Clean-up: consumer_add_metadata_stream always returns 0
[lttng-tools.git] / src / common / consumer / consumer.c
... / ...
CommitLineData
1/*
2 * Copyright (C) 2011 - Julien Desfossez <julien.desfossez@polymtl.ca>
3 * Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
4 * 2012 - David Goulet <dgoulet@efficios.com>
5 *
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.
9 *
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
13 * more details.
14 *
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.
18 */
19
20#define _LGPL_SOURCE
21#include <assert.h>
22#include <poll.h>
23#include <pthread.h>
24#include <stdlib.h>
25#include <string.h>
26#include <sys/mman.h>
27#include <sys/socket.h>
28#include <sys/types.h>
29#include <unistd.h>
30#include <inttypes.h>
31#include <signal.h>
32
33#include <bin/lttng-consumerd/health-consumerd.h>
34#include <common/common.h>
35#include <common/utils.h>
36#include <common/compat/poll.h>
37#include <common/compat/endian.h>
38#include <common/index/index.h>
39#include <common/kernel-ctl/kernel-ctl.h>
40#include <common/sessiond-comm/relayd.h>
41#include <common/sessiond-comm/sessiond-comm.h>
42#include <common/kernel-consumer/kernel-consumer.h>
43#include <common/relayd/relayd.h>
44#include <common/ust-consumer/ust-consumer.h>
45#include <common/consumer/consumer-timer.h>
46#include <common/consumer/consumer.h>
47#include <common/consumer/consumer-stream.h>
48#include <common/consumer/consumer-testpoint.h>
49#include <common/align.h>
50#include <common/consumer/consumer-metadata-cache.h>
51
52struct lttng_consumer_global_data consumer_data = {
53 .stream_count = 0,
54 .need_update = 1,
55 .type = LTTNG_CONSUMER_UNKNOWN,
56};
57
58enum consumer_channel_action {
59 CONSUMER_CHANNEL_ADD,
60 CONSUMER_CHANNEL_DEL,
61 CONSUMER_CHANNEL_QUIT,
62};
63
64struct consumer_channel_msg {
65 enum consumer_channel_action action;
66 struct lttng_consumer_channel *chan; /* add */
67 uint64_t key; /* del */
68};
69
70/* Flag used to temporarily pause data consumption from testpoints. */
71int data_consumption_paused;
72
73/*
74 * Flag to inform the polling thread to quit when all fd hung up. Updated by
75 * the consumer_thread_receive_fds when it notices that all fds has hung up.
76 * Also updated by the signal handler (consumer_should_exit()). Read by the
77 * polling threads.
78 */
79int consumer_quit;
80
81/*
82 * Global hash table containing respectively metadata and data streams. The
83 * stream element in this ht should only be updated by the metadata poll thread
84 * for the metadata and the data poll thread for the data.
85 */
86static struct lttng_ht *metadata_ht;
87static struct lttng_ht *data_ht;
88
89/*
90 * Notify a thread lttng pipe to poll back again. This usually means that some
91 * global state has changed so we just send back the thread in a poll wait
92 * call.
93 */
94static void notify_thread_lttng_pipe(struct lttng_pipe *pipe)
95{
96 struct lttng_consumer_stream *null_stream = NULL;
97
98 assert(pipe);
99
100 (void) lttng_pipe_write(pipe, &null_stream, sizeof(null_stream));
101}
102
103static void notify_health_quit_pipe(int *pipe)
104{
105 ssize_t ret;
106
107 ret = lttng_write(pipe[1], "4", 1);
108 if (ret < 1) {
109 PERROR("write consumer health quit");
110 }
111}
112
113static void notify_channel_pipe(struct lttng_consumer_local_data *ctx,
114 struct lttng_consumer_channel *chan,
115 uint64_t key,
116 enum consumer_channel_action action)
117{
118 struct consumer_channel_msg msg;
119 ssize_t ret;
120
121 memset(&msg, 0, sizeof(msg));
122
123 msg.action = action;
124 msg.chan = chan;
125 msg.key = key;
126 ret = lttng_write(ctx->consumer_channel_pipe[1], &msg, sizeof(msg));
127 if (ret < sizeof(msg)) {
128 PERROR("notify_channel_pipe write error");
129 }
130}
131
132void notify_thread_del_channel(struct lttng_consumer_local_data *ctx,
133 uint64_t key)
134{
135 notify_channel_pipe(ctx, NULL, key, CONSUMER_CHANNEL_DEL);
136}
137
138static int read_channel_pipe(struct lttng_consumer_local_data *ctx,
139 struct lttng_consumer_channel **chan,
140 uint64_t *key,
141 enum consumer_channel_action *action)
142{
143 struct consumer_channel_msg msg;
144 ssize_t ret;
145
146 ret = lttng_read(ctx->consumer_channel_pipe[0], &msg, sizeof(msg));
147 if (ret < sizeof(msg)) {
148 ret = -1;
149 goto error;
150 }
151 *action = msg.action;
152 *chan = msg.chan;
153 *key = msg.key;
154error:
155 return (int) ret;
156}
157
158/*
159 * Cleanup the stream list of a channel. Those streams are not yet globally
160 * visible
161 */
162static void clean_channel_stream_list(struct lttng_consumer_channel *channel)
163{
164 struct lttng_consumer_stream *stream, *stmp;
165
166 assert(channel);
167
168 /* Delete streams that might have been left in the stream list. */
169 cds_list_for_each_entry_safe(stream, stmp, &channel->streams.head,
170 send_node) {
171 cds_list_del(&stream->send_node);
172 /*
173 * Once a stream is added to this list, the buffers were created so we
174 * have a guarantee that this call will succeed. Setting the monitor
175 * mode to 0 so we don't lock nor try to delete the stream from the
176 * global hash table.
177 */
178 stream->monitor = 0;
179 consumer_stream_destroy(stream, NULL);
180 }
181}
182
183/*
184 * Find a stream. The consumer_data.lock must be locked during this
185 * call.
186 */
187static struct lttng_consumer_stream *find_stream(uint64_t key,
188 struct lttng_ht *ht)
189{
190 struct lttng_ht_iter iter;
191 struct lttng_ht_node_u64 *node;
192 struct lttng_consumer_stream *stream = NULL;
193
194 assert(ht);
195
196 /* -1ULL keys are lookup failures */
197 if (key == (uint64_t) -1ULL) {
198 return NULL;
199 }
200
201 rcu_read_lock();
202
203 lttng_ht_lookup(ht, &key, &iter);
204 node = lttng_ht_iter_get_node_u64(&iter);
205 if (node != NULL) {
206 stream = caa_container_of(node, struct lttng_consumer_stream, node);
207 }
208
209 rcu_read_unlock();
210
211 return stream;
212}
213
214static void steal_stream_key(uint64_t key, struct lttng_ht *ht)
215{
216 struct lttng_consumer_stream *stream;
217
218 rcu_read_lock();
219 stream = find_stream(key, ht);
220 if (stream) {
221 stream->key = (uint64_t) -1ULL;
222 /*
223 * We don't want the lookup to match, but we still need
224 * to iterate on this stream when iterating over the hash table. Just
225 * change the node key.
226 */
227 stream->node.key = (uint64_t) -1ULL;
228 }
229 rcu_read_unlock();
230}
231
232/*
233 * Return a channel object for the given key.
234 *
235 * RCU read side lock MUST be acquired before calling this function and
236 * protects the channel ptr.
237 */
238struct lttng_consumer_channel *consumer_find_channel(uint64_t key)
239{
240 struct lttng_ht_iter iter;
241 struct lttng_ht_node_u64 *node;
242 struct lttng_consumer_channel *channel = NULL;
243
244 /* -1ULL keys are lookup failures */
245 if (key == (uint64_t) -1ULL) {
246 return NULL;
247 }
248
249 lttng_ht_lookup(consumer_data.channel_ht, &key, &iter);
250 node = lttng_ht_iter_get_node_u64(&iter);
251 if (node != NULL) {
252 channel = caa_container_of(node, struct lttng_consumer_channel, node);
253 }
254
255 return channel;
256}
257
258/*
259 * There is a possibility that the consumer does not have enough time between
260 * the close of the channel on the session daemon and the cleanup in here thus
261 * once we have a channel add with an existing key, we know for sure that this
262 * channel will eventually get cleaned up by all streams being closed.
263 *
264 * This function just nullifies the already existing channel key.
265 */
266static void steal_channel_key(uint64_t key)
267{
268 struct lttng_consumer_channel *channel;
269
270 rcu_read_lock();
271 channel = consumer_find_channel(key);
272 if (channel) {
273 channel->key = (uint64_t) -1ULL;
274 /*
275 * We don't want the lookup to match, but we still need to iterate on
276 * this channel when iterating over the hash table. Just change the
277 * node key.
278 */
279 channel->node.key = (uint64_t) -1ULL;
280 }
281 rcu_read_unlock();
282}
283
284static void free_channel_rcu(struct rcu_head *head)
285{
286 struct lttng_ht_node_u64 *node =
287 caa_container_of(head, struct lttng_ht_node_u64, head);
288 struct lttng_consumer_channel *channel =
289 caa_container_of(node, struct lttng_consumer_channel, node);
290
291 switch (consumer_data.type) {
292 case LTTNG_CONSUMER_KERNEL:
293 break;
294 case LTTNG_CONSUMER32_UST:
295 case LTTNG_CONSUMER64_UST:
296 lttng_ustconsumer_free_channel(channel);
297 break;
298 default:
299 ERR("Unknown consumer_data type");
300 abort();
301 }
302 free(channel);
303}
304
305/*
306 * RCU protected relayd socket pair free.
307 */
308static void free_relayd_rcu(struct rcu_head *head)
309{
310 struct lttng_ht_node_u64 *node =
311 caa_container_of(head, struct lttng_ht_node_u64, head);
312 struct consumer_relayd_sock_pair *relayd =
313 caa_container_of(node, struct consumer_relayd_sock_pair, node);
314
315 /*
316 * Close all sockets. This is done in the call RCU since we don't want the
317 * socket fds to be reassigned thus potentially creating bad state of the
318 * relayd object.
319 *
320 * We do not have to lock the control socket mutex here since at this stage
321 * there is no one referencing to this relayd object.
322 */
323 (void) relayd_close(&relayd->control_sock);
324 (void) relayd_close(&relayd->data_sock);
325
326 free(relayd);
327}
328
329/*
330 * Destroy and free relayd socket pair object.
331 */
332void consumer_destroy_relayd(struct consumer_relayd_sock_pair *relayd)
333{
334 int ret;
335 struct lttng_ht_iter iter;
336
337 if (relayd == NULL) {
338 return;
339 }
340
341 DBG("Consumer destroy and close relayd socket pair");
342
343 iter.iter.node = &relayd->node.node;
344 ret = lttng_ht_del(consumer_data.relayd_ht, &iter);
345 if (ret != 0) {
346 /* We assume the relayd is being or is destroyed */
347 return;
348 }
349
350 /* RCU free() call */
351 call_rcu(&relayd->node.head, free_relayd_rcu);
352}
353
354/*
355 * Remove a channel from the global list protected by a mutex. This function is
356 * also responsible for freeing its data structures.
357 */
358void consumer_del_channel(struct lttng_consumer_channel *channel)
359{
360 int ret;
361 struct lttng_ht_iter iter;
362
363 DBG("Consumer delete channel key %" PRIu64, channel->key);
364
365 pthread_mutex_lock(&consumer_data.lock);
366 pthread_mutex_lock(&channel->lock);
367
368 /* Destroy streams that might have been left in the stream list. */
369 clean_channel_stream_list(channel);
370
371 if (channel->live_timer_enabled == 1) {
372 consumer_timer_live_stop(channel);
373 }
374 if (channel->monitor_timer_enabled == 1) {
375 consumer_timer_monitor_stop(channel);
376 }
377
378 switch (consumer_data.type) {
379 case LTTNG_CONSUMER_KERNEL:
380 break;
381 case LTTNG_CONSUMER32_UST:
382 case LTTNG_CONSUMER64_UST:
383 lttng_ustconsumer_del_channel(channel);
384 break;
385 default:
386 ERR("Unknown consumer_data type");
387 assert(0);
388 goto end;
389 }
390
391 rcu_read_lock();
392 iter.iter.node = &channel->node.node;
393 ret = lttng_ht_del(consumer_data.channel_ht, &iter);
394 assert(!ret);
395 rcu_read_unlock();
396
397 call_rcu(&channel->node.head, free_channel_rcu);
398end:
399 pthread_mutex_unlock(&channel->lock);
400 pthread_mutex_unlock(&consumer_data.lock);
401}
402
403/*
404 * Iterate over the relayd hash table and destroy each element. Finally,
405 * destroy the whole hash table.
406 */
407static void cleanup_relayd_ht(void)
408{
409 struct lttng_ht_iter iter;
410 struct consumer_relayd_sock_pair *relayd;
411
412 rcu_read_lock();
413
414 cds_lfht_for_each_entry(consumer_data.relayd_ht->ht, &iter.iter, relayd,
415 node.node) {
416 consumer_destroy_relayd(relayd);
417 }
418
419 rcu_read_unlock();
420
421 lttng_ht_destroy(consumer_data.relayd_ht);
422}
423
424/*
425 * Update the end point status of all streams having the given network sequence
426 * index (relayd index).
427 *
428 * It's atomically set without having the stream mutex locked which is fine
429 * because we handle the write/read race with a pipe wakeup for each thread.
430 */
431static void update_endpoint_status_by_netidx(uint64_t net_seq_idx,
432 enum consumer_endpoint_status status)
433{
434 struct lttng_ht_iter iter;
435 struct lttng_consumer_stream *stream;
436
437 DBG("Consumer set delete flag on stream by idx %" PRIu64, net_seq_idx);
438
439 rcu_read_lock();
440
441 /* Let's begin with metadata */
442 cds_lfht_for_each_entry(metadata_ht->ht, &iter.iter, stream, node.node) {
443 if (stream->net_seq_idx == net_seq_idx) {
444 uatomic_set(&stream->endpoint_status, status);
445 DBG("Delete flag set to metadata stream %d", stream->wait_fd);
446 }
447 }
448
449 /* Follow up by the data streams */
450 cds_lfht_for_each_entry(data_ht->ht, &iter.iter, stream, node.node) {
451 if (stream->net_seq_idx == net_seq_idx) {
452 uatomic_set(&stream->endpoint_status, status);
453 DBG("Delete flag set to data stream %d", stream->wait_fd);
454 }
455 }
456 rcu_read_unlock();
457}
458
459/*
460 * Cleanup a relayd object by flagging every associated streams for deletion,
461 * destroying the object meaning removing it from the relayd hash table,
462 * closing the sockets and freeing the memory in a RCU call.
463 *
464 * If a local data context is available, notify the threads that the streams'
465 * state have changed.
466 */
467static void cleanup_relayd(struct consumer_relayd_sock_pair *relayd,
468 struct lttng_consumer_local_data *ctx)
469{
470 uint64_t netidx;
471
472 assert(relayd);
473
474 DBG("Cleaning up relayd sockets");
475
476 /* Save the net sequence index before destroying the object */
477 netidx = relayd->net_seq_idx;
478
479 /*
480 * Delete the relayd from the relayd hash table, close the sockets and free
481 * the object in a RCU call.
482 */
483 consumer_destroy_relayd(relayd);
484
485 /* Set inactive endpoint to all streams */
486 update_endpoint_status_by_netidx(netidx, CONSUMER_ENDPOINT_INACTIVE);
487
488 /*
489 * With a local data context, notify the threads that the streams' state
490 * have changed. The write() action on the pipe acts as an "implicit"
491 * memory barrier ordering the updates of the end point status from the
492 * read of this status which happens AFTER receiving this notify.
493 */
494 if (ctx) {
495 notify_thread_lttng_pipe(ctx->consumer_data_pipe);
496 notify_thread_lttng_pipe(ctx->consumer_metadata_pipe);
497 }
498}
499
500/*
501 * Flag a relayd socket pair for destruction. Destroy it if the refcount
502 * reaches zero.
503 *
504 * RCU read side lock MUST be aquired before calling this function.
505 */
506void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair *relayd)
507{
508 assert(relayd);
509
510 /* Set destroy flag for this object */
511 uatomic_set(&relayd->destroy_flag, 1);
512
513 /* Destroy the relayd if refcount is 0 */
514 if (uatomic_read(&relayd->refcount) == 0) {
515 consumer_destroy_relayd(relayd);
516 }
517}
518
519/*
520 * Completly destroy stream from every visiable data structure and the given
521 * hash table if one.
522 *
523 * One this call returns, the stream object is not longer usable nor visible.
524 */
525void consumer_del_stream(struct lttng_consumer_stream *stream,
526 struct lttng_ht *ht)
527{
528 consumer_stream_destroy(stream, ht);
529}
530
531/*
532 * XXX naming of del vs destroy is all mixed up.
533 */
534void consumer_del_stream_for_data(struct lttng_consumer_stream *stream)
535{
536 consumer_stream_destroy(stream, data_ht);
537}
538
539void consumer_del_stream_for_metadata(struct lttng_consumer_stream *stream)
540{
541 consumer_stream_destroy(stream, metadata_ht);
542}
543
544struct lttng_consumer_stream *consumer_allocate_stream(uint64_t channel_key,
545 uint64_t stream_key,
546 enum lttng_consumer_stream_state state,
547 const char *channel_name,
548 uid_t uid,
549 gid_t gid,
550 uint64_t relayd_id,
551 uint64_t session_id,
552 int cpu,
553 int *alloc_ret,
554 enum consumer_channel_type type,
555 unsigned int monitor)
556{
557 int ret;
558 struct lttng_consumer_stream *stream;
559
560 stream = zmalloc(sizeof(*stream));
561 if (stream == NULL) {
562 PERROR("malloc struct lttng_consumer_stream");
563 ret = -ENOMEM;
564 goto end;
565 }
566
567 rcu_read_lock();
568
569 stream->key = stream_key;
570 stream->out_fd = -1;
571 stream->out_fd_offset = 0;
572 stream->output_written = 0;
573 stream->state = state;
574 stream->uid = uid;
575 stream->gid = gid;
576 stream->net_seq_idx = relayd_id;
577 stream->session_id = session_id;
578 stream->monitor = monitor;
579 stream->endpoint_status = CONSUMER_ENDPOINT_ACTIVE;
580 stream->index_file = NULL;
581 stream->last_sequence_number = -1ULL;
582 pthread_mutex_init(&stream->lock, NULL);
583 pthread_mutex_init(&stream->metadata_timer_lock, NULL);
584
585 /* If channel is the metadata, flag this stream as metadata. */
586 if (type == CONSUMER_CHANNEL_TYPE_METADATA) {
587 stream->metadata_flag = 1;
588 /* Metadata is flat out. */
589 strncpy(stream->name, DEFAULT_METADATA_NAME, sizeof(stream->name));
590 /* Live rendez-vous point. */
591 pthread_cond_init(&stream->metadata_rdv, NULL);
592 pthread_mutex_init(&stream->metadata_rdv_lock, NULL);
593 } else {
594 /* Format stream name to <channel_name>_<cpu_number> */
595 ret = snprintf(stream->name, sizeof(stream->name), "%s_%d",
596 channel_name, cpu);
597 if (ret < 0) {
598 PERROR("snprintf stream name");
599 goto error;
600 }
601 }
602
603 /* Key is always the wait_fd for streams. */
604 lttng_ht_node_init_u64(&stream->node, stream->key);
605
606 /* Init node per channel id key */
607 lttng_ht_node_init_u64(&stream->node_channel_id, channel_key);
608
609 /* Init session id node with the stream session id */
610 lttng_ht_node_init_u64(&stream->node_session_id, stream->session_id);
611
612 DBG3("Allocated stream %s (key %" PRIu64 ", chan_key %" PRIu64
613 " relayd_id %" PRIu64 ", session_id %" PRIu64,
614 stream->name, stream->key, channel_key,
615 stream->net_seq_idx, stream->session_id);
616
617 rcu_read_unlock();
618 return stream;
619
620error:
621 rcu_read_unlock();
622 free(stream);
623end:
624 if (alloc_ret) {
625 *alloc_ret = ret;
626 }
627 return NULL;
628}
629
630/*
631 * Add a stream to the global list protected by a mutex.
632 */
633void consumer_add_data_stream(struct lttng_consumer_stream *stream)
634{
635 struct lttng_ht *ht = data_ht;
636
637 assert(stream);
638 assert(ht);
639
640 DBG3("Adding consumer stream %" PRIu64, stream->key);
641
642 pthread_mutex_lock(&consumer_data.lock);
643 pthread_mutex_lock(&stream->chan->lock);
644 pthread_mutex_lock(&stream->chan->timer_lock);
645 pthread_mutex_lock(&stream->lock);
646 rcu_read_lock();
647
648 /* Steal stream identifier to avoid having streams with the same key */
649 steal_stream_key(stream->key, ht);
650
651 lttng_ht_add_unique_u64(ht, &stream->node);
652
653 lttng_ht_add_u64(consumer_data.stream_per_chan_id_ht,
654 &stream->node_channel_id);
655
656 /*
657 * Add stream to the stream_list_ht of the consumer data. No need to steal
658 * the key since the HT does not use it and we allow to add redundant keys
659 * into this table.
660 */
661 lttng_ht_add_u64(consumer_data.stream_list_ht, &stream->node_session_id);
662
663 /*
664 * When nb_init_stream_left reaches 0, we don't need to trigger any action
665 * in terms of destroying the associated channel, because the action that
666 * causes the count to become 0 also causes a stream to be added. The
667 * channel deletion will thus be triggered by the following removal of this
668 * stream.
669 */
670 if (uatomic_read(&stream->chan->nb_init_stream_left) > 0) {
671 /* Increment refcount before decrementing nb_init_stream_left */
672 cmm_smp_wmb();
673 uatomic_dec(&stream->chan->nb_init_stream_left);
674 }
675
676 /* Update consumer data once the node is inserted. */
677 consumer_data.stream_count++;
678 consumer_data.need_update = 1;
679
680 rcu_read_unlock();
681 pthread_mutex_unlock(&stream->lock);
682 pthread_mutex_unlock(&stream->chan->timer_lock);
683 pthread_mutex_unlock(&stream->chan->lock);
684 pthread_mutex_unlock(&consumer_data.lock);
685}
686
687void consumer_del_data_stream(struct lttng_consumer_stream *stream)
688{
689 consumer_del_stream(stream, data_ht);
690}
691
692/*
693 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
694 * be acquired before calling this.
695 */
696static int add_relayd(struct consumer_relayd_sock_pair *relayd)
697{
698 int ret = 0;
699 struct lttng_ht_node_u64 *node;
700 struct lttng_ht_iter iter;
701
702 assert(relayd);
703
704 lttng_ht_lookup(consumer_data.relayd_ht,
705 &relayd->net_seq_idx, &iter);
706 node = lttng_ht_iter_get_node_u64(&iter);
707 if (node != NULL) {
708 goto end;
709 }
710 lttng_ht_add_unique_u64(consumer_data.relayd_ht, &relayd->node);
711
712end:
713 return ret;
714}
715
716/*
717 * Allocate and return a consumer relayd socket.
718 */
719static struct consumer_relayd_sock_pair *consumer_allocate_relayd_sock_pair(
720 uint64_t net_seq_idx)
721{
722 struct consumer_relayd_sock_pair *obj = NULL;
723
724 /* net sequence index of -1 is a failure */
725 if (net_seq_idx == (uint64_t) -1ULL) {
726 goto error;
727 }
728
729 obj = zmalloc(sizeof(struct consumer_relayd_sock_pair));
730 if (obj == NULL) {
731 PERROR("zmalloc relayd sock");
732 goto error;
733 }
734
735 obj->net_seq_idx = net_seq_idx;
736 obj->refcount = 0;
737 obj->destroy_flag = 0;
738 obj->control_sock.sock.fd = -1;
739 obj->data_sock.sock.fd = -1;
740 lttng_ht_node_init_u64(&obj->node, obj->net_seq_idx);
741 pthread_mutex_init(&obj->ctrl_sock_mutex, NULL);
742
743error:
744 return obj;
745}
746
747/*
748 * Find a relayd socket pair in the global consumer data.
749 *
750 * Return the object if found else NULL.
751 * RCU read-side lock must be held across this call and while using the
752 * returned object.
753 */
754struct consumer_relayd_sock_pair *consumer_find_relayd(uint64_t key)
755{
756 struct lttng_ht_iter iter;
757 struct lttng_ht_node_u64 *node;
758 struct consumer_relayd_sock_pair *relayd = NULL;
759
760 /* Negative keys are lookup failures */
761 if (key == (uint64_t) -1ULL) {
762 goto error;
763 }
764
765 lttng_ht_lookup(consumer_data.relayd_ht, &key,
766 &iter);
767 node = lttng_ht_iter_get_node_u64(&iter);
768 if (node != NULL) {
769 relayd = caa_container_of(node, struct consumer_relayd_sock_pair, node);
770 }
771
772error:
773 return relayd;
774}
775
776/*
777 * Find a relayd and send the stream
778 *
779 * Returns 0 on success, < 0 on error
780 */
781int consumer_send_relayd_stream(struct lttng_consumer_stream *stream,
782 char *path)
783{
784 int ret = 0;
785 struct consumer_relayd_sock_pair *relayd;
786
787 assert(stream);
788 assert(stream->net_seq_idx != -1ULL);
789 assert(path);
790
791 /* The stream is not metadata. Get relayd reference if exists. */
792 rcu_read_lock();
793 relayd = consumer_find_relayd(stream->net_seq_idx);
794 if (relayd != NULL) {
795 /* Add stream on the relayd */
796 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
797 ret = relayd_add_stream(&relayd->control_sock, stream->name,
798 path, &stream->relayd_stream_id,
799 stream->chan->tracefile_size, stream->chan->tracefile_count);
800 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
801 if (ret < 0) {
802 goto end;
803 }
804
805 uatomic_inc(&relayd->refcount);
806 stream->sent_to_relayd = 1;
807 } else {
808 ERR("Stream %" PRIu64 " relayd ID %" PRIu64 " unknown. Can't send it.",
809 stream->key, stream->net_seq_idx);
810 ret = -1;
811 goto end;
812 }
813
814 DBG("Stream %s with key %" PRIu64 " sent to relayd id %" PRIu64,
815 stream->name, stream->key, stream->net_seq_idx);
816
817end:
818 rcu_read_unlock();
819 return ret;
820}
821
822/*
823 * Find a relayd and send the streams sent message
824 *
825 * Returns 0 on success, < 0 on error
826 */
827int consumer_send_relayd_streams_sent(uint64_t net_seq_idx)
828{
829 int ret = 0;
830 struct consumer_relayd_sock_pair *relayd;
831
832 assert(net_seq_idx != -1ULL);
833
834 /* The stream is not metadata. Get relayd reference if exists. */
835 rcu_read_lock();
836 relayd = consumer_find_relayd(net_seq_idx);
837 if (relayd != NULL) {
838 /* Add stream on the relayd */
839 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
840 ret = relayd_streams_sent(&relayd->control_sock);
841 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
842 if (ret < 0) {
843 goto end;
844 }
845 } else {
846 ERR("Relayd ID %" PRIu64 " unknown. Can't send streams_sent.",
847 net_seq_idx);
848 ret = -1;
849 goto end;
850 }
851
852 ret = 0;
853 DBG("All streams sent relayd id %" PRIu64, net_seq_idx);
854
855end:
856 rcu_read_unlock();
857 return ret;
858}
859
860/*
861 * Find a relayd and close the stream
862 */
863void close_relayd_stream(struct lttng_consumer_stream *stream)
864{
865 struct consumer_relayd_sock_pair *relayd;
866
867 /* The stream is not metadata. Get relayd reference if exists. */
868 rcu_read_lock();
869 relayd = consumer_find_relayd(stream->net_seq_idx);
870 if (relayd) {
871 consumer_stream_relayd_close(stream, relayd);
872 }
873 rcu_read_unlock();
874}
875
876/*
877 * Handle stream for relayd transmission if the stream applies for network
878 * streaming where the net sequence index is set.
879 *
880 * Return destination file descriptor or negative value on error.
881 */
882static int write_relayd_stream_header(struct lttng_consumer_stream *stream,
883 size_t data_size, unsigned long padding,
884 struct consumer_relayd_sock_pair *relayd)
885{
886 int outfd = -1, ret;
887 struct lttcomm_relayd_data_hdr data_hdr;
888
889 /* Safety net */
890 assert(stream);
891 assert(relayd);
892
893 /* Reset data header */
894 memset(&data_hdr, 0, sizeof(data_hdr));
895
896 if (stream->metadata_flag) {
897 /* Caller MUST acquire the relayd control socket lock */
898 ret = relayd_send_metadata(&relayd->control_sock, data_size);
899 if (ret < 0) {
900 goto error;
901 }
902
903 /* Metadata are always sent on the control socket. */
904 outfd = relayd->control_sock.sock.fd;
905 } else {
906 /* Set header with stream information */
907 data_hdr.stream_id = htobe64(stream->relayd_stream_id);
908 data_hdr.data_size = htobe32(data_size);
909 data_hdr.padding_size = htobe32(padding);
910 /*
911 * Note that net_seq_num below is assigned with the *current* value of
912 * next_net_seq_num and only after that the next_net_seq_num will be
913 * increment. This is why when issuing a command on the relayd using
914 * this next value, 1 should always be substracted in order to compare
915 * the last seen sequence number on the relayd side to the last sent.
916 */
917 data_hdr.net_seq_num = htobe64(stream->next_net_seq_num);
918 /* Other fields are zeroed previously */
919
920 ret = relayd_send_data_hdr(&relayd->data_sock, &data_hdr,
921 sizeof(data_hdr));
922 if (ret < 0) {
923 goto error;
924 }
925
926 ++stream->next_net_seq_num;
927
928 /* Set to go on data socket */
929 outfd = relayd->data_sock.sock.fd;
930 }
931
932error:
933 return outfd;
934}
935
936/*
937 * Allocate and return a new lttng_consumer_channel object using the given key
938 * to initialize the hash table node.
939 *
940 * On error, return NULL.
941 */
942struct lttng_consumer_channel *consumer_allocate_channel(uint64_t key,
943 uint64_t session_id,
944 const char *pathname,
945 const char *name,
946 uid_t uid,
947 gid_t gid,
948 uint64_t relayd_id,
949 enum lttng_event_output output,
950 uint64_t tracefile_size,
951 uint64_t tracefile_count,
952 uint64_t session_id_per_pid,
953 unsigned int monitor,
954 unsigned int live_timer_interval,
955 const char *root_shm_path,
956 const char *shm_path)
957{
958 struct lttng_consumer_channel *channel;
959
960 channel = zmalloc(sizeof(*channel));
961 if (channel == NULL) {
962 PERROR("malloc struct lttng_consumer_channel");
963 goto end;
964 }
965
966 channel->key = key;
967 channel->refcount = 0;
968 channel->session_id = session_id;
969 channel->session_id_per_pid = session_id_per_pid;
970 channel->uid = uid;
971 channel->gid = gid;
972 channel->relayd_id = relayd_id;
973 channel->tracefile_size = tracefile_size;
974 channel->tracefile_count = tracefile_count;
975 channel->monitor = monitor;
976 channel->live_timer_interval = live_timer_interval;
977 pthread_mutex_init(&channel->lock, NULL);
978 pthread_mutex_init(&channel->timer_lock, NULL);
979
980 switch (output) {
981 case LTTNG_EVENT_SPLICE:
982 channel->output = CONSUMER_CHANNEL_SPLICE;
983 break;
984 case LTTNG_EVENT_MMAP:
985 channel->output = CONSUMER_CHANNEL_MMAP;
986 break;
987 default:
988 assert(0);
989 free(channel);
990 channel = NULL;
991 goto end;
992 }
993
994 /*
995 * In monitor mode, the streams associated with the channel will be put in
996 * a special list ONLY owned by this channel. So, the refcount is set to 1
997 * here meaning that the channel itself has streams that are referenced.
998 *
999 * On a channel deletion, once the channel is no longer visible, the
1000 * refcount is decremented and checked for a zero value to delete it. With
1001 * streams in no monitor mode, it will now be safe to destroy the channel.
1002 */
1003 if (!channel->monitor) {
1004 channel->refcount = 1;
1005 }
1006
1007 strncpy(channel->pathname, pathname, sizeof(channel->pathname));
1008 channel->pathname[sizeof(channel->pathname) - 1] = '\0';
1009
1010 strncpy(channel->name, name, sizeof(channel->name));
1011 channel->name[sizeof(channel->name) - 1] = '\0';
1012
1013 if (root_shm_path) {
1014 strncpy(channel->root_shm_path, root_shm_path, sizeof(channel->root_shm_path));
1015 channel->root_shm_path[sizeof(channel->root_shm_path) - 1] = '\0';
1016 }
1017 if (shm_path) {
1018 strncpy(channel->shm_path, shm_path, sizeof(channel->shm_path));
1019 channel->shm_path[sizeof(channel->shm_path) - 1] = '\0';
1020 }
1021
1022 lttng_ht_node_init_u64(&channel->node, channel->key);
1023
1024 channel->wait_fd = -1;
1025
1026 CDS_INIT_LIST_HEAD(&channel->streams.head);
1027
1028 DBG("Allocated channel (key %" PRIu64 ")", channel->key);
1029
1030end:
1031 return channel;
1032}
1033
1034/*
1035 * Add a channel to the global list protected by a mutex.
1036 *
1037 * Always return 0 indicating success.
1038 */
1039int consumer_add_channel(struct lttng_consumer_channel *channel,
1040 struct lttng_consumer_local_data *ctx)
1041{
1042 pthread_mutex_lock(&consumer_data.lock);
1043 pthread_mutex_lock(&channel->lock);
1044 pthread_mutex_lock(&channel->timer_lock);
1045
1046 /*
1047 * This gives us a guarantee that the channel we are about to add to the
1048 * channel hash table will be unique. See this function comment on the why
1049 * we need to steel the channel key at this stage.
1050 */
1051 steal_channel_key(channel->key);
1052
1053 rcu_read_lock();
1054 lttng_ht_add_unique_u64(consumer_data.channel_ht, &channel->node);
1055 rcu_read_unlock();
1056
1057 pthread_mutex_unlock(&channel->timer_lock);
1058 pthread_mutex_unlock(&channel->lock);
1059 pthread_mutex_unlock(&consumer_data.lock);
1060
1061 if (channel->wait_fd != -1 && channel->type == CONSUMER_CHANNEL_TYPE_DATA) {
1062 notify_channel_pipe(ctx, channel, -1, CONSUMER_CHANNEL_ADD);
1063 }
1064
1065 return 0;
1066}
1067
1068/*
1069 * Allocate the pollfd structure and the local view of the out fds to avoid
1070 * doing a lookup in the linked list and concurrency issues when writing is
1071 * needed. Called with consumer_data.lock held.
1072 *
1073 * Returns the number of fds in the structures.
1074 */
1075static int update_poll_array(struct lttng_consumer_local_data *ctx,
1076 struct pollfd **pollfd, struct lttng_consumer_stream **local_stream,
1077 struct lttng_ht *ht)
1078{
1079 int i = 0;
1080 struct lttng_ht_iter iter;
1081 struct lttng_consumer_stream *stream;
1082
1083 assert(ctx);
1084 assert(ht);
1085 assert(pollfd);
1086 assert(local_stream);
1087
1088 DBG("Updating poll fd array");
1089 rcu_read_lock();
1090 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
1091 /*
1092 * Only active streams with an active end point can be added to the
1093 * poll set and local stream storage of the thread.
1094 *
1095 * There is a potential race here for endpoint_status to be updated
1096 * just after the check. However, this is OK since the stream(s) will
1097 * be deleted once the thread is notified that the end point state has
1098 * changed where this function will be called back again.
1099 */
1100 if (stream->state != LTTNG_CONSUMER_ACTIVE_STREAM ||
1101 stream->endpoint_status == CONSUMER_ENDPOINT_INACTIVE) {
1102 continue;
1103 }
1104 /*
1105 * This clobbers way too much the debug output. Uncomment that if you
1106 * need it for debugging purposes.
1107 *
1108 * DBG("Active FD %d", stream->wait_fd);
1109 */
1110 (*pollfd)[i].fd = stream->wait_fd;
1111 (*pollfd)[i].events = POLLIN | POLLPRI;
1112 local_stream[i] = stream;
1113 i++;
1114 }
1115 rcu_read_unlock();
1116
1117 /*
1118 * Insert the consumer_data_pipe at the end of the array and don't
1119 * increment i so nb_fd is the number of real FD.
1120 */
1121 (*pollfd)[i].fd = lttng_pipe_get_readfd(ctx->consumer_data_pipe);
1122 (*pollfd)[i].events = POLLIN | POLLPRI;
1123
1124 (*pollfd)[i + 1].fd = lttng_pipe_get_readfd(ctx->consumer_wakeup_pipe);
1125 (*pollfd)[i + 1].events = POLLIN | POLLPRI;
1126 return i;
1127}
1128
1129/*
1130 * Poll on the should_quit pipe and the command socket return -1 on
1131 * error, 1 if should exit, 0 if data is available on the command socket
1132 */
1133int lttng_consumer_poll_socket(struct pollfd *consumer_sockpoll)
1134{
1135 int num_rdy;
1136
1137restart:
1138 num_rdy = poll(consumer_sockpoll, 2, -1);
1139 if (num_rdy == -1) {
1140 /*
1141 * Restart interrupted system call.
1142 */
1143 if (errno == EINTR) {
1144 goto restart;
1145 }
1146 PERROR("Poll error");
1147 return -1;
1148 }
1149 if (consumer_sockpoll[0].revents & (POLLIN | POLLPRI)) {
1150 DBG("consumer_should_quit wake up");
1151 return 1;
1152 }
1153 return 0;
1154}
1155
1156/*
1157 * Set the error socket.
1158 */
1159void lttng_consumer_set_error_sock(struct lttng_consumer_local_data *ctx,
1160 int sock)
1161{
1162 ctx->consumer_error_socket = sock;
1163}
1164
1165/*
1166 * Set the command socket path.
1167 */
1168void lttng_consumer_set_command_sock_path(
1169 struct lttng_consumer_local_data *ctx, char *sock)
1170{
1171 ctx->consumer_command_sock_path = sock;
1172}
1173
1174/*
1175 * Send return code to the session daemon.
1176 * If the socket is not defined, we return 0, it is not a fatal error
1177 */
1178int lttng_consumer_send_error(struct lttng_consumer_local_data *ctx, int cmd)
1179{
1180 if (ctx->consumer_error_socket > 0) {
1181 return lttcomm_send_unix_sock(ctx->consumer_error_socket, &cmd,
1182 sizeof(enum lttcomm_sessiond_command));
1183 }
1184
1185 return 0;
1186}
1187
1188/*
1189 * Close all the tracefiles and stream fds and MUST be called when all
1190 * instances are destroyed i.e. when all threads were joined and are ended.
1191 */
1192void lttng_consumer_cleanup(void)
1193{
1194 struct lttng_ht_iter iter;
1195 struct lttng_consumer_channel *channel;
1196
1197 rcu_read_lock();
1198
1199 cds_lfht_for_each_entry(consumer_data.channel_ht->ht, &iter.iter, channel,
1200 node.node) {
1201 consumer_del_channel(channel);
1202 }
1203
1204 rcu_read_unlock();
1205
1206 lttng_ht_destroy(consumer_data.channel_ht);
1207
1208 cleanup_relayd_ht();
1209
1210 lttng_ht_destroy(consumer_data.stream_per_chan_id_ht);
1211
1212 /*
1213 * This HT contains streams that are freed by either the metadata thread or
1214 * the data thread so we do *nothing* on the hash table and simply destroy
1215 * it.
1216 */
1217 lttng_ht_destroy(consumer_data.stream_list_ht);
1218}
1219
1220/*
1221 * Called from signal handler.
1222 */
1223void lttng_consumer_should_exit(struct lttng_consumer_local_data *ctx)
1224{
1225 ssize_t ret;
1226
1227 CMM_STORE_SHARED(consumer_quit, 1);
1228 ret = lttng_write(ctx->consumer_should_quit[1], "4", 1);
1229 if (ret < 1) {
1230 PERROR("write consumer quit");
1231 }
1232
1233 DBG("Consumer flag that it should quit");
1234}
1235
1236
1237/*
1238 * Flush pending writes to trace output disk file.
1239 */
1240static
1241void lttng_consumer_sync_trace_file(struct lttng_consumer_stream *stream,
1242 off_t orig_offset)
1243{
1244 int ret;
1245 int outfd = stream->out_fd;
1246
1247 /*
1248 * This does a blocking write-and-wait on any page that belongs to the
1249 * subbuffer prior to the one we just wrote.
1250 * Don't care about error values, as these are just hints and ways to
1251 * limit the amount of page cache used.
1252 */
1253 if (orig_offset < stream->max_sb_size) {
1254 return;
1255 }
1256 lttng_sync_file_range(outfd, orig_offset - stream->max_sb_size,
1257 stream->max_sb_size,
1258 SYNC_FILE_RANGE_WAIT_BEFORE
1259 | SYNC_FILE_RANGE_WRITE
1260 | SYNC_FILE_RANGE_WAIT_AFTER);
1261 /*
1262 * Give hints to the kernel about how we access the file:
1263 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1264 * we write it.
1265 *
1266 * We need to call fadvise again after the file grows because the
1267 * kernel does not seem to apply fadvise to non-existing parts of the
1268 * file.
1269 *
1270 * Call fadvise _after_ having waited for the page writeback to
1271 * complete because the dirty page writeback semantic is not well
1272 * defined. So it can be expected to lead to lower throughput in
1273 * streaming.
1274 */
1275 ret = posix_fadvise(outfd, orig_offset - stream->max_sb_size,
1276 stream->max_sb_size, POSIX_FADV_DONTNEED);
1277 if (ret && ret != -ENOSYS) {
1278 errno = ret;
1279 PERROR("posix_fadvise on fd %i", outfd);
1280 }
1281}
1282
1283/*
1284 * Initialise the necessary environnement :
1285 * - create a new context
1286 * - create the poll_pipe
1287 * - create the should_quit pipe (for signal handler)
1288 * - create the thread pipe (for splice)
1289 *
1290 * Takes a function pointer as argument, this function is called when data is
1291 * available on a buffer. This function is responsible to do the
1292 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1293 * buffer configuration and then kernctl_put_next_subbuf at the end.
1294 *
1295 * Returns a pointer to the new context or NULL on error.
1296 */
1297struct lttng_consumer_local_data *lttng_consumer_create(
1298 enum lttng_consumer_type type,
1299 ssize_t (*buffer_ready)(struct lttng_consumer_stream *stream,
1300 struct lttng_consumer_local_data *ctx),
1301 int (*recv_channel)(struct lttng_consumer_channel *channel),
1302 int (*recv_stream)(struct lttng_consumer_stream *stream),
1303 int (*update_stream)(uint64_t stream_key, uint32_t state))
1304{
1305 int ret;
1306 struct lttng_consumer_local_data *ctx;
1307
1308 assert(consumer_data.type == LTTNG_CONSUMER_UNKNOWN ||
1309 consumer_data.type == type);
1310 consumer_data.type = type;
1311
1312 ctx = zmalloc(sizeof(struct lttng_consumer_local_data));
1313 if (ctx == NULL) {
1314 PERROR("allocating context");
1315 goto error;
1316 }
1317
1318 ctx->consumer_error_socket = -1;
1319 ctx->consumer_metadata_socket = -1;
1320 pthread_mutex_init(&ctx->metadata_socket_lock, NULL);
1321 /* assign the callbacks */
1322 ctx->on_buffer_ready = buffer_ready;
1323 ctx->on_recv_channel = recv_channel;
1324 ctx->on_recv_stream = recv_stream;
1325 ctx->on_update_stream = update_stream;
1326
1327 ctx->consumer_data_pipe = lttng_pipe_open(0);
1328 if (!ctx->consumer_data_pipe) {
1329 goto error_poll_pipe;
1330 }
1331
1332 ctx->consumer_wakeup_pipe = lttng_pipe_open(0);
1333 if (!ctx->consumer_wakeup_pipe) {
1334 goto error_wakeup_pipe;
1335 }
1336
1337 ret = pipe(ctx->consumer_should_quit);
1338 if (ret < 0) {
1339 PERROR("Error creating recv pipe");
1340 goto error_quit_pipe;
1341 }
1342
1343 ret = pipe(ctx->consumer_channel_pipe);
1344 if (ret < 0) {
1345 PERROR("Error creating channel pipe");
1346 goto error_channel_pipe;
1347 }
1348
1349 ctx->consumer_metadata_pipe = lttng_pipe_open(0);
1350 if (!ctx->consumer_metadata_pipe) {
1351 goto error_metadata_pipe;
1352 }
1353
1354 ctx->channel_monitor_pipe = -1;
1355
1356 return ctx;
1357
1358error_metadata_pipe:
1359 utils_close_pipe(ctx->consumer_channel_pipe);
1360error_channel_pipe:
1361 utils_close_pipe(ctx->consumer_should_quit);
1362error_quit_pipe:
1363 lttng_pipe_destroy(ctx->consumer_wakeup_pipe);
1364error_wakeup_pipe:
1365 lttng_pipe_destroy(ctx->consumer_data_pipe);
1366error_poll_pipe:
1367 free(ctx);
1368error:
1369 return NULL;
1370}
1371
1372/*
1373 * Iterate over all streams of the hashtable and free them properly.
1374 */
1375static void destroy_data_stream_ht(struct lttng_ht *ht)
1376{
1377 struct lttng_ht_iter iter;
1378 struct lttng_consumer_stream *stream;
1379
1380 if (ht == NULL) {
1381 return;
1382 }
1383
1384 rcu_read_lock();
1385 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
1386 /*
1387 * Ignore return value since we are currently cleaning up so any error
1388 * can't be handled.
1389 */
1390 (void) consumer_del_stream(stream, ht);
1391 }
1392 rcu_read_unlock();
1393
1394 lttng_ht_destroy(ht);
1395}
1396
1397/*
1398 * Iterate over all streams of the metadata hashtable and free them
1399 * properly.
1400 */
1401static void destroy_metadata_stream_ht(struct lttng_ht *ht)
1402{
1403 struct lttng_ht_iter iter;
1404 struct lttng_consumer_stream *stream;
1405
1406 if (ht == NULL) {
1407 return;
1408 }
1409
1410 rcu_read_lock();
1411 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
1412 /*
1413 * Ignore return value since we are currently cleaning up so any error
1414 * can't be handled.
1415 */
1416 (void) consumer_del_metadata_stream(stream, ht);
1417 }
1418 rcu_read_unlock();
1419
1420 lttng_ht_destroy(ht);
1421}
1422
1423/*
1424 * Close all fds associated with the instance and free the context.
1425 */
1426void lttng_consumer_destroy(struct lttng_consumer_local_data *ctx)
1427{
1428 int ret;
1429
1430 DBG("Consumer destroying it. Closing everything.");
1431
1432 if (!ctx) {
1433 return;
1434 }
1435
1436 destroy_data_stream_ht(data_ht);
1437 destroy_metadata_stream_ht(metadata_ht);
1438
1439 ret = close(ctx->consumer_error_socket);
1440 if (ret) {
1441 PERROR("close");
1442 }
1443 ret = close(ctx->consumer_metadata_socket);
1444 if (ret) {
1445 PERROR("close");
1446 }
1447 utils_close_pipe(ctx->consumer_channel_pipe);
1448 lttng_pipe_destroy(ctx->consumer_data_pipe);
1449 lttng_pipe_destroy(ctx->consumer_metadata_pipe);
1450 lttng_pipe_destroy(ctx->consumer_wakeup_pipe);
1451 utils_close_pipe(ctx->consumer_should_quit);
1452
1453 unlink(ctx->consumer_command_sock_path);
1454 free(ctx);
1455}
1456
1457/*
1458 * Write the metadata stream id on the specified file descriptor.
1459 */
1460static int write_relayd_metadata_id(int fd,
1461 struct lttng_consumer_stream *stream,
1462 struct consumer_relayd_sock_pair *relayd, unsigned long padding)
1463{
1464 ssize_t ret;
1465 struct lttcomm_relayd_metadata_payload hdr;
1466
1467 hdr.stream_id = htobe64(stream->relayd_stream_id);
1468 hdr.padding_size = htobe32(padding);
1469 ret = lttng_write(fd, (void *) &hdr, sizeof(hdr));
1470 if (ret < sizeof(hdr)) {
1471 /*
1472 * This error means that the fd's end is closed so ignore the PERROR
1473 * not to clubber the error output since this can happen in a normal
1474 * code path.
1475 */
1476 if (errno != EPIPE) {
1477 PERROR("write metadata stream id");
1478 }
1479 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno);
1480 /*
1481 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1482 * handle writting the missing part so report that as an error and
1483 * don't lie to the caller.
1484 */
1485 ret = -1;
1486 goto end;
1487 }
1488 DBG("Metadata stream id %" PRIu64 " with padding %lu written before data",
1489 stream->relayd_stream_id, padding);
1490
1491end:
1492 return (int) ret;
1493}
1494
1495/*
1496 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1497 * core function for writing trace buffers to either the local filesystem or
1498 * the network.
1499 *
1500 * It must be called with the stream lock held.
1501 *
1502 * Careful review MUST be put if any changes occur!
1503 *
1504 * Returns the number of bytes written
1505 */
1506ssize_t lttng_consumer_on_read_subbuffer_mmap(
1507 struct lttng_consumer_local_data *ctx,
1508 struct lttng_consumer_stream *stream, unsigned long len,
1509 unsigned long padding,
1510 struct ctf_packet_index *index)
1511{
1512 unsigned long mmap_offset;
1513 void *mmap_base;
1514 ssize_t ret = 0;
1515 off_t orig_offset = stream->out_fd_offset;
1516 /* Default is on the disk */
1517 int outfd = stream->out_fd;
1518 struct consumer_relayd_sock_pair *relayd = NULL;
1519 unsigned int relayd_hang_up = 0;
1520
1521 /* RCU lock for the relayd pointer */
1522 rcu_read_lock();
1523
1524 /* Flag that the current stream if set for network streaming. */
1525 if (stream->net_seq_idx != (uint64_t) -1ULL) {
1526 relayd = consumer_find_relayd(stream->net_seq_idx);
1527 if (relayd == NULL) {
1528 ret = -EPIPE;
1529 goto end;
1530 }
1531 }
1532
1533 /* get the offset inside the fd to mmap */
1534 switch (consumer_data.type) {
1535 case LTTNG_CONSUMER_KERNEL:
1536 mmap_base = stream->mmap_base;
1537 ret = kernctl_get_mmap_read_offset(stream->wait_fd, &mmap_offset);
1538 if (ret < 0) {
1539 PERROR("tracer ctl get_mmap_read_offset");
1540 goto end;
1541 }
1542 break;
1543 case LTTNG_CONSUMER32_UST:
1544 case LTTNG_CONSUMER64_UST:
1545 mmap_base = lttng_ustctl_get_mmap_base(stream);
1546 if (!mmap_base) {
1547 ERR("read mmap get mmap base for stream %s", stream->name);
1548 ret = -EPERM;
1549 goto end;
1550 }
1551 ret = lttng_ustctl_get_mmap_read_offset(stream, &mmap_offset);
1552 if (ret != 0) {
1553 PERROR("tracer ctl get_mmap_read_offset");
1554 ret = -EINVAL;
1555 goto end;
1556 }
1557 break;
1558 default:
1559 ERR("Unknown consumer_data type");
1560 assert(0);
1561 }
1562
1563 /* Handle stream on the relayd if the output is on the network */
1564 if (relayd) {
1565 unsigned long netlen = len;
1566
1567 /*
1568 * Lock the control socket for the complete duration of the function
1569 * since from this point on we will use the socket.
1570 */
1571 if (stream->metadata_flag) {
1572 /* Metadata requires the control socket. */
1573 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
1574 if (stream->reset_metadata_flag) {
1575 ret = relayd_reset_metadata(&relayd->control_sock,
1576 stream->relayd_stream_id,
1577 stream->metadata_version);
1578 if (ret < 0) {
1579 relayd_hang_up = 1;
1580 goto write_error;
1581 }
1582 stream->reset_metadata_flag = 0;
1583 }
1584 netlen += sizeof(struct lttcomm_relayd_metadata_payload);
1585 }
1586
1587 ret = write_relayd_stream_header(stream, netlen, padding, relayd);
1588 if (ret < 0) {
1589 relayd_hang_up = 1;
1590 goto write_error;
1591 }
1592 /* Use the returned socket. */
1593 outfd = ret;
1594
1595 /* Write metadata stream id before payload */
1596 if (stream->metadata_flag) {
1597 ret = write_relayd_metadata_id(outfd, stream, relayd, padding);
1598 if (ret < 0) {
1599 relayd_hang_up = 1;
1600 goto write_error;
1601 }
1602 }
1603 } else {
1604 /* No streaming, we have to set the len with the full padding */
1605 len += padding;
1606
1607 if (stream->metadata_flag && stream->reset_metadata_flag) {
1608 ret = utils_truncate_stream_file(stream->out_fd, 0);
1609 if (ret < 0) {
1610 ERR("Reset metadata file");
1611 goto end;
1612 }
1613 stream->reset_metadata_flag = 0;
1614 }
1615
1616 /*
1617 * Check if we need to change the tracefile before writing the packet.
1618 */
1619 if (stream->chan->tracefile_size > 0 &&
1620 (stream->tracefile_size_current + len) >
1621 stream->chan->tracefile_size) {
1622 ret = utils_rotate_stream_file(stream->chan->pathname,
1623 stream->name, stream->chan->tracefile_size,
1624 stream->chan->tracefile_count, stream->uid, stream->gid,
1625 stream->out_fd, &(stream->tracefile_count_current),
1626 &stream->out_fd);
1627 if (ret < 0) {
1628 ERR("Rotating output file");
1629 goto end;
1630 }
1631 outfd = stream->out_fd;
1632
1633 if (stream->index_file) {
1634 lttng_index_file_put(stream->index_file);
1635 stream->index_file = lttng_index_file_create(stream->chan->pathname,
1636 stream->name, stream->uid, stream->gid,
1637 stream->chan->tracefile_size,
1638 stream->tracefile_count_current,
1639 CTF_INDEX_MAJOR, CTF_INDEX_MINOR);
1640 if (!stream->index_file) {
1641 goto end;
1642 }
1643 }
1644
1645 /* Reset current size because we just perform a rotation. */
1646 stream->tracefile_size_current = 0;
1647 stream->out_fd_offset = 0;
1648 orig_offset = 0;
1649 }
1650 stream->tracefile_size_current += len;
1651 if (index) {
1652 index->offset = htobe64(stream->out_fd_offset);
1653 }
1654 }
1655
1656 /*
1657 * This call guarantee that len or less is returned. It's impossible to
1658 * receive a ret value that is bigger than len.
1659 */
1660 ret = lttng_write(outfd, mmap_base + mmap_offset, len);
1661 DBG("Consumer mmap write() ret %zd (len %lu)", ret, len);
1662 if (ret < 0 || ((size_t) ret != len)) {
1663 /*
1664 * Report error to caller if nothing was written else at least send the
1665 * amount written.
1666 */
1667 if (ret < 0) {
1668 ret = -errno;
1669 }
1670 relayd_hang_up = 1;
1671
1672 /* Socket operation failed. We consider the relayd dead */
1673 if (errno == EPIPE || errno == EINVAL || errno == EBADF) {
1674 /*
1675 * This is possible if the fd is closed on the other side
1676 * (outfd) or any write problem. It can be verbose a bit for a
1677 * normal execution if for instance the relayd is stopped
1678 * abruptly. This can happen so set this to a DBG statement.
1679 */
1680 DBG("Consumer mmap write detected relayd hang up");
1681 } else {
1682 /* Unhandled error, print it and stop function right now. */
1683 PERROR("Error in write mmap (ret %zd != len %lu)", ret, len);
1684 }
1685 goto write_error;
1686 }
1687 stream->output_written += ret;
1688
1689 /* This call is useless on a socket so better save a syscall. */
1690 if (!relayd) {
1691 /* This won't block, but will start writeout asynchronously */
1692 lttng_sync_file_range(outfd, stream->out_fd_offset, len,
1693 SYNC_FILE_RANGE_WRITE);
1694 stream->out_fd_offset += len;
1695 lttng_consumer_sync_trace_file(stream, orig_offset);
1696 }
1697
1698write_error:
1699 /*
1700 * This is a special case that the relayd has closed its socket. Let's
1701 * cleanup the relayd object and all associated streams.
1702 */
1703 if (relayd && relayd_hang_up) {
1704 cleanup_relayd(relayd, ctx);
1705 }
1706
1707end:
1708 /* Unlock only if ctrl socket used */
1709 if (relayd && stream->metadata_flag) {
1710 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
1711 }
1712
1713 rcu_read_unlock();
1714 return ret;
1715}
1716
1717/*
1718 * Splice the data from the ring buffer to the tracefile.
1719 *
1720 * It must be called with the stream lock held.
1721 *
1722 * Returns the number of bytes spliced.
1723 */
1724ssize_t lttng_consumer_on_read_subbuffer_splice(
1725 struct lttng_consumer_local_data *ctx,
1726 struct lttng_consumer_stream *stream, unsigned long len,
1727 unsigned long padding,
1728 struct ctf_packet_index *index)
1729{
1730 ssize_t ret = 0, written = 0, ret_splice = 0;
1731 loff_t offset = 0;
1732 off_t orig_offset = stream->out_fd_offset;
1733 int fd = stream->wait_fd;
1734 /* Default is on the disk */
1735 int outfd = stream->out_fd;
1736 struct consumer_relayd_sock_pair *relayd = NULL;
1737 int *splice_pipe;
1738 unsigned int relayd_hang_up = 0;
1739
1740 switch (consumer_data.type) {
1741 case LTTNG_CONSUMER_KERNEL:
1742 break;
1743 case LTTNG_CONSUMER32_UST:
1744 case LTTNG_CONSUMER64_UST:
1745 /* Not supported for user space tracing */
1746 return -ENOSYS;
1747 default:
1748 ERR("Unknown consumer_data type");
1749 assert(0);
1750 }
1751
1752 /* RCU lock for the relayd pointer */
1753 rcu_read_lock();
1754
1755 /* Flag that the current stream if set for network streaming. */
1756 if (stream->net_seq_idx != (uint64_t) -1ULL) {
1757 relayd = consumer_find_relayd(stream->net_seq_idx);
1758 if (relayd == NULL) {
1759 written = -ret;
1760 goto end;
1761 }
1762 }
1763 splice_pipe = stream->splice_pipe;
1764
1765 /* Write metadata stream id before payload */
1766 if (relayd) {
1767 unsigned long total_len = len;
1768
1769 if (stream->metadata_flag) {
1770 /*
1771 * Lock the control socket for the complete duration of the function
1772 * since from this point on we will use the socket.
1773 */
1774 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
1775
1776 if (stream->reset_metadata_flag) {
1777 ret = relayd_reset_metadata(&relayd->control_sock,
1778 stream->relayd_stream_id,
1779 stream->metadata_version);
1780 if (ret < 0) {
1781 relayd_hang_up = 1;
1782 goto write_error;
1783 }
1784 stream->reset_metadata_flag = 0;
1785 }
1786 ret = write_relayd_metadata_id(splice_pipe[1], stream, relayd,
1787 padding);
1788 if (ret < 0) {
1789 written = ret;
1790 relayd_hang_up = 1;
1791 goto write_error;
1792 }
1793
1794 total_len += sizeof(struct lttcomm_relayd_metadata_payload);
1795 }
1796
1797 ret = write_relayd_stream_header(stream, total_len, padding, relayd);
1798 if (ret < 0) {
1799 written = ret;
1800 relayd_hang_up = 1;
1801 goto write_error;
1802 }
1803 /* Use the returned socket. */
1804 outfd = ret;
1805 } else {
1806 /* No streaming, we have to set the len with the full padding */
1807 len += padding;
1808
1809 if (stream->metadata_flag && stream->reset_metadata_flag) {
1810 ret = utils_truncate_stream_file(stream->out_fd, 0);
1811 if (ret < 0) {
1812 ERR("Reset metadata file");
1813 goto end;
1814 }
1815 stream->reset_metadata_flag = 0;
1816 }
1817 /*
1818 * Check if we need to change the tracefile before writing the packet.
1819 */
1820 if (stream->chan->tracefile_size > 0 &&
1821 (stream->tracefile_size_current + len) >
1822 stream->chan->tracefile_size) {
1823 ret = utils_rotate_stream_file(stream->chan->pathname,
1824 stream->name, stream->chan->tracefile_size,
1825 stream->chan->tracefile_count, stream->uid, stream->gid,
1826 stream->out_fd, &(stream->tracefile_count_current),
1827 &stream->out_fd);
1828 if (ret < 0) {
1829 written = ret;
1830 ERR("Rotating output file");
1831 goto end;
1832 }
1833 outfd = stream->out_fd;
1834
1835 if (stream->index_file) {
1836 lttng_index_file_put(stream->index_file);
1837 stream->index_file = lttng_index_file_create(stream->chan->pathname,
1838 stream->name, stream->uid, stream->gid,
1839 stream->chan->tracefile_size,
1840 stream->tracefile_count_current,
1841 CTF_INDEX_MAJOR, CTF_INDEX_MINOR);
1842 if (!stream->index_file) {
1843 goto end;
1844 }
1845 }
1846
1847 /* Reset current size because we just perform a rotation. */
1848 stream->tracefile_size_current = 0;
1849 stream->out_fd_offset = 0;
1850 orig_offset = 0;
1851 }
1852 stream->tracefile_size_current += len;
1853 index->offset = htobe64(stream->out_fd_offset);
1854 }
1855
1856 while (len > 0) {
1857 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1858 (unsigned long)offset, len, fd, splice_pipe[1]);
1859 ret_splice = splice(fd, &offset, splice_pipe[1], NULL, len,
1860 SPLICE_F_MOVE | SPLICE_F_MORE);
1861 DBG("splice chan to pipe, ret %zd", ret_splice);
1862 if (ret_splice < 0) {
1863 ret = errno;
1864 written = -ret;
1865 PERROR("Error in relay splice");
1866 goto splice_error;
1867 }
1868
1869 /* Handle stream on the relayd if the output is on the network */
1870 if (relayd && stream->metadata_flag) {
1871 size_t metadata_payload_size =
1872 sizeof(struct lttcomm_relayd_metadata_payload);
1873
1874 /* Update counter to fit the spliced data */
1875 ret_splice += metadata_payload_size;
1876 len += metadata_payload_size;
1877 /*
1878 * We do this so the return value can match the len passed as
1879 * argument to this function.
1880 */
1881 written -= metadata_payload_size;
1882 }
1883
1884 /* Splice data out */
1885 ret_splice = splice(splice_pipe[0], NULL, outfd, NULL,
1886 ret_splice, SPLICE_F_MOVE | SPLICE_F_MORE);
1887 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
1888 outfd, ret_splice);
1889 if (ret_splice < 0) {
1890 ret = errno;
1891 written = -ret;
1892 relayd_hang_up = 1;
1893 goto write_error;
1894 } else if (ret_splice > len) {
1895 /*
1896 * We don't expect this code path to be executed but you never know
1897 * so this is an extra protection agains a buggy splice().
1898 */
1899 ret = errno;
1900 written += ret_splice;
1901 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice,
1902 len);
1903 goto splice_error;
1904 } else {
1905 /* All good, update current len and continue. */
1906 len -= ret_splice;
1907 }
1908
1909 /* This call is useless on a socket so better save a syscall. */
1910 if (!relayd) {
1911 /* This won't block, but will start writeout asynchronously */
1912 lttng_sync_file_range(outfd, stream->out_fd_offset, ret_splice,
1913 SYNC_FILE_RANGE_WRITE);
1914 stream->out_fd_offset += ret_splice;
1915 }
1916 stream->output_written += ret_splice;
1917 written += ret_splice;
1918 }
1919 if (!relayd) {
1920 lttng_consumer_sync_trace_file(stream, orig_offset);
1921 }
1922 goto end;
1923
1924write_error:
1925 /*
1926 * This is a special case that the relayd has closed its socket. Let's
1927 * cleanup the relayd object and all associated streams.
1928 */
1929 if (relayd && relayd_hang_up) {
1930 cleanup_relayd(relayd, ctx);
1931 /* Skip splice error so the consumer does not fail */
1932 goto end;
1933 }
1934
1935splice_error:
1936 /* send the appropriate error description to sessiond */
1937 switch (ret) {
1938 case EINVAL:
1939 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_EINVAL);
1940 break;
1941 case ENOMEM:
1942 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_ENOMEM);
1943 break;
1944 case ESPIPE:
1945 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_ESPIPE);
1946 break;
1947 }
1948
1949end:
1950 if (relayd && stream->metadata_flag) {
1951 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
1952 }
1953
1954 rcu_read_unlock();
1955 return written;
1956}
1957
1958/*
1959 * Take a snapshot for a specific fd
1960 *
1961 * Returns 0 on success, < 0 on error
1962 */
1963int lttng_consumer_take_snapshot(struct lttng_consumer_stream *stream)
1964{
1965 switch (consumer_data.type) {
1966 case LTTNG_CONSUMER_KERNEL:
1967 return lttng_kconsumer_take_snapshot(stream);
1968 case LTTNG_CONSUMER32_UST:
1969 case LTTNG_CONSUMER64_UST:
1970 return lttng_ustconsumer_take_snapshot(stream);
1971 default:
1972 ERR("Unknown consumer_data type");
1973 assert(0);
1974 return -ENOSYS;
1975 }
1976}
1977
1978/*
1979 * Get the produced position
1980 *
1981 * Returns 0 on success, < 0 on error
1982 */
1983int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream *stream,
1984 unsigned long *pos)
1985{
1986 switch (consumer_data.type) {
1987 case LTTNG_CONSUMER_KERNEL:
1988 return lttng_kconsumer_get_produced_snapshot(stream, pos);
1989 case LTTNG_CONSUMER32_UST:
1990 case LTTNG_CONSUMER64_UST:
1991 return lttng_ustconsumer_get_produced_snapshot(stream, pos);
1992 default:
1993 ERR("Unknown consumer_data type");
1994 assert(0);
1995 return -ENOSYS;
1996 }
1997}
1998
1999int lttng_consumer_recv_cmd(struct lttng_consumer_local_data *ctx,
2000 int sock, struct pollfd *consumer_sockpoll)
2001{
2002 switch (consumer_data.type) {
2003 case LTTNG_CONSUMER_KERNEL:
2004 return lttng_kconsumer_recv_cmd(ctx, sock, consumer_sockpoll);
2005 case LTTNG_CONSUMER32_UST:
2006 case LTTNG_CONSUMER64_UST:
2007 return lttng_ustconsumer_recv_cmd(ctx, sock, consumer_sockpoll);
2008 default:
2009 ERR("Unknown consumer_data type");
2010 assert(0);
2011 return -ENOSYS;
2012 }
2013}
2014
2015void lttng_consumer_close_all_metadata(void)
2016{
2017 switch (consumer_data.type) {
2018 case LTTNG_CONSUMER_KERNEL:
2019 /*
2020 * The Kernel consumer has a different metadata scheme so we don't
2021 * close anything because the stream will be closed by the session
2022 * daemon.
2023 */
2024 break;
2025 case LTTNG_CONSUMER32_UST:
2026 case LTTNG_CONSUMER64_UST:
2027 /*
2028 * Close all metadata streams. The metadata hash table is passed and
2029 * this call iterates over it by closing all wakeup fd. This is safe
2030 * because at this point we are sure that the metadata producer is
2031 * either dead or blocked.
2032 */
2033 lttng_ustconsumer_close_all_metadata(metadata_ht);
2034 break;
2035 default:
2036 ERR("Unknown consumer_data type");
2037 assert(0);
2038 }
2039}
2040
2041/*
2042 * Clean up a metadata stream and free its memory.
2043 */
2044void consumer_del_metadata_stream(struct lttng_consumer_stream *stream,
2045 struct lttng_ht *ht)
2046{
2047 struct lttng_consumer_channel *free_chan = NULL;
2048
2049 assert(stream);
2050 /*
2051 * This call should NEVER receive regular stream. It must always be
2052 * metadata stream and this is crucial for data structure synchronization.
2053 */
2054 assert(stream->metadata_flag);
2055
2056 DBG3("Consumer delete metadata stream %d", stream->wait_fd);
2057
2058 pthread_mutex_lock(&consumer_data.lock);
2059 pthread_mutex_lock(&stream->chan->lock);
2060 pthread_mutex_lock(&stream->lock);
2061 if (stream->chan->metadata_cache) {
2062 /* Only applicable to userspace consumers. */
2063 pthread_mutex_lock(&stream->chan->metadata_cache->lock);
2064 }
2065
2066 /* Remove any reference to that stream. */
2067 consumer_stream_delete(stream, ht);
2068
2069 /* Close down everything including the relayd if one. */
2070 consumer_stream_close(stream);
2071 /* Destroy tracer buffers of the stream. */
2072 consumer_stream_destroy_buffers(stream);
2073
2074 /* Atomically decrement channel refcount since other threads can use it. */
2075 if (!uatomic_sub_return(&stream->chan->refcount, 1)
2076 && !uatomic_read(&stream->chan->nb_init_stream_left)) {
2077 /* Go for channel deletion! */
2078 free_chan = stream->chan;
2079 }
2080
2081 /*
2082 * Nullify the stream reference so it is not used after deletion. The
2083 * channel lock MUST be acquired before being able to check for a NULL
2084 * pointer value.
2085 */
2086 stream->chan->metadata_stream = NULL;
2087
2088 if (stream->chan->metadata_cache) {
2089 pthread_mutex_unlock(&stream->chan->metadata_cache->lock);
2090 }
2091 pthread_mutex_unlock(&stream->lock);
2092 pthread_mutex_unlock(&stream->chan->lock);
2093 pthread_mutex_unlock(&consumer_data.lock);
2094
2095 if (free_chan) {
2096 consumer_del_channel(free_chan);
2097 }
2098
2099 consumer_stream_free(stream);
2100}
2101
2102/*
2103 * Action done with the metadata stream when adding it to the consumer internal
2104 * data structures to handle it.
2105 */
2106void consumer_add_metadata_stream(struct lttng_consumer_stream *stream)
2107{
2108 struct lttng_ht *ht = metadata_ht;
2109 struct lttng_ht_iter iter;
2110 struct lttng_ht_node_u64 *node;
2111
2112 assert(stream);
2113 assert(ht);
2114
2115 DBG3("Adding metadata stream %" PRIu64 " to hash table", stream->key);
2116
2117 pthread_mutex_lock(&consumer_data.lock);
2118 pthread_mutex_lock(&stream->chan->lock);
2119 pthread_mutex_lock(&stream->chan->timer_lock);
2120 pthread_mutex_lock(&stream->lock);
2121
2122 /*
2123 * From here, refcounts are updated so be _careful_ when returning an error
2124 * after this point.
2125 */
2126
2127 rcu_read_lock();
2128
2129 /*
2130 * Lookup the stream just to make sure it does not exist in our internal
2131 * state. This should NEVER happen.
2132 */
2133 lttng_ht_lookup(ht, &stream->key, &iter);
2134 node = lttng_ht_iter_get_node_u64(&iter);
2135 assert(!node);
2136
2137 /*
2138 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2139 * in terms of destroying the associated channel, because the action that
2140 * causes the count to become 0 also causes a stream to be added. The
2141 * channel deletion will thus be triggered by the following removal of this
2142 * stream.
2143 */
2144 if (uatomic_read(&stream->chan->nb_init_stream_left) > 0) {
2145 /* Increment refcount before decrementing nb_init_stream_left */
2146 cmm_smp_wmb();
2147 uatomic_dec(&stream->chan->nb_init_stream_left);
2148 }
2149
2150 lttng_ht_add_unique_u64(ht, &stream->node);
2151
2152 lttng_ht_add_unique_u64(consumer_data.stream_per_chan_id_ht,
2153 &stream->node_channel_id);
2154
2155 /*
2156 * Add stream to the stream_list_ht of the consumer data. No need to steal
2157 * the key since the HT does not use it and we allow to add redundant keys
2158 * into this table.
2159 */
2160 lttng_ht_add_u64(consumer_data.stream_list_ht, &stream->node_session_id);
2161
2162 rcu_read_unlock();
2163
2164 pthread_mutex_unlock(&stream->lock);
2165 pthread_mutex_unlock(&stream->chan->lock);
2166 pthread_mutex_unlock(&stream->chan->timer_lock);
2167 pthread_mutex_unlock(&consumer_data.lock);
2168}
2169
2170/*
2171 * Delete data stream that are flagged for deletion (endpoint_status).
2172 */
2173static void validate_endpoint_status_data_stream(void)
2174{
2175 struct lttng_ht_iter iter;
2176 struct lttng_consumer_stream *stream;
2177
2178 DBG("Consumer delete flagged data stream");
2179
2180 rcu_read_lock();
2181 cds_lfht_for_each_entry(data_ht->ht, &iter.iter, stream, node.node) {
2182 /* Validate delete flag of the stream */
2183 if (stream->endpoint_status == CONSUMER_ENDPOINT_ACTIVE) {
2184 continue;
2185 }
2186 /* Delete it right now */
2187 consumer_del_stream(stream, data_ht);
2188 }
2189 rcu_read_unlock();
2190}
2191
2192/*
2193 * Delete metadata stream that are flagged for deletion (endpoint_status).
2194 */
2195static void validate_endpoint_status_metadata_stream(
2196 struct lttng_poll_event *pollset)
2197{
2198 struct lttng_ht_iter iter;
2199 struct lttng_consumer_stream *stream;
2200
2201 DBG("Consumer delete flagged metadata stream");
2202
2203 assert(pollset);
2204
2205 rcu_read_lock();
2206 cds_lfht_for_each_entry(metadata_ht->ht, &iter.iter, stream, node.node) {
2207 /* Validate delete flag of the stream */
2208 if (stream->endpoint_status == CONSUMER_ENDPOINT_ACTIVE) {
2209 continue;
2210 }
2211 /*
2212 * Remove from pollset so the metadata thread can continue without
2213 * blocking on a deleted stream.
2214 */
2215 lttng_poll_del(pollset, stream->wait_fd);
2216
2217 /* Delete it right now */
2218 consumer_del_metadata_stream(stream, metadata_ht);
2219 }
2220 rcu_read_unlock();
2221}
2222
2223/*
2224 * Thread polls on metadata file descriptor and write them on disk or on the
2225 * network.
2226 */
2227void *consumer_thread_metadata_poll(void *data)
2228{
2229 int ret, i, pollfd, err = -1;
2230 uint32_t revents, nb_fd;
2231 struct lttng_consumer_stream *stream = NULL;
2232 struct lttng_ht_iter iter;
2233 struct lttng_ht_node_u64 *node;
2234 struct lttng_poll_event events;
2235 struct lttng_consumer_local_data *ctx = data;
2236 ssize_t len;
2237
2238 rcu_register_thread();
2239
2240 health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_METADATA);
2241
2242 if (testpoint(consumerd_thread_metadata)) {
2243 goto error_testpoint;
2244 }
2245
2246 health_code_update();
2247
2248 DBG("Thread metadata poll started");
2249
2250 /* Size is set to 1 for the consumer_metadata pipe */
2251 ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC);
2252 if (ret < 0) {
2253 ERR("Poll set creation failed");
2254 goto end_poll;
2255 }
2256
2257 ret = lttng_poll_add(&events,
2258 lttng_pipe_get_readfd(ctx->consumer_metadata_pipe), LPOLLIN);
2259 if (ret < 0) {
2260 goto end;
2261 }
2262
2263 /* Main loop */
2264 DBG("Metadata main loop started");
2265
2266 while (1) {
2267restart:
2268 health_code_update();
2269 health_poll_entry();
2270 DBG("Metadata poll wait");
2271 ret = lttng_poll_wait(&events, -1);
2272 DBG("Metadata poll return from wait with %d fd(s)",
2273 LTTNG_POLL_GETNB(&events));
2274 health_poll_exit();
2275 DBG("Metadata event caught in thread");
2276 if (ret < 0) {
2277 if (errno == EINTR) {
2278 ERR("Poll EINTR caught");
2279 goto restart;
2280 }
2281 if (LTTNG_POLL_GETNB(&events) == 0) {
2282 err = 0; /* All is OK */
2283 }
2284 goto end;
2285 }
2286
2287 nb_fd = ret;
2288
2289 /* From here, the event is a metadata wait fd */
2290 for (i = 0; i < nb_fd; i++) {
2291 health_code_update();
2292
2293 revents = LTTNG_POLL_GETEV(&events, i);
2294 pollfd = LTTNG_POLL_GETFD(&events, i);
2295
2296 if (!revents) {
2297 /* No activity for this FD (poll implementation). */
2298 continue;
2299 }
2300
2301 if (pollfd == lttng_pipe_get_readfd(ctx->consumer_metadata_pipe)) {
2302 if (revents & LPOLLIN) {
2303 ssize_t pipe_len;
2304
2305 pipe_len = lttng_pipe_read(ctx->consumer_metadata_pipe,
2306 &stream, sizeof(stream));
2307 if (pipe_len < sizeof(stream)) {
2308 if (pipe_len < 0) {
2309 PERROR("read metadata stream");
2310 }
2311 /*
2312 * Remove the pipe from the poll set and continue the loop
2313 * since their might be data to consume.
2314 */
2315 lttng_poll_del(&events,
2316 lttng_pipe_get_readfd(ctx->consumer_metadata_pipe));
2317 lttng_pipe_read_close(ctx->consumer_metadata_pipe);
2318 continue;
2319 }
2320
2321 /* A NULL stream means that the state has changed. */
2322 if (stream == NULL) {
2323 /* Check for deleted streams. */
2324 validate_endpoint_status_metadata_stream(&events);
2325 goto restart;
2326 }
2327
2328 DBG("Adding metadata stream %d to poll set",
2329 stream->wait_fd);
2330
2331 /* Add metadata stream to the global poll events list */
2332 lttng_poll_add(&events, stream->wait_fd,
2333 LPOLLIN | LPOLLPRI | LPOLLHUP);
2334 } else if (revents & (LPOLLERR | LPOLLHUP)) {
2335 DBG("Metadata thread pipe hung up");
2336 /*
2337 * Remove the pipe from the poll set and continue the loop
2338 * since their might be data to consume.
2339 */
2340 lttng_poll_del(&events,
2341 lttng_pipe_get_readfd(ctx->consumer_metadata_pipe));
2342 lttng_pipe_read_close(ctx->consumer_metadata_pipe);
2343 continue;
2344 } else {
2345 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
2346 goto end;
2347 }
2348
2349 /* Handle other stream */
2350 continue;
2351 }
2352
2353 rcu_read_lock();
2354 {
2355 uint64_t tmp_id = (uint64_t) pollfd;
2356
2357 lttng_ht_lookup(metadata_ht, &tmp_id, &iter);
2358 }
2359 node = lttng_ht_iter_get_node_u64(&iter);
2360 assert(node);
2361
2362 stream = caa_container_of(node, struct lttng_consumer_stream,
2363 node);
2364
2365 if (revents & (LPOLLIN | LPOLLPRI)) {
2366 /* Get the data out of the metadata file descriptor */
2367 DBG("Metadata available on fd %d", pollfd);
2368 assert(stream->wait_fd == pollfd);
2369
2370 do {
2371 health_code_update();
2372
2373 len = ctx->on_buffer_ready(stream, ctx);
2374 /*
2375 * We don't check the return value here since if we get
2376 * a negative len, it means an error occurred thus we
2377 * simply remove it from the poll set and free the
2378 * stream.
2379 */
2380 } while (len > 0);
2381
2382 /* It's ok to have an unavailable sub-buffer */
2383 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2384 /* Clean up stream from consumer and free it. */
2385 lttng_poll_del(&events, stream->wait_fd);
2386 consumer_del_metadata_stream(stream, metadata_ht);
2387 }
2388 } else if (revents & (LPOLLERR | LPOLLHUP)) {
2389 DBG("Metadata fd %d is hup|err.", pollfd);
2390 if (!stream->hangup_flush_done
2391 && (consumer_data.type == LTTNG_CONSUMER32_UST
2392 || consumer_data.type == LTTNG_CONSUMER64_UST)) {
2393 DBG("Attempting to flush and consume the UST buffers");
2394 lttng_ustconsumer_on_stream_hangup(stream);
2395
2396 /* We just flushed the stream now read it. */
2397 do {
2398 health_code_update();
2399
2400 len = ctx->on_buffer_ready(stream, ctx);
2401 /*
2402 * We don't check the return value here since if we get
2403 * a negative len, it means an error occurred thus we
2404 * simply remove it from the poll set and free the
2405 * stream.
2406 */
2407 } while (len > 0);
2408 }
2409
2410 lttng_poll_del(&events, stream->wait_fd);
2411 /*
2412 * This call update the channel states, closes file descriptors
2413 * and securely free the stream.
2414 */
2415 consumer_del_metadata_stream(stream, metadata_ht);
2416 } else {
2417 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
2418 rcu_read_unlock();
2419 goto end;
2420 }
2421 /* Release RCU lock for the stream looked up */
2422 rcu_read_unlock();
2423 }
2424 }
2425
2426 /* All is OK */
2427 err = 0;
2428end:
2429 DBG("Metadata poll thread exiting");
2430
2431 lttng_poll_clean(&events);
2432end_poll:
2433error_testpoint:
2434 if (err) {
2435 health_error();
2436 ERR("Health error occurred in %s", __func__);
2437 }
2438 health_unregister(health_consumerd);
2439 rcu_unregister_thread();
2440 return NULL;
2441}
2442
2443/*
2444 * This thread polls the fds in the set to consume the data and write
2445 * it to tracefile if necessary.
2446 */
2447void *consumer_thread_data_poll(void *data)
2448{
2449 int num_rdy, num_hup, high_prio, ret, i, err = -1;
2450 struct pollfd *pollfd = NULL;
2451 /* local view of the streams */
2452 struct lttng_consumer_stream **local_stream = NULL, *new_stream = NULL;
2453 /* local view of consumer_data.fds_count */
2454 int nb_fd = 0;
2455 struct lttng_consumer_local_data *ctx = data;
2456 ssize_t len;
2457
2458 rcu_register_thread();
2459
2460 health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_DATA);
2461
2462 if (testpoint(consumerd_thread_data)) {
2463 goto error_testpoint;
2464 }
2465
2466 health_code_update();
2467
2468 local_stream = zmalloc(sizeof(struct lttng_consumer_stream *));
2469 if (local_stream == NULL) {
2470 PERROR("local_stream malloc");
2471 goto end;
2472 }
2473
2474 while (1) {
2475 health_code_update();
2476
2477 high_prio = 0;
2478 num_hup = 0;
2479
2480 /*
2481 * the fds set has been updated, we need to update our
2482 * local array as well
2483 */
2484 pthread_mutex_lock(&consumer_data.lock);
2485 if (consumer_data.need_update) {
2486 free(pollfd);
2487 pollfd = NULL;
2488
2489 free(local_stream);
2490 local_stream = NULL;
2491
2492 /*
2493 * Allocate for all fds +1 for the consumer_data_pipe and +1 for
2494 * wake up pipe.
2495 */
2496 pollfd = zmalloc((consumer_data.stream_count + 2) * sizeof(struct pollfd));
2497 if (pollfd == NULL) {
2498 PERROR("pollfd malloc");
2499 pthread_mutex_unlock(&consumer_data.lock);
2500 goto end;
2501 }
2502
2503 local_stream = zmalloc((consumer_data.stream_count + 2) *
2504 sizeof(struct lttng_consumer_stream *));
2505 if (local_stream == NULL) {
2506 PERROR("local_stream malloc");
2507 pthread_mutex_unlock(&consumer_data.lock);
2508 goto end;
2509 }
2510 ret = update_poll_array(ctx, &pollfd, local_stream,
2511 data_ht);
2512 if (ret < 0) {
2513 ERR("Error in allocating pollfd or local_outfds");
2514 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
2515 pthread_mutex_unlock(&consumer_data.lock);
2516 goto end;
2517 }
2518 nb_fd = ret;
2519 consumer_data.need_update = 0;
2520 }
2521 pthread_mutex_unlock(&consumer_data.lock);
2522
2523 /* No FDs and consumer_quit, consumer_cleanup the thread */
2524 if (nb_fd == 0 && CMM_LOAD_SHARED(consumer_quit) == 1) {
2525 err = 0; /* All is OK */
2526 goto end;
2527 }
2528 /* poll on the array of fds */
2529 restart:
2530 DBG("polling on %d fd", nb_fd + 2);
2531 if (testpoint(consumerd_thread_data_poll)) {
2532 goto end;
2533 }
2534 health_poll_entry();
2535 num_rdy = poll(pollfd, nb_fd + 2, -1);
2536 health_poll_exit();
2537 DBG("poll num_rdy : %d", num_rdy);
2538 if (num_rdy == -1) {
2539 /*
2540 * Restart interrupted system call.
2541 */
2542 if (errno == EINTR) {
2543 goto restart;
2544 }
2545 PERROR("Poll error");
2546 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
2547 goto end;
2548 } else if (num_rdy == 0) {
2549 DBG("Polling thread timed out");
2550 goto end;
2551 }
2552
2553 if (caa_unlikely(data_consumption_paused)) {
2554 DBG("Data consumption paused, sleeping...");
2555 sleep(1);
2556 goto restart;
2557 }
2558
2559 /*
2560 * If the consumer_data_pipe triggered poll go directly to the
2561 * beginning of the loop to update the array. We want to prioritize
2562 * array update over low-priority reads.
2563 */
2564 if (pollfd[nb_fd].revents & (POLLIN | POLLPRI)) {
2565 ssize_t pipe_readlen;
2566
2567 DBG("consumer_data_pipe wake up");
2568 pipe_readlen = lttng_pipe_read(ctx->consumer_data_pipe,
2569 &new_stream, sizeof(new_stream));
2570 if (pipe_readlen < sizeof(new_stream)) {
2571 PERROR("Consumer data pipe");
2572 /* Continue so we can at least handle the current stream(s). */
2573 continue;
2574 }
2575
2576 /*
2577 * If the stream is NULL, just ignore it. It's also possible that
2578 * the sessiond poll thread changed the consumer_quit state and is
2579 * waking us up to test it.
2580 */
2581 if (new_stream == NULL) {
2582 validate_endpoint_status_data_stream();
2583 continue;
2584 }
2585
2586 /* Continue to update the local streams and handle prio ones */
2587 continue;
2588 }
2589
2590 /* Handle wakeup pipe. */
2591 if (pollfd[nb_fd + 1].revents & (POLLIN | POLLPRI)) {
2592 char dummy;
2593 ssize_t pipe_readlen;
2594
2595 pipe_readlen = lttng_pipe_read(ctx->consumer_wakeup_pipe, &dummy,
2596 sizeof(dummy));
2597 if (pipe_readlen < 0) {
2598 PERROR("Consumer data wakeup pipe");
2599 }
2600 /* We've been awakened to handle stream(s). */
2601 ctx->has_wakeup = 0;
2602 }
2603
2604 /* Take care of high priority channels first. */
2605 for (i = 0; i < nb_fd; i++) {
2606 health_code_update();
2607
2608 if (local_stream[i] == NULL) {
2609 continue;
2610 }
2611 if (pollfd[i].revents & POLLPRI) {
2612 DBG("Urgent read on fd %d", pollfd[i].fd);
2613 high_prio = 1;
2614 len = ctx->on_buffer_ready(local_stream[i], ctx);
2615 /* it's ok to have an unavailable sub-buffer */
2616 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2617 /* Clean the stream and free it. */
2618 consumer_del_stream(local_stream[i], data_ht);
2619 local_stream[i] = NULL;
2620 } else if (len > 0) {
2621 local_stream[i]->data_read = 1;
2622 }
2623 }
2624 }
2625
2626 /*
2627 * If we read high prio channel in this loop, try again
2628 * for more high prio data.
2629 */
2630 if (high_prio) {
2631 continue;
2632 }
2633
2634 /* Take care of low priority channels. */
2635 for (i = 0; i < nb_fd; i++) {
2636 health_code_update();
2637
2638 if (local_stream[i] == NULL) {
2639 continue;
2640 }
2641 if ((pollfd[i].revents & POLLIN) ||
2642 local_stream[i]->hangup_flush_done ||
2643 local_stream[i]->has_data) {
2644 DBG("Normal read on fd %d", pollfd[i].fd);
2645 len = ctx->on_buffer_ready(local_stream[i], ctx);
2646 /* it's ok to have an unavailable sub-buffer */
2647 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2648 /* Clean the stream and free it. */
2649 consumer_del_stream(local_stream[i], data_ht);
2650 local_stream[i] = NULL;
2651 } else if (len > 0) {
2652 local_stream[i]->data_read = 1;
2653 }
2654 }
2655 }
2656
2657 /* Handle hangup and errors */
2658 for (i = 0; i < nb_fd; i++) {
2659 health_code_update();
2660
2661 if (local_stream[i] == NULL) {
2662 continue;
2663 }
2664 if (!local_stream[i]->hangup_flush_done
2665 && (pollfd[i].revents & (POLLHUP | POLLERR | POLLNVAL))
2666 && (consumer_data.type == LTTNG_CONSUMER32_UST
2667 || consumer_data.type == LTTNG_CONSUMER64_UST)) {
2668 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2669 pollfd[i].fd);
2670 lttng_ustconsumer_on_stream_hangup(local_stream[i]);
2671 /* Attempt read again, for the data we just flushed. */
2672 local_stream[i]->data_read = 1;
2673 }
2674 /*
2675 * If the poll flag is HUP/ERR/NVAL and we have
2676 * read no data in this pass, we can remove the
2677 * stream from its hash table.
2678 */
2679 if ((pollfd[i].revents & POLLHUP)) {
2680 DBG("Polling fd %d tells it has hung up.", pollfd[i].fd);
2681 if (!local_stream[i]->data_read) {
2682 consumer_del_stream(local_stream[i], data_ht);
2683 local_stream[i] = NULL;
2684 num_hup++;
2685 }
2686 } else if (pollfd[i].revents & POLLERR) {
2687 ERR("Error returned in polling fd %d.", pollfd[i].fd);
2688 if (!local_stream[i]->data_read) {
2689 consumer_del_stream(local_stream[i], data_ht);
2690 local_stream[i] = NULL;
2691 num_hup++;
2692 }
2693 } else if (pollfd[i].revents & POLLNVAL) {
2694 ERR("Polling fd %d tells fd is not open.", pollfd[i].fd);
2695 if (!local_stream[i]->data_read) {
2696 consumer_del_stream(local_stream[i], data_ht);
2697 local_stream[i] = NULL;
2698 num_hup++;
2699 }
2700 }
2701 if (local_stream[i] != NULL) {
2702 local_stream[i]->data_read = 0;
2703 }
2704 }
2705 }
2706 /* All is OK */
2707 err = 0;
2708end:
2709 DBG("polling thread exiting");
2710 free(pollfd);
2711 free(local_stream);
2712
2713 /*
2714 * Close the write side of the pipe so epoll_wait() in
2715 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2716 * read side of the pipe. If we close them both, epoll_wait strangely does
2717 * not return and could create a endless wait period if the pipe is the
2718 * only tracked fd in the poll set. The thread will take care of closing
2719 * the read side.
2720 */
2721 (void) lttng_pipe_write_close(ctx->consumer_metadata_pipe);
2722
2723error_testpoint:
2724 if (err) {
2725 health_error();
2726 ERR("Health error occurred in %s", __func__);
2727 }
2728 health_unregister(health_consumerd);
2729
2730 rcu_unregister_thread();
2731 return NULL;
2732}
2733
2734/*
2735 * Close wake-up end of each stream belonging to the channel. This will
2736 * allow the poll() on the stream read-side to detect when the
2737 * write-side (application) finally closes them.
2738 */
2739static
2740void consumer_close_channel_streams(struct lttng_consumer_channel *channel)
2741{
2742 struct lttng_ht *ht;
2743 struct lttng_consumer_stream *stream;
2744 struct lttng_ht_iter iter;
2745
2746 ht = consumer_data.stream_per_chan_id_ht;
2747
2748 rcu_read_lock();
2749 cds_lfht_for_each_entry_duplicate(ht->ht,
2750 ht->hash_fct(&channel->key, lttng_ht_seed),
2751 ht->match_fct, &channel->key,
2752 &iter.iter, stream, node_channel_id.node) {
2753 /*
2754 * Protect against teardown with mutex.
2755 */
2756 pthread_mutex_lock(&stream->lock);
2757 if (cds_lfht_is_node_deleted(&stream->node.node)) {
2758 goto next;
2759 }
2760 switch (consumer_data.type) {
2761 case LTTNG_CONSUMER_KERNEL:
2762 break;
2763 case LTTNG_CONSUMER32_UST:
2764 case LTTNG_CONSUMER64_UST:
2765 if (stream->metadata_flag) {
2766 /* Safe and protected by the stream lock. */
2767 lttng_ustconsumer_close_metadata(stream->chan);
2768 } else {
2769 /*
2770 * Note: a mutex is taken internally within
2771 * liblttng-ust-ctl to protect timer wakeup_fd
2772 * use from concurrent close.
2773 */
2774 lttng_ustconsumer_close_stream_wakeup(stream);
2775 }
2776 break;
2777 default:
2778 ERR("Unknown consumer_data type");
2779 assert(0);
2780 }
2781 next:
2782 pthread_mutex_unlock(&stream->lock);
2783 }
2784 rcu_read_unlock();
2785}
2786
2787static void destroy_channel_ht(struct lttng_ht *ht)
2788{
2789 struct lttng_ht_iter iter;
2790 struct lttng_consumer_channel *channel;
2791 int ret;
2792
2793 if (ht == NULL) {
2794 return;
2795 }
2796
2797 rcu_read_lock();
2798 cds_lfht_for_each_entry(ht->ht, &iter.iter, channel, wait_fd_node.node) {
2799 ret = lttng_ht_del(ht, &iter);
2800 assert(ret != 0);
2801 }
2802 rcu_read_unlock();
2803
2804 lttng_ht_destroy(ht);
2805}
2806
2807/*
2808 * This thread polls the channel fds to detect when they are being
2809 * closed. It closes all related streams if the channel is detected as
2810 * closed. It is currently only used as a shim layer for UST because the
2811 * consumerd needs to keep the per-stream wakeup end of pipes open for
2812 * periodical flush.
2813 */
2814void *consumer_thread_channel_poll(void *data)
2815{
2816 int ret, i, pollfd, err = -1;
2817 uint32_t revents, nb_fd;
2818 struct lttng_consumer_channel *chan = NULL;
2819 struct lttng_ht_iter iter;
2820 struct lttng_ht_node_u64 *node;
2821 struct lttng_poll_event events;
2822 struct lttng_consumer_local_data *ctx = data;
2823 struct lttng_ht *channel_ht;
2824
2825 rcu_register_thread();
2826
2827 health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_CHANNEL);
2828
2829 if (testpoint(consumerd_thread_channel)) {
2830 goto error_testpoint;
2831 }
2832
2833 health_code_update();
2834
2835 channel_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
2836 if (!channel_ht) {
2837 /* ENOMEM at this point. Better to bail out. */
2838 goto end_ht;
2839 }
2840
2841 DBG("Thread channel poll started");
2842
2843 /* Size is set to 1 for the consumer_channel pipe */
2844 ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC);
2845 if (ret < 0) {
2846 ERR("Poll set creation failed");
2847 goto end_poll;
2848 }
2849
2850 ret = lttng_poll_add(&events, ctx->consumer_channel_pipe[0], LPOLLIN);
2851 if (ret < 0) {
2852 goto end;
2853 }
2854
2855 /* Main loop */
2856 DBG("Channel main loop started");
2857
2858 while (1) {
2859restart:
2860 health_code_update();
2861 DBG("Channel poll wait");
2862 health_poll_entry();
2863 ret = lttng_poll_wait(&events, -1);
2864 DBG("Channel poll return from wait with %d fd(s)",
2865 LTTNG_POLL_GETNB(&events));
2866 health_poll_exit();
2867 DBG("Channel event caught in thread");
2868 if (ret < 0) {
2869 if (errno == EINTR) {
2870 ERR("Poll EINTR caught");
2871 goto restart;
2872 }
2873 if (LTTNG_POLL_GETNB(&events) == 0) {
2874 err = 0; /* All is OK */
2875 }
2876 goto end;
2877 }
2878
2879 nb_fd = ret;
2880
2881 /* From here, the event is a channel wait fd */
2882 for (i = 0; i < nb_fd; i++) {
2883 health_code_update();
2884
2885 revents = LTTNG_POLL_GETEV(&events, i);
2886 pollfd = LTTNG_POLL_GETFD(&events, i);
2887
2888 if (!revents) {
2889 /* No activity for this FD (poll implementation). */
2890 continue;
2891 }
2892
2893 if (pollfd == ctx->consumer_channel_pipe[0]) {
2894 if (revents & LPOLLIN) {
2895 enum consumer_channel_action action;
2896 uint64_t key;
2897
2898 ret = read_channel_pipe(ctx, &chan, &key, &action);
2899 if (ret <= 0) {
2900 if (ret < 0) {
2901 ERR("Error reading channel pipe");
2902 }
2903 lttng_poll_del(&events, ctx->consumer_channel_pipe[0]);
2904 continue;
2905 }
2906
2907 switch (action) {
2908 case CONSUMER_CHANNEL_ADD:
2909 DBG("Adding channel %d to poll set",
2910 chan->wait_fd);
2911
2912 lttng_ht_node_init_u64(&chan->wait_fd_node,
2913 chan->wait_fd);
2914 rcu_read_lock();
2915 lttng_ht_add_unique_u64(channel_ht,
2916 &chan->wait_fd_node);
2917 rcu_read_unlock();
2918 /* Add channel to the global poll events list */
2919 lttng_poll_add(&events, chan->wait_fd,
2920 LPOLLERR | LPOLLHUP);
2921 break;
2922 case CONSUMER_CHANNEL_DEL:
2923 {
2924 /*
2925 * This command should never be called if the channel
2926 * has streams monitored by either the data or metadata
2927 * thread. The consumer only notify this thread with a
2928 * channel del. command if it receives a destroy
2929 * channel command from the session daemon that send it
2930 * if a command prior to the GET_CHANNEL failed.
2931 */
2932
2933 rcu_read_lock();
2934 chan = consumer_find_channel(key);
2935 if (!chan) {
2936 rcu_read_unlock();
2937 ERR("UST consumer get channel key %" PRIu64 " not found for del channel", key);
2938 break;
2939 }
2940 lttng_poll_del(&events, chan->wait_fd);
2941 iter.iter.node = &chan->wait_fd_node.node;
2942 ret = lttng_ht_del(channel_ht, &iter);
2943 assert(ret == 0);
2944
2945 switch (consumer_data.type) {
2946 case LTTNG_CONSUMER_KERNEL:
2947 break;
2948 case LTTNG_CONSUMER32_UST:
2949 case LTTNG_CONSUMER64_UST:
2950 health_code_update();
2951 /* Destroy streams that might have been left in the stream list. */
2952 clean_channel_stream_list(chan);
2953 break;
2954 default:
2955 ERR("Unknown consumer_data type");
2956 assert(0);
2957 }
2958
2959 /*
2960 * Release our own refcount. Force channel deletion even if
2961 * streams were not initialized.
2962 */
2963 if (!uatomic_sub_return(&chan->refcount, 1)) {
2964 consumer_del_channel(chan);
2965 }
2966 rcu_read_unlock();
2967 goto restart;
2968 }
2969 case CONSUMER_CHANNEL_QUIT:
2970 /*
2971 * Remove the pipe from the poll set and continue the loop
2972 * since their might be data to consume.
2973 */
2974 lttng_poll_del(&events, ctx->consumer_channel_pipe[0]);
2975 continue;
2976 default:
2977 ERR("Unknown action");
2978 break;
2979 }
2980 } else if (revents & (LPOLLERR | LPOLLHUP)) {
2981 DBG("Channel thread pipe hung up");
2982 /*
2983 * Remove the pipe from the poll set and continue the loop
2984 * since their might be data to consume.
2985 */
2986 lttng_poll_del(&events, ctx->consumer_channel_pipe[0]);
2987 continue;
2988 } else {
2989 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
2990 goto end;
2991 }
2992
2993 /* Handle other stream */
2994 continue;
2995 }
2996
2997 rcu_read_lock();
2998 {
2999 uint64_t tmp_id = (uint64_t) pollfd;
3000
3001 lttng_ht_lookup(channel_ht, &tmp_id, &iter);
3002 }
3003 node = lttng_ht_iter_get_node_u64(&iter);
3004 assert(node);
3005
3006 chan = caa_container_of(node, struct lttng_consumer_channel,
3007 wait_fd_node);
3008
3009 /* Check for error event */
3010 if (revents & (LPOLLERR | LPOLLHUP)) {
3011 DBG("Channel fd %d is hup|err.", pollfd);
3012
3013 lttng_poll_del(&events, chan->wait_fd);
3014 ret = lttng_ht_del(channel_ht, &iter);
3015 assert(ret == 0);
3016
3017 /*
3018 * This will close the wait fd for each stream associated to
3019 * this channel AND monitored by the data/metadata thread thus
3020 * will be clean by the right thread.
3021 */
3022 consumer_close_channel_streams(chan);
3023
3024 /* Release our own refcount */
3025 if (!uatomic_sub_return(&chan->refcount, 1)
3026 && !uatomic_read(&chan->nb_init_stream_left)) {
3027 consumer_del_channel(chan);
3028 }
3029 } else {
3030 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
3031 rcu_read_unlock();
3032 goto end;
3033 }
3034
3035 /* Release RCU lock for the channel looked up */
3036 rcu_read_unlock();
3037 }
3038 }
3039
3040 /* All is OK */
3041 err = 0;
3042end:
3043 lttng_poll_clean(&events);
3044end_poll:
3045 destroy_channel_ht(channel_ht);
3046end_ht:
3047error_testpoint:
3048 DBG("Channel poll thread exiting");
3049 if (err) {
3050 health_error();
3051 ERR("Health error occurred in %s", __func__);
3052 }
3053 health_unregister(health_consumerd);
3054 rcu_unregister_thread();
3055 return NULL;
3056}
3057
3058static int set_metadata_socket(struct lttng_consumer_local_data *ctx,
3059 struct pollfd *sockpoll, int client_socket)
3060{
3061 int ret;
3062
3063 assert(ctx);
3064 assert(sockpoll);
3065
3066 ret = lttng_consumer_poll_socket(sockpoll);
3067 if (ret) {
3068 goto error;
3069 }
3070 DBG("Metadata connection on client_socket");
3071
3072 /* Blocking call, waiting for transmission */
3073 ctx->consumer_metadata_socket = lttcomm_accept_unix_sock(client_socket);
3074 if (ctx->consumer_metadata_socket < 0) {
3075 WARN("On accept metadata");
3076 ret = -1;
3077 goto error;
3078 }
3079 ret = 0;
3080
3081error:
3082 return ret;
3083}
3084
3085/*
3086 * This thread listens on the consumerd socket and receives the file
3087 * descriptors from the session daemon.
3088 */
3089void *consumer_thread_sessiond_poll(void *data)
3090{
3091 int sock = -1, client_socket, ret, err = -1;
3092 /*
3093 * structure to poll for incoming data on communication socket avoids
3094 * making blocking sockets.
3095 */
3096 struct pollfd consumer_sockpoll[2];
3097 struct lttng_consumer_local_data *ctx = data;
3098
3099 rcu_register_thread();
3100
3101 health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_SESSIOND);
3102
3103 if (testpoint(consumerd_thread_sessiond)) {
3104 goto error_testpoint;
3105 }
3106
3107 health_code_update();
3108
3109 DBG("Creating command socket %s", ctx->consumer_command_sock_path);
3110 unlink(ctx->consumer_command_sock_path);
3111 client_socket = lttcomm_create_unix_sock(ctx->consumer_command_sock_path);
3112 if (client_socket < 0) {
3113 ERR("Cannot create command socket");
3114 goto end;
3115 }
3116
3117 ret = lttcomm_listen_unix_sock(client_socket);
3118 if (ret < 0) {
3119 goto end;
3120 }
3121
3122 DBG("Sending ready command to lttng-sessiond");
3123 ret = lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY);
3124 /* return < 0 on error, but == 0 is not fatal */
3125 if (ret < 0) {
3126 ERR("Error sending ready command to lttng-sessiond");
3127 goto end;
3128 }
3129
3130 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3131 consumer_sockpoll[0].fd = ctx->consumer_should_quit[0];
3132 consumer_sockpoll[0].events = POLLIN | POLLPRI;
3133 consumer_sockpoll[1].fd = client_socket;
3134 consumer_sockpoll[1].events = POLLIN | POLLPRI;
3135
3136 ret = lttng_consumer_poll_socket(consumer_sockpoll);
3137 if (ret) {
3138 if (ret > 0) {
3139 /* should exit */
3140 err = 0;
3141 }
3142 goto end;
3143 }
3144 DBG("Connection on client_socket");
3145
3146 /* Blocking call, waiting for transmission */
3147 sock = lttcomm_accept_unix_sock(client_socket);
3148 if (sock < 0) {
3149 WARN("On accept");
3150 goto end;
3151 }
3152
3153 /*
3154 * Setup metadata socket which is the second socket connection on the
3155 * command unix socket.
3156 */
3157 ret = set_metadata_socket(ctx, consumer_sockpoll, client_socket);
3158 if (ret) {
3159 if (ret > 0) {
3160 /* should exit */
3161 err = 0;
3162 }
3163 goto end;
3164 }
3165
3166 /* This socket is not useful anymore. */
3167 ret = close(client_socket);
3168 if (ret < 0) {
3169 PERROR("close client_socket");
3170 }
3171 client_socket = -1;
3172
3173 /* update the polling structure to poll on the established socket */
3174 consumer_sockpoll[1].fd = sock;
3175 consumer_sockpoll[1].events = POLLIN | POLLPRI;
3176
3177 while (1) {
3178 health_code_update();
3179
3180 health_poll_entry();
3181 ret = lttng_consumer_poll_socket(consumer_sockpoll);
3182 health_poll_exit();
3183 if (ret) {
3184 if (ret > 0) {
3185 /* should exit */
3186 err = 0;
3187 }
3188 goto end;
3189 }
3190 DBG("Incoming command on sock");
3191 ret = lttng_consumer_recv_cmd(ctx, sock, consumer_sockpoll);
3192 if (ret <= 0) {
3193 /*
3194 * This could simply be a session daemon quitting. Don't output
3195 * ERR() here.
3196 */
3197 DBG("Communication interrupted on command socket");
3198 err = 0;
3199 goto end;
3200 }
3201 if (CMM_LOAD_SHARED(consumer_quit)) {
3202 DBG("consumer_thread_receive_fds received quit from signal");
3203 err = 0; /* All is OK */
3204 goto end;
3205 }
3206 DBG("received command on sock");
3207 }
3208 /* All is OK */
3209 err = 0;
3210
3211end:
3212 DBG("Consumer thread sessiond poll exiting");
3213
3214 /*
3215 * Close metadata streams since the producer is the session daemon which
3216 * just died.
3217 *
3218 * NOTE: for now, this only applies to the UST tracer.
3219 */
3220 lttng_consumer_close_all_metadata();
3221
3222 /*
3223 * when all fds have hung up, the polling thread
3224 * can exit cleanly
3225 */
3226 CMM_STORE_SHARED(consumer_quit, 1);
3227
3228 /*
3229 * Notify the data poll thread to poll back again and test the
3230 * consumer_quit state that we just set so to quit gracefully.
3231 */
3232 notify_thread_lttng_pipe(ctx->consumer_data_pipe);
3233
3234 notify_channel_pipe(ctx, NULL, -1, CONSUMER_CHANNEL_QUIT);
3235
3236 notify_health_quit_pipe(health_quit_pipe);
3237
3238 /* Cleaning up possibly open sockets. */
3239 if (sock >= 0) {
3240 ret = close(sock);
3241 if (ret < 0) {
3242 PERROR("close sock sessiond poll");
3243 }
3244 }
3245 if (client_socket >= 0) {
3246 ret = close(client_socket);
3247 if (ret < 0) {
3248 PERROR("close client_socket sessiond poll");
3249 }
3250 }
3251
3252error_testpoint:
3253 if (err) {
3254 health_error();
3255 ERR("Health error occurred in %s", __func__);
3256 }
3257 health_unregister(health_consumerd);
3258
3259 rcu_unregister_thread();
3260 return NULL;
3261}
3262
3263ssize_t lttng_consumer_read_subbuffer(struct lttng_consumer_stream *stream,
3264 struct lttng_consumer_local_data *ctx)
3265{
3266 ssize_t ret;
3267
3268 pthread_mutex_lock(&stream->lock);
3269 if (stream->metadata_flag) {
3270 pthread_mutex_lock(&stream->metadata_rdv_lock);
3271 }
3272
3273 switch (consumer_data.type) {
3274 case LTTNG_CONSUMER_KERNEL:
3275 ret = lttng_kconsumer_read_subbuffer(stream, ctx);
3276 break;
3277 case LTTNG_CONSUMER32_UST:
3278 case LTTNG_CONSUMER64_UST:
3279 ret = lttng_ustconsumer_read_subbuffer(stream, ctx);
3280 break;
3281 default:
3282 ERR("Unknown consumer_data type");
3283 assert(0);
3284 ret = -ENOSYS;
3285 break;
3286 }
3287
3288 if (stream->metadata_flag) {
3289 pthread_cond_broadcast(&stream->metadata_rdv);
3290 pthread_mutex_unlock(&stream->metadata_rdv_lock);
3291 }
3292 pthread_mutex_unlock(&stream->lock);
3293 return ret;
3294}
3295
3296int lttng_consumer_on_recv_stream(struct lttng_consumer_stream *stream)
3297{
3298 switch (consumer_data.type) {
3299 case LTTNG_CONSUMER_KERNEL:
3300 return lttng_kconsumer_on_recv_stream(stream);
3301 case LTTNG_CONSUMER32_UST:
3302 case LTTNG_CONSUMER64_UST:
3303 return lttng_ustconsumer_on_recv_stream(stream);
3304 default:
3305 ERR("Unknown consumer_data type");
3306 assert(0);
3307 return -ENOSYS;
3308 }
3309}
3310
3311/*
3312 * Allocate and set consumer data hash tables.
3313 */
3314int lttng_consumer_init(void)
3315{
3316 consumer_data.channel_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
3317 if (!consumer_data.channel_ht) {
3318 goto error;
3319 }
3320
3321 consumer_data.relayd_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
3322 if (!consumer_data.relayd_ht) {
3323 goto error;
3324 }
3325
3326 consumer_data.stream_list_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
3327 if (!consumer_data.stream_list_ht) {
3328 goto error;
3329 }
3330
3331 consumer_data.stream_per_chan_id_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
3332 if (!consumer_data.stream_per_chan_id_ht) {
3333 goto error;
3334 }
3335
3336 data_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
3337 if (!data_ht) {
3338 goto error;
3339 }
3340
3341 metadata_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
3342 if (!metadata_ht) {
3343 goto error;
3344 }
3345
3346 return 0;
3347
3348error:
3349 return -1;
3350}
3351
3352/*
3353 * Process the ADD_RELAYD command receive by a consumer.
3354 *
3355 * This will create a relayd socket pair and add it to the relayd hash table.
3356 * The caller MUST acquire a RCU read side lock before calling it.
3357 */
3358 void consumer_add_relayd_socket(uint64_t net_seq_idx, int sock_type,
3359 struct lttng_consumer_local_data *ctx, int sock,
3360 struct pollfd *consumer_sockpoll,
3361 struct lttcomm_relayd_sock *relayd_sock, uint64_t sessiond_id,
3362 uint64_t relayd_session_id)
3363{
3364 int fd = -1, ret = -1, relayd_created = 0;
3365 enum lttcomm_return_code ret_code = LTTCOMM_CONSUMERD_SUCCESS;
3366 struct consumer_relayd_sock_pair *relayd = NULL;
3367
3368 assert(ctx);
3369 assert(relayd_sock);
3370
3371 DBG("Consumer adding relayd socket (idx: %" PRIu64 ")", net_seq_idx);
3372
3373 /* Get relayd reference if exists. */
3374 relayd = consumer_find_relayd(net_seq_idx);
3375 if (relayd == NULL) {
3376 assert(sock_type == LTTNG_STREAM_CONTROL);
3377 /* Not found. Allocate one. */
3378 relayd = consumer_allocate_relayd_sock_pair(net_seq_idx);
3379 if (relayd == NULL) {
3380 ret_code = LTTCOMM_CONSUMERD_ENOMEM;
3381 goto error;
3382 } else {
3383 relayd->sessiond_session_id = sessiond_id;
3384 relayd_created = 1;
3385 }
3386
3387 /*
3388 * This code path MUST continue to the consumer send status message to
3389 * we can notify the session daemon and continue our work without
3390 * killing everything.
3391 */
3392 } else {
3393 /*
3394 * relayd key should never be found for control socket.
3395 */
3396 assert(sock_type != LTTNG_STREAM_CONTROL);
3397 }
3398
3399 /* First send a status message before receiving the fds. */
3400 ret = consumer_send_status_msg(sock, LTTCOMM_CONSUMERD_SUCCESS);
3401 if (ret < 0) {
3402 /* Somehow, the session daemon is not responding anymore. */
3403 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_FATAL);
3404 goto error_nosignal;
3405 }
3406
3407 /* Poll on consumer socket. */
3408 ret = lttng_consumer_poll_socket(consumer_sockpoll);
3409 if (ret) {
3410 /* Needing to exit in the middle of a command: error. */
3411 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
3412 goto error_nosignal;
3413 }
3414
3415 /* Get relayd socket from session daemon */
3416 ret = lttcomm_recv_fds_unix_sock(sock, &fd, 1);
3417 if (ret != sizeof(fd)) {
3418 fd = -1; /* Just in case it gets set with an invalid value. */
3419
3420 /*
3421 * Failing to receive FDs might indicate a major problem such as
3422 * reaching a fd limit during the receive where the kernel returns a
3423 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3424 * don't take any chances and stop everything.
3425 *
3426 * XXX: Feature request #558 will fix that and avoid this possible
3427 * issue when reaching the fd limit.
3428 */
3429 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_ERROR_RECV_FD);
3430 ret_code = LTTCOMM_CONSUMERD_ERROR_RECV_FD;
3431 goto error;
3432 }
3433
3434 /* Copy socket information and received FD */
3435 switch (sock_type) {
3436 case LTTNG_STREAM_CONTROL:
3437 /* Copy received lttcomm socket */
3438 lttcomm_copy_sock(&relayd->control_sock.sock, &relayd_sock->sock);
3439 ret = lttcomm_create_sock(&relayd->control_sock.sock);
3440 /* Handle create_sock error. */
3441 if (ret < 0) {
3442 ret_code = LTTCOMM_CONSUMERD_ENOMEM;
3443 goto error;
3444 }
3445 /*
3446 * Close the socket created internally by
3447 * lttcomm_create_sock, so we can replace it by the one
3448 * received from sessiond.
3449 */
3450 if (close(relayd->control_sock.sock.fd)) {
3451 PERROR("close");
3452 }
3453
3454 /* Assign new file descriptor */
3455 relayd->control_sock.sock.fd = fd;
3456 fd = -1; /* For error path */
3457 /* Assign version values. */
3458 relayd->control_sock.major = relayd_sock->major;
3459 relayd->control_sock.minor = relayd_sock->minor;
3460
3461 relayd->relayd_session_id = relayd_session_id;
3462
3463 break;
3464 case LTTNG_STREAM_DATA:
3465 /* Copy received lttcomm socket */
3466 lttcomm_copy_sock(&relayd->data_sock.sock, &relayd_sock->sock);
3467 ret = lttcomm_create_sock(&relayd->data_sock.sock);
3468 /* Handle create_sock error. */
3469 if (ret < 0) {
3470 ret_code = LTTCOMM_CONSUMERD_ENOMEM;
3471 goto error;
3472 }
3473 /*
3474 * Close the socket created internally by
3475 * lttcomm_create_sock, so we can replace it by the one
3476 * received from sessiond.
3477 */
3478 if (close(relayd->data_sock.sock.fd)) {
3479 PERROR("close");
3480 }
3481
3482 /* Assign new file descriptor */
3483 relayd->data_sock.sock.fd = fd;
3484 fd = -1; /* for eventual error paths */
3485 /* Assign version values. */
3486 relayd->data_sock.major = relayd_sock->major;
3487 relayd->data_sock.minor = relayd_sock->minor;
3488 break;
3489 default:
3490 ERR("Unknown relayd socket type (%d)", sock_type);
3491 ret_code = LTTCOMM_CONSUMERD_FATAL;
3492 goto error;
3493 }
3494
3495 DBG("Consumer %s socket created successfully with net idx %" PRIu64 " (fd: %d)",
3496 sock_type == LTTNG_STREAM_CONTROL ? "control" : "data",
3497 relayd->net_seq_idx, fd);
3498
3499 /* We successfully added the socket. Send status back. */
3500 ret = consumer_send_status_msg(sock, ret_code);
3501 if (ret < 0) {
3502 /* Somehow, the session daemon is not responding anymore. */
3503 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_FATAL);
3504 goto error_nosignal;
3505 }
3506
3507 /*
3508 * Add relayd socket pair to consumer data hashtable. If object already
3509 * exists or on error, the function gracefully returns.
3510 */
3511 add_relayd(relayd);
3512
3513 /* All good! */
3514 return;
3515
3516error:
3517 if (consumer_send_status_msg(sock, ret_code) < 0) {
3518 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_FATAL);
3519 }
3520
3521error_nosignal:
3522 /* Close received socket if valid. */
3523 if (fd >= 0) {
3524 if (close(fd)) {
3525 PERROR("close received socket");
3526 }
3527 }
3528
3529 if (relayd_created) {
3530 free(relayd);
3531 }
3532}
3533
3534/*
3535 * Try to lock the stream mutex.
3536 *
3537 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
3538 */
3539static int stream_try_lock(struct lttng_consumer_stream *stream)
3540{
3541 int ret;
3542
3543 assert(stream);
3544
3545 /*
3546 * Try to lock the stream mutex. On failure, we know that the stream is
3547 * being used else where hence there is data still being extracted.
3548 */
3549 ret = pthread_mutex_trylock(&stream->lock);
3550 if (ret) {
3551 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
3552 ret = 0;
3553 goto end;
3554 }
3555
3556 ret = 1;
3557
3558end:
3559 return ret;
3560}
3561
3562/*
3563 * Search for a relayd associated to the session id and return the reference.
3564 *
3565 * A rcu read side lock MUST be acquire before calling this function and locked
3566 * until the relayd object is no longer necessary.
3567 */
3568static struct consumer_relayd_sock_pair *find_relayd_by_session_id(uint64_t id)
3569{
3570 struct lttng_ht_iter iter;
3571 struct consumer_relayd_sock_pair *relayd = NULL;
3572
3573 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3574 cds_lfht_for_each_entry(consumer_data.relayd_ht->ht, &iter.iter, relayd,
3575 node.node) {
3576 /*
3577 * Check by sessiond id which is unique here where the relayd session
3578 * id might not be when having multiple relayd.
3579 */
3580 if (relayd->sessiond_session_id == id) {
3581 /* Found the relayd. There can be only one per id. */
3582 goto found;
3583 }
3584 }
3585
3586 return NULL;
3587
3588found:
3589 return relayd;
3590}
3591
3592/*
3593 * Check if for a given session id there is still data needed to be extract
3594 * from the buffers.
3595 *
3596 * Return 1 if data is pending or else 0 meaning ready to be read.
3597 */
3598int consumer_data_pending(uint64_t id)
3599{
3600 int ret;
3601 struct lttng_ht_iter iter;
3602 struct lttng_ht *ht;
3603 struct lttng_consumer_stream *stream;
3604 struct consumer_relayd_sock_pair *relayd = NULL;
3605 int (*data_pending)(struct lttng_consumer_stream *);
3606
3607 DBG("Consumer data pending command on session id %" PRIu64, id);
3608
3609 rcu_read_lock();
3610 pthread_mutex_lock(&consumer_data.lock);
3611
3612 switch (consumer_data.type) {
3613 case LTTNG_CONSUMER_KERNEL:
3614 data_pending = lttng_kconsumer_data_pending;
3615 break;
3616 case LTTNG_CONSUMER32_UST:
3617 case LTTNG_CONSUMER64_UST:
3618 data_pending = lttng_ustconsumer_data_pending;
3619 break;
3620 default:
3621 ERR("Unknown consumer data type");
3622 assert(0);
3623 }
3624
3625 /* Ease our life a bit */
3626 ht = consumer_data.stream_list_ht;
3627
3628 relayd = find_relayd_by_session_id(id);
3629 if (relayd) {
3630 /* Send init command for data pending. */
3631 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
3632 ret = relayd_begin_data_pending(&relayd->control_sock,
3633 relayd->relayd_session_id);
3634 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
3635 if (ret < 0) {
3636 /* Communication error thus the relayd so no data pending. */
3637 goto data_not_pending;
3638 }
3639 }
3640
3641 cds_lfht_for_each_entry_duplicate(ht->ht,
3642 ht->hash_fct(&id, lttng_ht_seed),
3643 ht->match_fct, &id,
3644 &iter.iter, stream, node_session_id.node) {
3645 /* If this call fails, the stream is being used hence data pending. */
3646 ret = stream_try_lock(stream);
3647 if (!ret) {
3648 goto data_pending;
3649 }
3650
3651 /*
3652 * A removed node from the hash table indicates that the stream has
3653 * been deleted thus having a guarantee that the buffers are closed
3654 * on the consumer side. However, data can still be transmitted
3655 * over the network so don't skip the relayd check.
3656 */
3657 ret = cds_lfht_is_node_deleted(&stream->node.node);
3658 if (!ret) {
3659 /* Check the stream if there is data in the buffers. */
3660 ret = data_pending(stream);
3661 if (ret == 1) {
3662 pthread_mutex_unlock(&stream->lock);
3663 goto data_pending;
3664 }
3665 }
3666
3667 /* Relayd check */
3668 if (relayd) {
3669 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
3670 if (stream->metadata_flag) {
3671 ret = relayd_quiescent_control(&relayd->control_sock,
3672 stream->relayd_stream_id);
3673 } else {
3674 ret = relayd_data_pending(&relayd->control_sock,
3675 stream->relayd_stream_id,
3676 stream->next_net_seq_num - 1);
3677 }
3678 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
3679 if (ret == 1) {
3680 pthread_mutex_unlock(&stream->lock);
3681 goto data_pending;
3682 }
3683 }
3684 pthread_mutex_unlock(&stream->lock);
3685 }
3686
3687 if (relayd) {
3688 unsigned int is_data_inflight = 0;
3689
3690 /* Send init command for data pending. */
3691 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
3692 ret = relayd_end_data_pending(&relayd->control_sock,
3693 relayd->relayd_session_id, &is_data_inflight);
3694 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
3695 if (ret < 0) {
3696 goto data_not_pending;
3697 }
3698 if (is_data_inflight) {
3699 goto data_pending;
3700 }
3701 }
3702
3703 /*
3704 * Finding _no_ node in the hash table and no inflight data means that the
3705 * stream(s) have been removed thus data is guaranteed to be available for
3706 * analysis from the trace files.
3707 */
3708
3709data_not_pending:
3710 /* Data is available to be read by a viewer. */
3711 pthread_mutex_unlock(&consumer_data.lock);
3712 rcu_read_unlock();
3713 return 0;
3714
3715data_pending:
3716 /* Data is still being extracted from buffers. */
3717 pthread_mutex_unlock(&consumer_data.lock);
3718 rcu_read_unlock();
3719 return 1;
3720}
3721
3722/*
3723 * Send a ret code status message to the sessiond daemon.
3724 *
3725 * Return the sendmsg() return value.
3726 */
3727int consumer_send_status_msg(int sock, int ret_code)
3728{
3729 struct lttcomm_consumer_status_msg msg;
3730
3731 memset(&msg, 0, sizeof(msg));
3732 msg.ret_code = ret_code;
3733
3734 return lttcomm_send_unix_sock(sock, &msg, sizeof(msg));
3735}
3736
3737/*
3738 * Send a channel status message to the sessiond daemon.
3739 *
3740 * Return the sendmsg() return value.
3741 */
3742int consumer_send_status_channel(int sock,
3743 struct lttng_consumer_channel *channel)
3744{
3745 struct lttcomm_consumer_status_channel msg;
3746
3747 assert(sock >= 0);
3748
3749 memset(&msg, 0, sizeof(msg));
3750 if (!channel) {
3751 msg.ret_code = LTTCOMM_CONSUMERD_CHANNEL_FAIL;
3752 } else {
3753 msg.ret_code = LTTCOMM_CONSUMERD_SUCCESS;
3754 msg.key = channel->key;
3755 msg.stream_count = channel->streams.count;
3756 }
3757
3758 return lttcomm_send_unix_sock(sock, &msg, sizeof(msg));
3759}
3760
3761unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos,
3762 unsigned long produced_pos, uint64_t nb_packets_per_stream,
3763 uint64_t max_sb_size)
3764{
3765 unsigned long start_pos;
3766
3767 if (!nb_packets_per_stream) {
3768 return consumed_pos; /* Grab everything */
3769 }
3770 start_pos = produced_pos - offset_align_floor(produced_pos, max_sb_size);
3771 start_pos -= max_sb_size * nb_packets_per_stream;
3772 if ((long) (start_pos - consumed_pos) < 0) {
3773 return consumed_pos; /* Grab everything */
3774 }
3775 return start_pos;
3776}
This page took 0.036345 seconds and 4 git commands to generate.