Pass the consumerd stream's trace archive id to the relayd
[lttng-tools.git] / src / common / consumer / consumer.c
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
52 struct lttng_consumer_global_data consumer_data = {
53 .stream_count = 0,
54 .need_update = 1,
55 .type = LTTNG_CONSUMER_UNKNOWN,
56 };
57
58 enum consumer_channel_action {
59 CONSUMER_CHANNEL_ADD,
60 CONSUMER_CHANNEL_DEL,
61 CONSUMER_CHANNEL_QUIT,
62 };
63
64 struct 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. */
71 int 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 */
79 int 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 */
86 static struct lttng_ht *metadata_ht;
87 static 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 */
94 static 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
103 static 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
113 static 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
132 void 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
138 static 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;
154 error:
155 return (int) ret;
156 }
157
158 /*
159 * Cleanup the stream list of a channel. Those streams are not yet globally
160 * visible
161 */
162 static 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 */
187 static 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
214 static 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 */
238 struct 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 */
266 static 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
284 static 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 */
308 static 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 */
332 void 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 */
358 void 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);
398 end:
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 */
407 static 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 */
431 static 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 */
467 static 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 */
506 void 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 */
525 void 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 */
534 void consumer_del_stream_for_data(struct lttng_consumer_stream *stream)
535 {
536 consumer_stream_destroy(stream, data_ht);
537 }
538
539 void consumer_del_stream_for_metadata(struct lttng_consumer_stream *stream)
540 {
541 consumer_stream_destroy(stream, metadata_ht);
542 }
543
544 void consumer_stream_update_channel_attributes(
545 struct lttng_consumer_stream *stream,
546 struct lttng_consumer_channel *channel)
547 {
548 stream->channel_read_only_attributes.tracefile_size =
549 channel->tracefile_size;
550 memcpy(stream->channel_read_only_attributes.path, channel->pathname,
551 sizeof(stream->channel_read_only_attributes.path));
552 }
553
554 struct lttng_consumer_stream *consumer_allocate_stream(uint64_t channel_key,
555 uint64_t stream_key,
556 enum lttng_consumer_stream_state state,
557 const char *channel_name,
558 uid_t uid,
559 gid_t gid,
560 uint64_t relayd_id,
561 uint64_t session_id,
562 int cpu,
563 int *alloc_ret,
564 enum consumer_channel_type type,
565 unsigned int monitor,
566 uint64_t trace_archive_id)
567 {
568 int ret;
569 struct lttng_consumer_stream *stream;
570
571 stream = zmalloc(sizeof(*stream));
572 if (stream == NULL) {
573 PERROR("malloc struct lttng_consumer_stream");
574 ret = -ENOMEM;
575 goto end;
576 }
577
578 rcu_read_lock();
579
580 stream->key = stream_key;
581 stream->out_fd = -1;
582 stream->out_fd_offset = 0;
583 stream->output_written = 0;
584 stream->state = state;
585 stream->uid = uid;
586 stream->gid = gid;
587 stream->net_seq_idx = relayd_id;
588 stream->session_id = session_id;
589 stream->monitor = monitor;
590 stream->endpoint_status = CONSUMER_ENDPOINT_ACTIVE;
591 stream->index_file = NULL;
592 stream->last_sequence_number = -1ULL;
593 stream->trace_archive_id = trace_archive_id;
594 pthread_mutex_init(&stream->lock, NULL);
595 pthread_mutex_init(&stream->metadata_timer_lock, NULL);
596
597 /* If channel is the metadata, flag this stream as metadata. */
598 if (type == CONSUMER_CHANNEL_TYPE_METADATA) {
599 stream->metadata_flag = 1;
600 /* Metadata is flat out. */
601 strncpy(stream->name, DEFAULT_METADATA_NAME, sizeof(stream->name));
602 /* Live rendez-vous point. */
603 pthread_cond_init(&stream->metadata_rdv, NULL);
604 pthread_mutex_init(&stream->metadata_rdv_lock, NULL);
605 } else {
606 /* Format stream name to <channel_name>_<cpu_number> */
607 ret = snprintf(stream->name, sizeof(stream->name), "%s_%d",
608 channel_name, cpu);
609 if (ret < 0) {
610 PERROR("snprintf stream name");
611 goto error;
612 }
613 }
614
615 /* Key is always the wait_fd for streams. */
616 lttng_ht_node_init_u64(&stream->node, stream->key);
617
618 /* Init node per channel id key */
619 lttng_ht_node_init_u64(&stream->node_channel_id, channel_key);
620
621 /* Init session id node with the stream session id */
622 lttng_ht_node_init_u64(&stream->node_session_id, stream->session_id);
623
624 DBG3("Allocated stream %s (key %" PRIu64 ", chan_key %" PRIu64
625 " relayd_id %" PRIu64 ", session_id %" PRIu64,
626 stream->name, stream->key, channel_key,
627 stream->net_seq_idx, stream->session_id);
628
629 rcu_read_unlock();
630 return stream;
631
632 error:
633 rcu_read_unlock();
634 free(stream);
635 end:
636 if (alloc_ret) {
637 *alloc_ret = ret;
638 }
639 return NULL;
640 }
641
642 /*
643 * Add a stream to the global list protected by a mutex.
644 */
645 void consumer_add_data_stream(struct lttng_consumer_stream *stream)
646 {
647 struct lttng_ht *ht = data_ht;
648
649 assert(stream);
650 assert(ht);
651
652 DBG3("Adding consumer stream %" PRIu64, stream->key);
653
654 pthread_mutex_lock(&consumer_data.lock);
655 pthread_mutex_lock(&stream->chan->lock);
656 pthread_mutex_lock(&stream->chan->timer_lock);
657 pthread_mutex_lock(&stream->lock);
658 rcu_read_lock();
659
660 /* Steal stream identifier to avoid having streams with the same key */
661 steal_stream_key(stream->key, ht);
662
663 lttng_ht_add_unique_u64(ht, &stream->node);
664
665 lttng_ht_add_u64(consumer_data.stream_per_chan_id_ht,
666 &stream->node_channel_id);
667
668 /*
669 * Add stream to the stream_list_ht of the consumer data. No need to steal
670 * the key since the HT does not use it and we allow to add redundant keys
671 * into this table.
672 */
673 lttng_ht_add_u64(consumer_data.stream_list_ht, &stream->node_session_id);
674
675 /*
676 * When nb_init_stream_left reaches 0, we don't need to trigger any action
677 * in terms of destroying the associated channel, because the action that
678 * causes the count to become 0 also causes a stream to be added. The
679 * channel deletion will thus be triggered by the following removal of this
680 * stream.
681 */
682 if (uatomic_read(&stream->chan->nb_init_stream_left) > 0) {
683 /* Increment refcount before decrementing nb_init_stream_left */
684 cmm_smp_wmb();
685 uatomic_dec(&stream->chan->nb_init_stream_left);
686 }
687
688 /* Update consumer data once the node is inserted. */
689 consumer_data.stream_count++;
690 consumer_data.need_update = 1;
691
692 rcu_read_unlock();
693 pthread_mutex_unlock(&stream->lock);
694 pthread_mutex_unlock(&stream->chan->timer_lock);
695 pthread_mutex_unlock(&stream->chan->lock);
696 pthread_mutex_unlock(&consumer_data.lock);
697 }
698
699 void consumer_del_data_stream(struct lttng_consumer_stream *stream)
700 {
701 consumer_del_stream(stream, data_ht);
702 }
703
704 /*
705 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
706 * be acquired before calling this.
707 */
708 static int add_relayd(struct consumer_relayd_sock_pair *relayd)
709 {
710 int ret = 0;
711 struct lttng_ht_node_u64 *node;
712 struct lttng_ht_iter iter;
713
714 assert(relayd);
715
716 lttng_ht_lookup(consumer_data.relayd_ht,
717 &relayd->net_seq_idx, &iter);
718 node = lttng_ht_iter_get_node_u64(&iter);
719 if (node != NULL) {
720 goto end;
721 }
722 lttng_ht_add_unique_u64(consumer_data.relayd_ht, &relayd->node);
723
724 end:
725 return ret;
726 }
727
728 /*
729 * Allocate and return a consumer relayd socket.
730 */
731 static struct consumer_relayd_sock_pair *consumer_allocate_relayd_sock_pair(
732 uint64_t net_seq_idx)
733 {
734 struct consumer_relayd_sock_pair *obj = NULL;
735
736 /* net sequence index of -1 is a failure */
737 if (net_seq_idx == (uint64_t) -1ULL) {
738 goto error;
739 }
740
741 obj = zmalloc(sizeof(struct consumer_relayd_sock_pair));
742 if (obj == NULL) {
743 PERROR("zmalloc relayd sock");
744 goto error;
745 }
746
747 obj->net_seq_idx = net_seq_idx;
748 obj->refcount = 0;
749 obj->destroy_flag = 0;
750 obj->control_sock.sock.fd = -1;
751 obj->data_sock.sock.fd = -1;
752 lttng_ht_node_init_u64(&obj->node, obj->net_seq_idx);
753 pthread_mutex_init(&obj->ctrl_sock_mutex, NULL);
754
755 error:
756 return obj;
757 }
758
759 /*
760 * Find a relayd socket pair in the global consumer data.
761 *
762 * Return the object if found else NULL.
763 * RCU read-side lock must be held across this call and while using the
764 * returned object.
765 */
766 struct consumer_relayd_sock_pair *consumer_find_relayd(uint64_t key)
767 {
768 struct lttng_ht_iter iter;
769 struct lttng_ht_node_u64 *node;
770 struct consumer_relayd_sock_pair *relayd = NULL;
771
772 /* Negative keys are lookup failures */
773 if (key == (uint64_t) -1ULL) {
774 goto error;
775 }
776
777 lttng_ht_lookup(consumer_data.relayd_ht, &key,
778 &iter);
779 node = lttng_ht_iter_get_node_u64(&iter);
780 if (node != NULL) {
781 relayd = caa_container_of(node, struct consumer_relayd_sock_pair, node);
782 }
783
784 error:
785 return relayd;
786 }
787
788 /*
789 * Find a relayd and send the stream
790 *
791 * Returns 0 on success, < 0 on error
792 */
793 int consumer_send_relayd_stream(struct lttng_consumer_stream *stream,
794 char *path)
795 {
796 int ret = 0;
797 struct consumer_relayd_sock_pair *relayd;
798
799 assert(stream);
800 assert(stream->net_seq_idx != -1ULL);
801 assert(path);
802
803 /* The stream is not metadata. Get relayd reference if exists. */
804 rcu_read_lock();
805 relayd = consumer_find_relayd(stream->net_seq_idx);
806 if (relayd != NULL) {
807 /* Add stream on the relayd */
808 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
809 ret = relayd_add_stream(&relayd->control_sock, stream->name,
810 path, &stream->relayd_stream_id,
811 stream->chan->tracefile_size, stream->chan->tracefile_count,
812 stream->trace_archive_id);
813 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
814 if (ret < 0) {
815 goto end;
816 }
817
818 uatomic_inc(&relayd->refcount);
819 stream->sent_to_relayd = 1;
820 } else {
821 ERR("Stream %" PRIu64 " relayd ID %" PRIu64 " unknown. Can't send it.",
822 stream->key, stream->net_seq_idx);
823 ret = -1;
824 goto end;
825 }
826
827 DBG("Stream %s with key %" PRIu64 " sent to relayd id %" PRIu64,
828 stream->name, stream->key, stream->net_seq_idx);
829
830 end:
831 rcu_read_unlock();
832 return ret;
833 }
834
835 /*
836 * Find a relayd and send the streams sent message
837 *
838 * Returns 0 on success, < 0 on error
839 */
840 int consumer_send_relayd_streams_sent(uint64_t net_seq_idx)
841 {
842 int ret = 0;
843 struct consumer_relayd_sock_pair *relayd;
844
845 assert(net_seq_idx != -1ULL);
846
847 /* The stream is not metadata. Get relayd reference if exists. */
848 rcu_read_lock();
849 relayd = consumer_find_relayd(net_seq_idx);
850 if (relayd != NULL) {
851 /* Add stream on the relayd */
852 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
853 ret = relayd_streams_sent(&relayd->control_sock);
854 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
855 if (ret < 0) {
856 goto end;
857 }
858 } else {
859 ERR("Relayd ID %" PRIu64 " unknown. Can't send streams_sent.",
860 net_seq_idx);
861 ret = -1;
862 goto end;
863 }
864
865 ret = 0;
866 DBG("All streams sent relayd id %" PRIu64, net_seq_idx);
867
868 end:
869 rcu_read_unlock();
870 return ret;
871 }
872
873 /*
874 * Find a relayd and close the stream
875 */
876 void close_relayd_stream(struct lttng_consumer_stream *stream)
877 {
878 struct consumer_relayd_sock_pair *relayd;
879
880 /* The stream is not metadata. Get relayd reference if exists. */
881 rcu_read_lock();
882 relayd = consumer_find_relayd(stream->net_seq_idx);
883 if (relayd) {
884 consumer_stream_relayd_close(stream, relayd);
885 }
886 rcu_read_unlock();
887 }
888
889 /*
890 * Handle stream for relayd transmission if the stream applies for network
891 * streaming where the net sequence index is set.
892 *
893 * Return destination file descriptor or negative value on error.
894 */
895 static int write_relayd_stream_header(struct lttng_consumer_stream *stream,
896 size_t data_size, unsigned long padding,
897 struct consumer_relayd_sock_pair *relayd)
898 {
899 int outfd = -1, ret;
900 struct lttcomm_relayd_data_hdr data_hdr;
901
902 /* Safety net */
903 assert(stream);
904 assert(relayd);
905
906 /* Reset data header */
907 memset(&data_hdr, 0, sizeof(data_hdr));
908
909 if (stream->metadata_flag) {
910 /* Caller MUST acquire the relayd control socket lock */
911 ret = relayd_send_metadata(&relayd->control_sock, data_size);
912 if (ret < 0) {
913 goto error;
914 }
915
916 /* Metadata are always sent on the control socket. */
917 outfd = relayd->control_sock.sock.fd;
918 } else {
919 /* Set header with stream information */
920 data_hdr.stream_id = htobe64(stream->relayd_stream_id);
921 data_hdr.data_size = htobe32(data_size);
922 data_hdr.padding_size = htobe32(padding);
923 /*
924 * Note that net_seq_num below is assigned with the *current* value of
925 * next_net_seq_num and only after that the next_net_seq_num will be
926 * increment. This is why when issuing a command on the relayd using
927 * this next value, 1 should always be substracted in order to compare
928 * the last seen sequence number on the relayd side to the last sent.
929 */
930 data_hdr.net_seq_num = htobe64(stream->next_net_seq_num);
931 /* Other fields are zeroed previously */
932
933 ret = relayd_send_data_hdr(&relayd->data_sock, &data_hdr,
934 sizeof(data_hdr));
935 if (ret < 0) {
936 goto error;
937 }
938
939 ++stream->next_net_seq_num;
940
941 /* Set to go on data socket */
942 outfd = relayd->data_sock.sock.fd;
943 }
944
945 error:
946 return outfd;
947 }
948
949 /*
950 * Allocate and return a new lttng_consumer_channel object using the given key
951 * to initialize the hash table node.
952 *
953 * On error, return NULL.
954 */
955 struct lttng_consumer_channel *consumer_allocate_channel(uint64_t key,
956 uint64_t session_id,
957 const char *pathname,
958 const char *name,
959 uid_t uid,
960 gid_t gid,
961 uint64_t relayd_id,
962 enum lttng_event_output output,
963 uint64_t tracefile_size,
964 uint64_t tracefile_count,
965 uint64_t session_id_per_pid,
966 unsigned int monitor,
967 unsigned int live_timer_interval,
968 const char *root_shm_path,
969 const char *shm_path)
970 {
971 struct lttng_consumer_channel *channel;
972
973 channel = zmalloc(sizeof(*channel));
974 if (channel == NULL) {
975 PERROR("malloc struct lttng_consumer_channel");
976 goto end;
977 }
978
979 channel->key = key;
980 channel->refcount = 0;
981 channel->session_id = session_id;
982 channel->session_id_per_pid = session_id_per_pid;
983 channel->uid = uid;
984 channel->gid = gid;
985 channel->relayd_id = relayd_id;
986 channel->tracefile_size = tracefile_size;
987 channel->tracefile_count = tracefile_count;
988 channel->monitor = monitor;
989 channel->live_timer_interval = live_timer_interval;
990 pthread_mutex_init(&channel->lock, NULL);
991 pthread_mutex_init(&channel->timer_lock, NULL);
992
993 switch (output) {
994 case LTTNG_EVENT_SPLICE:
995 channel->output = CONSUMER_CHANNEL_SPLICE;
996 break;
997 case LTTNG_EVENT_MMAP:
998 channel->output = CONSUMER_CHANNEL_MMAP;
999 break;
1000 default:
1001 assert(0);
1002 free(channel);
1003 channel = NULL;
1004 goto end;
1005 }
1006
1007 /*
1008 * In monitor mode, the streams associated with the channel will be put in
1009 * a special list ONLY owned by this channel. So, the refcount is set to 1
1010 * here meaning that the channel itself has streams that are referenced.
1011 *
1012 * On a channel deletion, once the channel is no longer visible, the
1013 * refcount is decremented and checked for a zero value to delete it. With
1014 * streams in no monitor mode, it will now be safe to destroy the channel.
1015 */
1016 if (!channel->monitor) {
1017 channel->refcount = 1;
1018 }
1019
1020 strncpy(channel->pathname, pathname, sizeof(channel->pathname));
1021 channel->pathname[sizeof(channel->pathname) - 1] = '\0';
1022
1023 strncpy(channel->name, name, sizeof(channel->name));
1024 channel->name[sizeof(channel->name) - 1] = '\0';
1025
1026 if (root_shm_path) {
1027 strncpy(channel->root_shm_path, root_shm_path, sizeof(channel->root_shm_path));
1028 channel->root_shm_path[sizeof(channel->root_shm_path) - 1] = '\0';
1029 }
1030 if (shm_path) {
1031 strncpy(channel->shm_path, shm_path, sizeof(channel->shm_path));
1032 channel->shm_path[sizeof(channel->shm_path) - 1] = '\0';
1033 }
1034
1035 lttng_ht_node_init_u64(&channel->node, channel->key);
1036
1037 channel->wait_fd = -1;
1038
1039 CDS_INIT_LIST_HEAD(&channel->streams.head);
1040
1041 DBG("Allocated channel (key %" PRIu64 ")", channel->key);
1042
1043 end:
1044 return channel;
1045 }
1046
1047 /*
1048 * Add a channel to the global list protected by a mutex.
1049 *
1050 * Always return 0 indicating success.
1051 */
1052 int consumer_add_channel(struct lttng_consumer_channel *channel,
1053 struct lttng_consumer_local_data *ctx)
1054 {
1055 pthread_mutex_lock(&consumer_data.lock);
1056 pthread_mutex_lock(&channel->lock);
1057 pthread_mutex_lock(&channel->timer_lock);
1058
1059 /*
1060 * This gives us a guarantee that the channel we are about to add to the
1061 * channel hash table will be unique. See this function comment on the why
1062 * we need to steel the channel key at this stage.
1063 */
1064 steal_channel_key(channel->key);
1065
1066 rcu_read_lock();
1067 lttng_ht_add_unique_u64(consumer_data.channel_ht, &channel->node);
1068 rcu_read_unlock();
1069
1070 pthread_mutex_unlock(&channel->timer_lock);
1071 pthread_mutex_unlock(&channel->lock);
1072 pthread_mutex_unlock(&consumer_data.lock);
1073
1074 if (channel->wait_fd != -1 && channel->type == CONSUMER_CHANNEL_TYPE_DATA) {
1075 notify_channel_pipe(ctx, channel, -1, CONSUMER_CHANNEL_ADD);
1076 }
1077
1078 return 0;
1079 }
1080
1081 /*
1082 * Allocate the pollfd structure and the local view of the out fds to avoid
1083 * doing a lookup in the linked list and concurrency issues when writing is
1084 * needed. Called with consumer_data.lock held.
1085 *
1086 * Returns the number of fds in the structures.
1087 */
1088 static int update_poll_array(struct lttng_consumer_local_data *ctx,
1089 struct pollfd **pollfd, struct lttng_consumer_stream **local_stream,
1090 struct lttng_ht *ht, int *nb_inactive_fd)
1091 {
1092 int i = 0;
1093 struct lttng_ht_iter iter;
1094 struct lttng_consumer_stream *stream;
1095
1096 assert(ctx);
1097 assert(ht);
1098 assert(pollfd);
1099 assert(local_stream);
1100
1101 DBG("Updating poll fd array");
1102 *nb_inactive_fd = 0;
1103 rcu_read_lock();
1104 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
1105 /*
1106 * Only active streams with an active end point can be added to the
1107 * poll set and local stream storage of the thread.
1108 *
1109 * There is a potential race here for endpoint_status to be updated
1110 * just after the check. However, this is OK since the stream(s) will
1111 * be deleted once the thread is notified that the end point state has
1112 * changed where this function will be called back again.
1113 *
1114 * We track the number of inactive FDs because they still need to be
1115 * closed by the polling thread after a wakeup on the data_pipe or
1116 * metadata_pipe.
1117 */
1118 if (stream->state != LTTNG_CONSUMER_ACTIVE_STREAM ||
1119 stream->endpoint_status == CONSUMER_ENDPOINT_INACTIVE) {
1120 (*nb_inactive_fd)++;
1121 continue;
1122 }
1123 /*
1124 * This clobbers way too much the debug output. Uncomment that if you
1125 * need it for debugging purposes.
1126 *
1127 * DBG("Active FD %d", stream->wait_fd);
1128 */
1129 (*pollfd)[i].fd = stream->wait_fd;
1130 (*pollfd)[i].events = POLLIN | POLLPRI;
1131 local_stream[i] = stream;
1132 i++;
1133 }
1134 rcu_read_unlock();
1135
1136 /*
1137 * Insert the consumer_data_pipe at the end of the array and don't
1138 * increment i so nb_fd is the number of real FD.
1139 */
1140 (*pollfd)[i].fd = lttng_pipe_get_readfd(ctx->consumer_data_pipe);
1141 (*pollfd)[i].events = POLLIN | POLLPRI;
1142
1143 (*pollfd)[i + 1].fd = lttng_pipe_get_readfd(ctx->consumer_wakeup_pipe);
1144 (*pollfd)[i + 1].events = POLLIN | POLLPRI;
1145 return i;
1146 }
1147
1148 /*
1149 * Poll on the should_quit pipe and the command socket return -1 on
1150 * error, 1 if should exit, 0 if data is available on the command socket
1151 */
1152 int lttng_consumer_poll_socket(struct pollfd *consumer_sockpoll)
1153 {
1154 int num_rdy;
1155
1156 restart:
1157 num_rdy = poll(consumer_sockpoll, 2, -1);
1158 if (num_rdy == -1) {
1159 /*
1160 * Restart interrupted system call.
1161 */
1162 if (errno == EINTR) {
1163 goto restart;
1164 }
1165 PERROR("Poll error");
1166 return -1;
1167 }
1168 if (consumer_sockpoll[0].revents & (POLLIN | POLLPRI)) {
1169 DBG("consumer_should_quit wake up");
1170 return 1;
1171 }
1172 return 0;
1173 }
1174
1175 /*
1176 * Set the error socket.
1177 */
1178 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data *ctx,
1179 int sock)
1180 {
1181 ctx->consumer_error_socket = sock;
1182 }
1183
1184 /*
1185 * Set the command socket path.
1186 */
1187 void lttng_consumer_set_command_sock_path(
1188 struct lttng_consumer_local_data *ctx, char *sock)
1189 {
1190 ctx->consumer_command_sock_path = sock;
1191 }
1192
1193 /*
1194 * Send return code to the session daemon.
1195 * If the socket is not defined, we return 0, it is not a fatal error
1196 */
1197 int lttng_consumer_send_error(struct lttng_consumer_local_data *ctx, int cmd)
1198 {
1199 if (ctx->consumer_error_socket > 0) {
1200 return lttcomm_send_unix_sock(ctx->consumer_error_socket, &cmd,
1201 sizeof(enum lttcomm_sessiond_command));
1202 }
1203
1204 return 0;
1205 }
1206
1207 /*
1208 * Close all the tracefiles and stream fds and MUST be called when all
1209 * instances are destroyed i.e. when all threads were joined and are ended.
1210 */
1211 void lttng_consumer_cleanup(void)
1212 {
1213 struct lttng_ht_iter iter;
1214 struct lttng_consumer_channel *channel;
1215
1216 rcu_read_lock();
1217
1218 cds_lfht_for_each_entry(consumer_data.channel_ht->ht, &iter.iter, channel,
1219 node.node) {
1220 consumer_del_channel(channel);
1221 }
1222
1223 rcu_read_unlock();
1224
1225 lttng_ht_destroy(consumer_data.channel_ht);
1226
1227 cleanup_relayd_ht();
1228
1229 lttng_ht_destroy(consumer_data.stream_per_chan_id_ht);
1230
1231 /*
1232 * This HT contains streams that are freed by either the metadata thread or
1233 * the data thread so we do *nothing* on the hash table and simply destroy
1234 * it.
1235 */
1236 lttng_ht_destroy(consumer_data.stream_list_ht);
1237 }
1238
1239 /*
1240 * Called from signal handler.
1241 */
1242 void lttng_consumer_should_exit(struct lttng_consumer_local_data *ctx)
1243 {
1244 ssize_t ret;
1245
1246 CMM_STORE_SHARED(consumer_quit, 1);
1247 ret = lttng_write(ctx->consumer_should_quit[1], "4", 1);
1248 if (ret < 1) {
1249 PERROR("write consumer quit");
1250 }
1251
1252 DBG("Consumer flag that it should quit");
1253 }
1254
1255
1256 /*
1257 * Flush pending writes to trace output disk file.
1258 */
1259 static
1260 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream *stream,
1261 off_t orig_offset)
1262 {
1263 int ret;
1264 int outfd = stream->out_fd;
1265
1266 /*
1267 * This does a blocking write-and-wait on any page that belongs to the
1268 * subbuffer prior to the one we just wrote.
1269 * Don't care about error values, as these are just hints and ways to
1270 * limit the amount of page cache used.
1271 */
1272 if (orig_offset < stream->max_sb_size) {
1273 return;
1274 }
1275 lttng_sync_file_range(outfd, orig_offset - stream->max_sb_size,
1276 stream->max_sb_size,
1277 SYNC_FILE_RANGE_WAIT_BEFORE
1278 | SYNC_FILE_RANGE_WRITE
1279 | SYNC_FILE_RANGE_WAIT_AFTER);
1280 /*
1281 * Give hints to the kernel about how we access the file:
1282 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1283 * we write it.
1284 *
1285 * We need to call fadvise again after the file grows because the
1286 * kernel does not seem to apply fadvise to non-existing parts of the
1287 * file.
1288 *
1289 * Call fadvise _after_ having waited for the page writeback to
1290 * complete because the dirty page writeback semantic is not well
1291 * defined. So it can be expected to lead to lower throughput in
1292 * streaming.
1293 */
1294 ret = posix_fadvise(outfd, orig_offset - stream->max_sb_size,
1295 stream->max_sb_size, POSIX_FADV_DONTNEED);
1296 if (ret && ret != -ENOSYS) {
1297 errno = ret;
1298 PERROR("posix_fadvise on fd %i", outfd);
1299 }
1300 }
1301
1302 /*
1303 * Initialise the necessary environnement :
1304 * - create a new context
1305 * - create the poll_pipe
1306 * - create the should_quit pipe (for signal handler)
1307 * - create the thread pipe (for splice)
1308 *
1309 * Takes a function pointer as argument, this function is called when data is
1310 * available on a buffer. This function is responsible to do the
1311 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1312 * buffer configuration and then kernctl_put_next_subbuf at the end.
1313 *
1314 * Returns a pointer to the new context or NULL on error.
1315 */
1316 struct lttng_consumer_local_data *lttng_consumer_create(
1317 enum lttng_consumer_type type,
1318 ssize_t (*buffer_ready)(struct lttng_consumer_stream *stream,
1319 struct lttng_consumer_local_data *ctx),
1320 int (*recv_channel)(struct lttng_consumer_channel *channel),
1321 int (*recv_stream)(struct lttng_consumer_stream *stream),
1322 int (*update_stream)(uint64_t stream_key, uint32_t state))
1323 {
1324 int ret;
1325 struct lttng_consumer_local_data *ctx;
1326
1327 assert(consumer_data.type == LTTNG_CONSUMER_UNKNOWN ||
1328 consumer_data.type == type);
1329 consumer_data.type = type;
1330
1331 ctx = zmalloc(sizeof(struct lttng_consumer_local_data));
1332 if (ctx == NULL) {
1333 PERROR("allocating context");
1334 goto error;
1335 }
1336
1337 ctx->consumer_error_socket = -1;
1338 ctx->consumer_metadata_socket = -1;
1339 pthread_mutex_init(&ctx->metadata_socket_lock, NULL);
1340 /* assign the callbacks */
1341 ctx->on_buffer_ready = buffer_ready;
1342 ctx->on_recv_channel = recv_channel;
1343 ctx->on_recv_stream = recv_stream;
1344 ctx->on_update_stream = update_stream;
1345
1346 ctx->consumer_data_pipe = lttng_pipe_open(0);
1347 if (!ctx->consumer_data_pipe) {
1348 goto error_poll_pipe;
1349 }
1350
1351 ctx->consumer_wakeup_pipe = lttng_pipe_open(0);
1352 if (!ctx->consumer_wakeup_pipe) {
1353 goto error_wakeup_pipe;
1354 }
1355
1356 ret = pipe(ctx->consumer_should_quit);
1357 if (ret < 0) {
1358 PERROR("Error creating recv pipe");
1359 goto error_quit_pipe;
1360 }
1361
1362 ret = pipe(ctx->consumer_channel_pipe);
1363 if (ret < 0) {
1364 PERROR("Error creating channel pipe");
1365 goto error_channel_pipe;
1366 }
1367
1368 ctx->consumer_metadata_pipe = lttng_pipe_open(0);
1369 if (!ctx->consumer_metadata_pipe) {
1370 goto error_metadata_pipe;
1371 }
1372
1373 ctx->channel_monitor_pipe = -1;
1374
1375 return ctx;
1376
1377 error_metadata_pipe:
1378 utils_close_pipe(ctx->consumer_channel_pipe);
1379 error_channel_pipe:
1380 utils_close_pipe(ctx->consumer_should_quit);
1381 error_quit_pipe:
1382 lttng_pipe_destroy(ctx->consumer_wakeup_pipe);
1383 error_wakeup_pipe:
1384 lttng_pipe_destroy(ctx->consumer_data_pipe);
1385 error_poll_pipe:
1386 free(ctx);
1387 error:
1388 return NULL;
1389 }
1390
1391 /*
1392 * Iterate over all streams of the hashtable and free them properly.
1393 */
1394 static void destroy_data_stream_ht(struct lttng_ht *ht)
1395 {
1396 struct lttng_ht_iter iter;
1397 struct lttng_consumer_stream *stream;
1398
1399 if (ht == NULL) {
1400 return;
1401 }
1402
1403 rcu_read_lock();
1404 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
1405 /*
1406 * Ignore return value since we are currently cleaning up so any error
1407 * can't be handled.
1408 */
1409 (void) consumer_del_stream(stream, ht);
1410 }
1411 rcu_read_unlock();
1412
1413 lttng_ht_destroy(ht);
1414 }
1415
1416 /*
1417 * Iterate over all streams of the metadata hashtable and free them
1418 * properly.
1419 */
1420 static void destroy_metadata_stream_ht(struct lttng_ht *ht)
1421 {
1422 struct lttng_ht_iter iter;
1423 struct lttng_consumer_stream *stream;
1424
1425 if (ht == NULL) {
1426 return;
1427 }
1428
1429 rcu_read_lock();
1430 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
1431 /*
1432 * Ignore return value since we are currently cleaning up so any error
1433 * can't be handled.
1434 */
1435 (void) consumer_del_metadata_stream(stream, ht);
1436 }
1437 rcu_read_unlock();
1438
1439 lttng_ht_destroy(ht);
1440 }
1441
1442 /*
1443 * Close all fds associated with the instance and free the context.
1444 */
1445 void lttng_consumer_destroy(struct lttng_consumer_local_data *ctx)
1446 {
1447 int ret;
1448
1449 DBG("Consumer destroying it. Closing everything.");
1450
1451 if (!ctx) {
1452 return;
1453 }
1454
1455 destroy_data_stream_ht(data_ht);
1456 destroy_metadata_stream_ht(metadata_ht);
1457
1458 ret = close(ctx->consumer_error_socket);
1459 if (ret) {
1460 PERROR("close");
1461 }
1462 ret = close(ctx->consumer_metadata_socket);
1463 if (ret) {
1464 PERROR("close");
1465 }
1466 utils_close_pipe(ctx->consumer_channel_pipe);
1467 lttng_pipe_destroy(ctx->consumer_data_pipe);
1468 lttng_pipe_destroy(ctx->consumer_metadata_pipe);
1469 lttng_pipe_destroy(ctx->consumer_wakeup_pipe);
1470 utils_close_pipe(ctx->consumer_should_quit);
1471
1472 unlink(ctx->consumer_command_sock_path);
1473 free(ctx);
1474 }
1475
1476 /*
1477 * Write the metadata stream id on the specified file descriptor.
1478 */
1479 static int write_relayd_metadata_id(int fd,
1480 struct lttng_consumer_stream *stream,
1481 unsigned long padding)
1482 {
1483 ssize_t ret;
1484 struct lttcomm_relayd_metadata_payload hdr;
1485
1486 hdr.stream_id = htobe64(stream->relayd_stream_id);
1487 hdr.padding_size = htobe32(padding);
1488 ret = lttng_write(fd, (void *) &hdr, sizeof(hdr));
1489 if (ret < sizeof(hdr)) {
1490 /*
1491 * This error means that the fd's end is closed so ignore the PERROR
1492 * not to clubber the error output since this can happen in a normal
1493 * code path.
1494 */
1495 if (errno != EPIPE) {
1496 PERROR("write metadata stream id");
1497 }
1498 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno);
1499 /*
1500 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1501 * handle writting the missing part so report that as an error and
1502 * don't lie to the caller.
1503 */
1504 ret = -1;
1505 goto end;
1506 }
1507 DBG("Metadata stream id %" PRIu64 " with padding %lu written before data",
1508 stream->relayd_stream_id, padding);
1509
1510 end:
1511 return (int) ret;
1512 }
1513
1514 /*
1515 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1516 * core function for writing trace buffers to either the local filesystem or
1517 * the network.
1518 *
1519 * It must be called with the stream lock held.
1520 *
1521 * Careful review MUST be put if any changes occur!
1522 *
1523 * Returns the number of bytes written
1524 */
1525 ssize_t lttng_consumer_on_read_subbuffer_mmap(
1526 struct lttng_consumer_local_data *ctx,
1527 struct lttng_consumer_stream *stream, unsigned long len,
1528 unsigned long padding,
1529 struct ctf_packet_index *index)
1530 {
1531 unsigned long mmap_offset;
1532 void *mmap_base;
1533 ssize_t ret = 0;
1534 off_t orig_offset = stream->out_fd_offset;
1535 /* Default is on the disk */
1536 int outfd = stream->out_fd;
1537 struct consumer_relayd_sock_pair *relayd = NULL;
1538 unsigned int relayd_hang_up = 0;
1539
1540 /* RCU lock for the relayd pointer */
1541 rcu_read_lock();
1542
1543 /* Flag that the current stream if set for network streaming. */
1544 if (stream->net_seq_idx != (uint64_t) -1ULL) {
1545 relayd = consumer_find_relayd(stream->net_seq_idx);
1546 if (relayd == NULL) {
1547 ret = -EPIPE;
1548 goto end;
1549 }
1550 }
1551
1552 /* get the offset inside the fd to mmap */
1553 switch (consumer_data.type) {
1554 case LTTNG_CONSUMER_KERNEL:
1555 mmap_base = stream->mmap_base;
1556 ret = kernctl_get_mmap_read_offset(stream->wait_fd, &mmap_offset);
1557 if (ret < 0) {
1558 PERROR("tracer ctl get_mmap_read_offset");
1559 goto end;
1560 }
1561 break;
1562 case LTTNG_CONSUMER32_UST:
1563 case LTTNG_CONSUMER64_UST:
1564 mmap_base = lttng_ustctl_get_mmap_base(stream);
1565 if (!mmap_base) {
1566 ERR("read mmap get mmap base for stream %s", stream->name);
1567 ret = -EPERM;
1568 goto end;
1569 }
1570 ret = lttng_ustctl_get_mmap_read_offset(stream, &mmap_offset);
1571 if (ret != 0) {
1572 PERROR("tracer ctl get_mmap_read_offset");
1573 ret = -EINVAL;
1574 goto end;
1575 }
1576 break;
1577 default:
1578 ERR("Unknown consumer_data type");
1579 assert(0);
1580 }
1581
1582 /* Handle stream on the relayd if the output is on the network */
1583 if (relayd) {
1584 unsigned long netlen = len;
1585
1586 /*
1587 * Lock the control socket for the complete duration of the function
1588 * since from this point on we will use the socket.
1589 */
1590 if (stream->metadata_flag) {
1591 /* Metadata requires the control socket. */
1592 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
1593 if (stream->reset_metadata_flag) {
1594 ret = relayd_reset_metadata(&relayd->control_sock,
1595 stream->relayd_stream_id,
1596 stream->metadata_version);
1597 if (ret < 0) {
1598 relayd_hang_up = 1;
1599 goto write_error;
1600 }
1601 stream->reset_metadata_flag = 0;
1602 }
1603 netlen += sizeof(struct lttcomm_relayd_metadata_payload);
1604 }
1605
1606 ret = write_relayd_stream_header(stream, netlen, padding, relayd);
1607 if (ret < 0) {
1608 relayd_hang_up = 1;
1609 goto write_error;
1610 }
1611 /* Use the returned socket. */
1612 outfd = ret;
1613
1614 /* Write metadata stream id before payload */
1615 if (stream->metadata_flag) {
1616 ret = write_relayd_metadata_id(outfd, stream, padding);
1617 if (ret < 0) {
1618 relayd_hang_up = 1;
1619 goto write_error;
1620 }
1621 }
1622 } else {
1623 /* No streaming, we have to set the len with the full padding */
1624 len += padding;
1625
1626 if (stream->metadata_flag && stream->reset_metadata_flag) {
1627 ret = utils_truncate_stream_file(stream->out_fd, 0);
1628 if (ret < 0) {
1629 ERR("Reset metadata file");
1630 goto end;
1631 }
1632 stream->reset_metadata_flag = 0;
1633 }
1634
1635 /*
1636 * Check if we need to change the tracefile before writing the packet.
1637 */
1638 if (stream->chan->tracefile_size > 0 &&
1639 (stream->tracefile_size_current + len) >
1640 stream->chan->tracefile_size) {
1641 ret = utils_rotate_stream_file(stream->chan->pathname,
1642 stream->name, stream->chan->tracefile_size,
1643 stream->chan->tracefile_count, stream->uid, stream->gid,
1644 stream->out_fd, &(stream->tracefile_count_current),
1645 &stream->out_fd);
1646 if (ret < 0) {
1647 ERR("Rotating output file");
1648 goto end;
1649 }
1650 outfd = stream->out_fd;
1651
1652 if (stream->index_file) {
1653 lttng_index_file_put(stream->index_file);
1654 stream->index_file = lttng_index_file_create(stream->chan->pathname,
1655 stream->name, stream->uid, stream->gid,
1656 stream->chan->tracefile_size,
1657 stream->tracefile_count_current,
1658 CTF_INDEX_MAJOR, CTF_INDEX_MINOR);
1659 if (!stream->index_file) {
1660 goto end;
1661 }
1662 }
1663
1664 /* Reset current size because we just perform a rotation. */
1665 stream->tracefile_size_current = 0;
1666 stream->out_fd_offset = 0;
1667 orig_offset = 0;
1668 }
1669 stream->tracefile_size_current += len;
1670 if (index) {
1671 index->offset = htobe64(stream->out_fd_offset);
1672 }
1673 }
1674
1675 /*
1676 * This call guarantee that len or less is returned. It's impossible to
1677 * receive a ret value that is bigger than len.
1678 */
1679 ret = lttng_write(outfd, mmap_base + mmap_offset, len);
1680 DBG("Consumer mmap write() ret %zd (len %lu)", ret, len);
1681 if (ret < 0 || ((size_t) ret != len)) {
1682 /*
1683 * Report error to caller if nothing was written else at least send the
1684 * amount written.
1685 */
1686 if (ret < 0) {
1687 ret = -errno;
1688 }
1689 relayd_hang_up = 1;
1690
1691 /* Socket operation failed. We consider the relayd dead */
1692 if (errno == EPIPE || errno == EINVAL || errno == EBADF) {
1693 /*
1694 * This is possible if the fd is closed on the other side
1695 * (outfd) or any write problem. It can be verbose a bit for a
1696 * normal execution if for instance the relayd is stopped
1697 * abruptly. This can happen so set this to a DBG statement.
1698 */
1699 DBG("Consumer mmap write detected relayd hang up");
1700 } else {
1701 /* Unhandled error, print it and stop function right now. */
1702 PERROR("Error in write mmap (ret %zd != len %lu)", ret, len);
1703 }
1704 goto write_error;
1705 }
1706 stream->output_written += ret;
1707
1708 /* This call is useless on a socket so better save a syscall. */
1709 if (!relayd) {
1710 /* This won't block, but will start writeout asynchronously */
1711 lttng_sync_file_range(outfd, stream->out_fd_offset, len,
1712 SYNC_FILE_RANGE_WRITE);
1713 stream->out_fd_offset += len;
1714 lttng_consumer_sync_trace_file(stream, orig_offset);
1715 }
1716
1717 write_error:
1718 /*
1719 * This is a special case that the relayd has closed its socket. Let's
1720 * cleanup the relayd object and all associated streams.
1721 */
1722 if (relayd && relayd_hang_up) {
1723 cleanup_relayd(relayd, ctx);
1724 }
1725
1726 end:
1727 /* Unlock only if ctrl socket used */
1728 if (relayd && stream->metadata_flag) {
1729 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
1730 }
1731
1732 rcu_read_unlock();
1733 return ret;
1734 }
1735
1736 /*
1737 * Splice the data from the ring buffer to the tracefile.
1738 *
1739 * It must be called with the stream lock held.
1740 *
1741 * Returns the number of bytes spliced.
1742 */
1743 ssize_t lttng_consumer_on_read_subbuffer_splice(
1744 struct lttng_consumer_local_data *ctx,
1745 struct lttng_consumer_stream *stream, unsigned long len,
1746 unsigned long padding,
1747 struct ctf_packet_index *index)
1748 {
1749 ssize_t ret = 0, written = 0, ret_splice = 0;
1750 loff_t offset = 0;
1751 off_t orig_offset = stream->out_fd_offset;
1752 int fd = stream->wait_fd;
1753 /* Default is on the disk */
1754 int outfd = stream->out_fd;
1755 struct consumer_relayd_sock_pair *relayd = NULL;
1756 int *splice_pipe;
1757 unsigned int relayd_hang_up = 0;
1758
1759 switch (consumer_data.type) {
1760 case LTTNG_CONSUMER_KERNEL:
1761 break;
1762 case LTTNG_CONSUMER32_UST:
1763 case LTTNG_CONSUMER64_UST:
1764 /* Not supported for user space tracing */
1765 return -ENOSYS;
1766 default:
1767 ERR("Unknown consumer_data type");
1768 assert(0);
1769 }
1770
1771 /* RCU lock for the relayd pointer */
1772 rcu_read_lock();
1773
1774 /* Flag that the current stream if set for network streaming. */
1775 if (stream->net_seq_idx != (uint64_t) -1ULL) {
1776 relayd = consumer_find_relayd(stream->net_seq_idx);
1777 if (relayd == NULL) {
1778 written = -ret;
1779 goto end;
1780 }
1781 }
1782 splice_pipe = stream->splice_pipe;
1783
1784 /* Write metadata stream id before payload */
1785 if (relayd) {
1786 unsigned long total_len = len;
1787
1788 if (stream->metadata_flag) {
1789 /*
1790 * Lock the control socket for the complete duration of the function
1791 * since from this point on we will use the socket.
1792 */
1793 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
1794
1795 if (stream->reset_metadata_flag) {
1796 ret = relayd_reset_metadata(&relayd->control_sock,
1797 stream->relayd_stream_id,
1798 stream->metadata_version);
1799 if (ret < 0) {
1800 relayd_hang_up = 1;
1801 goto write_error;
1802 }
1803 stream->reset_metadata_flag = 0;
1804 }
1805 ret = write_relayd_metadata_id(splice_pipe[1], stream,
1806 padding);
1807 if (ret < 0) {
1808 written = ret;
1809 relayd_hang_up = 1;
1810 goto write_error;
1811 }
1812
1813 total_len += sizeof(struct lttcomm_relayd_metadata_payload);
1814 }
1815
1816 ret = write_relayd_stream_header(stream, total_len, padding, relayd);
1817 if (ret < 0) {
1818 written = ret;
1819 relayd_hang_up = 1;
1820 goto write_error;
1821 }
1822 /* Use the returned socket. */
1823 outfd = ret;
1824 } else {
1825 /* No streaming, we have to set the len with the full padding */
1826 len += padding;
1827
1828 if (stream->metadata_flag && stream->reset_metadata_flag) {
1829 ret = utils_truncate_stream_file(stream->out_fd, 0);
1830 if (ret < 0) {
1831 ERR("Reset metadata file");
1832 goto end;
1833 }
1834 stream->reset_metadata_flag = 0;
1835 }
1836 /*
1837 * Check if we need to change the tracefile before writing the packet.
1838 */
1839 if (stream->chan->tracefile_size > 0 &&
1840 (stream->tracefile_size_current + len) >
1841 stream->chan->tracefile_size) {
1842 ret = utils_rotate_stream_file(stream->chan->pathname,
1843 stream->name, stream->chan->tracefile_size,
1844 stream->chan->tracefile_count, stream->uid, stream->gid,
1845 stream->out_fd, &(stream->tracefile_count_current),
1846 &stream->out_fd);
1847 if (ret < 0) {
1848 written = ret;
1849 ERR("Rotating output file");
1850 goto end;
1851 }
1852 outfd = stream->out_fd;
1853
1854 if (stream->index_file) {
1855 lttng_index_file_put(stream->index_file);
1856 stream->index_file = lttng_index_file_create(stream->chan->pathname,
1857 stream->name, stream->uid, stream->gid,
1858 stream->chan->tracefile_size,
1859 stream->tracefile_count_current,
1860 CTF_INDEX_MAJOR, CTF_INDEX_MINOR);
1861 if (!stream->index_file) {
1862 goto end;
1863 }
1864 }
1865
1866 /* Reset current size because we just perform a rotation. */
1867 stream->tracefile_size_current = 0;
1868 stream->out_fd_offset = 0;
1869 orig_offset = 0;
1870 }
1871 stream->tracefile_size_current += len;
1872 index->offset = htobe64(stream->out_fd_offset);
1873 }
1874
1875 while (len > 0) {
1876 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1877 (unsigned long)offset, len, fd, splice_pipe[1]);
1878 ret_splice = splice(fd, &offset, splice_pipe[1], NULL, len,
1879 SPLICE_F_MOVE | SPLICE_F_MORE);
1880 DBG("splice chan to pipe, ret %zd", ret_splice);
1881 if (ret_splice < 0) {
1882 ret = errno;
1883 written = -ret;
1884 PERROR("Error in relay splice");
1885 goto splice_error;
1886 }
1887
1888 /* Handle stream on the relayd if the output is on the network */
1889 if (relayd && stream->metadata_flag) {
1890 size_t metadata_payload_size =
1891 sizeof(struct lttcomm_relayd_metadata_payload);
1892
1893 /* Update counter to fit the spliced data */
1894 ret_splice += metadata_payload_size;
1895 len += metadata_payload_size;
1896 /*
1897 * We do this so the return value can match the len passed as
1898 * argument to this function.
1899 */
1900 written -= metadata_payload_size;
1901 }
1902
1903 /* Splice data out */
1904 ret_splice = splice(splice_pipe[0], NULL, outfd, NULL,
1905 ret_splice, SPLICE_F_MOVE | SPLICE_F_MORE);
1906 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
1907 outfd, ret_splice);
1908 if (ret_splice < 0) {
1909 ret = errno;
1910 written = -ret;
1911 relayd_hang_up = 1;
1912 goto write_error;
1913 } else if (ret_splice > len) {
1914 /*
1915 * We don't expect this code path to be executed but you never know
1916 * so this is an extra protection agains a buggy splice().
1917 */
1918 ret = errno;
1919 written += ret_splice;
1920 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice,
1921 len);
1922 goto splice_error;
1923 } else {
1924 /* All good, update current len and continue. */
1925 len -= ret_splice;
1926 }
1927
1928 /* This call is useless on a socket so better save a syscall. */
1929 if (!relayd) {
1930 /* This won't block, but will start writeout asynchronously */
1931 lttng_sync_file_range(outfd, stream->out_fd_offset, ret_splice,
1932 SYNC_FILE_RANGE_WRITE);
1933 stream->out_fd_offset += ret_splice;
1934 }
1935 stream->output_written += ret_splice;
1936 written += ret_splice;
1937 }
1938 if (!relayd) {
1939 lttng_consumer_sync_trace_file(stream, orig_offset);
1940 }
1941 goto end;
1942
1943 write_error:
1944 /*
1945 * This is a special case that the relayd has closed its socket. Let's
1946 * cleanup the relayd object and all associated streams.
1947 */
1948 if (relayd && relayd_hang_up) {
1949 cleanup_relayd(relayd, ctx);
1950 /* Skip splice error so the consumer does not fail */
1951 goto end;
1952 }
1953
1954 splice_error:
1955 /* send the appropriate error description to sessiond */
1956 switch (ret) {
1957 case EINVAL:
1958 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_EINVAL);
1959 break;
1960 case ENOMEM:
1961 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_ENOMEM);
1962 break;
1963 case ESPIPE:
1964 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_ESPIPE);
1965 break;
1966 }
1967
1968 end:
1969 if (relayd && stream->metadata_flag) {
1970 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
1971 }
1972
1973 rcu_read_unlock();
1974 return written;
1975 }
1976
1977 /*
1978 * Sample the snapshot positions for a specific fd
1979 *
1980 * Returns 0 on success, < 0 on error
1981 */
1982 int lttng_consumer_sample_snapshot_positions(struct lttng_consumer_stream *stream)
1983 {
1984 switch (consumer_data.type) {
1985 case LTTNG_CONSUMER_KERNEL:
1986 return lttng_kconsumer_sample_snapshot_positions(stream);
1987 case LTTNG_CONSUMER32_UST:
1988 case LTTNG_CONSUMER64_UST:
1989 return lttng_ustconsumer_sample_snapshot_positions(stream);
1990 default:
1991 ERR("Unknown consumer_data type");
1992 assert(0);
1993 return -ENOSYS;
1994 }
1995 }
1996 /*
1997 * Take a snapshot for a specific fd
1998 *
1999 * Returns 0 on success, < 0 on error
2000 */
2001 int lttng_consumer_take_snapshot(struct lttng_consumer_stream *stream)
2002 {
2003 switch (consumer_data.type) {
2004 case LTTNG_CONSUMER_KERNEL:
2005 return lttng_kconsumer_take_snapshot(stream);
2006 case LTTNG_CONSUMER32_UST:
2007 case LTTNG_CONSUMER64_UST:
2008 return lttng_ustconsumer_take_snapshot(stream);
2009 default:
2010 ERR("Unknown consumer_data type");
2011 assert(0);
2012 return -ENOSYS;
2013 }
2014 }
2015
2016 /*
2017 * Get the produced position
2018 *
2019 * Returns 0 on success, < 0 on error
2020 */
2021 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream *stream,
2022 unsigned long *pos)
2023 {
2024 switch (consumer_data.type) {
2025 case LTTNG_CONSUMER_KERNEL:
2026 return lttng_kconsumer_get_produced_snapshot(stream, pos);
2027 case LTTNG_CONSUMER32_UST:
2028 case LTTNG_CONSUMER64_UST:
2029 return lttng_ustconsumer_get_produced_snapshot(stream, pos);
2030 default:
2031 ERR("Unknown consumer_data type");
2032 assert(0);
2033 return -ENOSYS;
2034 }
2035 }
2036
2037 /*
2038 * Get the consumed position (free-running counter position in bytes).
2039 *
2040 * Returns 0 on success, < 0 on error
2041 */
2042 int lttng_consumer_get_consumed_snapshot(struct lttng_consumer_stream *stream,
2043 unsigned long *pos)
2044 {
2045 switch (consumer_data.type) {
2046 case LTTNG_CONSUMER_KERNEL:
2047 return lttng_kconsumer_get_consumed_snapshot(stream, pos);
2048 case LTTNG_CONSUMER32_UST:
2049 case LTTNG_CONSUMER64_UST:
2050 return lttng_ustconsumer_get_consumed_snapshot(stream, pos);
2051 default:
2052 ERR("Unknown consumer_data type");
2053 assert(0);
2054 return -ENOSYS;
2055 }
2056 }
2057
2058 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data *ctx,
2059 int sock, struct pollfd *consumer_sockpoll)
2060 {
2061 switch (consumer_data.type) {
2062 case LTTNG_CONSUMER_KERNEL:
2063 return lttng_kconsumer_recv_cmd(ctx, sock, consumer_sockpoll);
2064 case LTTNG_CONSUMER32_UST:
2065 case LTTNG_CONSUMER64_UST:
2066 return lttng_ustconsumer_recv_cmd(ctx, sock, consumer_sockpoll);
2067 default:
2068 ERR("Unknown consumer_data type");
2069 assert(0);
2070 return -ENOSYS;
2071 }
2072 }
2073
2074 void lttng_consumer_close_all_metadata(void)
2075 {
2076 switch (consumer_data.type) {
2077 case LTTNG_CONSUMER_KERNEL:
2078 /*
2079 * The Kernel consumer has a different metadata scheme so we don't
2080 * close anything because the stream will be closed by the session
2081 * daemon.
2082 */
2083 break;
2084 case LTTNG_CONSUMER32_UST:
2085 case LTTNG_CONSUMER64_UST:
2086 /*
2087 * Close all metadata streams. The metadata hash table is passed and
2088 * this call iterates over it by closing all wakeup fd. This is safe
2089 * because at this point we are sure that the metadata producer is
2090 * either dead or blocked.
2091 */
2092 lttng_ustconsumer_close_all_metadata(metadata_ht);
2093 break;
2094 default:
2095 ERR("Unknown consumer_data type");
2096 assert(0);
2097 }
2098 }
2099
2100 /*
2101 * Clean up a metadata stream and free its memory.
2102 */
2103 void consumer_del_metadata_stream(struct lttng_consumer_stream *stream,
2104 struct lttng_ht *ht)
2105 {
2106 struct lttng_consumer_channel *free_chan = NULL;
2107
2108 assert(stream);
2109 /*
2110 * This call should NEVER receive regular stream. It must always be
2111 * metadata stream and this is crucial for data structure synchronization.
2112 */
2113 assert(stream->metadata_flag);
2114
2115 DBG3("Consumer delete metadata stream %d", stream->wait_fd);
2116
2117 pthread_mutex_lock(&consumer_data.lock);
2118 pthread_mutex_lock(&stream->chan->lock);
2119 pthread_mutex_lock(&stream->lock);
2120 if (stream->chan->metadata_cache) {
2121 /* Only applicable to userspace consumers. */
2122 pthread_mutex_lock(&stream->chan->metadata_cache->lock);
2123 }
2124
2125 /* Remove any reference to that stream. */
2126 consumer_stream_delete(stream, ht);
2127
2128 /* Close down everything including the relayd if one. */
2129 consumer_stream_close(stream);
2130 /* Destroy tracer buffers of the stream. */
2131 consumer_stream_destroy_buffers(stream);
2132
2133 /* Atomically decrement channel refcount since other threads can use it. */
2134 if (!uatomic_sub_return(&stream->chan->refcount, 1)
2135 && !uatomic_read(&stream->chan->nb_init_stream_left)) {
2136 /* Go for channel deletion! */
2137 free_chan = stream->chan;
2138 }
2139
2140 /*
2141 * Nullify the stream reference so it is not used after deletion. The
2142 * channel lock MUST be acquired before being able to check for a NULL
2143 * pointer value.
2144 */
2145 stream->chan->metadata_stream = NULL;
2146
2147 if (stream->chan->metadata_cache) {
2148 pthread_mutex_unlock(&stream->chan->metadata_cache->lock);
2149 }
2150 pthread_mutex_unlock(&stream->lock);
2151 pthread_mutex_unlock(&stream->chan->lock);
2152 pthread_mutex_unlock(&consumer_data.lock);
2153
2154 if (free_chan) {
2155 consumer_del_channel(free_chan);
2156 }
2157
2158 consumer_stream_free(stream);
2159 }
2160
2161 /*
2162 * Action done with the metadata stream when adding it to the consumer internal
2163 * data structures to handle it.
2164 */
2165 void consumer_add_metadata_stream(struct lttng_consumer_stream *stream)
2166 {
2167 struct lttng_ht *ht = metadata_ht;
2168 struct lttng_ht_iter iter;
2169 struct lttng_ht_node_u64 *node;
2170
2171 assert(stream);
2172 assert(ht);
2173
2174 DBG3("Adding metadata stream %" PRIu64 " to hash table", stream->key);
2175
2176 pthread_mutex_lock(&consumer_data.lock);
2177 pthread_mutex_lock(&stream->chan->lock);
2178 pthread_mutex_lock(&stream->chan->timer_lock);
2179 pthread_mutex_lock(&stream->lock);
2180
2181 /*
2182 * From here, refcounts are updated so be _careful_ when returning an error
2183 * after this point.
2184 */
2185
2186 rcu_read_lock();
2187
2188 /*
2189 * Lookup the stream just to make sure it does not exist in our internal
2190 * state. This should NEVER happen.
2191 */
2192 lttng_ht_lookup(ht, &stream->key, &iter);
2193 node = lttng_ht_iter_get_node_u64(&iter);
2194 assert(!node);
2195
2196 /*
2197 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2198 * in terms of destroying the associated channel, because the action that
2199 * causes the count to become 0 also causes a stream to be added. The
2200 * channel deletion will thus be triggered by the following removal of this
2201 * stream.
2202 */
2203 if (uatomic_read(&stream->chan->nb_init_stream_left) > 0) {
2204 /* Increment refcount before decrementing nb_init_stream_left */
2205 cmm_smp_wmb();
2206 uatomic_dec(&stream->chan->nb_init_stream_left);
2207 }
2208
2209 lttng_ht_add_unique_u64(ht, &stream->node);
2210
2211 lttng_ht_add_u64(consumer_data.stream_per_chan_id_ht,
2212 &stream->node_channel_id);
2213
2214 /*
2215 * Add stream to the stream_list_ht of the consumer data. No need to steal
2216 * the key since the HT does not use it and we allow to add redundant keys
2217 * into this table.
2218 */
2219 lttng_ht_add_u64(consumer_data.stream_list_ht, &stream->node_session_id);
2220
2221 rcu_read_unlock();
2222
2223 pthread_mutex_unlock(&stream->lock);
2224 pthread_mutex_unlock(&stream->chan->lock);
2225 pthread_mutex_unlock(&stream->chan->timer_lock);
2226 pthread_mutex_unlock(&consumer_data.lock);
2227 }
2228
2229 /*
2230 * Delete data stream that are flagged for deletion (endpoint_status).
2231 */
2232 static void validate_endpoint_status_data_stream(void)
2233 {
2234 struct lttng_ht_iter iter;
2235 struct lttng_consumer_stream *stream;
2236
2237 DBG("Consumer delete flagged data stream");
2238
2239 rcu_read_lock();
2240 cds_lfht_for_each_entry(data_ht->ht, &iter.iter, stream, node.node) {
2241 /* Validate delete flag of the stream */
2242 if (stream->endpoint_status == CONSUMER_ENDPOINT_ACTIVE) {
2243 continue;
2244 }
2245 /* Delete it right now */
2246 consumer_del_stream(stream, data_ht);
2247 }
2248 rcu_read_unlock();
2249 }
2250
2251 /*
2252 * Delete metadata stream that are flagged for deletion (endpoint_status).
2253 */
2254 static void validate_endpoint_status_metadata_stream(
2255 struct lttng_poll_event *pollset)
2256 {
2257 struct lttng_ht_iter iter;
2258 struct lttng_consumer_stream *stream;
2259
2260 DBG("Consumer delete flagged metadata stream");
2261
2262 assert(pollset);
2263
2264 rcu_read_lock();
2265 cds_lfht_for_each_entry(metadata_ht->ht, &iter.iter, stream, node.node) {
2266 /* Validate delete flag of the stream */
2267 if (stream->endpoint_status == CONSUMER_ENDPOINT_ACTIVE) {
2268 continue;
2269 }
2270 /*
2271 * Remove from pollset so the metadata thread can continue without
2272 * blocking on a deleted stream.
2273 */
2274 lttng_poll_del(pollset, stream->wait_fd);
2275
2276 /* Delete it right now */
2277 consumer_del_metadata_stream(stream, metadata_ht);
2278 }
2279 rcu_read_unlock();
2280 }
2281
2282 static
2283 int rotate_notify_sessiond(struct lttng_consumer_local_data *ctx,
2284 uint64_t key)
2285 {
2286 ssize_t ret;
2287
2288 do {
2289 ret = write(ctx->channel_rotate_pipe, &key, sizeof(key));
2290 } while (ret == -1 && errno == EINTR);
2291 if (ret == -1) {
2292 PERROR("Failed to write to the channel rotation pipe");
2293 } else {
2294 DBG("Sent channel rotation notification for channel key %"
2295 PRIu64, key);
2296 ret = 0;
2297 }
2298
2299 return (int) ret;
2300 }
2301
2302 /*
2303 * Perform operations that need to be done after a stream has
2304 * rotated and released the stream lock.
2305 *
2306 * Multiple rotations cannot occur simultaneously, so we know the state of the
2307 * "rotated" stream flag cannot change.
2308 *
2309 * This MUST be called WITHOUT the stream lock held.
2310 */
2311 static
2312 int consumer_post_rotation(struct lttng_consumer_stream *stream,
2313 struct lttng_consumer_local_data *ctx)
2314 {
2315 int ret = 0;
2316
2317 pthread_mutex_lock(&stream->chan->lock);
2318
2319 switch (consumer_data.type) {
2320 case LTTNG_CONSUMER_KERNEL:
2321 break;
2322 case LTTNG_CONSUMER32_UST:
2323 case LTTNG_CONSUMER64_UST:
2324 /*
2325 * The ust_metadata_pushed counter has been reset to 0, so now
2326 * we can wakeup the metadata thread so it dumps the metadata
2327 * cache to the new file.
2328 */
2329 if (stream->metadata_flag) {
2330 consumer_metadata_wakeup_pipe(stream->chan);
2331 }
2332 break;
2333 default:
2334 ERR("Unknown consumer_data type");
2335 abort();
2336 }
2337
2338 if (--stream->chan->nr_stream_rotate_pending == 0) {
2339 DBG("Rotation of channel \"%s\" completed, notifying the session daemon",
2340 stream->chan->name);
2341 ret = rotate_notify_sessiond(ctx, stream->chan->key);
2342 }
2343 assert(stream->chan->nr_stream_rotate_pending >= 0);
2344 pthread_mutex_unlock(&stream->chan->lock);
2345
2346 return ret;
2347 }
2348
2349 /*
2350 * Thread polls on metadata file descriptor and write them on disk or on the
2351 * network.
2352 */
2353 void *consumer_thread_metadata_poll(void *data)
2354 {
2355 int ret, i, pollfd, err = -1;
2356 uint32_t revents, nb_fd;
2357 struct lttng_consumer_stream *stream = NULL;
2358 struct lttng_ht_iter iter;
2359 struct lttng_ht_node_u64 *node;
2360 struct lttng_poll_event events;
2361 struct lttng_consumer_local_data *ctx = data;
2362 ssize_t len;
2363
2364 rcu_register_thread();
2365
2366 health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_METADATA);
2367
2368 if (testpoint(consumerd_thread_metadata)) {
2369 goto error_testpoint;
2370 }
2371
2372 health_code_update();
2373
2374 DBG("Thread metadata poll started");
2375
2376 /* Size is set to 1 for the consumer_metadata pipe */
2377 ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC);
2378 if (ret < 0) {
2379 ERR("Poll set creation failed");
2380 goto end_poll;
2381 }
2382
2383 ret = lttng_poll_add(&events,
2384 lttng_pipe_get_readfd(ctx->consumer_metadata_pipe), LPOLLIN);
2385 if (ret < 0) {
2386 goto end;
2387 }
2388
2389 /* Main loop */
2390 DBG("Metadata main loop started");
2391
2392 while (1) {
2393 restart:
2394 health_code_update();
2395 health_poll_entry();
2396 DBG("Metadata poll wait");
2397 ret = lttng_poll_wait(&events, -1);
2398 DBG("Metadata poll return from wait with %d fd(s)",
2399 LTTNG_POLL_GETNB(&events));
2400 health_poll_exit();
2401 DBG("Metadata event caught in thread");
2402 if (ret < 0) {
2403 if (errno == EINTR) {
2404 ERR("Poll EINTR caught");
2405 goto restart;
2406 }
2407 if (LTTNG_POLL_GETNB(&events) == 0) {
2408 err = 0; /* All is OK */
2409 }
2410 goto end;
2411 }
2412
2413 nb_fd = ret;
2414
2415 /* From here, the event is a metadata wait fd */
2416 for (i = 0; i < nb_fd; i++) {
2417 health_code_update();
2418
2419 revents = LTTNG_POLL_GETEV(&events, i);
2420 pollfd = LTTNG_POLL_GETFD(&events, i);
2421
2422 if (!revents) {
2423 /* No activity for this FD (poll implementation). */
2424 continue;
2425 }
2426
2427 if (pollfd == lttng_pipe_get_readfd(ctx->consumer_metadata_pipe)) {
2428 if (revents & LPOLLIN) {
2429 ssize_t pipe_len;
2430
2431 pipe_len = lttng_pipe_read(ctx->consumer_metadata_pipe,
2432 &stream, sizeof(stream));
2433 if (pipe_len < sizeof(stream)) {
2434 if (pipe_len < 0) {
2435 PERROR("read metadata stream");
2436 }
2437 /*
2438 * Remove the pipe from the poll set and continue the loop
2439 * since their might be data to consume.
2440 */
2441 lttng_poll_del(&events,
2442 lttng_pipe_get_readfd(ctx->consumer_metadata_pipe));
2443 lttng_pipe_read_close(ctx->consumer_metadata_pipe);
2444 continue;
2445 }
2446
2447 /* A NULL stream means that the state has changed. */
2448 if (stream == NULL) {
2449 /* Check for deleted streams. */
2450 validate_endpoint_status_metadata_stream(&events);
2451 goto restart;
2452 }
2453
2454 DBG("Adding metadata stream %d to poll set",
2455 stream->wait_fd);
2456
2457 /* Add metadata stream to the global poll events list */
2458 lttng_poll_add(&events, stream->wait_fd,
2459 LPOLLIN | LPOLLPRI | LPOLLHUP);
2460 } else if (revents & (LPOLLERR | LPOLLHUP)) {
2461 DBG("Metadata thread pipe hung up");
2462 /*
2463 * Remove the pipe from the poll set and continue the loop
2464 * since their might be data to consume.
2465 */
2466 lttng_poll_del(&events,
2467 lttng_pipe_get_readfd(ctx->consumer_metadata_pipe));
2468 lttng_pipe_read_close(ctx->consumer_metadata_pipe);
2469 continue;
2470 } else {
2471 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
2472 goto end;
2473 }
2474
2475 /* Handle other stream */
2476 continue;
2477 }
2478
2479 rcu_read_lock();
2480 {
2481 uint64_t tmp_id = (uint64_t) pollfd;
2482
2483 lttng_ht_lookup(metadata_ht, &tmp_id, &iter);
2484 }
2485 node = lttng_ht_iter_get_node_u64(&iter);
2486 assert(node);
2487
2488 stream = caa_container_of(node, struct lttng_consumer_stream,
2489 node);
2490
2491 if (revents & (LPOLLIN | LPOLLPRI)) {
2492 /* Get the data out of the metadata file descriptor */
2493 DBG("Metadata available on fd %d", pollfd);
2494 assert(stream->wait_fd == pollfd);
2495
2496 do {
2497 health_code_update();
2498
2499 len = ctx->on_buffer_ready(stream, ctx);
2500 /*
2501 * We don't check the return value here since if we get
2502 * a negative len, it means an error occurred thus we
2503 * simply remove it from the poll set and free the
2504 * stream.
2505 */
2506 } while (len > 0);
2507
2508 /* It's ok to have an unavailable sub-buffer */
2509 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2510 /* Clean up stream from consumer and free it. */
2511 lttng_poll_del(&events, stream->wait_fd);
2512 consumer_del_metadata_stream(stream, metadata_ht);
2513 }
2514 } else if (revents & (LPOLLERR | LPOLLHUP)) {
2515 DBG("Metadata fd %d is hup|err.", pollfd);
2516 if (!stream->hangup_flush_done
2517 && (consumer_data.type == LTTNG_CONSUMER32_UST
2518 || consumer_data.type == LTTNG_CONSUMER64_UST)) {
2519 DBG("Attempting to flush and consume the UST buffers");
2520 lttng_ustconsumer_on_stream_hangup(stream);
2521
2522 /* We just flushed the stream now read it. */
2523 do {
2524 health_code_update();
2525
2526 len = ctx->on_buffer_ready(stream, ctx);
2527 /*
2528 * We don't check the return value here since if we get
2529 * a negative len, it means an error occurred thus we
2530 * simply remove it from the poll set and free the
2531 * stream.
2532 */
2533 } while (len > 0);
2534 }
2535
2536 lttng_poll_del(&events, stream->wait_fd);
2537 /*
2538 * This call update the channel states, closes file descriptors
2539 * and securely free the stream.
2540 */
2541 consumer_del_metadata_stream(stream, metadata_ht);
2542 } else {
2543 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
2544 rcu_read_unlock();
2545 goto end;
2546 }
2547 /* Release RCU lock for the stream looked up */
2548 rcu_read_unlock();
2549 }
2550 }
2551
2552 /* All is OK */
2553 err = 0;
2554 end:
2555 DBG("Metadata poll thread exiting");
2556
2557 lttng_poll_clean(&events);
2558 end_poll:
2559 error_testpoint:
2560 if (err) {
2561 health_error();
2562 ERR("Health error occurred in %s", __func__);
2563 }
2564 health_unregister(health_consumerd);
2565 rcu_unregister_thread();
2566 return NULL;
2567 }
2568
2569 /*
2570 * This thread polls the fds in the set to consume the data and write
2571 * it to tracefile if necessary.
2572 */
2573 void *consumer_thread_data_poll(void *data)
2574 {
2575 int num_rdy, num_hup, high_prio, ret, i, err = -1;
2576 struct pollfd *pollfd = NULL;
2577 /* local view of the streams */
2578 struct lttng_consumer_stream **local_stream = NULL, *new_stream = NULL;
2579 /* local view of consumer_data.fds_count */
2580 int nb_fd = 0, nb_pipes_fd;
2581 /* Number of FDs with CONSUMER_ENDPOINT_INACTIVE but still open. */
2582 int nb_inactive_fd = 0;
2583 struct lttng_consumer_local_data *ctx = data;
2584 ssize_t len;
2585
2586 rcu_register_thread();
2587
2588 health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_DATA);
2589
2590 if (testpoint(consumerd_thread_data)) {
2591 goto error_testpoint;
2592 }
2593
2594 health_code_update();
2595
2596 local_stream = zmalloc(sizeof(struct lttng_consumer_stream *));
2597 if (local_stream == NULL) {
2598 PERROR("local_stream malloc");
2599 goto end;
2600 }
2601
2602 while (1) {
2603 health_code_update();
2604
2605 high_prio = 0;
2606 num_hup = 0;
2607
2608 /*
2609 * the fds set has been updated, we need to update our
2610 * local array as well
2611 */
2612 pthread_mutex_lock(&consumer_data.lock);
2613 if (consumer_data.need_update) {
2614 free(pollfd);
2615 pollfd = NULL;
2616
2617 free(local_stream);
2618 local_stream = NULL;
2619
2620 /*
2621 * Allocate for all fds + 2:
2622 * +1 for the consumer_data_pipe
2623 * +1 for wake up pipe
2624 */
2625 nb_pipes_fd = 2;
2626 pollfd = zmalloc((consumer_data.stream_count + nb_pipes_fd) * sizeof(struct pollfd));
2627 if (pollfd == NULL) {
2628 PERROR("pollfd malloc");
2629 pthread_mutex_unlock(&consumer_data.lock);
2630 goto end;
2631 }
2632
2633 local_stream = zmalloc((consumer_data.stream_count + nb_pipes_fd) *
2634 sizeof(struct lttng_consumer_stream *));
2635 if (local_stream == NULL) {
2636 PERROR("local_stream malloc");
2637 pthread_mutex_unlock(&consumer_data.lock);
2638 goto end;
2639 }
2640 ret = update_poll_array(ctx, &pollfd, local_stream,
2641 data_ht, &nb_inactive_fd);
2642 if (ret < 0) {
2643 ERR("Error in allocating pollfd or local_outfds");
2644 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
2645 pthread_mutex_unlock(&consumer_data.lock);
2646 goto end;
2647 }
2648 nb_fd = ret;
2649 consumer_data.need_update = 0;
2650 }
2651 pthread_mutex_unlock(&consumer_data.lock);
2652
2653 /* No FDs and consumer_quit, consumer_cleanup the thread */
2654 if (nb_fd == 0 && nb_inactive_fd == 0 &&
2655 CMM_LOAD_SHARED(consumer_quit) == 1) {
2656 err = 0; /* All is OK */
2657 goto end;
2658 }
2659 /* poll on the array of fds */
2660 restart:
2661 DBG("polling on %d fd", nb_fd + nb_pipes_fd);
2662 if (testpoint(consumerd_thread_data_poll)) {
2663 goto end;
2664 }
2665 health_poll_entry();
2666 num_rdy = poll(pollfd, nb_fd + nb_pipes_fd, -1);
2667 health_poll_exit();
2668 DBG("poll num_rdy : %d", num_rdy);
2669 if (num_rdy == -1) {
2670 /*
2671 * Restart interrupted system call.
2672 */
2673 if (errno == EINTR) {
2674 goto restart;
2675 }
2676 PERROR("Poll error");
2677 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
2678 goto end;
2679 } else if (num_rdy == 0) {
2680 DBG("Polling thread timed out");
2681 goto end;
2682 }
2683
2684 if (caa_unlikely(data_consumption_paused)) {
2685 DBG("Data consumption paused, sleeping...");
2686 sleep(1);
2687 goto restart;
2688 }
2689
2690 /*
2691 * If the consumer_data_pipe triggered poll go directly to the
2692 * beginning of the loop to update the array. We want to prioritize
2693 * array update over low-priority reads.
2694 */
2695 if (pollfd[nb_fd].revents & (POLLIN | POLLPRI)) {
2696 ssize_t pipe_readlen;
2697
2698 DBG("consumer_data_pipe wake up");
2699 pipe_readlen = lttng_pipe_read(ctx->consumer_data_pipe,
2700 &new_stream, sizeof(new_stream));
2701 if (pipe_readlen < sizeof(new_stream)) {
2702 PERROR("Consumer data pipe");
2703 /* Continue so we can at least handle the current stream(s). */
2704 continue;
2705 }
2706
2707 /*
2708 * If the stream is NULL, just ignore it. It's also possible that
2709 * the sessiond poll thread changed the consumer_quit state and is
2710 * waking us up to test it.
2711 */
2712 if (new_stream == NULL) {
2713 validate_endpoint_status_data_stream();
2714 continue;
2715 }
2716
2717 /* Continue to update the local streams and handle prio ones */
2718 continue;
2719 }
2720
2721 /* Handle wakeup pipe. */
2722 if (pollfd[nb_fd + 1].revents & (POLLIN | POLLPRI)) {
2723 char dummy;
2724 ssize_t pipe_readlen;
2725
2726 pipe_readlen = lttng_pipe_read(ctx->consumer_wakeup_pipe, &dummy,
2727 sizeof(dummy));
2728 if (pipe_readlen < 0) {
2729 PERROR("Consumer data wakeup pipe");
2730 }
2731 /* We've been awakened to handle stream(s). */
2732 ctx->has_wakeup = 0;
2733 }
2734
2735 /* Take care of high priority channels first. */
2736 for (i = 0; i < nb_fd; i++) {
2737 health_code_update();
2738
2739 if (local_stream[i] == NULL) {
2740 continue;
2741 }
2742 if (pollfd[i].revents & POLLPRI) {
2743 DBG("Urgent read on fd %d", pollfd[i].fd);
2744 high_prio = 1;
2745 len = ctx->on_buffer_ready(local_stream[i], ctx);
2746 /* it's ok to have an unavailable sub-buffer */
2747 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2748 /* Clean the stream and free it. */
2749 consumer_del_stream(local_stream[i], data_ht);
2750 local_stream[i] = NULL;
2751 } else if (len > 0) {
2752 local_stream[i]->data_read = 1;
2753 }
2754 }
2755 }
2756
2757 /*
2758 * If we read high prio channel in this loop, try again
2759 * for more high prio data.
2760 */
2761 if (high_prio) {
2762 continue;
2763 }
2764
2765 /* Take care of low priority channels. */
2766 for (i = 0; i < nb_fd; i++) {
2767 health_code_update();
2768
2769 if (local_stream[i] == NULL) {
2770 continue;
2771 }
2772 if ((pollfd[i].revents & POLLIN) ||
2773 local_stream[i]->hangup_flush_done ||
2774 local_stream[i]->has_data) {
2775 DBG("Normal read on fd %d", pollfd[i].fd);
2776 len = ctx->on_buffer_ready(local_stream[i], ctx);
2777 /* it's ok to have an unavailable sub-buffer */
2778 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2779 /* Clean the stream and free it. */
2780 consumer_del_stream(local_stream[i], data_ht);
2781 local_stream[i] = NULL;
2782 } else if (len > 0) {
2783 local_stream[i]->data_read = 1;
2784 }
2785 }
2786 }
2787
2788 /* Handle hangup and errors */
2789 for (i = 0; i < nb_fd; i++) {
2790 health_code_update();
2791
2792 if (local_stream[i] == NULL) {
2793 continue;
2794 }
2795 if (!local_stream[i]->hangup_flush_done
2796 && (pollfd[i].revents & (POLLHUP | POLLERR | POLLNVAL))
2797 && (consumer_data.type == LTTNG_CONSUMER32_UST
2798 || consumer_data.type == LTTNG_CONSUMER64_UST)) {
2799 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2800 pollfd[i].fd);
2801 lttng_ustconsumer_on_stream_hangup(local_stream[i]);
2802 /* Attempt read again, for the data we just flushed. */
2803 local_stream[i]->data_read = 1;
2804 }
2805 /*
2806 * If the poll flag is HUP/ERR/NVAL and we have
2807 * read no data in this pass, we can remove the
2808 * stream from its hash table.
2809 */
2810 if ((pollfd[i].revents & POLLHUP)) {
2811 DBG("Polling fd %d tells it has hung up.", pollfd[i].fd);
2812 if (!local_stream[i]->data_read) {
2813 consumer_del_stream(local_stream[i], data_ht);
2814 local_stream[i] = NULL;
2815 num_hup++;
2816 }
2817 } else if (pollfd[i].revents & POLLERR) {
2818 ERR("Error returned in polling fd %d.", pollfd[i].fd);
2819 if (!local_stream[i]->data_read) {
2820 consumer_del_stream(local_stream[i], data_ht);
2821 local_stream[i] = NULL;
2822 num_hup++;
2823 }
2824 } else if (pollfd[i].revents & POLLNVAL) {
2825 ERR("Polling fd %d tells fd is not open.", pollfd[i].fd);
2826 if (!local_stream[i]->data_read) {
2827 consumer_del_stream(local_stream[i], data_ht);
2828 local_stream[i] = NULL;
2829 num_hup++;
2830 }
2831 }
2832 if (local_stream[i] != NULL) {
2833 local_stream[i]->data_read = 0;
2834 }
2835 }
2836 }
2837 /* All is OK */
2838 err = 0;
2839 end:
2840 DBG("polling thread exiting");
2841 free(pollfd);
2842 free(local_stream);
2843
2844 /*
2845 * Close the write side of the pipe so epoll_wait() in
2846 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2847 * read side of the pipe. If we close them both, epoll_wait strangely does
2848 * not return and could create a endless wait period if the pipe is the
2849 * only tracked fd in the poll set. The thread will take care of closing
2850 * the read side.
2851 */
2852 (void) lttng_pipe_write_close(ctx->consumer_metadata_pipe);
2853
2854 error_testpoint:
2855 if (err) {
2856 health_error();
2857 ERR("Health error occurred in %s", __func__);
2858 }
2859 health_unregister(health_consumerd);
2860
2861 rcu_unregister_thread();
2862 return NULL;
2863 }
2864
2865 /*
2866 * Close wake-up end of each stream belonging to the channel. This will
2867 * allow the poll() on the stream read-side to detect when the
2868 * write-side (application) finally closes them.
2869 */
2870 static
2871 void consumer_close_channel_streams(struct lttng_consumer_channel *channel)
2872 {
2873 struct lttng_ht *ht;
2874 struct lttng_consumer_stream *stream;
2875 struct lttng_ht_iter iter;
2876
2877 ht = consumer_data.stream_per_chan_id_ht;
2878
2879 rcu_read_lock();
2880 cds_lfht_for_each_entry_duplicate(ht->ht,
2881 ht->hash_fct(&channel->key, lttng_ht_seed),
2882 ht->match_fct, &channel->key,
2883 &iter.iter, stream, node_channel_id.node) {
2884 /*
2885 * Protect against teardown with mutex.
2886 */
2887 pthread_mutex_lock(&stream->lock);
2888 if (cds_lfht_is_node_deleted(&stream->node.node)) {
2889 goto next;
2890 }
2891 switch (consumer_data.type) {
2892 case LTTNG_CONSUMER_KERNEL:
2893 break;
2894 case LTTNG_CONSUMER32_UST:
2895 case LTTNG_CONSUMER64_UST:
2896 if (stream->metadata_flag) {
2897 /* Safe and protected by the stream lock. */
2898 lttng_ustconsumer_close_metadata(stream->chan);
2899 } else {
2900 /*
2901 * Note: a mutex is taken internally within
2902 * liblttng-ust-ctl to protect timer wakeup_fd
2903 * use from concurrent close.
2904 */
2905 lttng_ustconsumer_close_stream_wakeup(stream);
2906 }
2907 break;
2908 default:
2909 ERR("Unknown consumer_data type");
2910 assert(0);
2911 }
2912 next:
2913 pthread_mutex_unlock(&stream->lock);
2914 }
2915 rcu_read_unlock();
2916 }
2917
2918 static void destroy_channel_ht(struct lttng_ht *ht)
2919 {
2920 struct lttng_ht_iter iter;
2921 struct lttng_consumer_channel *channel;
2922 int ret;
2923
2924 if (ht == NULL) {
2925 return;
2926 }
2927
2928 rcu_read_lock();
2929 cds_lfht_for_each_entry(ht->ht, &iter.iter, channel, wait_fd_node.node) {
2930 ret = lttng_ht_del(ht, &iter);
2931 assert(ret != 0);
2932 }
2933 rcu_read_unlock();
2934
2935 lttng_ht_destroy(ht);
2936 }
2937
2938 /*
2939 * This thread polls the channel fds to detect when they are being
2940 * closed. It closes all related streams if the channel is detected as
2941 * closed. It is currently only used as a shim layer for UST because the
2942 * consumerd needs to keep the per-stream wakeup end of pipes open for
2943 * periodical flush.
2944 */
2945 void *consumer_thread_channel_poll(void *data)
2946 {
2947 int ret, i, pollfd, err = -1;
2948 uint32_t revents, nb_fd;
2949 struct lttng_consumer_channel *chan = NULL;
2950 struct lttng_ht_iter iter;
2951 struct lttng_ht_node_u64 *node;
2952 struct lttng_poll_event events;
2953 struct lttng_consumer_local_data *ctx = data;
2954 struct lttng_ht *channel_ht;
2955
2956 rcu_register_thread();
2957
2958 health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_CHANNEL);
2959
2960 if (testpoint(consumerd_thread_channel)) {
2961 goto error_testpoint;
2962 }
2963
2964 health_code_update();
2965
2966 channel_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
2967 if (!channel_ht) {
2968 /* ENOMEM at this point. Better to bail out. */
2969 goto end_ht;
2970 }
2971
2972 DBG("Thread channel poll started");
2973
2974 /* Size is set to 1 for the consumer_channel pipe */
2975 ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC);
2976 if (ret < 0) {
2977 ERR("Poll set creation failed");
2978 goto end_poll;
2979 }
2980
2981 ret = lttng_poll_add(&events, ctx->consumer_channel_pipe[0], LPOLLIN);
2982 if (ret < 0) {
2983 goto end;
2984 }
2985
2986 /* Main loop */
2987 DBG("Channel main loop started");
2988
2989 while (1) {
2990 restart:
2991 health_code_update();
2992 DBG("Channel poll wait");
2993 health_poll_entry();
2994 ret = lttng_poll_wait(&events, -1);
2995 DBG("Channel poll return from wait with %d fd(s)",
2996 LTTNG_POLL_GETNB(&events));
2997 health_poll_exit();
2998 DBG("Channel event caught in thread");
2999 if (ret < 0) {
3000 if (errno == EINTR) {
3001 ERR("Poll EINTR caught");
3002 goto restart;
3003 }
3004 if (LTTNG_POLL_GETNB(&events) == 0) {
3005 err = 0; /* All is OK */
3006 }
3007 goto end;
3008 }
3009
3010 nb_fd = ret;
3011
3012 /* From here, the event is a channel wait fd */
3013 for (i = 0; i < nb_fd; i++) {
3014 health_code_update();
3015
3016 revents = LTTNG_POLL_GETEV(&events, i);
3017 pollfd = LTTNG_POLL_GETFD(&events, i);
3018
3019 if (!revents) {
3020 /* No activity for this FD (poll implementation). */
3021 continue;
3022 }
3023
3024 if (pollfd == ctx->consumer_channel_pipe[0]) {
3025 if (revents & LPOLLIN) {
3026 enum consumer_channel_action action;
3027 uint64_t key;
3028
3029 ret = read_channel_pipe(ctx, &chan, &key, &action);
3030 if (ret <= 0) {
3031 if (ret < 0) {
3032 ERR("Error reading channel pipe");
3033 }
3034 lttng_poll_del(&events, ctx->consumer_channel_pipe[0]);
3035 continue;
3036 }
3037
3038 switch (action) {
3039 case CONSUMER_CHANNEL_ADD:
3040 DBG("Adding channel %d to poll set",
3041 chan->wait_fd);
3042
3043 lttng_ht_node_init_u64(&chan->wait_fd_node,
3044 chan->wait_fd);
3045 rcu_read_lock();
3046 lttng_ht_add_unique_u64(channel_ht,
3047 &chan->wait_fd_node);
3048 rcu_read_unlock();
3049 /* Add channel to the global poll events list */
3050 lttng_poll_add(&events, chan->wait_fd,
3051 LPOLLERR | LPOLLHUP);
3052 break;
3053 case CONSUMER_CHANNEL_DEL:
3054 {
3055 /*
3056 * This command should never be called if the channel
3057 * has streams monitored by either the data or metadata
3058 * thread. The consumer only notify this thread with a
3059 * channel del. command if it receives a destroy
3060 * channel command from the session daemon that send it
3061 * if a command prior to the GET_CHANNEL failed.
3062 */
3063
3064 rcu_read_lock();
3065 chan = consumer_find_channel(key);
3066 if (!chan) {
3067 rcu_read_unlock();
3068 ERR("UST consumer get channel key %" PRIu64 " not found for del channel", key);
3069 break;
3070 }
3071 lttng_poll_del(&events, chan->wait_fd);
3072 iter.iter.node = &chan->wait_fd_node.node;
3073 ret = lttng_ht_del(channel_ht, &iter);
3074 assert(ret == 0);
3075
3076 switch (consumer_data.type) {
3077 case LTTNG_CONSUMER_KERNEL:
3078 break;
3079 case LTTNG_CONSUMER32_UST:
3080 case LTTNG_CONSUMER64_UST:
3081 health_code_update();
3082 /* Destroy streams that might have been left in the stream list. */
3083 clean_channel_stream_list(chan);
3084 break;
3085 default:
3086 ERR("Unknown consumer_data type");
3087 assert(0);
3088 }
3089
3090 /*
3091 * Release our own refcount. Force channel deletion even if
3092 * streams were not initialized.
3093 */
3094 if (!uatomic_sub_return(&chan->refcount, 1)) {
3095 consumer_del_channel(chan);
3096 }
3097 rcu_read_unlock();
3098 goto restart;
3099 }
3100 case CONSUMER_CHANNEL_QUIT:
3101 /*
3102 * Remove the pipe from the poll set and continue the loop
3103 * since their might be data to consume.
3104 */
3105 lttng_poll_del(&events, ctx->consumer_channel_pipe[0]);
3106 continue;
3107 default:
3108 ERR("Unknown action");
3109 break;
3110 }
3111 } else if (revents & (LPOLLERR | LPOLLHUP)) {
3112 DBG("Channel thread pipe hung up");
3113 /*
3114 * Remove the pipe from the poll set and continue the loop
3115 * since their might be data to consume.
3116 */
3117 lttng_poll_del(&events, ctx->consumer_channel_pipe[0]);
3118 continue;
3119 } else {
3120 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
3121 goto end;
3122 }
3123
3124 /* Handle other stream */
3125 continue;
3126 }
3127
3128 rcu_read_lock();
3129 {
3130 uint64_t tmp_id = (uint64_t) pollfd;
3131
3132 lttng_ht_lookup(channel_ht, &tmp_id, &iter);
3133 }
3134 node = lttng_ht_iter_get_node_u64(&iter);
3135 assert(node);
3136
3137 chan = caa_container_of(node, struct lttng_consumer_channel,
3138 wait_fd_node);
3139
3140 /* Check for error event */
3141 if (revents & (LPOLLERR | LPOLLHUP)) {
3142 DBG("Channel fd %d is hup|err.", pollfd);
3143
3144 lttng_poll_del(&events, chan->wait_fd);
3145 ret = lttng_ht_del(channel_ht, &iter);
3146 assert(ret == 0);
3147
3148 /*
3149 * This will close the wait fd for each stream associated to
3150 * this channel AND monitored by the data/metadata thread thus
3151 * will be clean by the right thread.
3152 */
3153 consumer_close_channel_streams(chan);
3154
3155 /* Release our own refcount */
3156 if (!uatomic_sub_return(&chan->refcount, 1)
3157 && !uatomic_read(&chan->nb_init_stream_left)) {
3158 consumer_del_channel(chan);
3159 }
3160 } else {
3161 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
3162 rcu_read_unlock();
3163 goto end;
3164 }
3165
3166 /* Release RCU lock for the channel looked up */
3167 rcu_read_unlock();
3168 }
3169 }
3170
3171 /* All is OK */
3172 err = 0;
3173 end:
3174 lttng_poll_clean(&events);
3175 end_poll:
3176 destroy_channel_ht(channel_ht);
3177 end_ht:
3178 error_testpoint:
3179 DBG("Channel poll thread exiting");
3180 if (err) {
3181 health_error();
3182 ERR("Health error occurred in %s", __func__);
3183 }
3184 health_unregister(health_consumerd);
3185 rcu_unregister_thread();
3186 return NULL;
3187 }
3188
3189 static int set_metadata_socket(struct lttng_consumer_local_data *ctx,
3190 struct pollfd *sockpoll, int client_socket)
3191 {
3192 int ret;
3193
3194 assert(ctx);
3195 assert(sockpoll);
3196
3197 ret = lttng_consumer_poll_socket(sockpoll);
3198 if (ret) {
3199 goto error;
3200 }
3201 DBG("Metadata connection on client_socket");
3202
3203 /* Blocking call, waiting for transmission */
3204 ctx->consumer_metadata_socket = lttcomm_accept_unix_sock(client_socket);
3205 if (ctx->consumer_metadata_socket < 0) {
3206 WARN("On accept metadata");
3207 ret = -1;
3208 goto error;
3209 }
3210 ret = 0;
3211
3212 error:
3213 return ret;
3214 }
3215
3216 /*
3217 * This thread listens on the consumerd socket and receives the file
3218 * descriptors from the session daemon.
3219 */
3220 void *consumer_thread_sessiond_poll(void *data)
3221 {
3222 int sock = -1, client_socket, ret, err = -1;
3223 /*
3224 * structure to poll for incoming data on communication socket avoids
3225 * making blocking sockets.