Fix: relayd control socket mutex is not destroyed
[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 pthread_mutex_destroy(&relayd->ctrl_sock_mutex);
327 free(relayd);
328 }
329
330 /*
331 * Destroy and free relayd socket pair object.
332 */
333 void consumer_destroy_relayd(struct consumer_relayd_sock_pair *relayd)
334 {
335 int ret;
336 struct lttng_ht_iter iter;
337
338 if (relayd == NULL) {
339 return;
340 }
341
342 DBG("Consumer destroy and close relayd socket pair");
343
344 iter.iter.node = &relayd->node.node;
345 ret = lttng_ht_del(consumer_data.relayd_ht, &iter);
346 if (ret != 0) {
347 /* We assume the relayd is being or is destroyed */
348 return;
349 }
350
351 /* RCU free() call */
352 call_rcu(&relayd->node.head, free_relayd_rcu);
353 }
354
355 /*
356 * Remove a channel from the global list protected by a mutex. This function is
357 * also responsible for freeing its data structures.
358 */
359 void consumer_del_channel(struct lttng_consumer_channel *channel)
360 {
361 int ret;
362 struct lttng_ht_iter iter;
363
364 DBG("Consumer delete channel key %" PRIu64, channel->key);
365
366 pthread_mutex_lock(&consumer_data.lock);
367 pthread_mutex_lock(&channel->lock);
368
369 /* Destroy streams that might have been left in the stream list. */
370 clean_channel_stream_list(channel);
371
372 if (channel->live_timer_enabled == 1) {
373 consumer_timer_live_stop(channel);
374 }
375 if (channel->monitor_timer_enabled == 1) {
376 consumer_timer_monitor_stop(channel);
377 }
378
379 switch (consumer_data.type) {
380 case LTTNG_CONSUMER_KERNEL:
381 break;
382 case LTTNG_CONSUMER32_UST:
383 case LTTNG_CONSUMER64_UST:
384 lttng_ustconsumer_del_channel(channel);
385 break;
386 default:
387 ERR("Unknown consumer_data type");
388 assert(0);
389 goto end;
390 }
391
392 rcu_read_lock();
393 iter.iter.node = &channel->node.node;
394 ret = lttng_ht_del(consumer_data.channel_ht, &iter);
395 assert(!ret);
396 rcu_read_unlock();
397
398 call_rcu(&channel->node.head, free_channel_rcu);
399 end:
400 pthread_mutex_unlock(&channel->lock);
401 pthread_mutex_unlock(&consumer_data.lock);
402 }
403
404 /*
405 * Iterate over the relayd hash table and destroy each element. Finally,
406 * destroy the whole hash table.
407 */
408 static void cleanup_relayd_ht(void)
409 {
410 struct lttng_ht_iter iter;
411 struct consumer_relayd_sock_pair *relayd;
412
413 rcu_read_lock();
414
415 cds_lfht_for_each_entry(consumer_data.relayd_ht->ht, &iter.iter, relayd,
416 node.node) {
417 consumer_destroy_relayd(relayd);
418 }
419
420 rcu_read_unlock();
421
422 lttng_ht_destroy(consumer_data.relayd_ht);
423 }
424
425 /*
426 * Update the end point status of all streams having the given network sequence
427 * index (relayd index).
428 *
429 * It's atomically set without having the stream mutex locked which is fine
430 * because we handle the write/read race with a pipe wakeup for each thread.
431 */
432 static void update_endpoint_status_by_netidx(uint64_t net_seq_idx,
433 enum consumer_endpoint_status status)
434 {
435 struct lttng_ht_iter iter;
436 struct lttng_consumer_stream *stream;
437
438 DBG("Consumer set delete flag on stream by idx %" PRIu64, net_seq_idx);
439
440 rcu_read_lock();
441
442 /* Let's begin with metadata */
443 cds_lfht_for_each_entry(metadata_ht->ht, &iter.iter, stream, node.node) {
444 if (stream->net_seq_idx == net_seq_idx) {
445 uatomic_set(&stream->endpoint_status, status);
446 DBG("Delete flag set to metadata stream %d", stream->wait_fd);
447 }
448 }
449
450 /* Follow up by the data streams */
451 cds_lfht_for_each_entry(data_ht->ht, &iter.iter, stream, node.node) {
452 if (stream->net_seq_idx == net_seq_idx) {
453 uatomic_set(&stream->endpoint_status, status);
454 DBG("Delete flag set to data stream %d", stream->wait_fd);
455 }
456 }
457 rcu_read_unlock();
458 }
459
460 /*
461 * Cleanup a relayd object by flagging every associated streams for deletion,
462 * destroying the object meaning removing it from the relayd hash table,
463 * closing the sockets and freeing the memory in a RCU call.
464 *
465 * If a local data context is available, notify the threads that the streams'
466 * state have changed.
467 */
468 static void cleanup_relayd(struct consumer_relayd_sock_pair *relayd,
469 struct lttng_consumer_local_data *ctx)
470 {
471 uint64_t netidx;
472
473 assert(relayd);
474
475 DBG("Cleaning up relayd sockets");
476
477 /* Save the net sequence index before destroying the object */
478 netidx = relayd->net_seq_idx;
479
480 /*
481 * Delete the relayd from the relayd hash table, close the sockets and free
482 * the object in a RCU call.
483 */
484 consumer_destroy_relayd(relayd);
485
486 /* Set inactive endpoint to all streams */
487 update_endpoint_status_by_netidx(netidx, CONSUMER_ENDPOINT_INACTIVE);
488
489 /*
490 * With a local data context, notify the threads that the streams' state
491 * have changed. The write() action on the pipe acts as an "implicit"
492 * memory barrier ordering the updates of the end point status from the
493 * read of this status which happens AFTER receiving this notify.
494 */
495 if (ctx) {
496 notify_thread_lttng_pipe(ctx->consumer_data_pipe);
497 notify_thread_lttng_pipe(ctx->consumer_metadata_pipe);
498 }
499 }
500
501 /*
502 * Flag a relayd socket pair for destruction. Destroy it if the refcount
503 * reaches zero.
504 *
505 * RCU read side lock MUST be aquired before calling this function.
506 */
507 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair *relayd)
508 {
509 assert(relayd);
510
511 /* Set destroy flag for this object */
512 uatomic_set(&relayd->destroy_flag, 1);
513
514 /* Destroy the relayd if refcount is 0 */
515 if (uatomic_read(&relayd->refcount) == 0) {
516 consumer_destroy_relayd(relayd);
517 }
518 }
519
520 /*
521 * Completly destroy stream from every visiable data structure and the given
522 * hash table if one.
523 *
524 * One this call returns, the stream object is not longer usable nor visible.
525 */
526 void consumer_del_stream(struct lttng_consumer_stream *stream,
527 struct lttng_ht *ht)
528 {
529 consumer_stream_destroy(stream, ht);
530 }
531
532 /*
533 * XXX naming of del vs destroy is all mixed up.
534 */
535 void consumer_del_stream_for_data(struct lttng_consumer_stream *stream)
536 {
537 consumer_stream_destroy(stream, data_ht);
538 }
539
540 void consumer_del_stream_for_metadata(struct lttng_consumer_stream *stream)
541 {
542 consumer_stream_destroy(stream, metadata_ht);
543 }
544
545 void consumer_stream_update_channel_attributes(
546 struct lttng_consumer_stream *stream,
547 struct lttng_consumer_channel *channel)
548 {
549 stream->channel_read_only_attributes.tracefile_size =
550 channel->tracefile_size;
551 memcpy(stream->channel_read_only_attributes.path, channel->pathname,
552 sizeof(stream->channel_read_only_attributes.path));
553 }
554
555 struct lttng_consumer_stream *consumer_allocate_stream(uint64_t channel_key,
556 uint64_t stream_key,
557 enum lttng_consumer_stream_state state,
558 const char *channel_name,
559 uid_t uid,
560 gid_t gid,
561 uint64_t relayd_id,
562 uint64_t session_id,
563 int cpu,
564 int *alloc_ret,
565 enum consumer_channel_type type,
566 unsigned int monitor,
567 uint64_t trace_archive_id)
568 {
569 int ret;
570 struct lttng_consumer_stream *stream;
571
572 stream = zmalloc(sizeof(*stream));
573 if (stream == NULL) {
574 PERROR("malloc struct lttng_consumer_stream");
575 ret = -ENOMEM;
576 goto end;
577 }
578
579 rcu_read_lock();
580
581 stream->key = stream_key;
582 stream->out_fd = -1;
583 stream->out_fd_offset = 0;
584 stream->output_written = 0;
585 stream->state = state;
586 stream->uid = uid;
587 stream->gid = gid;
588 stream->net_seq_idx = relayd_id;
589 stream->session_id = session_id;
590 stream->monitor = monitor;
591 stream->endpoint_status = CONSUMER_ENDPOINT_ACTIVE;
592 stream->index_file = NULL;
593 stream->last_sequence_number = -1ULL;
594 stream->trace_archive_id = trace_archive_id;
595 pthread_mutex_init(&stream->lock, NULL);
596 pthread_mutex_init(&stream->metadata_timer_lock, NULL);
597
598 /* If channel is the metadata, flag this stream as metadata. */
599 if (type == CONSUMER_CHANNEL_TYPE_METADATA) {
600 stream->metadata_flag = 1;
601 /* Metadata is flat out. */
602 strncpy(stream->name, DEFAULT_METADATA_NAME, sizeof(stream->name));
603 /* Live rendez-vous point. */
604 pthread_cond_init(&stream->metadata_rdv, NULL);
605 pthread_mutex_init(&stream->metadata_rdv_lock, NULL);
606 } else {
607 /* Format stream name to <channel_name>_<cpu_number> */
608 ret = snprintf(stream->name, sizeof(stream->name), "%s_%d",
609 channel_name, cpu);
610 if (ret < 0) {
611 PERROR("snprintf stream name");
612 goto error;
613 }
614 }
615
616 /* Key is always the wait_fd for streams. */
617 lttng_ht_node_init_u64(&stream->node, stream->key);
618
619 /* Init node per channel id key */
620 lttng_ht_node_init_u64(&stream->node_channel_id, channel_key);
621
622 /* Init session id node with the stream session id */
623 lttng_ht_node_init_u64(&stream->node_session_id, stream->session_id);
624
625 DBG3("Allocated stream %s (key %" PRIu64 ", chan_key %" PRIu64
626 " relayd_id %" PRIu64 ", session_id %" PRIu64,
627 stream->name, stream->key, channel_key,
628 stream->net_seq_idx, stream->session_id);
629
630 rcu_read_unlock();
631 return stream;
632
633 error:
634 rcu_read_unlock();
635 free(stream);
636 end:
637 if (alloc_ret) {
638 *alloc_ret = ret;
639 }
640 return NULL;
641 }
642
643 /*
644 * Add a stream to the global list protected by a mutex.
645 */
646 void consumer_add_data_stream(struct lttng_consumer_stream *stream)
647 {
648 struct lttng_ht *ht = data_ht;
649
650 assert(stream);
651 assert(ht);
652
653 DBG3("Adding consumer stream %" PRIu64, stream->key);
654
655 pthread_mutex_lock(&consumer_data.lock);
656 pthread_mutex_lock(&stream->chan->lock);
657 pthread_mutex_lock(&stream->chan->timer_lock);
658 pthread_mutex_lock(&stream->lock);
659 rcu_read_lock();
660
661 /* Steal stream identifier to avoid having streams with the same key */
662 steal_stream_key(stream->key, ht);
663
664 lttng_ht_add_unique_u64(ht, &stream->node);
665
666 lttng_ht_add_u64(consumer_data.stream_per_chan_id_ht,
667 &stream->node_channel_id);
668
669 /*
670 * Add stream to the stream_list_ht of the consumer data. No need to steal
671 * the key since the HT does not use it and we allow to add redundant keys
672 * into this table.
673 */
674 lttng_ht_add_u64(consumer_data.stream_list_ht, &stream->node_session_id);
675
676 /*
677 * When nb_init_stream_left reaches 0, we don't need to trigger any action
678 * in terms of destroying the associated channel, because the action that
679 * causes the count to become 0 also causes a stream to be added. The
680 * channel deletion will thus be triggered by the following removal of this
681 * stream.
682 */
683 if (uatomic_read(&stream->chan->nb_init_stream_left) > 0) {
684 /* Increment refcount before decrementing nb_init_stream_left */
685 cmm_smp_wmb();
686 uatomic_dec(&stream->chan->nb_init_stream_left);
687 }
688
689 /* Update consumer data once the node is inserted. */
690 consumer_data.stream_count++;
691 consumer_data.need_update = 1;
692
693 rcu_read_unlock();
694 pthread_mutex_unlock(&stream->lock);
695 pthread_mutex_unlock(&stream->chan->timer_lock);
696 pthread_mutex_unlock(&stream->chan->lock);
697 pthread_mutex_unlock(&consumer_data.lock);
698 }
699
700 void consumer_del_data_stream(struct lttng_consumer_stream *stream)
701 {
702 consumer_del_stream(stream, data_ht);
703 }
704
705 /*
706 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
707 * be acquired before calling this.
708 */
709 static int add_relayd(struct consumer_relayd_sock_pair *relayd)
710 {
711 int ret = 0;
712 struct lttng_ht_node_u64 *node;
713 struct lttng_ht_iter iter;
714
715 assert(relayd);
716
717 lttng_ht_lookup(consumer_data.relayd_ht,
718 &relayd->net_seq_idx, &iter);
719 node = lttng_ht_iter_get_node_u64(&iter);
720 if (node != NULL) {
721 goto end;
722 }
723 lttng_ht_add_unique_u64(consumer_data.relayd_ht, &relayd->node);
724
725 end:
726 return ret;
727 }
728
729 /*
730 * Allocate and return a consumer relayd socket.
731 */
732 static struct consumer_relayd_sock_pair *consumer_allocate_relayd_sock_pair(
733 uint64_t net_seq_idx)
734 {
735 struct consumer_relayd_sock_pair *obj = NULL;
736
737 /* net sequence index of -1 is a failure */
738 if (net_seq_idx == (uint64_t) -1ULL) {
739 goto error;
740 }
741
742 obj = zmalloc(sizeof(struct consumer_relayd_sock_pair));
743 if (obj == NULL) {
744 PERROR("zmalloc relayd sock");
745 goto error;
746 }
747
748 obj->net_seq_idx = net_seq_idx;
749 obj->refcount = 0;
750 obj->destroy_flag = 0;
751 obj->control_sock.sock.fd = -1;
752 obj->data_sock.sock.fd = -1;
753 lttng_ht_node_init_u64(&obj->node, obj->net_seq_idx);
754 pthread_mutex_init(&obj->ctrl_sock_mutex, NULL);
755
756 error:
757 return obj;
758 }
759
760 /*
761 * Find a relayd socket pair in the global consumer data.
762 *
763 * Return the object if found else NULL.
764 * RCU read-side lock must be held across this call and while using the
765 * returned object.
766 */
767 struct consumer_relayd_sock_pair *consumer_find_relayd(uint64_t key)
768 {
769 struct lttng_ht_iter iter;
770 struct lttng_ht_node_u64 *node;
771 struct consumer_relayd_sock_pair *relayd = NULL;
772
773 /* Negative keys are lookup failures */
774 if (key == (uint64_t) -1ULL) {
775 goto error;
776 }
777
778 lttng_ht_lookup(consumer_data.relayd_ht, &key,
779 &iter);
780 node = lttng_ht_iter_get_node_u64(&iter);
781 if (node != NULL) {
782 relayd = caa_container_of(node, struct consumer_relayd_sock_pair, node);
783 }
784
785 error:
786 return relayd;
787 }
788
789 /*
790 * Find a relayd and send the stream
791 *
792 * Returns 0 on success, < 0 on error
793 */
794 int consumer_send_relayd_stream(struct lttng_consumer_stream *stream,
795 char *path)
796 {
797 int ret = 0;
798 struct consumer_relayd_sock_pair *relayd;
799
800 assert(stream);
801 assert(stream->net_seq_idx != -1ULL);
802 assert(path);
803
804 /* The stream is not metadata. Get relayd reference if exists. */
805 rcu_read_lock();
806 relayd = consumer_find_relayd(stream->net_seq_idx);
807 if (relayd != NULL) {
808 /* Add stream on the relayd */
809 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
810 ret = relayd_add_stream(&relayd->control_sock, stream->name,
811 path, &stream->relayd_stream_id,
812 stream->chan->tracefile_size, stream->chan->tracefile_count,
813 stream->trace_archive_id);
814 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
815 if (ret < 0) {
816 goto end;
817 }
818
819 uatomic_inc(&relayd->refcount);
820 stream->sent_to_relayd = 1;
821 } else {
822 ERR("Stream %" PRIu64 " relayd ID %" PRIu64 " unknown. Can't send it.",
823 stream->key, stream->net_seq_idx);
824 ret = -1;
825 goto end;
826 }
827
828 DBG("Stream %s with key %" PRIu64 " sent to relayd id %" PRIu64,
829 stream->name, stream->key, stream->net_seq_idx);
830
831 end:
832 rcu_read_unlock();
833 return ret;
834 }
835
836 /*
837 * Find a relayd and send the streams sent message
838 *
839 * Returns 0 on success, < 0 on error
840 */
841 int consumer_send_relayd_streams_sent(uint64_t net_seq_idx)
842 {
843 int ret = 0;
844 struct consumer_relayd_sock_pair *relayd;
845
846 assert(net_seq_idx != -1ULL);
847
848 /* The stream is not metadata. Get relayd reference if exists. */
849 rcu_read_lock();
850 relayd = consumer_find_relayd(net_seq_idx);
851 if (relayd != NULL) {
852 /* Add stream on the relayd */
853 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
854 ret = relayd_streams_sent(&relayd->control_sock);
855 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
856 if (ret < 0) {
857 goto end;
858 }
859 } else {
860 ERR("Relayd ID %" PRIu64 " unknown. Can't send streams_sent.",
861 net_seq_idx);
862 ret = -1;
863 goto end;
864 }
865
866 ret = 0;
867 DBG("All streams sent relayd id %" PRIu64, net_seq_idx);
868
869 end:
870 rcu_read_unlock();
871 return ret;
872 }
873
874 /*
875 * Find a relayd and close the stream
876 */
877 void close_relayd_stream(struct lttng_consumer_stream *stream)
878 {
879 struct consumer_relayd_sock_pair *relayd;
880
881 /* The stream is not metadata. Get relayd reference if exists. */
882 rcu_read_lock();
883 relayd = consumer_find_relayd(stream->net_seq_idx);
884 if (relayd) {
885 consumer_stream_relayd_close(stream, relayd);
886 }
887 rcu_read_unlock();
888 }
889
890 /*
891 * Handle stream for relayd transmission if the stream applies for network
892 * streaming where the net sequence index is set.
893 *
894 * Return destination file descriptor or negative value on error.
895 */
896 static int write_relayd_stream_header(struct lttng_consumer_stream *stream,
897 size_t data_size, unsigned long padding,
898 struct consumer_relayd_sock_pair *relayd)
899 {
900 int outfd = -1, ret;
901 struct lttcomm_relayd_data_hdr data_hdr;
902
903 /* Safety net */
904 assert(stream);
905 assert(relayd);
906
907 /* Reset data header */
908 memset(&data_hdr, 0, sizeof(data_hdr));
909
910 if (stream->metadata_flag) {
911 /* Caller MUST acquire the relayd control socket lock */
912 ret = relayd_send_metadata(&relayd->control_sock, data_size);
913 if (ret < 0) {
914 goto error;
915 }
916
917 /* Metadata are always sent on the control socket. */
918 outfd = relayd->control_sock.sock.fd;
919 } else {
920 /* Set header with stream information */
921 data_hdr.stream_id = htobe64(stream->relayd_stream_id);
922 data_hdr.data_size = htobe32(data_size);
923 data_hdr.padding_size = htobe32(padding);
924 /*
925 * Note that net_seq_num below is assigned with the *current* value of
926 * next_net_seq_num and only after that the next_net_seq_num will be
927 * increment. This is why when issuing a command on the relayd using
928 * this next value, 1 should always be substracted in order to compare
929 * the last seen sequence number on the relayd side to the last sent.
930 */
931 data_hdr.net_seq_num = htobe64(stream->next_net_seq_num);
932 /* Other fields are zeroed previously */
933
934 ret = relayd_send_data_hdr(&relayd->data_sock, &data_hdr,
935 sizeof(data_hdr));
936 if (ret < 0) {
937 goto error;
938 }
939
940 ++stream->next_net_seq_num;
941
942 /* Set to go on data socket */
943 outfd = relayd->data_sock.sock.fd;
944 }
945
946 error:
947 return outfd;
948 }
949
950 /*
951 * Allocate and return a new lttng_consumer_channel object using the given key
952 * to initialize the hash table node.
953 *
954 * On error, return NULL.
955 */
956 struct lttng_consumer_channel *consumer_allocate_channel(uint64_t key,
957 uint64_t session_id,
958 const char *pathname,
959 const char *name,
960 uid_t uid,
961 gid_t gid,
962 uint64_t relayd_id,
963 enum lttng_event_output output,
964 uint64_t tracefile_size,
965 uint64_t tracefile_count,
966 uint64_t session_id_per_pid,
967 unsigned int monitor,
968 unsigned int live_timer_interval,
969 const char *root_shm_path,
970 const char *shm_path)
971 {
972 struct lttng_consumer_channel *channel;
973
974 channel = zmalloc(sizeof(*channel));
975 if (channel == NULL) {
976 PERROR("malloc struct lttng_consumer_channel");
977 goto end;
978 }
979
980 channel->key = key;
981 channel->refcount = 0;
982 channel->session_id = session_id;
983 channel->session_id_per_pid = session_id_per_pid;
984 channel->uid = uid;
985 channel->gid = gid;
986 channel->relayd_id = relayd_id;
987 channel->tracefile_size = tracefile_size;
988 channel->tracefile_count = tracefile_count;
989 channel->monitor = monitor;
990 channel->live_timer_interval = live_timer_interval;
991 pthread_mutex_init(&channel->lock, NULL);
992 pthread_mutex_init(&channel->timer_lock, NULL);
993
994 switch (output) {
995 case LTTNG_EVENT_SPLICE:
996 channel->output = CONSUMER_CHANNEL_SPLICE;
997 break;
998 case LTTNG_EVENT_MMAP:
999 channel->output = CONSUMER_CHANNEL_MMAP;
1000 break;
1001 default:
1002 assert(0);
1003 free(channel);
1004 channel = NULL;
1005 goto end;
1006 }
1007
1008 /*
1009 * In monitor mode, the streams associated with the channel will be put in
1010 * a special list ONLY owned by this channel. So, the refcount is set to 1
1011 * here meaning that the channel itself has streams that are referenced.
1012 *
1013 * On a channel deletion, once the channel is no longer visible, the
1014 * refcount is decremented and checked for a zero value to delete it. With
1015 * streams in no monitor mode, it will now be safe to destroy the channel.
1016 */
1017 if (!channel->monitor) {
1018 channel->refcount = 1;
1019 }
1020
1021 strncpy(channel->pathname, pathname, sizeof(channel->pathname));
1022 channel->pathname[sizeof(channel->pathname) - 1] = '\0';
1023
1024 strncpy(channel->name, name, sizeof(channel->name));
1025 channel->name[sizeof(channel->name) - 1] = '\0';
1026
1027 if (root_shm_path) {
1028 strncpy(channel->root_shm_path, root_shm_path, sizeof(channel->root_shm_path));
1029 channel->root_shm_path[sizeof(channel->root_shm_path) - 1] = '\0';
1030 }
1031 if (shm_path) {
1032 strncpy(channel->shm_path, shm_path, sizeof(channel->shm_path));
1033 channel->shm_path[sizeof(channel->shm_path) - 1] = '\0';
1034 }
1035
1036 lttng_ht_node_init_u64(&channel->node, channel->key);
1037
1038 channel->wait_fd = -1;
1039
1040 CDS_INIT_LIST_HEAD(&channel->streams.head);
1041
1042 DBG("Allocated channel (key %" PRIu64 ")", channel->key);
1043
1044 end:
1045 return channel;
1046 }
1047
1048 /*
1049 * Add a channel to the global list protected by a mutex.
1050 *
1051 * Always return 0 indicating success.
1052 */
1053 int consumer_add_channel(struct lttng_consumer_channel *channel,
1054 struct lttng_consumer_local_data *ctx)
1055 {
1056 pthread_mutex_lock(&consumer_data.lock);
1057 pthread_mutex_lock(&channel->lock);
1058 pthread_mutex_lock(&channel->timer_lock);
1059
1060 /*
1061 * This gives us a guarantee that the channel we are about to add to the
1062 * channel hash table will be unique. See this function comment on the why
1063 * we need to steel the channel key at this stage.
1064 */
1065 steal_channel_key(channel->key);
1066
1067 rcu_read_lock();
1068 lttng_ht_add_unique_u64(consumer_data.channel_ht, &channel->node);
1069 rcu_read_unlock();
1070
1071 pthread_mutex_unlock(&channel->timer_lock);
1072 pthread_mutex_unlock(&channel->lock);
1073 pthread_mutex_unlock(&consumer_data.lock);
1074
1075 if (channel->wait_fd != -1 && channel->type == CONSUMER_CHANNEL_TYPE_DATA) {
1076 notify_channel_pipe(ctx, channel, -1, CONSUMER_CHANNEL_ADD);
1077 }
1078
1079 return 0;
1080 }
1081
1082 /*
1083 * Allocate the pollfd structure and the local view of the out fds to avoid
1084 * doing a lookup in the linked list and concurrency issues when writing is
1085 * needed. Called with consumer_data.lock held.
1086 *
1087 * Returns the number of fds in the structures.
1088 */
1089 static int update_poll_array(struct lttng_consumer_local_data *ctx,
1090 struct pollfd **pollfd, struct lttng_consumer_stream **local_stream,
1091 struct lttng_ht *ht, int *nb_inactive_fd)
1092 {
1093 int i = 0;
1094 struct lttng_ht_iter iter;
1095 struct lttng_consumer_stream *stream;
1096
1097 assert(ctx);
1098 assert(ht);
1099 assert(pollfd);
1100 assert(local_stream);
1101
1102 DBG("Updating poll fd array");
1103 *nb_inactive_fd = 0;
1104 rcu_read_lock();
1105 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
1106 /*
1107 * Only active streams with an active end point can be added to the
1108 * poll set and local stream storage of the thread.
1109 *
1110 * There is a potential race here for endpoint_status to be updated
1111 * just after the check. However, this is OK since the stream(s) will
1112 * be deleted once the thread is notified that the end point state has
1113 * changed where this function will be called back again.
1114 *
1115 * We track the number of inactive FDs because they still need to be
1116 * closed by the polling thread after a wakeup on the data_pipe or
1117 * metadata_pipe.
1118 */
1119 if (stream->state != LTTNG_CONSUMER_ACTIVE_STREAM ||
1120 stream->endpoint_status == CONSUMER_ENDPOINT_INACTIVE) {
1121 (*nb_inactive_fd)++;
1122 continue;
1123 }
1124 /*
1125 * This clobbers way too much the debug output. Uncomment that if you
1126 * need it for debugging purposes.
1127 *
1128 * DBG("Active FD %d", stream->wait_fd);
1129 */
1130 (*pollfd)[i].fd = stream->wait_fd;
1131 (*pollfd)[i].events = POLLIN | POLLPRI;
1132 local_stream[i] = stream;
1133 i++;
1134 }
1135 rcu_read_unlock();
1136
1137 /*
1138 * Insert the consumer_data_pipe at the end of the array and don't
1139 * increment i so nb_fd is the number of real FD.
1140 */
1141 (*pollfd)[i].fd = lttng_pipe_get_readfd(ctx->consumer_data_pipe);
1142 (*pollfd)[i].events = POLLIN | POLLPRI;
1143
1144 (*pollfd)[i + 1].fd = lttng_pipe_get_readfd(ctx->consumer_wakeup_pipe);
1145 (*pollfd)[i + 1].events = POLLIN | POLLPRI;
1146 return i;
1147 }
1148
1149 /*
1150 * Poll on the should_quit pipe and the command socket return -1 on
1151 * error, 1 if should exit, 0 if data is available on the command socket
1152 */
1153 int lttng_consumer_poll_socket(struct pollfd *consumer_sockpoll)
1154 {
1155 int num_rdy;
1156
1157 restart:
1158 num_rdy = poll(consumer_sockpoll, 2, -1);
1159 if (num_rdy == -1) {
1160 /*
1161 * Restart interrupted system call.
1162 */
1163 if (errno == EINTR) {
1164 goto restart;
1165 }
1166 PERROR("Poll error");
1167 return -1;
1168 }
1169 if (consumer_sockpoll[0].revents & (POLLIN | POLLPRI)) {
1170 DBG("consumer_should_quit wake up");
1171 return 1;
1172 }
1173 return 0;
1174 }
1175
1176 /*
1177 * Set the error socket.
1178 */
1179 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data *ctx,
1180 int sock)
1181 {
1182 ctx->consumer_error_socket = sock;
1183 }
1184
1185 /*
1186 * Set the command socket path.
1187 */
1188 void lttng_consumer_set_command_sock_path(
1189 struct lttng_consumer_local_data *ctx, char *sock)
1190 {
1191 ctx->consumer_command_sock_path = sock;
1192 }
1193
1194 /*
1195 * Send return code to the session daemon.
1196 * If the socket is not defined, we return 0, it is not a fatal error
1197 */
1198 int lttng_consumer_send_error(struct lttng_consumer_local_data *ctx, int cmd)
1199 {
1200 if (ctx->consumer_error_socket > 0) {
1201 return lttcomm_send_unix_sock(ctx->consumer_error_socket, &cmd,
1202 sizeof(enum lttcomm_sessiond_command));
1203 }
1204
1205 return 0;
1206 }
1207
1208 /*
1209 * Close all the tracefiles and stream fds and MUST be called when all
1210 * instances are destroyed i.e. when all threads were joined and are ended.
1211 */
1212 void lttng_consumer_cleanup(void)
1213 {
1214 struct lttng_ht_iter iter;
1215 struct lttng_consumer_channel *channel;
1216
1217 rcu_read_lock();
1218
1219 cds_lfht_for_each_entry(consumer_data.channel_ht->ht, &iter.iter, channel,
1220 node.node) {
1221 consumer_del_channel(channel);
1222 }
1223
1224 rcu_read_unlock();
1225
1226 lttng_ht_destroy(consumer_data.channel_ht);
1227
1228 cleanup_relayd_ht();
1229
1230 lttng_ht_destroy(consumer_data.stream_per_chan_id_ht);
1231
1232 /*
1233 * This HT contains streams that are freed by either the metadata thread or
1234 * the data thread so we do *nothing* on the hash table and simply destroy
1235 * it.
1236 */
1237 lttng_ht_destroy(consumer_data.stream_list_ht);
1238 }
1239
1240 /*
1241 * Called from signal handler.
1242 */
1243 void lttng_consumer_should_exit(struct lttng_consumer_local_data *ctx)
1244 {
1245 ssize_t ret;
1246
1247 CMM_STORE_SHARED(consumer_quit, 1);
1248 ret = lttng_write(ctx->consumer_should_quit[1], "4", 1);
1249 if (ret < 1) {
1250 PERROR("write consumer quit");
1251 }
1252
1253 DBG("Consumer flag that it should quit");
1254 }
1255
1256
1257 /*
1258 * Flush pending writes to trace output disk file.
1259 */
1260 static
1261 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream *stream,
1262 off_t orig_offset)
1263 {
1264 int ret;
1265 int outfd = stream->out_fd;
1266
1267 /*
1268 * This does a blocking write-and-wait on any page that belongs to the
1269 * subbuffer prior to the one we just wrote.
1270 * Don't care about error values, as these are just hints and ways to
1271 * limit the amount of page cache used.
1272 */
1273 if (orig_offset < stream->max_sb_size) {
1274 return;
1275 }
1276 lttng_sync_file_range(outfd, orig_offset - stream->max_sb_size,
1277 stream->max_sb_size,
1278 SYNC_FILE_RANGE_WAIT_BEFORE
1279 | SYNC_FILE_RANGE_WRITE
1280 | SYNC_FILE_RANGE_WAIT_AFTER);
1281 /*
1282 * Give hints to the kernel about how we access the file:
1283 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1284 * we write it.
1285 *
1286 * We need to call fadvise again after the file grows because the
1287 * kernel does not seem to apply fadvise to non-existing parts of the
1288 * file.
1289 *
1290 * Call fadvise _after_ having waited for the page writeback to
1291 * complete because the dirty page writeback semantic is not well
1292 * defined. So it can be expected to lead to lower throughput in
1293 * streaming.
1294 */
1295 ret = posix_fadvise(outfd, orig_offset - stream->max_sb_size,
1296 stream->max_sb_size, POSIX_FADV_DONTNEED);
1297 if (ret && ret != -ENOSYS) {
1298 errno = ret;
1299 PERROR("posix_fadvise on fd %i", outfd);
1300 }
1301 }
1302
1303 /*
1304 * Initialise the necessary environnement :
1305 * - create a new context
1306 * - create the poll_pipe
1307 * - create the should_quit pipe (for signal handler)
1308 * - create the thread pipe (for splice)
1309 *
1310 * Takes a function pointer as argument, this function is called when data is
1311 * available on a buffer. This function is responsible to do the
1312 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1313 * buffer configuration and then kernctl_put_next_subbuf at the end.
1314 *
1315 * Returns a pointer to the new context or NULL on error.
1316 */
1317 struct lttng_consumer_local_data *lttng_consumer_create(
1318 enum lttng_consumer_type type,
1319 ssize_t (*buffer_ready)(struct lttng_consumer_stream *stream,
1320 struct lttng_consumer_local_data *ctx),
1321 int (*recv_channel)(struct lttng_consumer_channel *channel),
1322 int (*recv_stream)(struct lttng_consumer_stream *stream),
1323 int (*update_stream)(uint64_t stream_key, uint32_t state))
1324 {
1325 int ret;
1326 struct lttng_consumer_local_data *ctx;
1327
1328 assert(consumer_data.type == LTTNG_CONSUMER_UNKNOWN ||
1329 consumer_data.type == type);
1330 consumer_data.type = type;
1331
1332 ctx = zmalloc(sizeof(struct lttng_consumer_local_data));
1333 if (ctx == NULL) {
1334 PERROR("allocating context");
1335 goto error;
1336 }
1337
1338 ctx->consumer_error_socket = -1;
1339 ctx->consumer_metadata_socket = -1;
1340 pthread_mutex_init(&ctx->metadata_socket_lock, NULL);
1341 /* assign the callbacks */
1342 ctx->on_buffer_ready = buffer_ready;
1343 ctx->on_recv_channel = recv_channel;
1344 ctx->on_recv_stream = recv_stream;
1345 ctx->on_update_stream = update_stream;
1346
1347 ctx->consumer_data_pipe = lttng_pipe_open(0);
1348 if (!ctx->consumer_data_pipe) {
1349 goto error_poll_pipe;
1350 }
1351
1352 ctx->consumer_wakeup_pipe = lttng_pipe_open(0);
1353 if (!ctx->consumer_wakeup_pipe) {
1354 goto error_wakeup_pipe;
1355 }
1356
1357 ret = pipe(ctx->consumer_should_quit);
1358 if (ret < 0) {
1359 PERROR("Error creating recv pipe");
1360 goto error_quit_pipe;
1361 }
1362
1363 ret = pipe(ctx->consumer_channel_pipe);
1364 if (ret < 0) {
1365 PERROR("Error creating channel pipe");
1366 goto error_channel_pipe;
1367 }
1368
1369 ctx->consumer_metadata_pipe = lttng_pipe_open(0);
1370 if (!ctx->consumer_metadata_pipe) {
1371 goto error_metadata_pipe;
1372 }
1373
1374 ctx->channel_monitor_pipe = -1;
1375
1376 return ctx;
1377
1378 error_metadata_pipe:
1379 utils_close_pipe(ctx->consumer_channel_pipe);
1380 error_channel_pipe:
1381 utils_close_pipe(ctx->consumer_should_quit);
1382 error_quit_pipe:
1383 lttng_pipe_destroy(ctx->consumer_wakeup_pipe);
1384 error_wakeup_pipe:
1385 lttng_pipe_destroy(ctx->consumer_data_pipe);
1386 error_poll_pipe:
1387 free(ctx);
1388 error:
1389 return NULL;
1390 }
1391
1392 /*
1393 * Iterate over all streams of the hashtable and free them properly.
1394 */
1395 static void destroy_data_stream_ht(struct lttng_ht *ht)
1396 {
1397 struct lttng_ht_iter iter;
1398 struct lttng_consumer_stream *stream;
1399
1400 if (ht == NULL) {
1401 return;
1402 }
1403
1404 rcu_read_lock();
1405 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
1406 /*
1407 * Ignore return value since we are currently cleaning up so any error
1408 * can't be handled.
1409 */
1410 (void) consumer_del_stream(stream, ht);
1411 }
1412 rcu_read_unlock();
1413
1414 lttng_ht_destroy(ht);
1415 }
1416
1417 /*
1418 * Iterate over all streams of the metadata hashtable and free them
1419 * properly.
1420 */
1421 static void destroy_metadata_stream_ht(struct lttng_ht *ht)
1422 {
1423 struct lttng_ht_iter iter;
1424 struct lttng_consumer_stream *stream;
1425
1426 if (ht == NULL) {
1427 return;
1428 }
1429
1430 rcu_read_lock();
1431 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
1432 /*
1433 * Ignore return value since we are currently cleaning up so any error
1434 * can't be handled.
1435 */
1436 (void) consumer_del_metadata_stream(stream, ht);
1437 }
1438 rcu_read_unlock();
1439
1440 lttng_ht_destroy(ht);
1441 }
1442
1443 /*
1444 * Close all fds associated with the instance and free the context.
1445 */
1446 void lttng_consumer_destroy(struct lttng_consumer_local_data *ctx)
1447 {
1448 int ret;
1449
1450 DBG("Consumer destroying it. Closing everything.");
1451
1452 if (!ctx) {
1453 return;
1454 }
1455
1456 destroy_data_stream_ht(data_ht);
1457 destroy_metadata_stream_ht(metadata_ht);
1458
1459 ret = close(ctx->consumer_error_socket);
1460 if (ret) {
1461 PERROR("close");
1462 }
1463 ret = close(ctx->consumer_metadata_socket);
1464 if (ret) {
1465 PERROR("close");
1466 }
1467 utils_close_pipe(ctx->consumer_channel_pipe);
1468 lttng_pipe_destroy(ctx->consumer_data_pipe);
1469 lttng_pipe_destroy(ctx->consumer_metadata_pipe);
1470 lttng_pipe_destroy(ctx->consumer_wakeup_pipe);
1471 utils_close_pipe(ctx->consumer_should_quit);
1472
1473 unlink(ctx->consumer_command_sock_path);
1474 free(ctx);
1475 }
1476
1477 /*
1478 * Write the metadata stream id on the specified file descriptor.
1479 */
1480 static int write_relayd_metadata_id(int fd,
1481 struct lttng_consumer_stream *stream,
1482 unsigned long padding)
1483 {
1484 ssize_t ret;
1485 struct lttcomm_relayd_metadata_payload hdr;
1486
1487 hdr.stream_id = htobe64(stream->relayd_stream_id);
1488 hdr.padding_size = htobe32(padding);
1489 ret = lttng_write(fd, (void *) &hdr, sizeof(hdr));
1490 if (ret < sizeof(hdr)) {
1491 /*
1492 * This error means that the fd's end is closed so ignore the PERROR
1493 * not to clubber the error output since this can happen in a normal
1494 * code path.
1495 */
1496 if (errno != EPIPE) {
1497 PERROR("write metadata stream id");
1498 }
1499 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno);
1500 /*
1501 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1502 * handle writting the missing part so report that as an error and
1503 * don't lie to the caller.
1504 */
1505 ret = -1;
1506 goto end;
1507 }
1508 DBG("Metadata stream id %" PRIu64 " with padding %lu written before data",
1509 stream->relayd_stream_id, padding);
1510
1511 end:
1512 return (int) ret;
1513 }
1514
1515 /*
1516 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1517 * core function for writing trace buffers to either the local filesystem or
1518 * the network.
1519 *
1520 * It must be called with the stream lock held.
1521 *
1522 * Careful review MUST be put if any changes occur!
1523 *
1524 * Returns the number of bytes written
1525 */
1526 ssize_t lttng_consumer_on_read_subbuffer_mmap(
1527 struct lttng_consumer_local_data *ctx,
1528 struct lttng_consumer_stream *stream, unsigned long len,
1529 unsigned long padding,
1530 struct ctf_packet_index *index)
1531 {
1532 unsigned long mmap_offset;
1533 void *mmap_base;
1534 ssize_t ret = 0;
1535 off_t orig_offset = stream->out_fd_offset;
1536 /* Default is on the disk */
1537 int outfd = stream->out_fd;
1538 struct consumer_relayd_sock_pair *relayd = NULL;
1539 unsigned int relayd_hang_up = 0;
1540
1541 /* RCU lock for the relayd pointer */
1542 rcu_read_lock();
1543
1544 /* Flag that the current stream if set for network streaming. */
1545 if (stream->net_seq_idx != (uint64_t) -1ULL) {
1546 relayd = consumer_find_relayd(stream->net_seq_idx);
1547 if (relayd == NULL) {
1548 ret = -EPIPE;
1549 goto end;
1550 }
1551 }
1552
1553 /* get the offset inside the fd to mmap */
1554 switch (consumer_data.type) {
1555 case LTTNG_CONSUMER_KERNEL:
1556 mmap_base = stream->mmap_base;
1557 ret = kernctl_get_mmap_read_offset(stream->wait_fd, &mmap_offset);
1558 if (ret < 0) {
1559 PERROR("tracer ctl get_mmap_read_offset");
1560 goto end;
1561 }
1562 break;
1563 case LTTNG_CONSUMER32_UST:
1564 case LTTNG_CONSUMER64_UST:
1565 mmap_base = lttng_ustctl_get_mmap_base(stream);
1566 if (!mmap_base) {
1567 ERR("read mmap get mmap base for stream %s", stream->name);
1568 ret = -EPERM;
1569 goto end;
1570 }
1571 ret = lttng_ustctl_get_mmap_read_offset(stream, &mmap_offset);
1572 if (ret != 0) {
1573 PERROR("tracer ctl get_mmap_read_offset");
1574 ret = -EINVAL;
1575 goto end;
1576 }
1577 break;
1578 default:
1579 ERR("Unknown consumer_data type");
1580 assert(0);
1581 }
1582
1583 /* Handle stream on the relayd if the output is on the network */
1584 if (relayd) {
1585 unsigned long netlen = len;
1586
1587 /*
1588 * Lock the control socket for the complete duration of the function
1589 * since from this point on we will use the socket.
1590 */
1591 if (stream->metadata_flag) {
1592 /* Metadata requires the control socket. */
1593 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
1594 if (stream->reset_metadata_flag) {
1595 ret = relayd_reset_metadata(&relayd->control_sock,
1596 stream->relayd_stream_id,
1597 stream->metadata_version);
1598 if (ret < 0) {
1599 relayd_hang_up = 1;
1600 goto write_error;
1601 }
1602 stream->reset_metadata_flag = 0;
1603 }
1604 netlen += sizeof(struct lttcomm_relayd_metadata_payload);
1605 }
1606
1607 ret = write_relayd_stream_header(stream, netlen, padding, relayd);
1608 if (ret < 0) {
1609 relayd_hang_up = 1;
1610 goto write_error;
1611 }
1612 /* Use the returned socket. */
1613 outfd = ret;
1614
1615 /* Write metadata stream id before payload */
1616 if (stream->metadata_flag) {
1617 ret = write_relayd_metadata_id(outfd, stream, padding);
1618 if (ret < 0) {
1619 relayd_hang_up = 1;
1620 goto write_error;
1621 }
1622 }
1623 } else {
1624 /* No streaming, we have to set the len with the full padding */
1625 len += padding;
1626
1627 if (stream->metadata_flag && stream->reset_metadata_flag) {
1628 ret = utils_truncate_stream_file(stream->out_fd, 0);
1629 if (ret < 0) {
1630 ERR("Reset metadata file");
1631 goto end;
1632 }
1633 stream->reset_metadata_flag = 0;
1634 }
1635
1636 /*
1637 * Check if we need to change the tracefile before writing the packet.
1638 */
1639 if (stream->chan->tracefile_size > 0 &&
1640 (stream->tracefile_size_current + len) >
1641 stream->chan->tracefile_size) {
1642 ret = utils_rotate_stream_file(stream->chan->pathname,
1643 stream->name, stream->chan->tracefile_size,
1644 stream->chan->tracefile_count, stream->uid, stream->gid,
1645 stream->out_fd, &(stream->tracefile_count_current),
1646 &stream->out_fd);
1647 if (ret < 0) {
1648 ERR("Rotating output file");
1649 goto end;
1650 }
1651 outfd = stream->out_fd;
1652
1653 if (stream->index_file) {
1654 lttng_index_file_put(stream->index_file);
1655 stream->index_file = lttng_index_file_create(stream->chan->pathname,
1656 stream->name, stream->uid, stream->gid,
1657 stream->chan->tracefile_size,
1658 stream->tracefile_count_current,
1659 CTF_INDEX_MAJOR, CTF_INDEX_MINOR);
1660 if (!stream->index_file) {
1661 goto end;
1662 }
1663 }
1664
1665 /* Reset current size because we just perform a rotation. */
1666 stream->tracefile_size_current = 0;
1667 stream->out_fd_offset = 0;
1668 orig_offset = 0;
1669 }
1670 stream->tracefile_size_current += len;
1671 if (index) {
1672 index->offset = htobe64(stream->out_fd_offset);
1673 }
1674 }
1675
1676 /*
1677 * This call guarantee that len or less is returned. It's impossible to
1678 * receive a ret value that is bigger than len.
1679 */
1680 ret = lttng_write(outfd, mmap_base + mmap_offset, len);
1681 DBG("Consumer mmap write() ret %zd (len %lu)", ret, len);
1682 if (ret < 0 || ((size_t) ret != len)) {
1683 /*
1684 * Report error to caller if nothing was written else at least send the
1685 * amount written.
1686 */
1687 if (ret < 0) {
1688 ret = -errno;
1689 }
1690 relayd_hang_up = 1;
1691
1692 /* Socket operation failed. We consider the relayd dead */
1693 if (errno == EPIPE || errno == EINVAL || errno == EBADF) {
1694 /*
1695 * This is possible if the fd is closed on the other side
1696 * (outfd) or any write problem. It can be verbose a bit for a
1697 * normal execution if for instance the relayd is stopped
1698 * abruptly. This can happen so set this to a DBG statement.
1699 */
1700 DBG("Consumer mmap write detected relayd hang up");
1701 } else {
1702 /* Unhandled error, print it and stop function right now. */
1703 PERROR("Error in write mmap (ret %zd != len %lu)", ret, len);
1704 }
1705 goto write_error;
1706 }
1707 stream->output_written += ret;
1708
1709 /* This call is useless on a socket so better save a syscall. */
1710 if (!relayd) {
1711 /* This won't block, but will start writeout asynchronously */
1712 lttng_sync_file_range(outfd, stream->out_fd_offset, len,
1713 SYNC_FILE_RANGE_WRITE);
1714 stream->out_fd_offset += len;
1715 lttng_consumer_sync_trace_file(stream, orig_offset);
1716 }
1717
1718 write_error:
1719 /*
1720 * This is a special case that the relayd has closed its socket. Let's
1721 * cleanup the relayd object and all associated streams.
1722 */
1723 if (relayd && relayd_hang_up) {
1724 cleanup_relayd(relayd, ctx);
1725 }
1726
1727 end:
1728 /* Unlock only if ctrl socket used */
1729 if (relayd && stream->metadata_flag) {
1730 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
1731 }
1732
1733 rcu_read_unlock();
1734 return ret;
1735 }
1736
1737 /*
1738 * Splice the data from the ring buffer to the tracefile.
1739 *
1740 * It must be called with the stream lock held.
1741 *
1742 * Returns the number of bytes spliced.
1743 */
1744 ssize_t lttng_consumer_on_read_subbuffer_splice(
1745 struct lttng_consumer_local_data *ctx,
1746 struct lttng_consumer_stream *stream, unsigned long len,
1747 unsigned long padding,
1748 struct ctf_packet_index *index)
1749 {
1750 ssize_t ret = 0, written = 0, ret_splice = 0;
1751 loff_t offset = 0;
1752 off_t orig_offset = stream->out_fd_offset;
1753 int fd = stream->wait_fd;
1754 /* Default is on the disk */
1755 int outfd = stream->out_fd;
1756 struct consumer_relayd_sock_pair *relayd = NULL;
1757 int *splice_pipe;
1758 unsigned int relayd_hang_up = 0;
1759
1760 switch (consumer_data.type) {
1761 case LTTNG_CONSUMER_KERNEL:
1762 break;
1763 case LTTNG_CONSUMER32_UST:
1764 case LTTNG_CONSUMER64_UST:
1765 /* Not supported for user space tracing */
1766 return -ENOSYS;
1767 default:
1768 ERR("Unknown consumer_data type");
1769 assert(0);
1770 }
1771
1772 /* RCU lock for the relayd pointer */
1773 rcu_read_lock();
1774
1775 /* Flag that the current stream if set for network streaming. */
1776 if (stream->net_seq_idx != (uint64_t) -1ULL) {
1777 relayd = consumer_find_relayd(stream->net_seq_idx);
1778 if (relayd == NULL) {
1779 written = -ret;
1780 goto end;
1781 }
1782 }
1783 splice_pipe = stream->splice_pipe;
1784
1785 /* Write metadata stream id before payload */
1786 if (relayd) {
1787 unsigned long total_len = len;
1788
1789 if (stream->metadata_flag) {
1790 /*
1791 * Lock the control socket for the complete duration of the function
1792 * since from this point on we will use the socket.
1793 */
1794 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
1795
1796 if (stream->reset_metadata_flag) {
1797 ret = relayd_reset_metadata(&relayd->control_sock,
1798 stream->relayd_stream_id,
1799 stream->metadata_version);
1800 if (ret < 0) {
1801 relayd_hang_up = 1;
1802 goto write_error;
1803 }
1804 stream->reset_metadata_flag = 0;
1805 }
1806 ret = write_relayd_metadata_id(splice_pipe[1], stream,
1807 padding);
1808 if (ret < 0) {
1809 written = ret;
1810 relayd_hang_up = 1;
1811 goto write_error;
1812 }
1813
1814 total_len += sizeof(struct lttcomm_relayd_metadata_payload);
1815 }
1816
1817 ret = write_relayd_stream_header(stream, total_len, padding, relayd);
1818 if (ret < 0) {
1819 written = ret;
1820 relayd_hang_up = 1;
1821 goto write_error;
1822 }
1823 /* Use the returned socket. */
1824 outfd = ret;
1825 } else {
1826 /* No streaming, we have to set the len with the full padding */
1827 len += padding;
1828
1829 if (stream->metadata_flag && stream->reset_metadata_flag) {
1830 ret = utils_truncate_stream_file(stream->out_fd, 0);
1831 if (ret < 0) {
1832 ERR("Reset metadata file");
1833 goto end;
1834 }
1835 stream->reset_metadata_flag = 0;
1836 }
1837 /*
1838 * Check if we need to change the tracefile before writing the packet.
1839 */
1840 if (stream->chan->tracefile_size > 0 &&
1841 (stream->tracefile_size_current + len) >
1842 stream->chan->tracefile_size) {
1843 ret = utils_rotate_stream_file(stream->chan->pathname,
1844 stream->name, stream->chan->tracefile_size,
1845 stream->chan->tracefile_count, stream->uid, stream->gid,
1846 stream->out_fd, &(stream->tracefile_count_current),
1847 &stream->out_fd);
1848 if (ret < 0) {
1849 written = ret;
1850 ERR("Rotating output file");
1851 goto end;
1852 }
1853 outfd = stream->out_fd;
1854
1855 if (stream->index_file) {
1856 lttng_index_file_put(stream->index_file);
1857 stream->index_file = lttng_index_file_create(stream->chan->pathname,
1858 stream->name, stream->uid, stream->gid,
1859 stream->chan->tracefile_size,
1860 stream->tracefile_count_current,
1861 CTF_INDEX_MAJOR, CTF_INDEX_MINOR);
1862 if (!stream->index_file) {
1863 goto end;
1864 }
1865 }
1866
1867 /* Reset current size because we just perform a rotation. */
1868 stream->tracefile_size_current = 0;
1869 stream->out_fd_offset = 0;
1870 orig_offset = 0;
1871 }
1872 stream->tracefile_size_current += len;
1873 index->offset = htobe64(stream->out_fd_offset);
1874 }
1875
1876 while (len > 0) {
1877 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1878 (unsigned long)offset, len, fd, splice_pipe[1]);
1879 ret_splice = splice(fd, &offset, splice_pipe[1], NULL, len,
1880 SPLICE_F_MOVE | SPLICE_F_MORE);
1881 DBG("splice chan to pipe, ret %zd", ret_splice);
1882 if (ret_splice < 0) {
1883 ret = errno;
1884 written = -ret;
1885 PERROR("Error in relay splice");
1886 goto splice_error;
1887 }
1888
1889 /* Handle stream on the relayd if the output is on the network */
1890 if (relayd && stream->metadata_flag) {
1891 size_t metadata_payload_size =
1892 sizeof(struct lttcomm_relayd_metadata_payload);
1893
1894 /* Update counter to fit the spliced data */
1895 ret_splice += metadata_payload_size;
1896 len += metadata_payload_size;
1897 /*
1898 * We do this so the return value can match the len passed as
1899 * argument to this function.
1900 */
1901 written -= metadata_payload_size;
1902 }
1903
1904 /* Splice data out */
1905 ret_splice = splice(splice_pipe[0], NULL, outfd, NULL,
1906 ret_splice, SPLICE_F_MOVE | SPLICE_F_MORE);
1907 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
1908 outfd, ret_splice);
1909 if (ret_splice < 0) {
1910 ret = errno;
1911 written = -ret;
1912 relayd_hang_up = 1;
1913 goto write_error;
1914 } else if (ret_splice > len) {
1915 /*
1916 * We don't expect this code path to be executed but you never know
1917 * so this is an extra protection agains a buggy splice().
1918 */
1919 ret = errno;
1920 written += ret_splice;
1921 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice,
1922 len);
1923 goto splice_error;
1924 } else {
1925 /* All good, update current len and continue. */
1926 len -= ret_splice;
1927 }
1928
1929 /* This call is useless on a socket so better save a syscall. */
1930 if (!relayd) {
1931 /* This won't block, but will start writeout asynchronously */
1932 lttng_sync_file_range(outfd, stream->out_fd_offset, ret_splice,
1933 SYNC_FILE_RANGE_WRITE);
1934 stream->out_fd_offset += ret_splice;
1935 }
1936 stream->output_written += ret_splice;
1937 written += ret_splice;
1938 }
1939 if (!relayd) {
1940 lttng_consumer_sync_trace_file(stream, orig_offset);
1941 }
1942 goto end;
1943
1944 write_error:
1945 /*
1946 * This is a special case that the relayd has closed its socket. Let's
1947 * cleanup the relayd object and all associated streams.
1948 */
1949 if (relayd && relayd_hang_up) {
1950 cleanup_relayd(relayd, ctx);
1951 /* Skip splice error so the consumer does not fail */
1952 goto end;
1953 }
1954
1955 splice_error:
1956 /* send the appropriate error description to sessiond */
1957 switch (ret) {
1958 case EINVAL:
1959 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_EINVAL);
1960 break;
1961 case ENOMEM:
1962 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_ENOMEM);
1963 break;
1964 case ESPIPE:
1965 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_ESPIPE);
1966 break;
1967 }
1968
1969 end:
1970 if (relayd && stream->metadata_flag) {
1971 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
1972 }
1973
1974 rcu_read_unlock();
1975 return written;
1976 }
1977
1978 /*
1979 * Sample the snapshot positions for a specific fd
1980 *
1981 * Returns 0 on success, < 0 on error
1982 */
1983 int lttng_consumer_sample_snapshot_positions(struct lttng_consumer_stream *stream)
1984 {
1985 switch (consumer_data.type) {
1986 case LTTNG_CONSUMER_KERNEL:
1987 return lttng_kconsumer_sample_snapshot_positions(stream);
1988 case LTTNG_CONSUMER32_UST:
1989 case LTTNG_CONSUMER64_UST:
1990 return lttng_ustconsumer_sample_snapshot_positions(stream);
1991 default:
1992 ERR("Unknown consumer_data type");
1993 assert(0);
1994 return -ENOSYS;
1995 }
1996 }
1997 /*
1998 * Take a snapshot for a specific fd
1999 *
2000 * Returns 0 on success, < 0 on error
2001 */
2002 int lttng_consumer_take_snapshot(struct lttng_consumer_stream *stream)
2003 {
2004 switch (consumer_data.type) {
2005 case LTTNG_CONSUMER_KERNEL:
2006 return lttng_kconsumer_take_snapshot(stream);
2007 case LTTNG_CONSUMER32_UST:
2008 case LTTNG_CONSUMER64_UST:
2009 return lttng_ustconsumer_take_snapshot(stream);
2010 default:
2011 ERR("Unknown consumer_data type");
2012 assert(0);
2013 return -ENOSYS;
2014 }
2015 }
2016
2017 /*
2018 * Get the produced position
2019 *
2020 * Returns 0 on success, < 0 on error
2021 */
2022 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream *stream,
2023 unsigned long *pos)
2024 {
2025 switch (consumer_data.type) {
2026 case LTTNG_CONSUMER_KERNEL:
2027 return lttng_kconsumer_get_produced_snapshot(stream, pos);
2028 case LTTNG_CONSUMER32_UST:
2029 case LTTNG_CONSUMER64_UST:
2030 return lttng_ustconsumer_get_produced_snapshot(stream, pos);
2031 default:
2032 ERR("Unknown consumer_data type");
2033 assert(0);
2034 return -ENOSYS;
2035 }
2036 }
2037
2038 /*
2039 * Get the consumed position (free-running counter position in bytes).
2040 *
2041 * Returns 0 on success, < 0 on error
2042 */
2043 int lttng_consumer_get_consumed_snapshot(struct lttng_consumer_stream *stream,
2044 unsigned long *pos)
2045 {
2046 switch (consumer_data.type) {
2047 case LTTNG_CONSUMER_KERNEL:
2048 return lttng_kconsumer_get_consumed_snapshot(stream, pos);
2049 case LTTNG_CONSUMER32_UST:
2050 case LTTNG_CONSUMER64_UST:
2051 return lttng_ustconsumer_get_consumed_snapshot(stream, pos);
2052 default:
2053 ERR("Unknown consumer_data type");
2054 assert(0);
2055 return -ENOSYS;
2056 }
2057 }
2058
2059 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data *ctx,
2060 int sock, struct pollfd *consumer_sockpoll)
2061 {
2062 switch (consumer_data.type) {
2063 case LTTNG_CONSUMER_KERNEL:
2064 return lttng_kconsumer_recv_cmd(ctx, sock, consumer_sockpoll);
2065 case LTTNG_CONSUMER32_UST:
2066 case LTTNG_CONSUMER64_UST:
2067 return lttng_ustconsumer_recv_cmd(ctx, sock, consumer_sockpoll);
2068 default:
2069 ERR("Unknown consumer_data type");
2070 assert(0);
2071 return -ENOSYS;
2072 }
2073 }
2074
2075 void lttng_consumer_close_all_metadata(void)
2076 {
2077 switch (consumer_data.type) {
2078 case LTTNG_CONSUMER_KERNEL:
2079 /*
2080 * The Kernel consumer has a different metadata scheme so we don't
2081 * close anything because the stream will be closed by the session
2082 * daemon.
2083 */
2084 break;
2085 case LTTNG_CONSUMER32_UST:
2086 case LTTNG_CONSUMER64_UST:
2087 /*
2088 * Close all metadata streams. The metadata hash table is passed and
2089 * this call iterates over it by closing all wakeup fd. This is safe
2090 * because at this point we are sure that the metadata producer is
2091 * either dead or blocked.
2092 */
2093 lttng_ustconsumer_close_all_metadata(metadata_ht);
2094 break;
2095 default:
2096 ERR("Unknown consumer_data type");
2097 assert(0);
2098 }
2099 }
2100
2101 /*
2102 * Clean up a metadata stream and free its memory.
2103 */
2104 void consumer_del_metadata_stream(struct lttng_consumer_stream *stream,
2105 struct lttng_ht *ht)
2106 {
2107 struct lttng_consumer_channel *free_chan = NULL;
2108
2109 assert(stream);
2110 /*
2111 * This call should NEVER receive regular stream. It must always be
2112 * metadata stream and this is crucial for data structure synchronization.
2113 */
2114 assert(stream->metadata_flag);
2115
2116 DBG3("Consumer delete metadata stream %d", stream->wait_fd);
2117
2118 pthread_mutex_lock(&consumer_data.lock);
2119 pthread_mutex_lock(&stream->chan->lock);
2120 pthread_mutex_lock(&stream->lock);
2121 if (stream->chan->metadata_cache) {
2122 /* Only applicable to userspace consumers. */
2123 pthread_mutex_lock(&stream->chan->metadata_cache->lock);
2124 }
2125
2126 /* Remove any reference to that stream. */
2127 consumer_stream_delete(stream, ht);
2128
2129 /* Close down everything including the relayd if one. */
2130 consumer_stream_close(stream);
2131 /* Destroy tracer buffers of the stream. */
2132 consumer_stream_destroy_buffers(stream);
2133
2134 /* Atomically decrement channel refcount since other threads can use it. */
2135 if (!uatomic_sub_return(&stream->chan->refcount, 1)
2136 && !uatomic_read(&stream->chan->nb_init_stream_left)) {
2137 /* Go for channel deletion! */
2138 free_chan = stream->chan;
2139 }
2140
2141 /*
2142 * Nullify the stream reference so it is not used after deletion. The
2143 * channel lock MUST be acquired before being able to check for a NULL
2144 * pointer value.
2145 */
2146 stream->chan->metadata_stream = NULL;
2147
2148 if (stream->chan->metadata_cache) {
2149 pthread_mutex_unlock(&stream->chan->metadata_cache->lock);
2150 }
2151 pthread_mutex_unlock(&stream->lock);
2152 pthread_mutex_unlock(&stream->chan->lock);
2153 pthread_mutex_unlock(&consumer_data.lock);
2154
2155 if (free_chan) {
2156 consumer_del_channel(free_chan);
2157 }
2158
2159 consumer_stream_free(stream);
2160 }
2161
2162 /*
2163 * Action done with the metadata stream when adding it to the consumer internal
2164 * data structures to handle it.
2165 */
2166 void consumer_add_metadata_stream(struct lttng_consumer_stream *stream)
2167 {
2168 struct lttng_ht *ht = metadata_ht;
2169 struct lttng_ht_iter iter;
2170 struct lttng_ht_node_u64 *node;
2171
2172 assert(stream);
2173 assert(ht);
2174
2175 DBG3("Adding metadata stream %" PRIu64 " to hash table", stream->key);
2176
2177 pthread_mutex_lock(&consumer_data.lock);
2178 pthread_mutex_lock(&stream->chan->lock);
2179 pthread_mutex_lock(&stream->chan->timer_lock);
2180 pthread_mutex_lock(&stream->lock);
2181
2182 /*
2183 * From here, refcounts are updated so be _careful_ when returning an error
2184 * after this point.
2185 */
2186
2187 rcu_read_lock();
2188
2189 /*
2190 * Lookup the stream just to make sure it does not exist in our internal
2191 * state. This should NEVER happen.
2192 */
2193 lttng_ht_lookup(ht, &stream->key, &iter);
2194 node = lttng_ht_iter_get_node_u64(&iter);
2195 assert(!node);
2196
2197 /*
2198 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2199 * in terms of destroying the associated channel, because the action that
2200 * causes the count to become 0 also causes a stream to be added. The
2201 * channel deletion will thus be triggered by the following removal of this
2202 * stream.
2203 */
2204 if (uatomic_read(&stream->chan->nb_init_stream_left) > 0) {
2205 /* Increment refcount before decrementing nb_init_stream_left */
2206 cmm_smp_wmb();
2207 uatomic_dec(&stream->chan->nb_init_stream_left);
2208 }
2209
2210 lttng_ht_add_unique_u64(ht, &stream->node);
2211
2212 lttng_ht_add_u64(consumer_data.stream_per_chan_id_ht,
2213 &stream->node_channel_id);
2214
2215 /*
2216 * Add stream to the stream_list_ht of the consumer data. No need to steal
2217 * the key since the HT does not use it and we allow to add redundant keys
2218 * into this table.
2219 */
2220 lttng_ht_add_u64(consumer_data.stream_list_ht, &stream->node_session_id);
2221
2222 rcu_read_unlock();
2223
2224 pthread_mutex_unlock(&stream->lock);
2225 pthread_mutex_unlock(&stream->chan->lock);
2226 pthread_mutex_unlock(&stream->chan->timer_lock);
2227 pthread_mutex_unlock(&consumer_data.lock);
2228 }
2229
2230 /*
2231 * Delete data stream that are flagged for deletion (endpoint_status).
2232 */
2233 static void validate_endpoint_status_data_stream(void)
2234 {
2235 struct lttng_ht_iter iter;
2236 struct lttng_consumer_stream *stream;
2237
2238 DBG("Consumer delete flagged data stream");
2239
2240 rcu_read_lock();
2241 cds_lfht_for_each_entry(data_ht->ht, &iter.iter, stream, node.node) {
2242 /* Validate delete flag of the stream */
2243 if (stream->endpoint_status == CONSUMER_ENDPOINT_ACTIVE) {
2244 continue;
2245 }
2246 /* Delete it right now */
2247 consumer_del_stream(stream, data_ht);
2248 }
2249 rcu_read_unlock();
2250 }
2251
2252 /*
2253 * Delete metadata stream that are flagged for deletion (endpoint_status).
2254 */
2255 static void validate_endpoint_status_metadata_stream(
2256 struct lttng_poll_event *pollset)
2257 {
2258 struct lttng_ht_iter iter;
2259 struct lttng_consumer_stream *stream;
2260
2261 DBG("Consumer delete flagged metadata stream");
2262
2263 assert(pollset);
2264
2265 rcu_read_lock();
2266 cds_lfht_for_each_entry(metadata_ht->ht, &iter.iter, stream, node.node) {
2267 /* Validate delete flag of the stream */
2268 if (stream->endpoint_status == CONSUMER_ENDPOINT_ACTIVE) {
2269 continue;
2270 }
2271 /*
2272 * Remove from pollset so the metadata thread can continue without
2273 * blocking on a deleted stream.
2274 */
2275 lttng_poll_del(pollset, stream->wait_fd);
2276
2277 /* Delete it right now */
2278 consumer_del_metadata_stream(stream, metadata_ht);
2279 }
2280 rcu_read_unlock();
2281 }
2282
2283 static
2284 int rotate_notify_sessiond(struct lttng_consumer_local_data *ctx,
2285 uint64_t key)
2286 {
2287 ssize_t ret;
2288
2289 do {
2290 ret = write(ctx->channel_rotate_pipe, &key, sizeof(key));
2291 } while (ret == -1 && errno == EINTR);
2292 if (ret == -1) {
2293 PERROR("Failed to write to the channel rotation pipe");
2294 } else {
2295 DBG("Sent channel rotation notification for channel key %"
2296 PRIu64, key);
2297 ret = 0;
2298 }
2299
2300 return (int) ret;
2301 }
2302
2303 /*
2304 * Perform operations that need to be done after a stream has
2305 * rotated and released the stream lock.
2306 *
2307 * Multiple rotations cannot occur simultaneously, so we know the state of the
2308 * "rotated" stream flag cannot change.
2309 *
2310 * This MUST be called WITHOUT the stream lock held.
2311 */
2312 static
2313 int consumer_post_rotation(struct lttng_consumer_stream *stream,
2314 struct lttng_consumer_local_data *ctx)
2315 {
2316 int ret = 0;
2317
2318 pthread_mutex_lock(&stream->chan->lock);
2319
2320 switch (consumer_data.type) {
2321 case LTTNG_CONSUMER_KERNEL:
2322 break;
2323 case LTTNG_CONSUMER32_UST:
2324 case LTTNG_CONSUMER64_UST:
2325 /*
2326 * The ust_metadata_pushed counter has been reset to 0, so now
2327 * we can wakeup the metadata thread so it dumps the metadata
2328 * cache to the new file.
2329 */
2330 if (stream->metadata_flag) {
2331 consumer_metadata_wakeup_pipe(stream->chan);
2332 }
2333 break;
2334 default:
2335 ERR("Unknown consumer_data type");
2336 abort();
2337 }
2338
2339 if (--stream->chan->nr_stream_rotate_pending == 0) {
2340 DBG("Rotation of channel \"%s\" completed, notifying the session daemon",
2341 stream->chan->name);
2342 ret = rotate_notify_sessiond(ctx, stream->chan->key);
2343 }
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;
2581 /* 2 for the consumer_data_pipe and wake up pipe */
2582 const int nb_pipes_fd = 2;
2583 /* Number of FDs with CONSUMER_ENDPOINT_INACTIVE but still open. */
2584 int nb_inactive_fd = 0;
2585 struct lttng_consumer_local_data *ctx = data;
2586 ssize_t len;
2587
2588 rcu_register_thread();
2589
2590 health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_DATA);
2591
2592 if (testpoint(consumerd_thread_data)) {
2593 goto error_testpoint;
2594 }
2595
2596 health_code_update();
2597
2598 local_stream = zmalloc(sizeof(struct lttng_consumer_stream *));
2599 if (local_stream == NULL) {
2600 PERROR("local_stream malloc");
2601 goto end;
2602 }
2603
2604 while (1) {
2605 health_code_update();
2606
2607 high_prio = 0;
2608 num_hup = 0;
2609
2610 /*
2611 * the fds set has been updated, we need to update our
2612 * local array as well
2613 */
2614 pthread_mutex_lock(&consumer_data.lock);
2615 if (consumer_data.need_update) {
2616 free(pollfd);
2617 pollfd = NULL;
2618
2619 free(local_stream);
2620 local_stream = NULL;
2621
2622 /* Allocate for all fds */
2623 pollfd = zmalloc((consumer_data.stream_count + nb_pipes_fd) * sizeof(struct pollfd));
2624 if (pollfd == NULL) {
2625 PERROR("pollfd malloc");
2626 pthread_mutex_unlock(&consumer_data.lock);
2627 goto end;
2628 }
2629
2630 local_stream = zmalloc((consumer_data.stream_count + nb_pipes_fd) *
2631 sizeof(struct lttng_consumer_stream *));
2632 if (local_stream == NULL) {
2633 PERROR("local_stream malloc");
2634 pthread_mutex_unlock(&consumer_data.lock);
2635 goto end;
2636 }
2637 ret = update_poll_array(ctx, &pollfd, local_stream,
2638 data_ht, &nb_inactive_fd);
2639 if (ret < 0) {
2640 ERR("Error in allocating pollfd or local_outfds");
2641 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
2642 pthread_mutex_unlock(&consumer_data.lock);
2643 goto end;
2644 }
2645 nb_fd = ret;
2646 consumer_data.need_update = 0;
2647 }
2648 pthread_mutex_unlock(&consumer_data.lock);
2649
2650 /* No FDs and consumer_quit, consumer_cleanup the thread */
2651 if (nb_fd == 0 && nb_inactive_fd == 0 &&
2652 CMM_LOAD_SHARED(consumer_quit) == 1) {
2653 err = 0; /* All is OK */
2654 goto end;
2655 }
2656 /* poll on the array of fds */
2657 restart:
2658 DBG("polling on %d fd", nb_fd + nb_pipes_fd);
2659 if (testpoint(consumerd_thread_data_poll)) {
2660 goto end;
2661 }
2662 health_poll_entry();
2663 num_rdy = poll(pollfd, nb_fd + nb_pipes_fd, -1);
2664 health_poll_exit();
2665 DBG("poll num_rdy : %d", num_rdy);
2666 if (num_rdy == -1) {
2667 /*
2668 * Restart interrupted system call.
2669 */
2670 if (errno == EINTR) {
2671 goto restart;
2672 }
2673 PERROR("Poll error");
2674 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
2675 goto end;
2676 } else if (num_rdy == 0) {
2677 DBG("Polling thread timed out");
2678 goto end;
2679 }
2680
2681 if (caa_unlikely(data_consumption_paused)) {
2682 DBG("Data consumption paused, sleeping...");
2683 sleep(1);
2684 goto restart;
2685 }
2686
2687 /*
2688 * If the consumer_data_pipe triggered poll go directly to the
2689 * beginning of the loop to update the array. We want to prioritize
2690 * array update over low-priority reads.
2691 */
2692 if (pollfd[nb_fd].revents & (POLLIN | POLLPRI)) {
2693 ssize_t pipe_readlen;
2694
2695 DBG("consumer_data_pipe wake up");
2696 pipe_readlen = lttng_pipe_read(ctx->consumer_data_pipe,
2697 &new_stream, sizeof(new_stream));
2698 if (pipe_readlen < sizeof(new_stream)) {
2699 PERROR("Consumer data pipe");
2700 /* Continue so we can at least handle the current stream(s). */
2701 continue;
2702 }
2703
2704 /*
2705 * If the stream is NULL, just ignore it. It's also possible that
2706 * the sessiond poll thread changed the consumer_quit state and is
2707 * waking us up to test it.
2708 */
2709 if (new_stream == NULL) {
2710 validate_endpoint_status_data_stream();
2711 continue;
2712 }
2713
2714 /* Continue to update the local streams and handle prio ones */
2715 continue;
2716 }
2717
2718 /* Handle wakeup pipe. */
2719 if (pollfd[nb_fd + 1].revents & (POLLIN | POLLPRI)) {
2720 char dummy;
2721 ssize_t pipe_readlen;
2722
2723 pipe_readlen = lttng_pipe_read(ctx->consumer_wakeup_pipe, &dummy,
2724 sizeof(dummy));
2725 if (pipe_readlen < 0) {
2726 PERROR("Consumer data wakeup pipe");
2727 }
2728 /* We've been awakened to handle stream(s). */
2729 ctx->has_wakeup = 0;
2730 }
2731
2732 /* Take care of high priority channels first. */
2733 for (i = 0; i < nb_fd; i++) {
2734 health_code_update();
2735
2736 if (local_stream[i] == NULL) {
2737 continue;
2738 }
2739 if (pollfd[i].revents & POLLPRI) {
2740 DBG("Urgent read on fd %d", pollfd[i].fd);
2741 high_prio = 1;
2742 len = ctx->on_buffer_ready(local_stream[i], ctx);
2743 /* it's ok to have an unavailable sub-buffer */
2744 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2745 /* Clean the stream and free it. */
2746 consumer_del_stream(local_stream[i], data_ht);
2747 local_stream[i] = NULL;
2748 } else if (len > 0) {
2749 local_stream[i]->data_read = 1;
2750 }
2751 }
2752 }
2753
2754 /*
2755 * If we read high prio channel in this loop, try again
2756 * for more high prio data.
2757 */
2758 if (high_prio) {
2759 continue;
2760 }
2761
2762 /* Take care of low priority channels. */
2763 for (i = 0; i < nb_fd; i++) {
2764 health_code_update();
2765
2766 if (local_stream[i] == NULL) {
2767 continue;
2768 }
2769 if ((pollfd[i].revents & POLLIN) ||
2770 local_stream[i]->hangup_flush_done ||
2771 local_stream[i]->has_data) {
2772 DBG("Normal read on fd %d", pollfd[i].fd);
2773 len = ctx->on_buffer_ready(local_stream[i], ctx);
2774 /* it's ok to have an unavailable sub-buffer */
2775 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2776 /* Clean the stream and free it. */
2777 consumer_del_stream(local_stream[i], data_ht);
2778 local_stream[i] = NULL;
2779 } else if (len > 0) {
2780 local_stream[i]->data_read = 1;
2781 }
2782 }
2783 }
2784
2785 /* Handle hangup and errors */
2786 for (i = 0; i < nb_fd; i++) {
2787 health_code_update();
2788
2789 if (local_stream[i] == NULL) {
2790 continue;
2791 }
2792 if (!local_stream[i]->hangup_flush_done
2793 && (pollfd[i].revents & (POLLHUP | POLLERR | POLLNVAL))
2794 && (consumer_data.type == LTTNG_CONSUMER32_UST
2795 || consumer_data.type == LTTNG_CONSUMER64_UST)) {
2796 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2797 pollfd[i].fd);
2798 lttng_ustconsumer_on_stream_hangup(local_stream[i]);
2799 /* Attempt read again, for the data we just flushed. */
2800 local_stream[i]->data_read = 1;
2801 }
2802 /*
2803 * If the poll flag is HUP/ERR/NVAL and we have
2804 * read no data in this pass, we can remove the
2805 * stream from its hash table.
2806 */
2807 if ((pollfd[i].revents & POLLHUP)) {
2808 DBG("Polling fd %d tells it has hung up.", pollfd[i].fd);
2809 if (!local_stream[i]->data_read) {
2810 consumer_del_stream(local_stream[i], data_ht);
2811 local_stream[i] = NULL;
2812 num_hup++;
2813 }
2814 } else if (pollfd[i].revents & POLLERR) {
2815 ERR("Error returned in polling fd %d.", pollfd[i].fd);
2816 if (!local_stream[i]->data_read) {
2817 consumer_del_stream(local_stream[i], data_ht);
2818 local_stream[i] = NULL;
2819 num_hup++;
2820 }
2821 } else if (pollfd[i].revents & POLLNVAL) {
2822 ERR("Polling fd %d tells fd is not open.", pollfd[i].fd);
2823 if (!local_stream[i]->data_read) {
2824 consumer_del_stream(local_stream[i], data_ht);
2825 local_stream[i] = NULL;
2826 num_hup++;
2827 }
2828 }
2829 if (local_stream[i] != NULL) {
2830 local_stream[i]->data_read = 0;
2831 }
2832 }
2833 }
2834 /* All is OK */
2835 err = 0;
2836 end:
2837 DBG("polling thread exiting");
2838 free(pollfd);
2839 free(local_stream);
2840
2841 /*
2842 * Close the write side of the pipe so epoll_wait() in
2843 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2844 * read side of the pipe. If we close them both, epoll_wait strangely does
2845 * not return and could create a endless wait period if the pipe is the
2846 * only tracked fd in the poll set. The thread will take care of closing
2847 * the read side.
2848 */
2849 (void) lttng_pipe_write_close(ctx->consumer_metadata_pipe);
2850
2851 error_testpoint:
2852 if (err) {
2853 health_error();
2854 ERR("Health error occurred in %s", __func__);
2855 }
2856 health_unregister(health_consumerd);
2857
2858 rcu_unregister_thread();
2859 return NULL;
2860 }
2861
2862 /*
2863 * Close wake-up end of each stream belonging to the channel. This will
2864 * allow the poll() on the stream read-side to detect when the
2865 * write-side (application) finally closes them.
2866 */
2867 static
2868 void consumer_close_channel_streams(struct lttng_consumer_channel *channel)
2869 {
2870 struct lttng_ht *ht;
2871 struct lttng_consumer_stream *stream;
2872 struct lttng_ht_iter iter;
2873
2874 ht = consumer_data.stream_per_chan_id_ht;
2875
2876 rcu_read_lock();
2877 cds_lfht_for_each_entry_duplicate(ht->ht,
2878 ht->hash_fct(&channel->key, lttng_ht_seed),
2879 ht->match_fct, &channel->key,
2880 &iter.iter, stream, node_channel_id.node) {
2881 /*
2882 * Protect against teardown with mutex.
2883 */
2884 pthread_mutex_lock(&stream->lock);
2885 if (cds_lfht_is_node_deleted(&stream->node.node)) {
2886 goto next;
2887 }
2888 switch (consumer_data.type) {
2889 case LTTNG_CONSUMER_KERNEL:
2890 break;
2891 case LTTNG_CONSUMER32_UST:
2892 case LTTNG_CONSUMER64_UST:
2893 if (stream->metadata_flag) {
2894 /* Safe and protected by the stream lock. */
2895 lttng_ustconsumer_close_metadata(stream->chan);
2896 } else {
2897 /*
2898 * Note: a mutex is taken internally within
2899 * liblttng-ust-ctl to protect timer wakeup_fd
2900 * use from concurrent close.
2901 */
2902 lttng_ustconsumer_close_stream_wakeup(stream);
2903 }
2904 break;
2905 default:
2906 ERR("Unknown consumer_data type");
2907 assert(0);
2908 }
2909 next:
2910 pthread_mutex_unlock(&stream->lock);
2911 }
2912 rcu_read_unlock();
2913 }
2914
2915 static void destroy_channel_ht(struct lttng_ht *ht)
2916 {
2917 struct lttng_ht_iter iter;
2918 struct lttng_consumer_channel *channel;
2919 int ret;
2920
2921 if (ht == NULL) {
2922 return;
2923 }
2924
2925 rcu_read_lock();
2926 cds_lfht_for_each_entry(ht->ht, &iter.iter, channel, wait_fd_node.node) {
2927 ret = lttng_ht_del(ht, &iter);
2928 assert(ret != 0);
2929 }
2930 rcu_read_unlock();
2931
2932 lttng_ht_destroy(ht);
2933 }
2934
2935 /*
2936 * This thread polls the channel fds to detect when they are being
2937 * closed. It closes all related streams if the channel is detected as
2938 * closed. It is currently only used as a shim layer for UST because the
2939 * consumerd needs to keep the per-stream wakeup end of pipes open for
2940 * periodical flush.
2941 */
2942 void *consumer_thread_channel_poll(void *data)
2943 {
2944 int ret, i, pollfd, err = -1;
2945 uint32_t revents, nb_fd;
2946 struct lttng_consumer_channel *chan = NULL;
2947 struct lttng_ht_iter iter;
2948 struct lttng_ht_node_u64 *node;
2949 struct lttng_poll_event events;
2950 struct lttng_consumer_local_data *ctx = data;
2951 struct lttng_ht *channel_ht;
2952
2953 rcu_register_thread();
2954
2955 health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_CHANNEL);
2956
2957 if (testpoint(consumerd_thread_channel)) {
2958 goto error_testpoint;
2959 }
2960
2961 health_code_update();
2962
2963 channel_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
2964 if (!channel_ht) {
2965 /* ENOMEM at this point. Better to bail out. */
2966 goto end_ht;
2967 }
2968
2969 DBG("Thread channel poll started");
2970
2971 /* Size is set to 1 for the consumer_channel pipe */
2972 ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC);
2973 if (ret < 0) {
2974 ERR("Poll set creation failed");
2975 goto end_poll;
2976 }
2977
2978 ret = lttng_poll_add(&events, ctx->consumer_channel_pipe[0], LPOLLIN);
2979 if (ret < 0) {
2980 goto end;
2981 }
2982
2983 /* Main loop */
2984 DBG("Channel main loop started");
2985
2986 while (1) {
2987 restart:
2988 health_code_update();
2989 DBG("Channel poll wait");
2990 health_poll_entry();
2991 ret = lttng_poll_wait(&events, -1);
2992 DBG("Channel poll return from wait with %d fd(s)",
2993 LTTNG_POLL_GETNB(&events));
2994 health_poll_exit();
2995 DBG("Channel event caught in thread");
2996 if (ret < 0) {
2997 if (errno == EINTR) {
2998 ERR("Poll EINTR caught");
2999 goto restart;
3000 }
3001 if (LTTNG_POLL_GETNB(&events) == 0) {
3002 err = 0; /* All is OK */
3003 }
3004 goto end;
3005 }
3006
3007 nb_fd = ret;
3008
3009 /* From here, the event is a channel wait fd */
3010 for (i = 0; i < nb_fd; i++) {
3011 health_code_update();
3012
3013 revents = LTTNG_POLL_GETEV(&events, i);
3014 pollfd = LTTNG_POLL_GETFD(&events, i);
3015
3016 if (!revents) {
3017 /* No activity for this FD (poll implementation). */
3018 continue;
3019 }
3020
3021 if (pollfd == ctx->consumer_channel_pipe[0]) {
3022 if (revents & LPOLLIN) {
3023 enum consumer_channel_action action;
3024 uint64_t key;
3025
3026 ret = read_channel_pipe(ctx, &chan, &key, &action);
3027 if (ret <= 0) {
3028 if (ret < 0) {
3029 ERR("Error reading channel pipe");
3030 }
3031 lttng_poll_del(&events, ctx->consumer_channel_pipe[0]);
3032 continue;
3033 }
3034
3035 switch (action) {
3036 case CONSUMER_CHANNEL_ADD:
3037 DBG("Adding channel %d to poll set",
3038 chan->wait_fd);
3039
3040 lttng_ht_node_init_u64(&chan->wait_fd_node,
3041 chan->wait_fd);
3042 rcu_read_lock();
3043 lttng_ht_add_unique_u64(channel_ht,
3044 &chan->wait_fd_node);
3045 rcu_read_unlock();
3046 /* Add channel to the global poll events list */
3047 lttng_poll_add(&events, chan->wait_fd,
3048 LPOLLERR | LPOLLHUP);
3049 break;
3050 case CONSUMER_CHANNEL_DEL:
3051 {
3052 /*
3053 * This command should never be called if the channel
3054 * has streams monitored by either the data or metadata
3055 * thread. The consumer only notify this thread with a
3056 * channel del. command if it receives a destroy
3057 * channel command from the session daemon that send it
3058 * if a command prior to the GET_CHANNEL failed.
3059 */
3060
3061 rcu_read_lock();
3062 chan = consumer_find_channel(key);
3063 if (!chan) {
3064 rcu_read_unlock();
3065 ERR("UST consumer get channel key %" PRIu64 " not found for del channel", key);
3066 break;
3067 }
3068 lttng_poll_del(&events, chan->wait_fd);
3069 iter.iter.node = &chan->wait_fd_node.node;
3070 ret = lttng_ht_del(channel_ht, &iter);
3071 assert(ret == 0);
3072
3073 switch (consumer_data.type) {
3074 case LTTNG_CONSUMER_KERNEL:
3075 break;
3076 case LTTNG_CONSUMER32_UST:
3077 case LTTNG_CONSUMER64_UST:
3078 health_code_update();
3079 /* Destroy streams that might have been left in the stream list. */
3080 clean_channel_stream_list(chan);
3081 break;
3082 default:
3083 ERR("Unknown consumer_data type");
3084 assert(0);
3085 }
3086
3087 /*
3088 * Release our own refcount. Force channel deletion even if
3089 * streams were not initialized.
3090 */
3091 if (!uatomic_sub_return(&chan->refcount, 1)) {
3092 consumer_del_channel(chan);
3093 }
3094 rcu_read_unlock();
3095 goto restart;
3096 }
3097 case CONSUMER_CHANNEL_QUIT:
3098 /*
3099 * Remove the pipe from the poll set and continue the loop
3100 * since their might be data to consume.
3101 */
3102 lttng_poll_del(&events, ctx->consumer_channel_pipe[0]);
3103 continue;
3104 default:
3105 ERR("Unknown action");
3106 break;
3107 }
3108 } else if (revents & (LPOLLERR | LPOLLHUP)) {
3109 DBG("Channel thread pipe hung up");
3110 /*
3111 * Remove the pipe from the poll set and continue the loop
3112 * since their might be data to consume.
3113 */
3114 lttng_poll_del(&events, ctx->consumer_channel_pipe[0]);
3115 continue;
3116 } else {
3117 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
3118 goto end;
3119 }
3120
3121 /* Handle other stream */
3122 continue;
3123 }
3124
3125 rcu_read_lock();
3126 {
3127 uint64_t tmp_id = (uint64_t) pollfd;
3128
3129 lttng_ht_lookup(channel_ht, &tmp_id, &iter);
3130 }
3131 node = lttng_ht_iter_get_node_u64(&iter);
3132 assert(node);
3133
3134 chan = caa_container_of(node, struct lttng_consumer_channel,
3135 wait_fd_node);
3136
3137 /* Check for error event */
3138 if (revents & (LPOLLERR | LPOLLHUP)) {
3139 DBG("Channel fd %d is hup|err.", pollfd);
3140
3141 lttng_poll_del(&events, chan->wait_fd);
3142 ret = lttng_ht_del(channel_ht, &iter);
3143 assert(ret == 0);
3144
3145 /*
3146 * This will close the wait fd for each stream associated to
3147 * this channel AND monitored by the data/metadata thread thus
3148 * will be clean by the right thread.
3149 */
3150 consumer_close_channel_streams(chan);
3151
3152 /* Release our own refcount */
3153 if (!uatomic_sub_return(&chan->refcount, 1)
3154 && !uatomic_read(&chan->nb_init_stream_left)) {
3155 consumer_del_channel(chan);
3156 }
3157 } else {
3158 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
3159 rcu_read_unlock();
3160 goto end;
3161 }
3162
3163 /* Release RCU lock for the channel looked up */
3164 rcu_read_unlock();
3165 }
3166 }
3167
3168 /* All is OK */
3169 err = 0;
3170 end:
3171 lttng_poll_clean(&events);
3172 end_poll:
3173 destroy_channel_ht(channel_ht);
3174 end_ht:
3175 error_testpoint:
3176 DBG("Channel poll thread exiting");
3177 if (err) {
3178 health_error();
3179 ERR("Health error occurred in %s", __func__);
3180 }
3181 health_unregister(health_consumerd);
3182 rcu_unregister_thread();
3183 return NULL;
3184 }
3185
3186 static int set_metadata_socket(struct lttng_consumer_local_data *ctx,
3187 struct pollfd *sockpoll, int client_socket)
3188 {
3189 int ret;
3190
3191 assert(ctx);
3192 assert(sockpoll);
3193
3194 ret = lttng_consumer_poll_socket(sockpoll);
3195 if (ret) {
3196 goto error;
3197 }
3198 DBG("Metadata connection on client_socket");
3199
3200 /* Blocking call, waiting for transmission */
3201 ctx->consumer_metadata_socket = lttcomm_accept_unix_sock(client_socket);
3202 if (ctx->consumer_metadata_socket < 0) {
3203 WARN("On accept metadata");
3204 ret = -1;
3205 goto error;
3206 }
3207 ret = 0;
3208
3209 error:
3210 return ret;
3211 }
3212
3213 /*
3214 * This thread listens on the consumerd socket and receives the file
3215 * descriptors from the session daemon.
3216 */
3217 void *consumer_thread_sessiond_poll(void *data)
3218 {
3219 int sock = -1, client_socket, ret, err = -1;
3220 /*
3221 * structure to poll for incoming data on communication socket avoids
3222 * making blocking sockets.
3223 */
3224 struct pollfd consumer_sockpoll[2];
3225 struct lttng_consumer_local_data *ctx = data;