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
,
50 /* timeout parameter, to control the polling thread grace period. */
51 int consumer_poll_timeout
= -1;
54 * Flag to inform the polling thread to quit when all fd hung up. Updated by
55 * the consumer_thread_receive_fds when it notices that all fds has hung up.
56 * Also updated by the signal handler (consumer_should_exit()). Read by the
59 volatile int consumer_quit
= 0;
62 * Find a stream. The consumer_data.lock must be locked during this
65 static struct lttng_consumer_stream
*consumer_find_stream(int key
)
67 struct lttng_ht_iter iter
;
68 struct lttng_ht_node_ulong
*node
;
69 struct lttng_consumer_stream
*stream
= NULL
;
71 /* Negative keys are lookup failures */
78 lttng_ht_lookup(consumer_data
.stream_ht
, (void *)((unsigned long) key
),
80 node
= lttng_ht_iter_get_node_ulong(&iter
);
82 stream
= caa_container_of(node
, struct lttng_consumer_stream
, node
);
90 static void consumer_steal_stream_key(int key
)
92 struct lttng_consumer_stream
*stream
;
95 stream
= consumer_find_stream(key
);
99 * We don't want the lookup to match, but we still need
100 * to iterate on this stream when iterating over the hash table. Just
101 * change the node key.
103 stream
->node
.key
= -1;
108 static struct lttng_consumer_channel
*consumer_find_channel(int key
)
110 struct lttng_ht_iter iter
;
111 struct lttng_ht_node_ulong
*node
;
112 struct lttng_consumer_channel
*channel
= NULL
;
114 /* Negative keys are lookup failures */
121 lttng_ht_lookup(consumer_data
.channel_ht
, (void *)((unsigned long) key
),
123 node
= lttng_ht_iter_get_node_ulong(&iter
);
125 channel
= caa_container_of(node
, struct lttng_consumer_channel
, node
);
133 static void consumer_steal_channel_key(int key
)
135 struct lttng_consumer_channel
*channel
;
138 channel
= consumer_find_channel(key
);
142 * We don't want the lookup to match, but we still need
143 * to iterate on this channel when iterating over the hash table. Just
144 * change the node key.
146 channel
->node
.key
= -1;
152 void consumer_free_stream(struct rcu_head
*head
)
154 struct lttng_ht_node_ulong
*node
=
155 caa_container_of(head
, struct lttng_ht_node_ulong
, head
);
156 struct lttng_consumer_stream
*stream
=
157 caa_container_of(node
, struct lttng_consumer_stream
, node
);
163 * RCU protected relayd socket pair free.
165 static void consumer_rcu_free_relayd(struct rcu_head
*head
)
167 struct lttng_ht_node_ulong
*node
=
168 caa_container_of(head
, struct lttng_ht_node_ulong
, head
);
169 struct consumer_relayd_sock_pair
*relayd
=
170 caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
176 * Destroy and free relayd socket pair object.
178 * This function MUST be called with the consumer_data lock acquired.
180 static void destroy_relayd(struct consumer_relayd_sock_pair
*relayd
)
183 struct lttng_ht_iter iter
;
185 if (relayd
== NULL
) {
189 DBG("Consumer destroy and close relayd socket pair");
191 iter
.iter
.node
= &relayd
->node
.node
;
192 ret
= lttng_ht_del(consumer_data
.relayd_ht
, &iter
);
194 /* We assume the relayd was already destroyed */
198 /* Close all sockets */
199 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
200 (void) relayd_close(&relayd
->control_sock
);
201 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
202 (void) relayd_close(&relayd
->data_sock
);
204 /* RCU free() call */
205 call_rcu(&relayd
->node
.head
, consumer_rcu_free_relayd
);
209 * Flag a relayd socket pair for destruction. Destroy it if the refcount
212 * RCU read side lock MUST be aquired before calling this function.
214 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair
*relayd
)
218 /* Set destroy flag for this object */
219 uatomic_set(&relayd
->destroy_flag
, 1);
221 /* Destroy the relayd if refcount is 0 */
222 if (uatomic_read(&relayd
->refcount
) == 0) {
223 destroy_relayd(relayd
);
228 * Remove a stream from the global list protected by a mutex. This
229 * function is also responsible for freeing its data structures.
231 void consumer_del_stream(struct lttng_consumer_stream
*stream
)
234 struct lttng_ht_iter iter
;
235 struct lttng_consumer_channel
*free_chan
= NULL
;
236 struct consumer_relayd_sock_pair
*relayd
;
240 pthread_mutex_lock(&consumer_data
.lock
);
242 switch (consumer_data
.type
) {
243 case LTTNG_CONSUMER_KERNEL
:
244 if (stream
->mmap_base
!= NULL
) {
245 ret
= munmap(stream
->mmap_base
, stream
->mmap_len
);
251 case LTTNG_CONSUMER32_UST
:
252 case LTTNG_CONSUMER64_UST
:
253 lttng_ustconsumer_del_stream(stream
);
256 ERR("Unknown consumer_data type");
262 iter
.iter
.node
= &stream
->node
.node
;
263 ret
= lttng_ht_del(consumer_data
.stream_ht
, &iter
);
268 if (consumer_data
.stream_count
<= 0) {
271 consumer_data
.stream_count
--;
275 if (stream
->out_fd
>= 0) {
276 ret
= close(stream
->out_fd
);
281 if (stream
->wait_fd
>= 0 && !stream
->wait_fd_is_copy
) {
282 ret
= close(stream
->wait_fd
);
287 if (stream
->shm_fd
>= 0 && stream
->wait_fd
!= stream
->shm_fd
) {
288 ret
= close(stream
->shm_fd
);
294 /* Check and cleanup relayd */
296 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
297 if (relayd
!= NULL
) {
298 uatomic_dec(&relayd
->refcount
);
299 assert(uatomic_read(&relayd
->refcount
) >= 0);
301 /* Closing streams requires to lock the control socket. */
302 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
303 ret
= relayd_send_close_stream(&relayd
->control_sock
,
304 stream
->relayd_stream_id
,
305 stream
->next_net_seq_num
- 1);
306 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
308 DBG("Unable to close stream on the relayd. Continuing");
310 * Continue here. There is nothing we can do for the relayd.
311 * Chances are that the relayd has closed the socket so we just
312 * continue cleaning up.
316 /* Both conditions are met, we destroy the relayd. */
317 if (uatomic_read(&relayd
->refcount
) == 0 &&
318 uatomic_read(&relayd
->destroy_flag
)) {
319 destroy_relayd(relayd
);
324 uatomic_dec(&stream
->chan
->refcount
);
325 if (!uatomic_read(&stream
->chan
->refcount
)
326 && !uatomic_read(&stream
->chan
->nb_init_streams
)) {
327 free_chan
= stream
->chan
;
330 call_rcu(&stream
->node
.head
, consumer_free_stream
);
332 consumer_data
.need_update
= 1;
333 pthread_mutex_unlock(&consumer_data
.lock
);
336 consumer_del_channel(free_chan
);
340 struct lttng_consumer_stream
*consumer_allocate_stream(
341 int channel_key
, int stream_key
,
342 int shm_fd
, int wait_fd
,
343 enum lttng_consumer_stream_state state
,
345 enum lttng_event_output output
,
346 const char *path_name
,
353 struct lttng_consumer_stream
*stream
;
356 stream
= zmalloc(sizeof(*stream
));
357 if (stream
== NULL
) {
358 PERROR("malloc struct lttng_consumer_stream");
359 *alloc_ret
= -ENOMEM
;
364 * Get stream's channel reference. Needed when adding the stream to the
367 stream
->chan
= consumer_find_channel(channel_key
);
369 *alloc_ret
= -ENOENT
;
370 ERR("Unable to find channel for stream %d", stream_key
);
373 stream
->chan
->refcount
++;
374 stream
->key
= stream_key
;
375 stream
->shm_fd
= shm_fd
;
376 stream
->wait_fd
= wait_fd
;
378 stream
->out_fd_offset
= 0;
379 stream
->state
= state
;
380 stream
->mmap_len
= mmap_len
;
381 stream
->mmap_base
= NULL
;
382 stream
->output
= output
;
385 stream
->net_seq_idx
= net_index
;
386 stream
->metadata_flag
= metadata_flag
;
387 strncpy(stream
->path_name
, path_name
, sizeof(stream
->path_name
));
388 stream
->path_name
[sizeof(stream
->path_name
) - 1] = '\0';
389 lttng_ht_node_init_ulong(&stream
->node
, stream
->key
);
390 lttng_ht_node_init_ulong(&stream
->waitfd_node
, stream
->wait_fd
);
392 switch (consumer_data
.type
) {
393 case LTTNG_CONSUMER_KERNEL
:
395 case LTTNG_CONSUMER32_UST
:
396 case LTTNG_CONSUMER64_UST
:
397 stream
->cpu
= stream
->chan
->cpucount
++;
398 ret
= lttng_ustconsumer_allocate_stream(stream
);
400 *alloc_ret
= -EINVAL
;
405 ERR("Unknown consumer_data type");
406 *alloc_ret
= -EINVAL
;
411 * When nb_init_streams reaches 0, we don't need to trigger any action in
412 * terms of destroying the associated channel, because the action that
413 * causes the count to become 0 also causes a stream to be added. The
414 * channel deletion will thus be triggered by the following removal of this
417 if (uatomic_read(&stream
->chan
->nb_init_streams
) > 0) {
418 uatomic_dec(&stream
->chan
->nb_init_streams
);
421 DBG3("Allocated stream %s (key %d, shm_fd %d, wait_fd %d, mmap_len %llu,"
422 " out_fd %d, net_seq_idx %d)", stream
->path_name
, stream
->key
,
423 stream
->shm_fd
, stream
->wait_fd
,
424 (unsigned long long) stream
->mmap_len
, stream
->out_fd
,
425 stream
->net_seq_idx
);
435 * Add a stream to the global list protected by a mutex.
437 int consumer_add_stream(struct lttng_consumer_stream
*stream
)
440 struct lttng_ht_node_ulong
*node
;
441 struct lttng_ht_iter iter
;
442 struct consumer_relayd_sock_pair
*relayd
;
444 pthread_mutex_lock(&consumer_data
.lock
);
445 /* Steal stream identifier, for UST */
446 consumer_steal_stream_key(stream
->key
);
449 lttng_ht_lookup(consumer_data
.stream_ht
,
450 (void *)((unsigned long) stream
->key
), &iter
);
451 node
= lttng_ht_iter_get_node_ulong(&iter
);
454 /* Stream already exist. Ignore the insertion */
458 lttng_ht_add_unique_ulong(consumer_data
.stream_ht
, &stream
->node
);
460 /* Check and cleanup relayd */
461 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
462 if (relayd
!= NULL
) {
463 uatomic_inc(&relayd
->refcount
);
467 /* Update consumer data */
468 consumer_data
.stream_count
++;
469 consumer_data
.need_update
= 1;
472 pthread_mutex_unlock(&consumer_data
.lock
);
478 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
479 * be acquired before calling this.
481 static int add_relayd(struct consumer_relayd_sock_pair
*relayd
)
484 struct lttng_ht_node_ulong
*node
;
485 struct lttng_ht_iter iter
;
487 if (relayd
== NULL
) {
492 lttng_ht_lookup(consumer_data
.relayd_ht
,
493 (void *)((unsigned long) relayd
->net_seq_idx
), &iter
);
494 node
= lttng_ht_iter_get_node_ulong(&iter
);
496 /* Relayd already exist. Ignore the insertion */
499 lttng_ht_add_unique_ulong(consumer_data
.relayd_ht
, &relayd
->node
);
506 * Allocate and return a consumer relayd socket.
508 struct consumer_relayd_sock_pair
*consumer_allocate_relayd_sock_pair(
511 struct consumer_relayd_sock_pair
*obj
= NULL
;
513 /* Negative net sequence index is a failure */
514 if (net_seq_idx
< 0) {
518 obj
= zmalloc(sizeof(struct consumer_relayd_sock_pair
));
520 PERROR("zmalloc relayd sock");
524 obj
->net_seq_idx
= net_seq_idx
;
526 obj
->destroy_flag
= 0;
527 lttng_ht_node_init_ulong(&obj
->node
, obj
->net_seq_idx
);
528 pthread_mutex_init(&obj
->ctrl_sock_mutex
, NULL
);
535 * Find a relayd socket pair in the global consumer data.
537 * Return the object if found else NULL.
538 * RCU read-side lock must be held across this call and while using the
541 struct consumer_relayd_sock_pair
*consumer_find_relayd(int key
)
543 struct lttng_ht_iter iter
;
544 struct lttng_ht_node_ulong
*node
;
545 struct consumer_relayd_sock_pair
*relayd
= NULL
;
547 /* Negative keys are lookup failures */
552 lttng_ht_lookup(consumer_data
.relayd_ht
, (void *)((unsigned long) key
),
554 node
= lttng_ht_iter_get_node_ulong(&iter
);
556 relayd
= caa_container_of(node
, struct consumer_relayd_sock_pair
, node
);
564 * Handle stream for relayd transmission if the stream applies for network
565 * streaming where the net sequence index is set.
567 * Return destination file descriptor or negative value on error.
569 static int write_relayd_stream_header(struct lttng_consumer_stream
*stream
,
570 size_t data_size
, unsigned long padding
,
571 struct consumer_relayd_sock_pair
*relayd
)
574 struct lttcomm_relayd_data_hdr data_hdr
;
580 /* Reset data header */
581 memset(&data_hdr
, 0, sizeof(data_hdr
));
583 if (stream
->metadata_flag
) {
584 /* Caller MUST acquire the relayd control socket lock */
585 ret
= relayd_send_metadata(&relayd
->control_sock
, data_size
);
590 /* Metadata are always sent on the control socket. */
591 outfd
= relayd
->control_sock
.fd
;
593 /* Set header with stream information */
594 data_hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
595 data_hdr
.data_size
= htobe32(data_size
);
596 data_hdr
.padding_size
= htobe32(padding
);
597 data_hdr
.net_seq_num
= htobe64(stream
->next_net_seq_num
++);
598 /* Other fields are zeroed previously */
600 ret
= relayd_send_data_hdr(&relayd
->data_sock
, &data_hdr
,
606 /* Set to go on data socket */
607 outfd
= relayd
->data_sock
.fd
;
615 * Update a stream according to what we just received.
617 void consumer_change_stream_state(int stream_key
,
618 enum lttng_consumer_stream_state state
)
620 struct lttng_consumer_stream
*stream
;
622 pthread_mutex_lock(&consumer_data
.lock
);
623 stream
= consumer_find_stream(stream_key
);
625 stream
->state
= state
;
627 consumer_data
.need_update
= 1;
628 pthread_mutex_unlock(&consumer_data
.lock
);
632 void consumer_free_channel(struct rcu_head
*head
)
634 struct lttng_ht_node_ulong
*node
=
635 caa_container_of(head
, struct lttng_ht_node_ulong
, head
);
636 struct lttng_consumer_channel
*channel
=
637 caa_container_of(node
, struct lttng_consumer_channel
, node
);
643 * Remove a channel from the global list protected by a mutex. This
644 * function is also responsible for freeing its data structures.
646 void consumer_del_channel(struct lttng_consumer_channel
*channel
)
649 struct lttng_ht_iter iter
;
651 pthread_mutex_lock(&consumer_data
.lock
);
653 switch (consumer_data
.type
) {
654 case LTTNG_CONSUMER_KERNEL
:
656 case LTTNG_CONSUMER32_UST
:
657 case LTTNG_CONSUMER64_UST
:
658 lttng_ustconsumer_del_channel(channel
);
661 ERR("Unknown consumer_data type");
667 iter
.iter
.node
= &channel
->node
.node
;
668 ret
= lttng_ht_del(consumer_data
.channel_ht
, &iter
);
672 if (channel
->mmap_base
!= NULL
) {
673 ret
= munmap(channel
->mmap_base
, channel
->mmap_len
);
678 if (channel
->wait_fd
>= 0 && !channel
->wait_fd_is_copy
) {
679 ret
= close(channel
->wait_fd
);
684 if (channel
->shm_fd
>= 0 && channel
->wait_fd
!= channel
->shm_fd
) {
685 ret
= close(channel
->shm_fd
);
691 call_rcu(&channel
->node
.head
, consumer_free_channel
);
693 pthread_mutex_unlock(&consumer_data
.lock
);
696 struct lttng_consumer_channel
*consumer_allocate_channel(
698 int shm_fd
, int wait_fd
,
700 uint64_t max_sb_size
,
701 unsigned int nb_init_streams
)
703 struct lttng_consumer_channel
*channel
;
706 channel
= zmalloc(sizeof(*channel
));
707 if (channel
== NULL
) {
708 PERROR("malloc struct lttng_consumer_channel");
711 channel
->key
= channel_key
;
712 channel
->shm_fd
= shm_fd
;
713 channel
->wait_fd
= wait_fd
;
714 channel
->mmap_len
= mmap_len
;
715 channel
->max_sb_size
= max_sb_size
;
716 channel
->refcount
= 0;
717 channel
->nb_init_streams
= nb_init_streams
;
718 lttng_ht_node_init_ulong(&channel
->node
, channel
->key
);
720 switch (consumer_data
.type
) {
721 case LTTNG_CONSUMER_KERNEL
:
722 channel
->mmap_base
= NULL
;
723 channel
->mmap_len
= 0;
725 case LTTNG_CONSUMER32_UST
:
726 case LTTNG_CONSUMER64_UST
:
727 ret
= lttng_ustconsumer_allocate_channel(channel
);
734 ERR("Unknown consumer_data type");
738 DBG("Allocated channel (key %d, shm_fd %d, wait_fd %d, mmap_len %llu, max_sb_size %llu)",
739 channel
->key
, channel
->shm_fd
, channel
->wait_fd
,
740 (unsigned long long) channel
->mmap_len
,
741 (unsigned long long) channel
->max_sb_size
);
747 * Add a channel to the global list protected by a mutex.
749 int consumer_add_channel(struct lttng_consumer_channel
*channel
)
751 struct lttng_ht_node_ulong
*node
;
752 struct lttng_ht_iter iter
;
754 pthread_mutex_lock(&consumer_data
.lock
);
755 /* Steal channel identifier, for UST */
756 consumer_steal_channel_key(channel
->key
);
759 lttng_ht_lookup(consumer_data
.channel_ht
,
760 (void *)((unsigned long) channel
->key
), &iter
);
761 node
= lttng_ht_iter_get_node_ulong(&iter
);
763 /* Channel already exist. Ignore the insertion */
767 lttng_ht_add_unique_ulong(consumer_data
.channel_ht
, &channel
->node
);
771 pthread_mutex_unlock(&consumer_data
.lock
);
777 * Allocate the pollfd structure and the local view of the out fds to avoid
778 * doing a lookup in the linked list and concurrency issues when writing is
779 * needed. Called with consumer_data.lock held.
781 * Returns the number of fds in the structures.
783 int consumer_update_poll_array(
784 struct lttng_consumer_local_data
*ctx
, struct pollfd
**pollfd
,
785 struct lttng_consumer_stream
**local_stream
)
788 struct lttng_ht_iter iter
;
789 struct lttng_consumer_stream
*stream
;
791 DBG("Updating poll fd array");
793 cds_lfht_for_each_entry(consumer_data
.stream_ht
->ht
, &iter
.iter
, stream
,
795 if (stream
->state
!= LTTNG_CONSUMER_ACTIVE_STREAM
) {
798 DBG("Active FD %d", stream
->wait_fd
);
799 (*pollfd
)[i
].fd
= stream
->wait_fd
;
800 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
801 local_stream
[i
] = stream
;
807 * Insert the consumer_poll_pipe at the end of the array and don't
808 * increment i so nb_fd is the number of real FD.
810 (*pollfd
)[i
].fd
= ctx
->consumer_poll_pipe
[0];
811 (*pollfd
)[i
].events
= POLLIN
| POLLPRI
;
816 * Poll on the should_quit pipe and the command socket return -1 on error and
817 * should exit, 0 if data is available on the command socket
819 int lttng_consumer_poll_socket(struct pollfd
*consumer_sockpoll
)
824 num_rdy
= poll(consumer_sockpoll
, 2, -1);
827 * Restart interrupted system call.
829 if (errno
== EINTR
) {
832 PERROR("Poll error");
835 if (consumer_sockpoll
[0].revents
& (POLLIN
| POLLPRI
)) {
836 DBG("consumer_should_quit wake up");
846 * Set the error socket.
848 void lttng_consumer_set_error_sock(
849 struct lttng_consumer_local_data
*ctx
, int sock
)
851 ctx
->consumer_error_socket
= sock
;
855 * Set the command socket path.
857 void lttng_consumer_set_command_sock_path(
858 struct lttng_consumer_local_data
*ctx
, char *sock
)
860 ctx
->consumer_command_sock_path
= sock
;
864 * Send return code to the session daemon.
865 * If the socket is not defined, we return 0, it is not a fatal error
867 int lttng_consumer_send_error(
868 struct lttng_consumer_local_data
*ctx
, int cmd
)
870 if (ctx
->consumer_error_socket
> 0) {
871 return lttcomm_send_unix_sock(ctx
->consumer_error_socket
, &cmd
,
872 sizeof(enum lttcomm_sessiond_command
));
879 * Close all the tracefiles and stream fds, should be called when all instances
882 void lttng_consumer_cleanup(void)
884 struct lttng_ht_iter iter
;
885 struct lttng_ht_node_ulong
*node
;
890 * close all outfd. Called when there are no more threads running (after
891 * joining on the threads), no need to protect list iteration with mutex.
893 cds_lfht_for_each_entry(consumer_data
.stream_ht
->ht
, &iter
.iter
, node
,
895 struct lttng_consumer_stream
*stream
=
896 caa_container_of(node
, struct lttng_consumer_stream
, node
);
897 consumer_del_stream(stream
);
900 cds_lfht_for_each_entry(consumer_data
.channel_ht
->ht
, &iter
.iter
, node
,
902 struct lttng_consumer_channel
*channel
=
903 caa_container_of(node
, struct lttng_consumer_channel
, node
);
904 consumer_del_channel(channel
);
909 lttng_ht_destroy(consumer_data
.stream_ht
);
910 lttng_ht_destroy(consumer_data
.channel_ht
);
914 * Called from signal handler.
916 void lttng_consumer_should_exit(struct lttng_consumer_local_data
*ctx
)
921 ret
= write(ctx
->consumer_should_quit
[1], "4", 1);
922 } while (ret
< 0 && errno
== EINTR
);
924 PERROR("write consumer quit");
928 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream
*stream
,
931 int outfd
= stream
->out_fd
;
934 * This does a blocking write-and-wait on any page that belongs to the
935 * subbuffer prior to the one we just wrote.
936 * Don't care about error values, as these are just hints and ways to
937 * limit the amount of page cache used.
939 if (orig_offset
< stream
->chan
->max_sb_size
) {
942 lttng_sync_file_range(outfd
, orig_offset
- stream
->chan
->max_sb_size
,
943 stream
->chan
->max_sb_size
,
944 SYNC_FILE_RANGE_WAIT_BEFORE
945 | SYNC_FILE_RANGE_WRITE
946 | SYNC_FILE_RANGE_WAIT_AFTER
);
948 * Give hints to the kernel about how we access the file:
949 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
952 * We need to call fadvise again after the file grows because the
953 * kernel does not seem to apply fadvise to non-existing parts of the
956 * Call fadvise _after_ having waited for the page writeback to
957 * complete because the dirty page writeback semantic is not well
958 * defined. So it can be expected to lead to lower throughput in
961 posix_fadvise(outfd
, orig_offset
- stream
->chan
->max_sb_size
,
962 stream
->chan
->max_sb_size
, POSIX_FADV_DONTNEED
);
966 * Initialise the necessary environnement :
967 * - create a new context
968 * - create the poll_pipe
969 * - create the should_quit pipe (for signal handler)
970 * - create the thread pipe (for splice)
972 * Takes a function pointer as argument, this function is called when data is
973 * available on a buffer. This function is responsible to do the
974 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
975 * buffer configuration and then kernctl_put_next_subbuf at the end.
977 * Returns a pointer to the new context or NULL on error.
979 struct lttng_consumer_local_data
*lttng_consumer_create(
980 enum lttng_consumer_type type
,
981 ssize_t (*buffer_ready
)(struct lttng_consumer_stream
*stream
,
982 struct lttng_consumer_local_data
*ctx
),
983 int (*recv_channel
)(struct lttng_consumer_channel
*channel
),
984 int (*recv_stream
)(struct lttng_consumer_stream
*stream
),
985 int (*update_stream
)(int stream_key
, uint32_t state
))
988 struct lttng_consumer_local_data
*ctx
;
990 assert(consumer_data
.type
== LTTNG_CONSUMER_UNKNOWN
||
991 consumer_data
.type
== type
);
992 consumer_data
.type
= type
;
994 ctx
= zmalloc(sizeof(struct lttng_consumer_local_data
));
996 PERROR("allocating context");
1000 ctx
->consumer_error_socket
= -1;
1001 /* assign the callbacks */
1002 ctx
->on_buffer_ready
= buffer_ready
;
1003 ctx
->on_recv_channel
= recv_channel
;
1004 ctx
->on_recv_stream
= recv_stream
;
1005 ctx
->on_update_stream
= update_stream
;
1007 ret
= pipe(ctx
->consumer_poll_pipe
);
1009 PERROR("Error creating poll pipe");
1010 goto error_poll_pipe
;
1013 /* set read end of the pipe to non-blocking */
1014 ret
= fcntl(ctx
->consumer_poll_pipe
[0], F_SETFL
, O_NONBLOCK
);
1016 PERROR("fcntl O_NONBLOCK");
1017 goto error_poll_fcntl
;
1020 /* set write end of the pipe to non-blocking */
1021 ret
= fcntl(ctx
->consumer_poll_pipe
[1], F_SETFL
, O_NONBLOCK
);
1023 PERROR("fcntl O_NONBLOCK");
1024 goto error_poll_fcntl
;
1027 ret
= pipe(ctx
->consumer_should_quit
);
1029 PERROR("Error creating recv pipe");
1030 goto error_quit_pipe
;
1033 ret
= pipe(ctx
->consumer_thread_pipe
);
1035 PERROR("Error creating thread pipe");
1036 goto error_thread_pipe
;
1039 ret
= utils_create_pipe(ctx
->consumer_metadata_pipe
);
1041 goto error_metadata_pipe
;
1044 ret
= utils_create_pipe(ctx
->consumer_splice_metadata_pipe
);
1046 goto error_splice_pipe
;
1052 utils_close_pipe(ctx
->consumer_metadata_pipe
);
1053 error_metadata_pipe
:
1054 utils_close_pipe(ctx
->consumer_thread_pipe
);
1056 for (i
= 0; i
< 2; i
++) {
1059 err
= close(ctx
->consumer_should_quit
[i
]);
1066 for (i
= 0; i
< 2; i
++) {
1069 err
= close(ctx
->consumer_poll_pipe
[i
]);
1081 * Close all fds associated with the instance and free the context.
1083 void lttng_consumer_destroy(struct lttng_consumer_local_data
*ctx
)
1087 ret
= close(ctx
->consumer_error_socket
);
1091 ret
= close(ctx
->consumer_thread_pipe
[0]);
1095 ret
= close(ctx
->consumer_thread_pipe
[1]);
1099 ret
= close(ctx
->consumer_poll_pipe
[0]);
1103 ret
= close(ctx
->consumer_poll_pipe
[1]);
1107 ret
= close(ctx
->consumer_should_quit
[0]);
1111 ret
= close(ctx
->consumer_should_quit
[1]);
1115 utils_close_pipe(ctx
->consumer_splice_metadata_pipe
);
1117 unlink(ctx
->consumer_command_sock_path
);
1122 * Write the metadata stream id on the specified file descriptor.
1124 static int write_relayd_metadata_id(int fd
,
1125 struct lttng_consumer_stream
*stream
,
1126 struct consumer_relayd_sock_pair
*relayd
,
1127 unsigned long padding
)
1130 struct lttcomm_relayd_metadata_payload hdr
;
1132 hdr
.stream_id
= htobe64(stream
->relayd_stream_id
);
1133 hdr
.padding_size
= htobe32(padding
);
1135 ret
= write(fd
, (void *) &hdr
, sizeof(hdr
));
1136 } while (ret
< 0 && errno
== EINTR
);
1138 PERROR("write metadata stream id");
1141 DBG("Metadata stream id %" PRIu64
" with padding %lu written before data",
1142 stream
->relayd_stream_id
, padding
);
1149 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1150 * core function for writing trace buffers to either the local filesystem or
1153 * Careful review MUST be put if any changes occur!
1155 * Returns the number of bytes written
1157 ssize_t
lttng_consumer_on_read_subbuffer_mmap(
1158 struct lttng_consumer_local_data
*ctx
,
1159 struct lttng_consumer_stream
*stream
, unsigned long len
,
1160 unsigned long padding
)
1162 unsigned long mmap_offset
;
1163 ssize_t ret
= 0, written
= 0;
1164 off_t orig_offset
= stream
->out_fd_offset
;
1165 /* Default is on the disk */
1166 int outfd
= stream
->out_fd
;
1167 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1169 /* RCU lock for the relayd pointer */
1172 /* Flag that the current stream if set for network streaming. */
1173 if (stream
->net_seq_idx
!= -1) {
1174 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1175 if (relayd
== NULL
) {
1180 /* get the offset inside the fd to mmap */
1181 switch (consumer_data
.type
) {
1182 case LTTNG_CONSUMER_KERNEL
:
1183 ret
= kernctl_get_mmap_read_offset(stream
->wait_fd
, &mmap_offset
);
1185 case LTTNG_CONSUMER32_UST
:
1186 case LTTNG_CONSUMER64_UST
:
1187 ret
= lttng_ustctl_get_mmap_read_offset(stream
->chan
->handle
,
1188 stream
->buf
, &mmap_offset
);
1191 ERR("Unknown consumer_data type");
1196 PERROR("tracer ctl get_mmap_read_offset");
1201 /* Handle stream on the relayd if the output is on the network */
1203 unsigned long netlen
= len
;
1206 * Lock the control socket for the complete duration of the function
1207 * since from this point on we will use the socket.
1209 if (stream
->metadata_flag
) {
1210 /* Metadata requires the control socket. */
1211 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1212 netlen
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1215 ret
= write_relayd_stream_header(stream
, netlen
, padding
, relayd
);
1217 /* Use the returned socket. */
1220 /* Write metadata stream id before payload */
1221 if (stream
->metadata_flag
) {
1222 ret
= write_relayd_metadata_id(outfd
, stream
, relayd
, padding
);
1229 /* Else, use the default set before which is the filesystem. */
1231 /* No streaming, we have to set the len with the full padding */
1237 ret
= write(outfd
, stream
->mmap_base
+ mmap_offset
, len
);
1238 } while (ret
< 0 && errno
== EINTR
);
1239 DBG("Consumer mmap write() ret %zd (len %lu)", ret
, len
);
1241 PERROR("Error in file write");
1246 } else if (ret
> len
) {
1247 PERROR("Error in file write (ret %zd > len %lu)", ret
, len
);
1255 /* This call is useless on a socket so better save a syscall. */
1257 /* This won't block, but will start writeout asynchronously */
1258 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret
,
1259 SYNC_FILE_RANGE_WRITE
);
1260 stream
->out_fd_offset
+= ret
;
1264 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1267 /* Unlock only if ctrl socket used */
1268 if (relayd
&& stream
->metadata_flag
) {
1269 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1277 * Splice the data from the ring buffer to the tracefile.
1279 * Returns the number of bytes spliced.
1281 ssize_t
lttng_consumer_on_read_subbuffer_splice(
1282 struct lttng_consumer_local_data
*ctx
,
1283 struct lttng_consumer_stream
*stream
, unsigned long len
,
1284 unsigned long padding
)
1286 ssize_t ret
= 0, written
= 0, ret_splice
= 0;
1288 off_t orig_offset
= stream
->out_fd_offset
;
1289 int fd
= stream
->wait_fd
;
1290 /* Default is on the disk */
1291 int outfd
= stream
->out_fd
;
1292 struct consumer_relayd_sock_pair
*relayd
= NULL
;
1295 switch (consumer_data
.type
) {
1296 case LTTNG_CONSUMER_KERNEL
:
1298 case LTTNG_CONSUMER32_UST
:
1299 case LTTNG_CONSUMER64_UST
:
1300 /* Not supported for user space tracing */
1303 ERR("Unknown consumer_data type");
1307 /* RCU lock for the relayd pointer */
1310 /* Flag that the current stream if set for network streaming. */
1311 if (stream
->net_seq_idx
!= -1) {
1312 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1313 if (relayd
== NULL
) {
1319 * Choose right pipe for splice. Metadata and trace data are handled by
1320 * different threads hence the use of two pipes in order not to race or
1321 * corrupt the written data.
1323 if (stream
->metadata_flag
) {
1324 splice_pipe
= ctx
->consumer_splice_metadata_pipe
;
1326 splice_pipe
= ctx
->consumer_thread_pipe
;
1329 /* Write metadata stream id before payload */
1331 int total_len
= len
;
1333 if (stream
->metadata_flag
) {
1335 * Lock the control socket for the complete duration of the function
1336 * since from this point on we will use the socket.
1338 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1340 ret
= write_relayd_metadata_id(splice_pipe
[1], stream
, relayd
,
1347 total_len
+= sizeof(struct lttcomm_relayd_metadata_payload
);
1350 ret
= write_relayd_stream_header(stream
, total_len
, padding
, relayd
);
1352 /* Use the returned socket. */
1355 ERR("Remote relayd disconnected. Stopping");
1359 /* No streaming, we have to set the len with the full padding */
1364 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1365 (unsigned long)offset
, len
, fd
, splice_pipe
[1]);
1366 ret_splice
= splice(fd
, &offset
, splice_pipe
[1], NULL
, len
,
1367 SPLICE_F_MOVE
| SPLICE_F_MORE
);
1368 DBG("splice chan to pipe, ret %zd", ret_splice
);
1369 if (ret_splice
< 0) {
1370 PERROR("Error in relay splice");
1372 written
= ret_splice
;
1378 /* Handle stream on the relayd if the output is on the network */
1380 if (stream
->metadata_flag
) {
1381 size_t metadata_payload_size
=
1382 sizeof(struct lttcomm_relayd_metadata_payload
);
1384 /* Update counter to fit the spliced data */
1385 ret_splice
+= metadata_payload_size
;
1386 len
+= metadata_payload_size
;
1388 * We do this so the return value can match the len passed as
1389 * argument to this function.
1391 written
-= metadata_payload_size
;
1395 /* Splice data out */
1396 ret_splice
= splice(splice_pipe
[0], NULL
, outfd
, NULL
,
1397 ret_splice
, SPLICE_F_MOVE
| SPLICE_F_MORE
);
1398 DBG("Consumer splice pipe to file, ret %zd", ret_splice
);
1399 if (ret_splice
< 0) {
1400 PERROR("Error in file splice");
1402 written
= ret_splice
;
1406 } else if (ret_splice
> len
) {
1408 PERROR("Wrote more data than requested %zd (len: %lu)",
1410 written
+= ret_splice
;
1416 /* This call is useless on a socket so better save a syscall. */
1418 /* This won't block, but will start writeout asynchronously */
1419 lttng_sync_file_range(outfd
, stream
->out_fd_offset
, ret_splice
,
1420 SYNC_FILE_RANGE_WRITE
);
1421 stream
->out_fd_offset
+= ret_splice
;
1423 written
+= ret_splice
;
1425 lttng_consumer_sync_trace_file(stream
, orig_offset
);
1432 /* send the appropriate error description to sessiond */
1435 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EBADF
);
1438 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_EINVAL
);
1441 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ENOMEM
);
1444 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_SPLICE_ESPIPE
);
1449 if (relayd
&& stream
->metadata_flag
) {
1450 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1458 * Take a snapshot for a specific fd
1460 * Returns 0 on success, < 0 on error
1462 int lttng_consumer_take_snapshot(struct lttng_consumer_local_data
*ctx
,
1463 struct lttng_consumer_stream
*stream
)
1465 switch (consumer_data
.type
) {
1466 case LTTNG_CONSUMER_KERNEL
:
1467 return lttng_kconsumer_take_snapshot(ctx
, stream
);
1468 case LTTNG_CONSUMER32_UST
:
1469 case LTTNG_CONSUMER64_UST
:
1470 return lttng_ustconsumer_take_snapshot(ctx
, stream
);
1472 ERR("Unknown consumer_data type");
1480 * Get the produced position
1482 * Returns 0 on success, < 0 on error
1484 int lttng_consumer_get_produced_snapshot(
1485 struct lttng_consumer_local_data
*ctx
,
1486 struct lttng_consumer_stream
*stream
,
1489 switch (consumer_data
.type
) {
1490 case LTTNG_CONSUMER_KERNEL
:
1491 return lttng_kconsumer_get_produced_snapshot(ctx
, stream
, pos
);
1492 case LTTNG_CONSUMER32_UST
:
1493 case LTTNG_CONSUMER64_UST
:
1494 return lttng_ustconsumer_get_produced_snapshot(ctx
, stream
, pos
);
1496 ERR("Unknown consumer_data type");
1502 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data
*ctx
,
1503 int sock
, struct pollfd
*consumer_sockpoll
)
1505 switch (consumer_data
.type
) {
1506 case LTTNG_CONSUMER_KERNEL
:
1507 return lttng_kconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1508 case LTTNG_CONSUMER32_UST
:
1509 case LTTNG_CONSUMER64_UST
:
1510 return lttng_ustconsumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
1512 ERR("Unknown consumer_data type");
1519 * Iterate over all stream element of the hashtable and free them. This is race
1520 * free since the hashtable received MUST be in a race free synchronization
1521 * state. It's the caller responsability to make sure of that.
1523 static void destroy_stream_ht(struct lttng_ht
*ht
)
1526 struct lttng_ht_iter iter
;
1527 struct lttng_consumer_stream
*stream
;
1534 cds_lfht_for_each_entry(ht
->ht
, &iter
.iter
, stream
, node
.node
) {
1535 ret
= lttng_ht_del(ht
, &iter
);
1542 lttng_ht_destroy(ht
);
1546 * Clean up a metadata stream and free its memory.
1548 static void consumer_del_metadata_stream(struct lttng_consumer_stream
*stream
)
1551 struct consumer_relayd_sock_pair
*relayd
;
1555 * This call should NEVER receive regular stream. It must always be
1556 * metadata stream and this is crucial for data structure synchronization.
1558 assert(stream
->metadata_flag
);
1560 pthread_mutex_lock(&consumer_data
.lock
);
1561 switch (consumer_data
.type
) {
1562 case LTTNG_CONSUMER_KERNEL
:
1563 if (stream
->mmap_base
!= NULL
) {
1564 ret
= munmap(stream
->mmap_base
, stream
->mmap_len
);
1566 PERROR("munmap metadata stream");
1570 case LTTNG_CONSUMER32_UST
:
1571 case LTTNG_CONSUMER64_UST
:
1572 lttng_ustconsumer_del_stream(stream
);
1575 ERR("Unknown consumer_data type");
1578 pthread_mutex_unlock(&consumer_data
.lock
);
1580 if (stream
->out_fd
>= 0) {
1581 ret
= close(stream
->out_fd
);
1587 if (stream
->wait_fd
>= 0 && !stream
->wait_fd_is_copy
) {
1588 ret
= close(stream
->wait_fd
);
1594 if (stream
->shm_fd
>= 0 && stream
->wait_fd
!= stream
->shm_fd
) {
1595 ret
= close(stream
->shm_fd
);
1601 /* Check and cleanup relayd */
1603 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1604 if (relayd
!= NULL
) {
1605 uatomic_dec(&relayd
->refcount
);
1606 assert(uatomic_read(&relayd
->refcount
) >= 0);
1608 /* Closing streams requires to lock the control socket. */
1609 pthread_mutex_lock(&relayd
->ctrl_sock_mutex
);
1610 ret
= relayd_send_close_stream(&relayd
->control_sock
,
1611 stream
->relayd_stream_id
, stream
->next_net_seq_num
- 1);
1612 pthread_mutex_unlock(&relayd
->ctrl_sock_mutex
);
1614 DBG("Unable to close stream on the relayd. Continuing");
1616 * Continue here. There is nothing we can do for the relayd.
1617 * Chances are that the relayd has closed the socket so we just
1618 * continue cleaning up.
1622 /* Both conditions are met, we destroy the relayd. */
1623 if (uatomic_read(&relayd
->refcount
) == 0 &&
1624 uatomic_read(&relayd
->destroy_flag
)) {
1625 destroy_relayd(relayd
);
1630 /* Atomically decrement channel refcount since other threads can use it. */
1631 uatomic_dec(&stream
->chan
->refcount
);
1632 if (!uatomic_read(&stream
->chan
->refcount
)
1633 && !uatomic_read(&stream
->chan
->nb_init_streams
)) {
1634 /* Go for channel deletion! */
1635 consumer_del_channel(stream
->chan
);
1642 * Action done with the metadata stream when adding it to the consumer internal
1643 * data structures to handle it.
1645 static void consumer_add_metadata_stream(struct lttng_consumer_stream
*stream
)
1647 struct consumer_relayd_sock_pair
*relayd
;
1649 /* Find relayd and, if one is found, increment refcount. */
1651 relayd
= consumer_find_relayd(stream
->net_seq_idx
);
1652 if (relayd
!= NULL
) {
1653 uatomic_inc(&relayd
->refcount
);
1659 * Thread polls on metadata file descriptor and write them on disk or on the
1662 void *lttng_consumer_thread_poll_metadata(void *data
)
1665 uint32_t revents
, nb_fd
;
1666 struct lttng_consumer_stream
*stream
;
1667 struct lttng_ht_iter iter
;
1668 struct lttng_ht_node_ulong
*node
;
1669 struct lttng_ht
*metadata_ht
= NULL
;
1670 struct lttng_poll_event events
;
1671 struct lttng_consumer_local_data
*ctx
= data
;
1674 rcu_register_thread();
1676 DBG("Thread metadata poll started");
1678 metadata_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_ULONG
);
1679 if (metadata_ht
== NULL
) {
1683 /* Size is set to 1 for the consumer_metadata pipe */
1684 ret
= lttng_poll_create(&events
, 2, LTTNG_CLOEXEC
);
1686 ERR("Poll set creation failed");
1690 ret
= lttng_poll_add(&events
, ctx
->consumer_metadata_pipe
[0], LPOLLIN
);
1696 DBG("Metadata main loop started");
1699 lttng_poll_reset(&events
);
1701 nb_fd
= LTTNG_POLL_GETNB(&events
);
1703 /* Only the metadata pipe is set */
1704 if (nb_fd
== 0 && consumer_quit
== 1) {
1709 DBG("Metadata poll wait with %d fd(s)", nb_fd
);
1710 ret
= lttng_poll_wait(&events
, -1);
1711 DBG("Metadata event catched in thread");
1713 if (errno
== EINTR
) {
1719 for (i
= 0; i
< nb_fd
; i
++) {
1720 revents
= LTTNG_POLL_GETEV(&events
, i
);
1721 pollfd
= LTTNG_POLL_GETFD(&events
, i
);
1723 /* Check the metadata pipe for incoming metadata. */
1724 if (pollfd
== ctx
->consumer_metadata_pipe
[0]) {
1725 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
1726 DBG("Metadata thread pipe hung up");
1728 * Remove the pipe from the poll set and continue the loop
1729 * since their might be data to consume.
1731 lttng_poll_del(&events
, ctx
->consumer_metadata_pipe
[0]);
1732 close(ctx
->consumer_metadata_pipe
[0]);
1734 } else if (revents
& LPOLLIN
) {
1736 /* Get the stream pointer received */
1737 ret
= read(pollfd
, &stream
, sizeof(stream
));
1738 } while (ret
< 0 && errno
== EINTR
);
1740 ret
< sizeof(struct lttng_consumer_stream
*)) {
1741 PERROR("read metadata stream");
1743 * Let's continue here and hope we can still work
1744 * without stopping the consumer. XXX: Should we?
1749 DBG("Adding metadata stream %d to poll set",
1753 /* The node should be init at this point */
1754 lttng_ht_add_unique_ulong(metadata_ht
,
1755 &stream
->waitfd_node
);
1758 /* Add metadata stream to the global poll events list */
1759 lttng_poll_add(&events
, stream
->wait_fd
,
1760 LPOLLIN
| LPOLLPRI
);
1762 consumer_add_metadata_stream(stream
);
1765 /* Metadata pipe handled. Continue handling the others */
1769 /* From here, the event is a metadata wait fd */
1772 lttng_ht_lookup(metadata_ht
, (void *)((unsigned long) pollfd
),
1774 node
= lttng_ht_iter_get_node_ulong(&iter
);
1776 /* FD not found, continue loop */
1781 stream
= caa_container_of(node
, struct lttng_consumer_stream
,
1784 /* Get the data out of the metadata file descriptor */
1785 if (revents
& (LPOLLIN
| LPOLLPRI
)) {
1786 DBG("Metadata available on fd %d", pollfd
);
1787 assert(stream
->wait_fd
== pollfd
);
1789 len
= ctx
->on_buffer_ready(stream
, ctx
);
1790 /* It's ok to have an unavailable sub-buffer */
1791 if (len
< 0 && len
!= -EAGAIN
) {
1794 } else if (len
> 0) {
1795 stream
->data_read
= 1;
1800 * Remove the stream from the hash table since there is no data
1801 * left on the fd because we previously did a read on the buffer.
1803 if (revents
& (LPOLLERR
| LPOLLHUP
)) {
1804 DBG("Metadata fd %d is hup|err|nval.", pollfd
);
1805 if (!stream
->hangup_flush_done
1806 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
1807 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
1808 DBG("Attempting to flush and consume the UST buffers");
1809 lttng_ustconsumer_on_stream_hangup(stream
);
1811 /* We just flushed the stream now read it. */
1812 len
= ctx
->on_buffer_ready(stream
, ctx
);
1813 /* It's ok to have an unavailable sub-buffer */
1814 if (len
< 0 && len
!= -EAGAIN
) {
1820 /* Removing it from hash table, poll set and free memory */
1821 lttng_ht_del(metadata_ht
, &iter
);
1823 lttng_poll_del(&events
, stream
->wait_fd
);
1824 consumer_del_metadata_stream(stream
);
1832 DBG("Metadata poll thread exiting");
1833 lttng_poll_clean(&events
);
1836 destroy_stream_ht(metadata_ht
);
1839 rcu_unregister_thread();
1844 * This thread polls the fds in the set to consume the data and write
1845 * it to tracefile if necessary.
1847 void *lttng_consumer_thread_poll_fds(void *data
)
1849 int num_rdy
, num_hup
, high_prio
, ret
, i
;
1850 struct pollfd
*pollfd
= NULL
;
1851 /* local view of the streams */
1852 struct lttng_consumer_stream
**local_stream
= NULL
;
1853 /* local view of consumer_data.fds_count */
1855 struct lttng_consumer_local_data
*ctx
= data
;
1857 pthread_t metadata_thread
;
1860 rcu_register_thread();
1862 /* Start metadata polling thread */
1863 ret
= pthread_create(&metadata_thread
, NULL
,
1864 lttng_consumer_thread_poll_metadata
, (void *) ctx
);
1866 PERROR("pthread_create metadata thread");
1870 local_stream
= zmalloc(sizeof(struct lttng_consumer_stream
));
1877 * the fds set has been updated, we need to update our
1878 * local array as well
1880 pthread_mutex_lock(&consumer_data
.lock
);
1881 if (consumer_data
.need_update
) {
1882 if (pollfd
!= NULL
) {
1886 if (local_stream
!= NULL
) {
1888 local_stream
= NULL
;
1891 /* allocate for all fds + 1 for the consumer_poll_pipe */
1892 pollfd
= zmalloc((consumer_data
.stream_count
+ 1) * sizeof(struct pollfd
));
1893 if (pollfd
== NULL
) {
1894 PERROR("pollfd malloc");
1895 pthread_mutex_unlock(&consumer_data
.lock
);
1899 /* allocate for all fds + 1 for the consumer_poll_pipe */
1900 local_stream
= zmalloc((consumer_data
.stream_count
+ 1) *
1901 sizeof(struct lttng_consumer_stream
));
1902 if (local_stream
== NULL
) {
1903 PERROR("local_stream malloc");
1904 pthread_mutex_unlock(&consumer_data
.lock
);
1907 ret
= consumer_update_poll_array(ctx
, &pollfd
, local_stream
);
1909 ERR("Error in allocating pollfd or local_outfds");
1910 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
1911 pthread_mutex_unlock(&consumer_data
.lock
);
1915 consumer_data
.need_update
= 0;
1917 pthread_mutex_unlock(&consumer_data
.lock
);
1919 /* No FDs and consumer_quit, consumer_cleanup the thread */
1920 if (nb_fd
== 0 && consumer_quit
== 1) {
1923 /* poll on the array of fds */
1925 DBG("polling on %d fd", nb_fd
+ 1);
1926 num_rdy
= poll(pollfd
, nb_fd
+ 1, consumer_poll_timeout
);
1927 DBG("poll num_rdy : %d", num_rdy
);
1928 if (num_rdy
== -1) {
1930 * Restart interrupted system call.
1932 if (errno
== EINTR
) {
1935 PERROR("Poll error");
1936 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_POLL_ERROR
);
1938 } else if (num_rdy
== 0) {
1939 DBG("Polling thread timed out");
1944 * If the consumer_poll_pipe triggered poll go directly to the
1945 * beginning of the loop to update the array. We want to prioritize
1946 * array update over low-priority reads.
1948 if (pollfd
[nb_fd
].revents
& (POLLIN
| POLLPRI
)) {
1949 size_t pipe_readlen
;
1952 DBG("consumer_poll_pipe wake up");
1953 /* Consume 1 byte of pipe data */
1955 pipe_readlen
= read(ctx
->consumer_poll_pipe
[0], &tmp
, 1);
1956 } while (pipe_readlen
== -1 && errno
== EINTR
);
1960 /* Take care of high priority channels first. */
1961 for (i
= 0; i
< nb_fd
; i
++) {
1962 if (pollfd
[i
].revents
& POLLPRI
) {
1963 DBG("Urgent read on fd %d", pollfd
[i
].fd
);
1965 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
1966 /* it's ok to have an unavailable sub-buffer */
1967 if (len
< 0 && len
!= -EAGAIN
) {
1969 } else if (len
> 0) {
1970 local_stream
[i
]->data_read
= 1;
1976 * If we read high prio channel in this loop, try again
1977 * for more high prio data.
1983 /* Take care of low priority channels. */
1984 for (i
= 0; i
< nb_fd
; i
++) {
1985 if ((pollfd
[i
].revents
& POLLIN
) ||
1986 local_stream
[i
]->hangup_flush_done
) {
1987 DBG("Normal read on fd %d", pollfd
[i
].fd
);
1988 len
= ctx
->on_buffer_ready(local_stream
[i
], ctx
);
1989 /* it's ok to have an unavailable sub-buffer */
1990 if (len
< 0 && len
!= -EAGAIN
) {
1992 } else if (len
> 0) {
1993 local_stream
[i
]->data_read
= 1;
1998 /* Handle hangup and errors */
1999 for (i
= 0; i
< nb_fd
; i
++) {
2000 if (!local_stream
[i
]->hangup_flush_done
2001 && (pollfd
[i
].revents
& (POLLHUP
| POLLERR
| POLLNVAL
))
2002 && (consumer_data
.type
== LTTNG_CONSUMER32_UST
2003 || consumer_data
.type
== LTTNG_CONSUMER64_UST
)) {
2004 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2006 lttng_ustconsumer_on_stream_hangup(local_stream
[i
]);
2007 /* Attempt read again, for the data we just flushed. */
2008 local_stream
[i
]->data_read
= 1;
2011 * If the poll flag is HUP/ERR/NVAL and we have
2012 * read no data in this pass, we can remove the
2013 * stream from its hash table.
2015 if ((pollfd
[i
].revents
& POLLHUP
)) {
2016 DBG("Polling fd %d tells it has hung up.", pollfd
[i
].fd
);
2017 if (!local_stream
[i
]->data_read
) {
2018 consumer_del_stream(local_stream
[i
]);
2021 } else if (pollfd
[i
].revents
& POLLERR
) {
2022 ERR("Error returned in polling fd %d.", pollfd
[i
].fd
);
2023 if (!local_stream
[i
]->data_read
) {
2024 consumer_del_stream(local_stream
[i
]);
2027 } else if (pollfd
[i
].revents
& POLLNVAL
) {
2028 ERR("Polling fd %d tells fd is not open.", pollfd
[i
].fd
);
2029 if (!local_stream
[i
]->data_read
) {
2030 consumer_del_stream(local_stream
[i
]);
2034 local_stream
[i
]->data_read
= 0;
2038 DBG("polling thread exiting");
2039 if (pollfd
!= NULL
) {
2043 if (local_stream
!= NULL
) {
2045 local_stream
= NULL
;
2049 * Close the write side of the pipe so epoll_wait() in
2050 * lttng_consumer_thread_poll_metadata can catch it. The thread is
2051 * monitoring the read side of the pipe. If we close them both, epoll_wait
2052 * strangely does not return and could create a endless wait period if the
2053 * pipe is the only tracked fd in the poll set. The thread will take care
2054 * of closing the read side.
2056 close(ctx
->consumer_metadata_pipe
[1]);
2058 ret
= pthread_join(metadata_thread
, &status
);
2060 PERROR("pthread_join metadata thread");
2064 rcu_unregister_thread();
2069 * This thread listens on the consumerd socket and receives the file
2070 * descriptors from the session daemon.
2072 void *lttng_consumer_thread_receive_fds(void *data
)
2074 int sock
, client_socket
, ret
;
2076 * structure to poll for incoming data on communication socket avoids
2077 * making blocking sockets.
2079 struct pollfd consumer_sockpoll
[2];
2080 struct lttng_consumer_local_data
*ctx
= data
;
2082 rcu_register_thread();
2084 DBG("Creating command socket %s", ctx
->consumer_command_sock_path
);
2085 unlink(ctx
->consumer_command_sock_path
);
2086 client_socket
= lttcomm_create_unix_sock(ctx
->consumer_command_sock_path
);
2087 if (client_socket
< 0) {
2088 ERR("Cannot create command socket");
2092 ret
= lttcomm_listen_unix_sock(client_socket
);
2097 DBG("Sending ready command to lttng-sessiond");
2098 ret
= lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY
);
2099 /* return < 0 on error, but == 0 is not fatal */
2101 ERR("Error sending ready command to lttng-sessiond");
2105 ret
= fcntl(client_socket
, F_SETFL
, O_NONBLOCK
);
2107 PERROR("fcntl O_NONBLOCK");
2111 /* prepare the FDs to poll : to client socket and the should_quit pipe */
2112 consumer_sockpoll
[0].fd
= ctx
->consumer_should_quit
[0];
2113 consumer_sockpoll
[0].events
= POLLIN
| POLLPRI
;
2114 consumer_sockpoll
[1].fd
= client_socket
;
2115 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
2117 if (lttng_consumer_poll_socket(consumer_sockpoll
) < 0) {
2120 DBG("Connection on client_socket");
2122 /* Blocking call, waiting for transmission */
2123 sock
= lttcomm_accept_unix_sock(client_socket
);
2128 ret
= fcntl(sock
, F_SETFL
, O_NONBLOCK
);
2130 PERROR("fcntl O_NONBLOCK");
2134 /* update the polling structure to poll on the established socket */
2135 consumer_sockpoll
[1].fd
= sock
;
2136 consumer_sockpoll
[1].events
= POLLIN
| POLLPRI
;
2139 if (lttng_consumer_poll_socket(consumer_sockpoll
) < 0) {
2142 DBG("Incoming command on sock");
2143 ret
= lttng_consumer_recv_cmd(ctx
, sock
, consumer_sockpoll
);
2144 if (ret
== -ENOENT
) {
2145 DBG("Received STOP command");
2150 * This could simply be a session daemon quitting. Don't output
2153 DBG("Communication interrupted on command socket");
2156 if (consumer_quit
) {
2157 DBG("consumer_thread_receive_fds received quit from signal");
2160 DBG("received fds on sock");
2163 DBG("consumer_thread_receive_fds exiting");
2166 * when all fds have hung up, the polling thread
2172 * 2s of grace period, if no polling events occur during
2173 * this period, the polling thread will exit even if there
2174 * are still open FDs (should not happen, but safety mechanism).
2176 consumer_poll_timeout
= LTTNG_CONSUMER_POLL_TIMEOUT
;
2179 * Wake-up the other end by writing a null byte in the pipe
2180 * (non-blocking). Important note: Because writing into the
2181 * pipe is non-blocking (and therefore we allow dropping wakeup
2182 * data, as long as there is wakeup data present in the pipe
2183 * buffer to wake up the other end), the other end should
2184 * perform the following sequence for waiting:
2185 * 1) empty the pipe (reads).
2186 * 2) perform update operation.
2187 * 3) wait on the pipe (poll).
2190 ret
= write(ctx
->consumer_poll_pipe
[1], "", 1);
2191 } while (ret
< 0 && errno
== EINTR
);
2192 rcu_unregister_thread();
2196 ssize_t
lttng_consumer_read_subbuffer(struct lttng_consumer_stream
*stream
,
2197 struct lttng_consumer_local_data
*ctx
)
2199 switch (consumer_data
.type
) {
2200 case LTTNG_CONSUMER_KERNEL
:
2201 return lttng_kconsumer_read_subbuffer(stream
, ctx
);
2202 case LTTNG_CONSUMER32_UST
:
2203 case LTTNG_CONSUMER64_UST
:
2204 return lttng_ustconsumer_read_subbuffer(stream
, ctx
);
2206 ERR("Unknown consumer_data type");
2212 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream
*stream
)
2214 switch (consumer_data
.type
) {
2215 case LTTNG_CONSUMER_KERNEL
:
2216 return lttng_kconsumer_on_recv_stream(stream
);
2217 case LTTNG_CONSUMER32_UST
:
2218 case LTTNG_CONSUMER64_UST
:
2219 return lttng_ustconsumer_on_recv_stream(stream
);
2221 ERR("Unknown consumer_data type");
2228 * Allocate and set consumer data hash tables.
2230 void lttng_consumer_init(void)
2232 consumer_data
.stream_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_ULONG
);
2233 consumer_data
.channel_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_ULONG
);
2234 consumer_data
.relayd_ht
= lttng_ht_new(0, LTTNG_HT_TYPE_ULONG
);
2238 * Process the ADD_RELAYD command receive by a consumer.
2240 * This will create a relayd socket pair and add it to the relayd hash table.
2241 * The caller MUST acquire a RCU read side lock before calling it.
2243 int consumer_add_relayd_socket(int net_seq_idx
, int sock_type
,
2244 struct lttng_consumer_local_data
*ctx
, int sock
,
2245 struct pollfd
*consumer_sockpoll
, struct lttcomm_sock
*relayd_sock
)
2248 struct consumer_relayd_sock_pair
*relayd
;
2250 DBG("Consumer adding relayd socket (idx: %d)", net_seq_idx
);
2252 /* Get relayd reference if exists. */
2253 relayd
= consumer_find_relayd(net_seq_idx
);
2254 if (relayd
== NULL
) {
2255 /* Not found. Allocate one. */
2256 relayd
= consumer_allocate_relayd_sock_pair(net_seq_idx
);
2257 if (relayd
== NULL
) {
2258 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_OUTFD_ERROR
);
2263 /* Poll on consumer socket. */
2264 if (lttng_consumer_poll_socket(consumer_sockpoll
) < 0) {
2269 /* Get relayd socket from session daemon */
2270 ret
= lttcomm_recv_fds_unix_sock(sock
, &fd
, 1);
2271 if (ret
!= sizeof(fd
)) {
2272 lttng_consumer_send_error(ctx
, LTTCOMM_CONSUMERD_ERROR_RECV_FD
);
2277 /* Copy socket information and received FD */
2278 switch (sock_type
) {
2279 case LTTNG_STREAM_CONTROL
:
2280 /* Copy received lttcomm socket */
2281 lttcomm_copy_sock(&relayd
->control_sock
, relayd_sock
);
2282 ret
= lttcomm_create_sock(&relayd
->control_sock
);
2287 /* Close the created socket fd which is useless */
2288 close(relayd
->control_sock
.fd
);
2290 /* Assign new file descriptor */
2291 relayd
->control_sock
.fd
= fd
;
2293 case LTTNG_STREAM_DATA
:
2294 /* Copy received lttcomm socket */
2295 lttcomm_copy_sock(&relayd
->data_sock
, relayd_sock
);
2296 ret
= lttcomm_create_sock(&relayd
->data_sock
);
2301 /* Close the created socket fd which is useless */
2302 close(relayd
->data_sock
.fd
);
2304 /* Assign new file descriptor */
2305 relayd
->data_sock
.fd
= fd
;
2308 ERR("Unknown relayd socket type (%d)", sock_type
);
2312 DBG("Consumer %s socket created successfully with net idx %d (fd: %d)",
2313 sock_type
== LTTNG_STREAM_CONTROL
? "control" : "data",
2314 relayd
->net_seq_idx
, fd
);
2317 * Add relayd socket pair to consumer data hashtable. If object already
2318 * exists or on error, the function gracefully returns.