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