 struct lttng_consumer_channel
*consumer_allocate_channel(unsigned long key
,
775 const char *pathname
,
780 enum lttng_event_output output
)
782 struct lttng_consumer_channel
*channel
;
784 channel
= zmalloc(sizeof(*channel
));
785 if (channel
== NULL
) {
786 PERROR("malloc struct lttng_consumer_channel");
791 channel
->refcount
= 0;
792 channel
->session_id
= session_id
;
795 channel
->relayd_id
= relayd_id
;
796 channel
->output
= output
;
798 strncpy(channel
->pathname
, pathname
, sizeof(channel
->pathname
));
799 channel
->pathname
[sizeof(channel
->pathname
) - 1] = '\0';
801 strncpy(channel
->name
, name
, sizeof(channel
->name
));
802 channel
->name
[sizeof(channel
->name
) - 1] = '\0';
804 lttng_ht_node_init_ulong(&channel
->node
, channel
->key
);
805 CDS_INIT_LIST_HEAD(&channel
->streams
.head
);
807 DBG("Allocated channel (key %d)", channel
->key
)
814 * Add a channel to the global list protected by a mutex.
816 int consumer_add_channel(struct lttng_consumer_channel
*channel
)
819 struct lttng_ht_node_ulong
*node
;
820 struct lttng_ht_iter iter
;
822 pthread_mutex_lock(&consumer_data
.lock
);
825 lttng_ht_lookup(consumer_data
.channel_ht
,
826 (void *)((unsigned long) channel
->key
), &iter
);
827 node
= lttng_ht_iter_get_node_ulong(&iter
);
829 /* Channel already exist. Ignore the insertion */
830 ERR("Consumer add channel key %d already exists!", channel
->key
);
835 lttng_ht_add_unique_ulong(consumer_data
.channel_ht
, &channel
->node
);
839 pthread_mutex_unlock(&consumer_data
.lock
);
845 * Allocate the pollfd structure and the local view of the out fds to avoid
846 * doing a lookup in the linked list and concurrency issues when writing is
847 * needed. Called with consumer_data.lock held.
849 * Returns the number of fds in the structures.
851 static int update_poll_array(struct lttng_consumer_local_data
*ctx
,
852 struct pollfd
**pollfd
, struct lttng_consumer_stream
**local_stream
,
856 struct lttng_ht_iter iter
;
857 struct lttng_consumer_stream
*stream
;
862 assert(local_stream
);
864 DBG("Updating poll fd array");
866 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
868 * Only active streams with an active end point can be added to the
869 * poll set and local stream storage of the thread.
871 * There is a potential race here for endpoint_status to be updated
872 * just after the check. However, this is OK since the stream(s) will
873 * be deleted once the thread is notified that the end point state has
874 * changed where this function will be called back again.
876 if (stream
->state
!= LTTNG_CONSUMER_ACTIVE_STREAM
||
877 stream
->endpoint_status
== CONSUMER_ENDPOINT_INACTIVE
) {
880 DBG("Active FD %d", stream
->wait_fd
);
881 (*pollfd
)[i
].fd
= stream
->wait_fd
;
882 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
883 local_stream
[i
] = stream
;
889 * Insert the consumer_data_pipe at the end of the array and don't
890 * increment i so nb_fd is the number of real FD.
892 (*pollfd
)[i
].fd
= ctx
->consumer_data_pipe
[0];
893 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
898 * Poll on the should_quit pipe and the command socket return -1 on error and
899 * should exit, 0 if data is available on the command socket
901 int lttng_consumer_poll_socket(struct pollfd
*consumer_sockpoll
)
906 num_rdy
= poll(consumer_sockpoll
, 2, -1);
909 * Restart interrupted system call.
911 if (errno
== EINTR
) {
914 PERROR("Poll error");
917 if (consumer_sockpoll
[0].revents
& (POLLIN
| POLLPRI
)) {
918 DBG("consumer_should_quit wake up");
928 * Set the error socket.
930 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data
*ctx
,
933 ctx
->consumer_error_socket
= sock
;
937 * Set the command socket path.
939 void lttng_consumer_set_command_sock_path(
940 struct lttng_consumer_local_data
*ctx
, char *sock
)
942 ctx
->consumer_command_sock_path
= sock
;
946 * Send return code to the session daemon.
947 * If the socket is not defined, we return 0, it is not a fatal error
949 int lttng_consumer_send_error(struct lttng_consumer_local_data
*ctx
, int cmd
)
951 if (ctx
->consumer_error_socket
> 0) {
952 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
953 sizeof(enum lttcomm_sessiond_command
));
960 * Close all the tracefiles and stream fds and MUST be called when all
961 * instances are destroyed i.e. when all threads were joined and are ended.
963 void lttng_consumer_cleanup(void)
965 struct lttng_ht_iter iter
;
966 struct lttng_consumer_channel
*channel
;
970 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
, channel
,
972 consumer_del_channel(channel
);
977 lttng_ht_destroy(consumer_data
.channel_ht
);
982 * This HT contains streams that are freed by either the metadata thread or
983 * the data thread so we do *nothing* on the hash table and simply destroy
986 lttng_ht_destroy(consumer_data
.stream_list_ht
);
990 * Called from signal handler.
992 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
997 ret
= write(ctx
->consumer_should_quit
[1], "4", 1);
998 } while (ret
< 0 && errno
== EINTR
);
999 if (ret
< 0 || ret
!= 1) {
1000 PERROR("write consumer quit");
1003 DBG("Consumer flag that it should quit");
1006 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
1009 int outfd
= stream
->out_fd
;
1012 * This does a blocking write-and-wait on any page that belongs to the
1013 * subbuffer prior to the one we just wrote.
1014 * Don't care about error values, as these are just hints and ways to
1015 * limit the amount of page cache used.
1017 if (orig_offset
< stream
->max_sb_size
) {
1020 lttng_sync_file_range(outfd
, orig_offset
- stream
->max_sb_size
,
1021 stream
->max_sb_size
,
1022 SYNC_FILE_RANGE_WAIT_BEFORE
1023 | SYNC_FILE_RANGE_WRITE
1024 | SYNC_FILE_RANGE_WAIT_AFTER
);
1026 * Give hints to the kernel about how we access the file:
1027 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1030 * We need to call fadvise again after the file grows because the
1031 * kernel does not seem to apply fadvise to non-existing parts of the
1034 * Call fadvise _after_ having waited for the page writeback to
1035 * complete because the dirty page writeback semantic is not well
1036 * defined. So it can be expected to lead to lower throughput in
1039 posix_fadvise(outfd
, orig_offset
- stream
->max_sb_size
,
1040 stream
->max_sb_size
, POSIX_FADV_DONTNEED
);
1044 * Initialise the necessary environnement :
1045 * - create a new context
1046 * - create the poll_pipe
1047 * - create the should_quit pipe (for signal handler)
1048 * - create the thread pipe (for splice)
1050 * Takes a function pointer as argument, this function is called when data is
1051 * available on a buffer. This function is responsible to do the
1052 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1053 * buffer configuration and then kernctl_put_next_subbuf at the end.
1055 * Returns a pointer to the new context or NULL on error.
1057 struct lttng_consumer_local_data
*lttng_consumer_create(
1058 enum lttng_consumer_type type
,
1059 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
1060 struct lttng_consumer_local_data
*ctx
),
1061 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
1062 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
1063 int (*update_stream
)(int stream_key
, uint32_t state
))
1066 struct lttng_consumer_local_data
*ctx
;
1068 assert(consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
1069 consumer_data
.type
== type
);
1070 consumer_data
.type
= type
;
1072 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
1074 PERROR("allocating context");
1078 ctx
->consumer_error_socket
= -1;
1079 /* assign the callbacks */
1080 ctx
->on_buffer_ready
= buffer_ready
;
1081 ctx
->on_recv_channel
= recv_channel
;
1082 ctx
->on_recv_stream
= recv_stream
;
1083 ctx
->on_update_stream
= update_stream
;
1085 ret
= pipe(ctx
->consumer_data_pipe
);
1087 PERROR("Error creating poll pipe");
1088 goto error_poll_pipe
;
1091 /* set read end of the pipe to non-blocking */
1092 ret
= fcntl(ctx
->consumer_data_pipe
[0], F_SETFL
, O_NONBLOCK
);
1094 PERROR("fcntl O_NONBLOCK");
1095 goto error_poll_fcntl
;
1098 /* set write end of the pipe to non-blocking */
1099 ret
= fcntl(ctx
->consumer_data_pipe
[1], F_SETFL
, O_NONBLOCK
);
1101 PERROR("fcntl O_NONBLOCK");
1102 goto error_poll_fcntl
;
1105 ret
= pipe(ctx
->consumer_should_quit
);
1107 PERROR("Error creating recv pipe");
1108 goto error_quit_pipe
;
1111 ret
= pipe(ctx
->consumer_thread_pipe
);
1113 PERROR("Error creating thread pipe");
1114 goto error_thread_pipe
;
1117 ret
= utils_create_pipe(ctx
->consumer_metadata_pipe
);
1119 goto error_metadata_pipe
;
1122 ret
= utils_create_pipe(ctx
->consumer_splice_metadata_pipe
);
1124 goto error_splice_pipe
;
1130 utils_close_pipe(ctx
->consumer_metadata_pipe
);
1131 error_metadata_pipe
:
1132 utils_close_pipe(ctx
->consumer_thread_pipe
);
1134 for (i
= 0; i
< 2; i
++) {
1137 err
= close(ctx
->consumer_should_quit
[i
]);
1144 for (i
= 0; i
< 2; i
++) {
1147 err
= close(ctx
->consumer_data_pipe
[i
]);
1159 * Close all fds associated with the instance and free the context.
1161 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1165 DBG("Consumer destroying it. Closing everything.");
1167 ret
= close(ctx
->consumer_error_socket
);
1171 ret
= close(ctx
->consumer_thread_pipe
[0]);
1175 ret
= close(ctx
->consumer_thread_pipe
[1]);
1179 ret
= close(ctx
->consumer_data_pipe
[0]);
1183 ret
= close(ctx
->consumer_data_pipe
[1]);
1187 ret
= close(ctx
->consumer_should_quit
[0]);
1191 ret
= close(ctx
->consumer_should_quit
[1]);
1195 utils_close_pipe(ctx
->consumer_splice_metadata_pipe
);
1197 unlink(ctx
->consumer_command_sock_path
);
1202 * Write the metadata stream id on the specified file descriptor.
1204 static int write_relayd_metadata_id(int fd
,
1205 struct lttng_consumer_stream
*stream
,
1206 struct consumer_relayd_sock_pair
*relayd
, unsigned long padding
)
1209 struct lttcomm_relayd_metadata_payload hdr
;
1211 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1212 hdr
.padding_size
= htobe32(padding
);
1214 ret
= write(fd
, (void *) &hdr
, sizeof(hdr
));
1215 } while (ret
< 0 && errno
== EINTR
);
1216 if (ret
< 0 || ret
!= sizeof(hdr
)) {
1218 * This error means that the fd's end is closed so ignore the perror
1219 * not to clubber the error output since this can happen in a normal
1222 if (errno
!= EPIPE
) {
1223 PERROR("write metadata stream id");
1225 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno
);
1227 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1228 * handle writting the missing part so report that as an error and
1229 * don't lie to the caller.
1234 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1235 stream
->relayd_stream_id
, padding
);
1242 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1243 * core function for writing trace buffers to either the local filesystem or
1246 * It must be called with the stream lock held.
1248 * Careful review MUST be put if any changes occur!
1250 * Returns the number of bytes written
1252 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1253 struct lttng_consumer_local_data
*ctx
,
1254 struct lttng_consumer_stream
*stream
, unsigned long len
,
1255 unsigned long padding
)
1257 unsigned long mmap_offset
;
1259 ssize_t ret
= 0, written
= 0;
1260 off_t orig_offset
= stream
->out_fd_offset
;
1261 /* Default is on the disk */
1262 int outfd
= stream
->out_fd
;
1263 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1264 unsigned int relayd_hang_up
= 0;
1266 /* RCU lock for the relayd pointer */
1269 /* Flag that the current stream if set for network streaming. */
1270 if (stream
->net_seq_idx
!= -1) {
1271 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1272 if (relayd
== NULL
) {
1277 /* get the offset inside the fd to mmap */
1278 switch (consumer_data
.type
) {
1279 case LTTNG_CONSUMER_KERNEL
:
1280 mmap_base
= stream
->mmap_base
;
1281 ret
= kernctl_get_mmap_read_offset(stream
->wait_fd
, &mmap_offset
);
1283 case LTTNG_CONSUMER32_UST
:
1284 case LTTNG_CONSUMER64_UST
:
1285 mmap_base
= lttng_ustctl_get_mmap_base(stream
);
1287 ERR("read mmap get mmap base for stream %s", stream
->name
);
1291 ret
= lttng_ustctl_get_mmap_read_offset(stream
, &mmap_offset
);
1294 ERR("Unknown consumer_data type");
1299 PERROR("tracer ctl get_mmap_read_offset");
1304 /* Handle stream on the relayd if the output is on the network */
1306 unsigned long netlen
= len
;
1309 * Lock the control socket for the complete duration of the function
1310 * since from this point on we will use the socket.
1312 if (stream
->metadata_flag
) {
1313 /* Metadata requires the control socket. */
1314 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1315 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1318 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1320 /* Use the returned socket. */
1323 /* Write metadata stream id before payload */
1324 if (stream
->metadata_flag
) {
1325 ret
= write_relayd_metadata_id(outfd
, stream
, relayd
, padding
);
1328 /* Socket operation failed. We consider the relayd dead */
1329 if (ret
== -EPIPE
|| ret
== -EINVAL
) {
1337 /* Socket operation failed. We consider the relayd dead */
1338 if (ret
== -EPIPE
|| ret
== -EINVAL
) {
1342 /* Else, use the default set before which is the filesystem. */
1345 /* No streaming, we have to set the len with the full padding */
1351 ret
= write(outfd
, mmap_base
+ mmap_offset
, len
);
1352 } while (ret
< 0 && errno
== EINTR
);
1353 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, len
);
1356 * This is possible if the fd is closed on the other side (outfd)
1357 * or any write problem. It can be verbose a bit for a normal
1358 * execution if for instance the relayd is stopped abruptly. This
1359 * can happen so set this to a DBG statement.
1361 DBG("Error in file write mmap");
1365 /* Socket operation failed. We consider the relayd dead */
1366 if (errno
== EPIPE
|| errno
== EINVAL
) {
1371 } else if (ret
> len
) {
1372 PERROR("Error in file write (ret %zd > len %lu)", ret
, len
);
1380 /* This call is useless on a socket so better save a syscall. */
1382 /* This won't block, but will start writeout asynchronously */
1383 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret
,
1384 SYNC_FILE_RANGE_WRITE
);
1385 stream
->out_fd_offset
+= ret
;
1389 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1393 * This is a special case that the relayd has closed its socket. Let's
1394 * cleanup the relayd object and all associated streams.
1396 if (relayd
&& relayd_hang_up
) {
1397 cleanup_relayd(relayd
, ctx
);
1401 /* Unlock only if ctrl socket used */
1402 if (relayd
&& stream
->metadata_flag
) {
1403 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1411 * Splice the data from the ring buffer to the tracefile.
1413 * It must be called with the stream lock held.
1415 * Returns the number of bytes spliced.
1417 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1418 struct lttng_consumer_local_data
*ctx
,
1419 struct lttng_consumer_stream
*stream
, unsigned long len
,
1420 unsigned long padding
)
1422 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1424 off_t orig_offset
= stream
->out_fd_offset
;
1425 int fd
= stream
->wait_fd
;
1426 /* Default is on the disk */
1427 int outfd
= stream
->out_fd
;
1428 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1430 unsigned int relayd_hang_up
= 0;
1432 switch (consumer_data
.type
) {
1433 case LTTNG_CONSUMER_KERNEL
:
1435 case LTTNG_CONSUMER32_UST
:
1436 case LTTNG_CONSUMER64_UST
:
1437 /* Not supported for user space tracing */
1440 ERR("Unknown consumer_data type");
1444 /* RCU lock for the relayd pointer */
1447 /* Flag that the current stream if set for network streaming. */
1448 if (stream
->net_seq_idx
!= -1) {
1449 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1450 if (relayd
== NULL
) {
1456 * Choose right pipe for splice. Metadata and trace data are handled by
1457 * different threads hence the use of two pipes in order not to race or
1458 * corrupt the written data.
1460 if (stream
->metadata_flag
) {
1461 splice_pipe
= ctx
->consumer_splice_metadata_pipe
;
1463 splice_pipe
= ctx
->consumer_thread_pipe
;
1466 /* Write metadata stream id before payload */
1468 int total_len
= len
;
1470 if (stream
->metadata_flag
) {
1472 * Lock the control socket for the complete duration of the function
1473 * since from this point on we will use the socket.
1475 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1477 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
, relayd
,
1481 /* Socket operation failed. We consider the relayd dead */
1482 if (ret
== -EBADF
) {
1483 WARN("Remote relayd disconnected. Stopping");
1490 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1493 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1495 /* Use the returned socket. */
1498 /* Socket operation failed. We consider the relayd dead */
1499 if (ret
== -EBADF
) {
1500 WARN("Remote relayd disconnected. Stopping");
1507 /* No streaming, we have to set the len with the full padding */
1512 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1513 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1514 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1515 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1516 DBG("splice chan to pipe, ret %zd", ret_splice
);
1517 if (ret_splice
< 0) {
1518 PERROR("Error in relay splice");
1520 written
= ret_splice
;
1526 /* Handle stream on the relayd if the output is on the network */
1528 if (stream
->metadata_flag
) {
1529 size_t metadata_payload_size
=
1530 sizeof(struct lttcomm_relayd_metadata_payload
);
1532 /* Update counter to fit the spliced data */
1533 ret_splice
+= metadata_payload_size
;
1534 len
+= metadata_payload_size
;
1536 * We do this so the return value can match the len passed as
1537 * argument to this function.
1539 written
-= metadata_payload_size
;
1543 /* Splice data out */
1544 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1545 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1546 DBG("Consumer splice pipe to file, ret %zd", ret_splice
);
1547 if (ret_splice
< 0) {
1548 PERROR("Error in file splice");
1550 written
= ret_splice
;
1552 /* Socket operation failed. We consider the relayd dead */
1553 if (errno
== EBADF
|| errno
== EPIPE
) {
1554 WARN("Remote relayd disconnected. Stopping");
1560 } else if (ret_splice
> len
) {
1562 PERROR("Wrote more data than requested %zd (len: %lu)",
1564 written
+= ret_splice
;
1570 /* This call is useless on a socket so better save a syscall. */
1572 /* This won't block, but will start writeout asynchronously */
1573 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1574 SYNC_FILE_RANGE_WRITE
);
1575 stream
->out_fd_offset
+= ret_splice
;
1577 written
+= ret_splice
;
1579 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1587 * This is a special case that the relayd has closed its socket. Let's
1588 * cleanup the relayd object and all associated streams.
1590 if (relayd
&& relayd_hang_up
) {
1591 cleanup_relayd(relayd
, ctx
);
1592 /* Skip splice error so the consumer does not fail */
1597 /* send the appropriate error description to sessiond */
1600 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1603 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1606 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
1611 if (relayd
&& stream
->metadata_flag
) {
1612 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1620 * Take a snapshot for a specific fd
1622 * Returns 0 on success, < 0 on error
1624 int lttng_consumer_take_snapshot(struct lttng_consumer_stream
*stream
)
1626 switch (consumer_data
.type
) {
1627 case LTTNG_CONSUMER_KERNEL
:
1628 return lttng_kconsumer_take_snapshot(stream
);
1629 case LTTNG_CONSUMER32_UST
:
1630 case LTTNG_CONSUMER64_UST
:
1631 return lttng_ustconsumer_take_snapshot(stream
);
1633 ERR("Unknown consumer_data type");
1641 * Get the produced position
1643 * Returns 0 on success, < 0 on error
1645 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream
*stream
,
1648 switch (consumer_data
.type
) {
1649 case LTTNG_CONSUMER_KERNEL
:
1650 return lttng_kconsumer_get_produced_snapshot(stream
, pos
);
1651 case LTTNG_CONSUMER32_UST
:
1652 case LTTNG_CONSUMER64_UST
:
1653 return lttng_ustconsumer_get_produced_snapshot(stream
, pos
);
1655 ERR("Unknown consumer_data type");
1661 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
1662 int sock
, struct pollfd
*consumer_sockpoll
)
1664 switch (consumer_data
.type
) {
1665 case LTTNG_CONSUMER_KERNEL
:
1666 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1667 case LTTNG_CONSUMER32_UST
:
1668 case LTTNG_CONSUMER64_UST
:
1669 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1671 ERR("Unknown consumer_data type");
1678 * Iterate over all streams of the hashtable and free them properly.
1680 * WARNING: *MUST* be used with data stream only.
1682 static void destroy_data_stream_ht(struct lttng_ht
*ht
)
1684 struct lttng_ht_iter iter
;
1685 struct lttng_consumer_stream
*stream
;
1692 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1694 * Ignore return value since we are currently cleaning up so any error
1697 (void) consumer_del_stream(stream
, ht
);
1701 lttng_ht_destroy(ht
);
1705 * Iterate over all streams of the hashtable and free them properly.
1707 * XXX: Should not be only for metadata stream or else use an other name.
1709 static void destroy_stream_ht(struct lttng_ht
*ht
)
1711 struct lttng_ht_iter iter
;
1712 struct lttng_consumer_stream
*stream
;
1719 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1721 * Ignore return value since we are currently cleaning up so any error
1724 (void) consumer_del_metadata_stream(stream
, ht
);
1728 lttng_ht_destroy(ht
);
1732 * Clean up a metadata stream and free its memory.
1734 void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
,
1735 struct lttng_ht
*ht
)
1738 struct lttng_ht_iter iter
;
1739 struct lttng_consumer_channel
*free_chan
= NULL
;
1740 struct consumer_relayd_sock_pair
*relayd
;
1744 * This call should NEVER receive regular stream. It must always be
1745 * metadata stream and this is crucial for data structure synchronization.
1747 assert(stream
->metadata_flag
);
1749 DBG3("Consumer delete metadata stream %d", stream
->wait_fd
);
1752 /* Means the stream was allocated but not successfully added */
1753 goto free_stream_rcu
;
1756 pthread_mutex_lock(&consumer_data
.lock
);
1757 pthread_mutex_lock(&stream
->lock
);
1759 switch (consumer_data
.type
) {
1760 case LTTNG_CONSUMER_KERNEL
:
1761 if (stream
->mmap_base
!= NULL
) {
1762 ret
= munmap(stream
->mmap_base
, stream
->mmap_len
);
1764 PERROR("munmap metadata stream");
1768 case LTTNG_CONSUMER32_UST
:
1769 case LTTNG_CONSUMER64_UST
:
1770 lttng_ustconsumer_del_stream(stream
);
1773 ERR("Unknown consumer_data type");
1779 iter
.iter
.node
= &stream
->node
.node
;
1780 ret
= lttng_ht_del(ht
, &iter
);
1783 /* Remove node session id from the consumer_data stream ht */
1784 iter
.iter
.node
= &stream
->node_session_id
.node
;
1785 ret
= lttng_ht_del(consumer_data
.stream_list_ht
, &iter
);
1789 if (stream
->out_fd
>= 0) {
1790 ret
= close(stream
->out_fd
);
1796 /* Check and cleanup relayd */
1798 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1799 if (relayd
!= NULL
) {
1800 uatomic_dec(&relayd
->refcount
);
1801 assert(uatomic_read(&relayd
->refcount
) >= 0);
1803 /* Closing streams requires to lock the control socket. */
1804 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1805 ret
= relayd_send_close_stream(&relayd
->control_sock
,
1806 stream
->relayd_stream_id
, stream
->next_net_seq_num
- 1);
1807 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1809 DBG("Unable to close stream on the relayd. Continuing");
1811 * Continue here. There is nothing we can do for the relayd.
1812 * Chances are that the relayd has closed the socket so we just
1813 * continue cleaning up.
1817 /* Both conditions are met, we destroy the relayd. */
1818 if (uatomic_read(&relayd
->refcount
) == 0 &&
1819 uatomic_read(&relayd
->destroy_flag
)) {
1820 destroy_relayd(relayd
);
1825 /* Atomically decrement channel refcount since other threads can use it. */
1826 uatomic_dec(&stream
->chan
->refcount
);
1827 if (!uatomic_read(&stream
->chan
->refcount
)
1828 && !uatomic_read(&stream
->chan
->nb_init_stream_left
)) {
1829 /* Go for channel deletion! */
1830 free_chan
= stream
->chan
;
1834 pthread_mutex_unlock(&stream
->lock
);
1835 pthread_mutex_unlock(&consumer_data
.lock
);
1838 consumer_del_channel(free_chan
);
1842 call_rcu(&stream
->node
.head
, free_stream_rcu
);
1846 * Action done with the metadata stream when adding it to the consumer internal
1847 * data structures to handle it.
1849 static int add_metadata_stream(struct lttng_consumer_stream
*stream
,
1850 struct lttng_ht
*ht
)
1853 struct consumer_relayd_sock_pair
*relayd
;
1854 struct lttng_ht_iter iter
;
1855 struct lttng_ht_node_ulong
*node
;
1860 DBG3("Adding metadata stream %d to hash table", stream
->key
);
1862 pthread_mutex_lock(&consumer_data
.lock
);
1863 pthread_mutex_lock(&stream
->lock
);
1866 * From here, refcounts are updated so be _careful_ when returning an error
1873 * Lookup the stream just to make sure it does not exist in our internal
1874 * state. This should NEVER happen.
1876 lttng_ht_lookup(ht
, (void *)((unsigned long) stream
->key
), &iter
);
1877 node
= lttng_ht_iter_get_node_ulong(&iter
);
1880 /* Find relayd and, if one is found, increment refcount. */
1881 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1882 if (relayd
!= NULL
) {
1883 uatomic_inc(&relayd
->refcount
);
1886 /* Update channel refcount once added without error(s). */
1887 uatomic_inc(&stream
->chan
->refcount
);
1890 * When nb_init_stream_left reaches 0, we don't need to trigger any action
1891 * in terms of destroying the associated channel, because the action that
1892 * causes the count to become 0 also causes a stream to be added. The
1893 * channel deletion will thus be triggered by the following removal of this
1896 if (uatomic_read(&stream
->chan
->nb_init_stream_left
) > 0) {
1897 uatomic_dec(&stream
->chan
->nb_init_stream_left
);
1900 lttng_ht_add_unique_ulong(ht
, &stream
->node
);
1903 * Add stream to the stream_list_ht of the consumer data. No need to steal
1904 * the key since the HT does not use it and we allow to add redundant keys
1907 lttng_ht_add_ulong(consumer_data
.stream_list_ht
, &stream
->node_session_id
);
1911 pthread_mutex_unlock(&stream
->lock
);
1912 pthread_mutex_unlock(&consumer_data
.lock
);
1917 * Delete data stream that are flagged for deletion (endpoint_status).
1919 static void validate_endpoint_status_data_stream(void)
1921 struct lttng_ht_iter iter
;
1922 struct lttng_consumer_stream
*stream
;
1924 DBG("Consumer delete flagged data stream");
1927 cds_lfht_for_each_entry(data_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1928 /* Validate delete flag of the stream */
1929 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
1932 /* Delete it right now */
1933 consumer_del_stream(stream
, data_ht
);
1939 * Delete metadata stream that are flagged for deletion (endpoint_status).
1941 static void validate_endpoint_status_metadata_stream(
1942 struct lttng_poll_event
*pollset
)
1944 struct lttng_ht_iter iter
;
1945 struct lttng_consumer_stream
*stream
;
1947 DBG("Consumer delete flagged metadata stream");
1952 cds_lfht_for_each_entry(metadata_ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1953 /* Validate delete flag of the stream */
1954 if (stream
->endpoint_status
== CONSUMER_ENDPOINT_ACTIVE
) {
1958 * Remove from pollset so the metadata thread can continue without
1959 * blocking on a deleted stream.
1961 lttng_poll_del(pollset
, stream
->wait_fd
);
1963 /* Delete it right now */
1964 consumer_del_metadata_stream(stream
, metadata_ht
);
1970 * Thread polls on metadata file descriptor and write them on disk or on the
1973 void *consumer_thread_metadata_poll(void *data
)
1976 uint32_t revents
, nb_fd
;
1977 struct lttng_consumer_stream
*stream
= NULL
;
1978 struct lttng_ht_iter iter
;
1979 struct lttng_ht_node_ulong
*node
;
1980 struct lttng_poll_event events
;
1981 struct lttng_consumer_local_data
*ctx
= data
;
1984 rcu_register_thread();
1986 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_ULONG
);
1988 /* ENOMEM at this point. Better to bail out. */
1992 DBG("Thread metadata poll started");
1994 /* Size is set to 1 for the consumer_metadata pipe */
1995 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
1997 ERR("Poll set creation failed");
2001 ret
= lttng_poll_add(&events
, ctx
->consumer_metadata_pipe
[0], LPOLLIN
);
2007 DBG("Metadata main loop started");
2010 /* Only the metadata pipe is set */
2011 if (LTTNG_POLL_GETNB(&events
) == 0 && consumer_quit
== 1) {
2016 DBG("Metadata poll wait with %d fd(s)", LTTNG_POLL_GETNB(&events
));
2017 ret
= lttng_poll_wait(&events
, -1);
2018 DBG("Metadata event catched in thread");
2020 if (errno
== EINTR
) {
2021 ERR("Poll EINTR catched");
2029 /* From here, the event is a metadata wait fd */
2030 for (i
= 0; i
< nb_fd
; i
++) {
2031 revents
= LTTNG_POLL_GETEV(&events
, i
);
2032 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
2034 /* Just don't waste time if no returned events for the fd */
2039 if (pollfd
== ctx
->consumer_metadata_pipe
[0]) {
2040 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2041 DBG("Metadata thread pipe hung up");
2043 * Remove the pipe from the poll set and continue the loop
2044 * since their might be data to consume.
2046 lttng_poll_del(&events
, ctx
->consumer_metadata_pipe
[0]);
2047 ret
= close(ctx
->consumer_metadata_pipe
[0]);
2049 PERROR("close metadata pipe");
2052 } else if (revents
& LPOLLIN
) {
2054 /* Get the stream pointer received */
2055 ret
= read(pollfd
, &stream
, sizeof(stream
));
2056 } while (ret
< 0 && errno
== EINTR
);
2058 ret
< sizeof(struct lttng_consumer_stream
*)) {
2059 PERROR("read metadata stream");
2061 * Let's continue here and hope we can still work
2062 * without stopping the consumer. XXX: Should we?
2067 /* A NULL stream means that the state has changed. */
2068 if (stream
== NULL
) {
2069 /* Check for deleted streams. */
2070 validate_endpoint_status_metadata_stream(&events
);
2074 DBG("Adding metadata stream %d to poll set",
2077 ret
= add_metadata_stream(stream
, metadata_ht
);
2079 ERR("Unable to add metadata stream");
2080 /* Stream was not setup properly. Continuing. */
2081 consumer_del_metadata_stream(stream
, NULL
);
2085 /* Add metadata stream to the global poll events list */
2086 lttng_poll_add(&events
, stream
->wait_fd
,
2087 LPOLLIN
| LPOLLPRI
);
2090 /* Handle other stream */
2095 lttng_ht_lookup(metadata_ht
, (void *)((unsigned long) pollfd
),
2097 node
= lttng_ht_iter_get_node_ulong(&iter
);
2100 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
2103 /* Check for error event */
2104 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
2105 DBG("Metadata fd %d is hup|err.", pollfd
);
2106 if (!stream
->hangup_flush_done
2107 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2108 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2109 DBG("Attempting to flush and consume the UST buffers");
2110 lttng_ustconsumer_on_stream_hangup(stream
);
2112 /* We just flushed the stream now read it. */
2114 len
= ctx
->on_buffer_ready(stream
, ctx
);
2116 * We don't check the return value here since if we get
2117 * a negative len, it means an error occured thus we
2118 * simply remove it from the poll set and free the
2124 lttng_poll_del(&events
, stream
->wait_fd
);
2126 * This call update the channel states, closes file descriptors
2127 * and securely free the stream.
2129 consumer_del_metadata_stream(stream
, metadata_ht
);
2130 } else if (revents
& (LPOLLIN
| LPOLLPRI
)) {
2131 /* Get the data out of the metadata file descriptor */
2132 DBG("Metadata available on fd %d", pollfd
);
2133 assert(stream
->wait_fd
== pollfd
);
2135 len
= ctx
->on_buffer_ready(stream
, ctx
);
2136 /* It's ok to have an unavailable sub-buffer */
2137 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2138 /* Clean up stream from consumer and free it. */
2139 lttng_poll_del(&events
, stream
->wait_fd
);
2140 consumer_del_metadata_stream(stream
, metadata_ht
);
2141 } else if (len
> 0) {
2142 stream
->data_read
= 1;
2146 /* Release RCU lock for the stream looked up */
2153 DBG("Metadata poll thread exiting");
2154 lttng_poll_clean(&events
);
2156 destroy_stream_ht(metadata_ht
);
2158 rcu_unregister_thread();
2163 * This thread polls the fds in the set to consume the data and write
2164 * it to tracefile if necessary.
2166 void *consumer_thread_data_poll(void *data
)
2168 int num_rdy
, num_hup
, high_prio
, ret
, i
;
2169 struct pollfd
*pollfd
= NULL
;
2170 /* local view of the streams */
2171 struct lttng_consumer_stream
**local_stream
= NULL
, *new_stream
= NULL
;
2172 /* local view of consumer_data.fds_count */
2174 struct lttng_consumer_local_data
*ctx
= data
;
2177 rcu_register_thread();
2179 data_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_ULONG
);
2180 if (data_ht
== NULL
) {
2181 /* ENOMEM at this point. Better to bail out. */
2185 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
));
2192 * the fds set has been updated, we need to update our
2193 * local array as well
2195 pthread_mutex_lock(&consumer_data
.lock
);
2196 if (consumer_data
.need_update
) {
2201 local_stream
= NULL
;
2203 /* allocate for all fds + 1 for the consumer_data_pipe */
2204 pollfd
= zmalloc((consumer_data
.stream_count
+ 1) * sizeof(struct pollfd
));
2205 if (pollfd
== NULL
) {
2206 PERROR("pollfd malloc");
2207 pthread_mutex_unlock(&consumer_data
.lock
);
2211 /* allocate for all fds + 1 for the consumer_data_pipe */
2212 local_stream
= zmalloc((consumer_data
.stream_count
+ 1) *
2213 sizeof(struct lttng_consumer_stream
));
2214 if (local_stream
== NULL
) {
2215 PERROR("local_stream malloc");
2216 pthread_mutex_unlock(&consumer_data
.lock
);
2219 ret
= update_poll_array(ctx
, &pollfd
, local_stream
,
2222 ERR("Error in allocating pollfd or local_outfds");
2223 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2224 pthread_mutex_unlock(&consumer_data
.lock
);
2228 consumer_data
.need_update
= 0;
2230 pthread_mutex_unlock(&consumer_data
.lock
);
2232 /* No FDs and consumer_quit, consumer_cleanup the thread */
2233 if (nb_fd
== 0 && consumer_quit
== 1) {
2236 /* poll on the array of fds */
2238 DBG("polling on %d fd", nb_fd
+ 1);
2239 num_rdy
= poll(pollfd
, nb_fd
+ 1, -1);
2240 DBG("poll num_rdy : %d", num_rdy
);
2241 if (num_rdy
== -1) {
2243 * Restart interrupted system call.
2245 if (errno
== EINTR
) {
2248 PERROR("Poll error");
2249 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
2251 } else if (num_rdy
== 0) {
2252 DBG("Polling thread timed out");
2257 * If the consumer_data_pipe triggered poll go directly to the
2258 * beginning of the loop to update the array. We want to prioritize
2259 * array update over low-priority reads.
2261 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
2262 ssize_t pipe_readlen
;
2264 DBG("consumer_data_pipe wake up");
2265 /* Consume 1 byte of pipe data */
2267 pipe_readlen
= read(ctx
->consumer_data_pipe
[0], &new_stream
,
2268 sizeof(new_stream
));
2269 } while (pipe_readlen
== -1 && errno
== EINTR
);
2270 if (pipe_readlen
< 0) {
2271 PERROR("read consumer data pipe");
2272 /* Continue so we can at least handle the current stream(s). */
2277 * If the stream is NULL, just ignore it. It's also possible that
2278 * the sessiond poll thread changed the consumer_quit state and is
2279 * waking us up to test it.
2281 if (new_stream
== NULL
) {
2282 validate_endpoint_status_data_stream();
2286 ret
= add_stream(new_stream
, data_ht
);
2288 ERR("Consumer add stream %d failed. Continuing",
2291 * At this point, if the add_stream fails, it is not in the
2292 * hash table thus passing the NULL value here.
2294 consumer_del_stream(new_stream
, NULL
);
2297 /* Continue to update the local streams and handle prio ones */
2301 /* Take care of high priority channels first. */
2302 for (i
= 0; i
< nb_fd
; i
++) {
2303 if (local_stream
[i
] == NULL
) {
2306 if (pollfd
[i
].revents
& POLLPRI
) {
2307 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
2309 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2310 /* it's ok to have an unavailable sub-buffer */
2311 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2312 /* Clean the stream and free it. */
2313 consumer_del_stream(local_stream
[i
], data_ht
);
2314 local_stream
[i
] = NULL
;
2315 } else if (len
> 0) {
2316 local_stream
[i
]->data_read
= 1;
2322 * If we read high prio channel in this loop, try again
2323 * for more high prio data.
2329 /* Take care of low priority channels. */
2330 for (i
= 0; i
< nb_fd
; i
++) {
2331 if (local_stream
[i
] == NULL
) {
2334 if ((pollfd
[i
].revents
& POLLIN
) ||
2335 local_stream
[i
]->hangup_flush_done
) {
2336 DBG("Normal read on fd %d", pollfd
[i
].fd
);
2337 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
2338 /* it's ok to have an unavailable sub-buffer */
2339 if (len
< 0 && len
!= -EAGAIN
&& len
!= -ENODATA
) {
2340 /* Clean the stream and free it. */
2341 consumer_del_stream(local_stream
[i
], data_ht
);
2342 local_stream
[i
] = NULL
;
2343 } else if (len
> 0) {
2344 local_stream
[i
]->data_read
= 1;
2349 /* Handle hangup and errors */
2350 for (i
= 0; i
< nb_fd
; i
++) {
2351 if (local_stream
[i
] == NULL
) {
2354 if (!local_stream
[i
]->hangup_flush_done
2355 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2356 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2357 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2358 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2360 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2361 /* Attempt read again, for the data we just flushed. */
2362 local_stream
[i
]->data_read
= 1;
2365 * If the poll flag is HUP/ERR/NVAL and we have
2366 * read no data in this pass, we can remove the
2367 * stream from its hash table.
2369 if ((pollfd
[i
].revents
& POLLHUP
)) {
2370 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2371 if (!local_stream
[i
]->data_read
) {
2372 consumer_del_stream(local_stream
[i
], data_ht
);
2373 local_stream
[i
] = NULL
;
2376 } else if (pollfd
[i
].revents
& POLLERR
) {
2377 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2378 if (!local_stream
[i
]->data_read
) {
2379 consumer_del_stream(local_stream
[i
], data_ht
);
2380 local_stream
[i
] = NULL
;
2383 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2384 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2385 if (!local_stream
[i
]->data_read
) {
2386 consumer_del_stream(local_stream
[i
], data_ht
);
2387 local_stream
[i
] = NULL
;
2391 if (local_stream
[i
] != NULL
) {
2392 local_stream
[i
]->data_read
= 0;
2397 DBG("polling thread exiting");
2402 * Close the write side of the pipe so epoll_wait() in
2403 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2404 * read side of the pipe. If we close them both, epoll_wait strangely does
2405 * not return and could create a endless wait period if the pipe is the
2406 * only tracked fd in the poll set. The thread will take care of closing
2409 ret
= close(ctx
->consumer_metadata_pipe
[1]);
2411 PERROR("close data pipe");
2414 destroy_data_stream_ht(data_ht
);
2416 rcu_unregister_thread();
2421 * This thread listens on the consumerd socket and receives the file
2422 * descriptors from the session daemon.
2424 void *consumer_thread_sessiond_poll(void *data
)
2426 int sock
= -1, client_socket
, ret
;
2428 * structure to poll for incoming data on communication socket avoids
2429 * making blocking sockets.
2431 struct pollfd consumer_sockpoll
[2];
2432 struct lttng_consumer_local_data
*ctx
= data
;
2434 rcu_register_thread();
2436 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
2437 unlink(ctx
->consumer_command_sock_path
);
2438 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
2439 if (client_socket
< 0) {
2440 ERR("Cannot create command socket");
2444 ret
= lttcomm_listen_unix_sock(client_socket
);
2449 DBG("Sending ready command to lttng-sessiond");
2450 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
2451 /* return < 0 on error, but == 0 is not fatal */
2453 ERR("Error sending ready command to lttng-sessiond");
2457 ret
= fcntl(client_socket
, F_SETFL
, O_NONBLOCK
);
2459 PERROR("fcntl O_NONBLOCK");
2463 /* prepare the FDs to poll : to client socket and the should_quit pipe */
2464 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
2465 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
2466 consumer_sockpoll
[1].fd
= client_socket
;
2467 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
2469 if (lttng_consumer_poll_socket(consumer_sockpoll
) < 0) {
2472 DBG("Connection on client_socket");
2474 /* Blocking call, waiting for transmission */
2475 sock
= lttcomm_accept_unix_sock(client_socket
);
2480 ret
= fcntl(sock
, F_SETFL
, O_NONBLOCK
);
2482 PERROR("fcntl O_NONBLOCK");
2486 /* This socket is not useful anymore. */
2487 ret
= close(client_socket
);
2489 PERROR("close client_socket");
2493 /* update the polling structure to poll on the established socket */
2494 consumer_sockpoll
[1].fd
= sock
;
2495 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
2498 if (lttng_consumer_poll_socket(consumer_sockpoll
) < 0) {
2501 DBG("Incoming command on sock");
2502 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2503 if (ret
== -ENOENT
) {
2504 DBG("Received STOP command");
2509 * This could simply be a session daemon quitting. Don't output
2512 DBG("Communication interrupted on command socket");
2515 if (consumer_quit
) {
2516 DBG("consumer_thread_receive_fds received quit from signal");
2519 DBG("received command on sock");
2522 DBG("Consumer thread sessiond poll exiting");
2525 * when all fds have hung up, the polling thread
2531 * Notify the data poll thread to poll back again and test the
2532 * consumer_quit state that we just set so to quit gracefully.
2534 notify_thread_pipe(ctx
->consumer_data_pipe
[1]);
2536 /* Cleaning up possibly open sockets. */
2540 PERROR("close sock sessiond poll");
2543 if (client_socket
>= 0) {
2546 PERROR("close client_socket sessiond poll");
2550 rcu_unregister_thread();
2554 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
2555 struct lttng_consumer_local_data
*ctx
)
2559 pthread_mutex_lock(&stream
->lock
);
2561 switch (consumer_data
.type
) {
2562 case LTTNG_CONSUMER_KERNEL
:
2563 ret
= lttng_kconsumer_read_subbuffer(stream
, ctx
);
2565 case LTTNG_CONSUMER32_UST
:
2566 case LTTNG_CONSUMER64_UST
:
2567 ret
= lttng_ustconsumer_read_subbuffer(stream
, ctx
);
2570 ERR("Unknown consumer_data type");
2576 pthread_mutex_unlock(&stream
->lock
);
2580 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
2582 switch (consumer_data
.type
) {
2583 case LTTNG_CONSUMER_KERNEL
:
2584 return lttng_kconsumer_on_recv_stream(stream
);
2585 case LTTNG_CONSUMER32_UST
:
2586 case LTTNG_CONSUMER64_UST
:
2587 return lttng_ustconsumer_on_recv_stream(stream
);
2589 ERR("Unknown consumer_data type");
2596 * Allocate and set consumer data hash tables.
2598 void lttng_consumer_init(void)
2600 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_ULONG
);
2601 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_ULONG
);
2602 consumer_data
.stream_list_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_ULONG
);
2606 * Process the ADD_RELAYD command receive by a consumer.
2608 * This will create a relayd socket pair and add it to the relayd hash table.
2609 * The caller MUST acquire a RCU read side lock before calling it.
2611 int consumer_add_relayd_socket(int net_seq_idx
, int sock_type
,
2612 struct lttng_consumer_local_data
*ctx
, int sock
,
2613 struct pollfd
*consumer_sockpoll
, struct lttcomm_sock
*relayd_sock
,
2614 unsigned int sessiond_id
)
2616 int fd
= -1, ret
= -1, relayd_created
= 0;
2617 enum lttng_error_code ret_code
= LTTNG_OK
;
2618 struct consumer_relayd_sock_pair
*relayd
;
2620 DBG("Consumer adding relayd socket (idx: %d)", net_seq_idx
);
2622 /* First send a status message before receiving the fds. */
2623 ret
= consumer_send_status_msg(sock
, ret_code
);
2625 /* Somehow, the session daemon is not responding anymore. */
2629 /* Get relayd reference if exists. */
2630 relayd
= consumer_find_relayd(net_seq_idx
);
2631 if (relayd
== NULL
) {
2632 /* Not found. Allocate one. */
2633 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
2634 if (relayd
== NULL
) {
2635 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_OUTFD_ERROR
);
2639 relayd
->sessiond_session_id
= (uint64_t) sessiond_id
;
2643 /* Poll on consumer socket. */
2644 if (lttng_consumer_poll_socket(consumer_sockpoll
) < 0) {
2649 /* Get relayd socket from session daemon */
2650 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
2651 if (ret
!= sizeof(fd
)) {
2652 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
2654 fd
= -1; /* Just in case it gets set with an invalid value. */
2658 /* We have the fds without error. Send status back. */
2659 ret
= consumer_send_status_msg(sock
, ret_code
);
2661 /* Somehow, the session daemon is not responding anymore. */
2665 /* Copy socket information and received FD */
2666 switch (sock_type
) {
2667 case LTTNG_STREAM_CONTROL
:
2668 /* Copy received lttcomm socket */
2669 lttcomm_copy_sock(&relayd
->control_sock
, relayd_sock
);
2670 ret
= lttcomm_create_sock(&relayd
->control_sock
);
2671 /* Immediately try to close the created socket if valid. */
2672 if (relayd
->control_sock
.fd
>= 0) {
2673 if (close(relayd
->control_sock
.fd
)) {
2674 PERROR("close relayd control socket");
2677 /* Handle create_sock error. */
2682 /* Assign new file descriptor */
2683 relayd
->control_sock
.fd
= fd
;
2686 * Create a session on the relayd and store the returned id. Lock the
2687 * control socket mutex if the relayd was NOT created before.
2689 if (!relayd_created
) {
2690 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
2692 ret
= relayd_create_session(&relayd
->control_sock
,
2693 &relayd
->relayd_session_id
);
2694 if (!relayd_created
) {
2695 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
2699 * Close all sockets of a relayd object. It will be freed if it was
2700 * created at the error code path or else it will be garbage
2703 (void) relayd_close(&relayd
->control_sock
);
2704 (void) relayd_close(&relayd
->data_sock
);
2709 case LTTNG_STREAM_DATA
:
2710 /* Copy received lttcomm socket */
2711 lttcomm_copy_sock(&relayd
->data_sock
, relayd_sock
);
2712 ret
= lttcomm_create_sock(&relayd
->data_sock
);
2713 /* Immediately try to close the created socket if valid. */
2714 if (relayd
->data_sock
.fd
>= 0) {
2715 if (close(relayd
->data_sock
.fd
)) {
2716 PERROR("close relayd data socket");
2719 /* Handle create_sock error. */
2724 /* Assign new file descriptor */
2725 relayd
->data_sock
.fd
= fd
;
2728 ERR("Unknown relayd socket type (%d)", sock_type
);
2733 DBG("Consumer %s socket created successfully with net idx %d (fd: %d)",
2734 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
2735 relayd
->net_seq_idx
, fd
);
2738 * Add relayd socket pair to consumer data hashtable. If object already
2739 * exists or on error, the function gracefully returns.
2747 /* Close received socket if valid. */
2750 PERROR("close received socket");
2755 if (relayd_created
) {
2763 * Try to lock the stream mutex.
2765 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
2767 static int stream_try_lock(struct lttng_consumer_stream
*stream
)
2774 * Try to lock the stream mutex. On failure, we know that the stream is
2775 * being used else where hence there is data still being extracted.
2777 ret
= pthread_mutex_trylock(&stream
->lock
);
2779 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
2791 * Search for a relayd associated to the session id and return the reference.
2793 * A rcu read side lock MUST be acquire before calling this function and locked
2794 * until the relayd object is no longer necessary.
2796 static struct consumer_relayd_sock_pair
*find_relayd_by_session_id(uint64_t id
)
2798 struct lttng_ht_iter iter
;
2799 struct consumer_relayd_sock_pair
*relayd
= NULL
;
2801 /* Iterate over all relayd since they are indexed by net_seq_idx. */
2802 cds_lfht_for_each_entry(consumer_data
.relayd_ht
->ht
, &iter
.iter
, relayd
,
2805 * Check by sessiond id which is unique here where the relayd session
2806 * id might not be when having multiple relayd.
2808 if (relayd
->sessiond_session_id
== id
) {
2809 /* Found the relayd. There can be only one per id. */
2821 * Check if for a given session id there is still data needed to be extract
2824 * Return 1 if data is pending or else 0 meaning ready to be read.
2826 int consumer_data_pending(uint64_t id
)
2829 struct lttng_ht_iter iter
;
2830 struct lttng_ht
*ht
;
2831 struct lttng_consumer_stream
*stream
;
2832 struct consumer_relayd_sock_pair
*relayd
= NULL
;
2833 int (*data_pending
)(struct lttng_consumer_stream
*);
2835 DBG("Consumer data pending command on session id %" PRIu64
, id
);
2838 pthread_mutex_lock(&consumer_data
.lock
);
2840 switch (consumer_data
.type
) {
2841 case LTTNG_CONSUMER_KERNEL
:
2842 data_pending
= lttng_kconsumer_data_pending
;
2844 case LTTNG_CONSUMER32_UST
:
2845 case LTTNG_CONSUMER64_UST
:
2846 data_pending
= lttng_ustconsumer_data_pending
;
2849 ERR("Unknown consumer data type");
2853 /* Ease our life a bit */
2854 ht
= consumer_data
.stream_list_ht
;
2856 relayd
= find_relayd_by_session_id(id
);
2858 /* Send init command for data pending. */
2859 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
2860 ret
= relayd_begin_data_pending(&relayd
->control_sock
,
2861 relayd
->relayd_session_id
);
2862 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
2864 /* Communication error thus the relayd so no data pending. */
2865 goto data_not_pending
;
2869 cds_lfht_for_each_entry_duplicate(ht
->ht
,
2870 ht
->hash_fct((void *)((unsigned long) id
), lttng_ht_seed
),
2871 ht
->match_fct
, (void *)((unsigned long) id
),
2872 &iter
.iter
, stream
, node_session_id
.node
) {
2873 /* If this call fails, the stream is being used hence data pending. */
2874 ret
= stream_try_lock(stream
);
2880 * A removed node from the hash table indicates that the stream has
2881 * been deleted thus having a guarantee that the buffers are closed
2882 * on the consumer side. However, data can still be transmitted
2883 * over the network so don't skip the relayd check.
2885 ret
= cds_lfht_is_node_deleted(&stream
->node
.node
);
2887 /* Check the stream if there is data in the buffers. */
2888 ret
= data_pending(stream
);
2890 pthread_mutex_unlock(&stream
->lock
);
2897 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
2898 if (stream
->metadata_flag
) {
2899 ret
= relayd_quiescent_control(&relayd
->control_sock
,
2900 stream
->relayd_stream_id
);
2902 ret
= relayd_data_pending(&relayd
->control_sock
,
2903 stream
->relayd_stream_id
,
2904 stream
->next_net_seq_num
- 1);
2906 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
2908 pthread_mutex_unlock(&stream
->lock
);
2912 pthread_mutex_unlock(&stream
->lock
);
2916 unsigned int is_data_inflight
= 0;
2918 /* Send init command for data pending. */
2919 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
2920 ret
= relayd_end_data_pending(&relayd
->control_sock
,
2921 relayd
->relayd_session_id
, &is_data_inflight
);
2922 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
2924 goto data_not_pending
;
2926 if (is_data_inflight
) {
2932 * Finding _no_ node in the hash table and no inflight data means that the
2933 * stream(s) have been removed thus data is guaranteed to be available for
2934 * analysis from the trace files.
2938 /* Data is available to be read by a viewer. */
2939 pthread_mutex_unlock(&consumer_data
.lock
);
2944 /* Data is still being extracted from buffers. */
2945 pthread_mutex_unlock(&consumer_data
.lock
);
2951 * Send a ret code status message to the sessiond daemon.
2953 * Return the sendmsg() return value.
2955 int consumer_send_status_msg(int sock
, int ret_code
)
2957 struct lttcomm_consumer_status_msg msg
;
2959 msg
.ret_code
= ret_code
;
2961 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));
2965 * Send a channel status message to the sessiond daemon.
2967 * Return the sendmsg() return value.
2969 int consumer_send_status_channel(int sock
,
2970 struct lttng_consumer_channel
*channel
)
2972 struct lttcomm_consumer_status_channel msg
;
2977 msg
.ret_code
= -LTTNG_ERR_UST_CHAN_FAIL
;
2979 msg
.ret_code
= LTTNG_OK
;
2980 msg
.key
= channel
->key
;
2981 msg
.stream_count
= channel
->streams
.count
;
2984 return lttcomm_send_unix_sock(sock
, &msg
, sizeof(msg
));