Fix: perform relayd socket pair cleanup on control socket error
[lttng-tools.git] / src / common / consumer / consumer.c
1 /*
2 * Copyright (C) 2011 - Julien Desfossez <julien.desfossez@polymtl.ca>
3 * Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
4 * 2012 - David Goulet <dgoulet@efficios.com>
5 *
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License, version 2 only,
8 * as published by the Free Software Foundation.
9 *
10 * This program is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
13 * more details.
14 *
15 * You should have received a copy of the GNU General Public License along
16 * with this program; if not, write to the Free Software Foundation, Inc.,
17 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
18 */
19
20 #define _LGPL_SOURCE
21 #include <assert.h>
22 #include <poll.h>
23 #include <pthread.h>
24 #include <stdlib.h>
25 #include <string.h>
26 #include <sys/mman.h>
27 #include <sys/socket.h>
28 #include <sys/types.h>
29 #include <unistd.h>
30 #include <inttypes.h>
31 #include <signal.h>
32
33 #include <bin/lttng-consumerd/health-consumerd.h>
34 #include <common/common.h>
35 #include <common/utils.h>
36 #include <common/compat/poll.h>
37 #include <common/compat/endian.h>
38 #include <common/index/index.h>
39 #include <common/kernel-ctl/kernel-ctl.h>
40 #include <common/sessiond-comm/relayd.h>
41 #include <common/sessiond-comm/sessiond-comm.h>
42 #include <common/kernel-consumer/kernel-consumer.h>
43 #include <common/relayd/relayd.h>
44 #include <common/ust-consumer/ust-consumer.h>
45 #include <common/consumer/consumer-timer.h>
46 #include <common/consumer/consumer.h>
47 #include <common/consumer/consumer-stream.h>
48 #include <common/consumer/consumer-testpoint.h>
49 #include <common/align.h>
50 #include <common/consumer/consumer-metadata-cache.h>
51
52 struct lttng_consumer_global_data consumer_data = {
53 .stream_count = 0,
54 .need_update = 1,
55 .type = LTTNG_CONSUMER_UNKNOWN,
56 };
57
58 enum consumer_channel_action {
59 CONSUMER_CHANNEL_ADD,
60 CONSUMER_CHANNEL_DEL,
61 CONSUMER_CHANNEL_QUIT,
62 };
63
64 struct consumer_channel_msg {
65 enum consumer_channel_action action;
66 struct lttng_consumer_channel *chan; /* add */
67 uint64_t key; /* del */
68 };
69
70 /* Flag used to temporarily pause data consumption from testpoints. */
71 int data_consumption_paused;
72
73 /*
74 * Flag to inform the polling thread to quit when all fd hung up. Updated by
75 * the consumer_thread_receive_fds when it notices that all fds has hung up.
76 * Also updated by the signal handler (consumer_should_exit()). Read by the
77 * polling threads.
78 */
79 int consumer_quit;
80
81 /*
82 * Global hash table containing respectively metadata and data streams. The
83 * stream element in this ht should only be updated by the metadata poll thread
84 * for the metadata and the data poll thread for the data.
85 */
86 static struct lttng_ht *metadata_ht;
87 static struct lttng_ht *data_ht;
88
89 /*
90 * Notify a thread lttng pipe to poll back again. This usually means that some
91 * global state has changed so we just send back the thread in a poll wait
92 * call.
93 */
94 static void notify_thread_lttng_pipe(struct lttng_pipe *pipe)
95 {
96 struct lttng_consumer_stream *null_stream = NULL;
97
98 assert(pipe);
99
100 (void) lttng_pipe_write(pipe, &null_stream, sizeof(null_stream));
101 }
102
103 static void notify_health_quit_pipe(int *pipe)
104 {
105 ssize_t ret;
106
107 ret = lttng_write(pipe[1], "4", 1);
108 if (ret < 1) {
109 PERROR("write consumer health quit");
110 }
111 }
112
113 static void notify_channel_pipe(struct lttng_consumer_local_data *ctx,
114 struct lttng_consumer_channel *chan,
115 uint64_t key,
116 enum consumer_channel_action action)
117 {
118 struct consumer_channel_msg msg;
119 ssize_t ret;
120
121 memset(&msg, 0, sizeof(msg));
122
123 msg.action = action;
124 msg.chan = chan;
125 msg.key = key;
126 ret = lttng_write(ctx->consumer_channel_pipe[1], &msg, sizeof(msg));
127 if (ret < sizeof(msg)) {
128 PERROR("notify_channel_pipe write error");
129 }
130 }
131
132 void notify_thread_del_channel(struct lttng_consumer_local_data *ctx,
133 uint64_t key)
134 {
135 notify_channel_pipe(ctx, NULL, key, CONSUMER_CHANNEL_DEL);
136 }
137
138 static int read_channel_pipe(struct lttng_consumer_local_data *ctx,
139 struct lttng_consumer_channel **chan,
140 uint64_t *key,
141 enum consumer_channel_action *action)
142 {
143 struct consumer_channel_msg msg;
144 ssize_t ret;
145
146 ret = lttng_read(ctx->consumer_channel_pipe[0], &msg, sizeof(msg));
147 if (ret < sizeof(msg)) {
148 ret = -1;
149 goto error;
150 }
151 *action = msg.action;
152 *chan = msg.chan;
153 *key = msg.key;
154 error:
155 return (int) ret;
156 }
157
158 /*
159 * Cleanup the stream list of a channel. Those streams are not yet globally
160 * visible
161 */
162 static void clean_channel_stream_list(struct lttng_consumer_channel *channel)
163 {
164 struct lttng_consumer_stream *stream, *stmp;
165
166 assert(channel);
167
168 /* Delete streams that might have been left in the stream list. */
169 cds_list_for_each_entry_safe(stream, stmp, &channel->streams.head,
170 send_node) {
171 cds_list_del(&stream->send_node);
172 /*
173 * Once a stream is added to this list, the buffers were created so we
174 * have a guarantee that this call will succeed. Setting the monitor
175 * mode to 0 so we don't lock nor try to delete the stream from the
176 * global hash table.
177 */
178 stream->monitor = 0;
179 consumer_stream_destroy(stream, NULL);
180 }
181 }
182
183 /*
184 * Find a stream. The consumer_data.lock must be locked during this
185 * call.
186 */
187 static struct lttng_consumer_stream *find_stream(uint64_t key,
188 struct lttng_ht *ht)
189 {
190 struct lttng_ht_iter iter;
191 struct lttng_ht_node_u64 *node;
192 struct lttng_consumer_stream *stream = NULL;
193
194 assert(ht);
195
196 /* -1ULL keys are lookup failures */
197 if (key == (uint64_t) -1ULL) {
198 return NULL;
199 }
200
201 rcu_read_lock();
202
203 lttng_ht_lookup(ht, &key, &iter);
204 node = lttng_ht_iter_get_node_u64(&iter);
205 if (node != NULL) {
206 stream = caa_container_of(node, struct lttng_consumer_stream, node);
207 }
208
209 rcu_read_unlock();
210
211 return stream;
212 }
213
214 static void steal_stream_key(uint64_t key, struct lttng_ht *ht)
215 {
216 struct lttng_consumer_stream *stream;
217
218 rcu_read_lock();
219 stream = find_stream(key, ht);
220 if (stream) {
221 stream->key = (uint64_t) -1ULL;
222 /*
223 * We don't want the lookup to match, but we still need
224 * to iterate on this stream when iterating over the hash table. Just
225 * change the node key.
226 */
227 stream->node.key = (uint64_t) -1ULL;
228 }
229 rcu_read_unlock();
230 }
231
232 /*
233 * Return a channel object for the given key.
234 *
235 * RCU read side lock MUST be acquired before calling this function and
236 * protects the channel ptr.
237 */
238 struct lttng_consumer_channel *consumer_find_channel(uint64_t key)
239 {
240 struct lttng_ht_iter iter;
241 struct lttng_ht_node_u64 *node;
242 struct lttng_consumer_channel *channel = NULL;
243
244 /* -1ULL keys are lookup failures */
245 if (key == (uint64_t) -1ULL) {
246 return NULL;
247 }
248
249 lttng_ht_lookup(consumer_data.channel_ht, &key, &iter);
250 node = lttng_ht_iter_get_node_u64(&iter);
251 if (node != NULL) {
252 channel = caa_container_of(node, struct lttng_consumer_channel, node);
253 }
254
255 return channel;
256 }
257
258 /*
259 * There is a possibility that the consumer does not have enough time between
260 * the close of the channel on the session daemon and the cleanup in here thus
261 * once we have a channel add with an existing key, we know for sure that this
262 * channel will eventually get cleaned up by all streams being closed.
263 *
264 * This function just nullifies the already existing channel key.
265 */
266 static void steal_channel_key(uint64_t key)
267 {
268 struct lttng_consumer_channel *channel;
269
270 rcu_read_lock();
271 channel = consumer_find_channel(key);
272 if (channel) {
273 channel->key = (uint64_t) -1ULL;
274 /*
275 * We don't want the lookup to match, but we still need to iterate on
276 * this channel when iterating over the hash table. Just change the
277 * node key.
278 */
279 channel->node.key = (uint64_t) -1ULL;
280 }
281 rcu_read_unlock();
282 }
283
284 static void free_channel_rcu(struct rcu_head *head)
285 {
286 struct lttng_ht_node_u64 *node =
287 caa_container_of(head, struct lttng_ht_node_u64, head);
288 struct lttng_consumer_channel *channel =
289 caa_container_of(node, struct lttng_consumer_channel, node);
290
291 switch (consumer_data.type) {
292 case LTTNG_CONSUMER_KERNEL:
293 break;
294 case LTTNG_CONSUMER32_UST:
295 case LTTNG_CONSUMER64_UST:
296 lttng_ustconsumer_free_channel(channel);
297 break;
298 default:
299 ERR("Unknown consumer_data type");
300 abort();
301 }
302 free(channel);
303 }
304
305 /*
306 * RCU protected relayd socket pair free.
307 */
308 static void free_relayd_rcu(struct rcu_head *head)
309 {
310 struct lttng_ht_node_u64 *node =
311 caa_container_of(head, struct lttng_ht_node_u64, head);
312 struct consumer_relayd_sock_pair *relayd =
313 caa_container_of(node, struct consumer_relayd_sock_pair, node);
314
315 /*
316 * Close all sockets. This is done in the call RCU since we don't want the
317 * socket fds to be reassigned thus potentially creating bad state of the
318 * relayd object.
319 *
320 * We do not have to lock the control socket mutex here since at this stage
321 * there is no one referencing to this relayd object.
322 */
323 (void) relayd_close(&relayd->control_sock);
324 (void) relayd_close(&relayd->data_sock);
325
326 pthread_mutex_destroy(&relayd->ctrl_sock_mutex);
327 free(relayd);
328 }
329
330 /*
331 * Destroy and free relayd socket pair object.
332 */
333 void consumer_destroy_relayd(struct consumer_relayd_sock_pair *relayd)
334 {
335 int ret;
336 struct lttng_ht_iter iter;
337
338 if (relayd == NULL) {
339 return;
340 }
341
342 DBG("Consumer destroy and close relayd socket pair");
343
344 iter.iter.node = &relayd->node.node;
345 ret = lttng_ht_del(consumer_data.relayd_ht, &iter);
346 if (ret != 0) {
347 /* We assume the relayd is being or is destroyed */
348 return;
349 }
350
351 /* RCU free() call */
352 call_rcu(&relayd->node.head, free_relayd_rcu);
353 }
354
355 /*
356 * Remove a channel from the global list protected by a mutex. This function is
357 * also responsible for freeing its data structures.
358 */
359 void consumer_del_channel(struct lttng_consumer_channel *channel)
360 {
361 int ret;
362 struct lttng_ht_iter iter;
363
364 DBG("Consumer delete channel key %" PRIu64, channel->key);
365
366 pthread_mutex_lock(&consumer_data.lock);
367 pthread_mutex_lock(&channel->lock);
368
369 /* Destroy streams that might have been left in the stream list. */
370 clean_channel_stream_list(channel);
371
372 if (channel->live_timer_enabled == 1) {
373 consumer_timer_live_stop(channel);
374 }
375 if (channel->monitor_timer_enabled == 1) {
376 consumer_timer_monitor_stop(channel);
377 }
378
379 switch (consumer_data.type) {
380 case LTTNG_CONSUMER_KERNEL:
381 break;
382 case LTTNG_CONSUMER32_UST:
383 case LTTNG_CONSUMER64_UST:
384 lttng_ustconsumer_del_channel(channel);
385 break;
386 default:
387 ERR("Unknown consumer_data type");
388 assert(0);
389 goto end;
390 }
391
392 rcu_read_lock();
393 iter.iter.node = &channel->node.node;
394 ret = lttng_ht_del(consumer_data.channel_ht, &iter);
395 assert(!ret);
396 rcu_read_unlock();
397
398 call_rcu(&channel->node.head, free_channel_rcu);
399 end:
400 pthread_mutex_unlock(&channel->lock);
401 pthread_mutex_unlock(&consumer_data.lock);
402 }
403
404 /*
405 * Iterate over the relayd hash table and destroy each element. Finally,
406 * destroy the whole hash table.
407 */
408 static void cleanup_relayd_ht(void)
409 {
410 struct lttng_ht_iter iter;
411 struct consumer_relayd_sock_pair *relayd;
412
413 rcu_read_lock();
414
415 cds_lfht_for_each_entry(consumer_data.relayd_ht->ht, &iter.iter, relayd,
416 node.node) {
417 consumer_destroy_relayd(relayd);
418 }
419
420 rcu_read_unlock();
421
422 lttng_ht_destroy(consumer_data.relayd_ht);
423 }
424
425 /*
426 * Update the end point status of all streams having the given network sequence
427 * index (relayd index).
428 *
429 * It's atomically set without having the stream mutex locked which is fine
430 * because we handle the write/read race with a pipe wakeup for each thread.
431 */
432 static void update_endpoint_status_by_netidx(uint64_t net_seq_idx,
433 enum consumer_endpoint_status status)
434 {
435 struct lttng_ht_iter iter;
436 struct lttng_consumer_stream *stream;
437
438 DBG("Consumer set delete flag on stream by idx %" PRIu64, net_seq_idx);
439
440 rcu_read_lock();
441
442 /* Let's begin with metadata */
443 cds_lfht_for_each_entry(metadata_ht->ht, &iter.iter, stream, node.node) {
444 if (stream->net_seq_idx == net_seq_idx) {
445 uatomic_set(&stream->endpoint_status, status);
446 DBG("Delete flag set to metadata stream %d", stream->wait_fd);
447 }
448 }
449
450 /* Follow up by the data streams */
451 cds_lfht_for_each_entry(data_ht->ht, &iter.iter, stream, node.node) {
452 if (stream->net_seq_idx == net_seq_idx) {
453 uatomic_set(&stream->endpoint_status, status);
454 DBG("Delete flag set to data stream %d", stream->wait_fd);
455 }
456 }
457 rcu_read_unlock();
458 }
459
460 /*
461 * Cleanup a relayd object by flagging every associated streams for deletion,
462 * destroying the object meaning removing it from the relayd hash table,
463 * closing the sockets and freeing the memory in a RCU call.
464 *
465 * If a local data context is available, notify the threads that the streams'
466 * state have changed.
467 */
468 void lttng_consumer_cleanup_relayd(struct consumer_relayd_sock_pair *relayd)
469 {
470 uint64_t netidx;
471
472 assert(relayd);
473
474 DBG("Cleaning up relayd object ID %"PRIu64, relayd->net_seq_idx);
475
476 /* Save the net sequence index before destroying the object */
477 netidx = relayd->net_seq_idx;
478
479 /*
480 * Delete the relayd from the relayd hash table, close the sockets and free
481 * the object in a RCU call.
482 */
483 consumer_destroy_relayd(relayd);
484
485 /* Set inactive endpoint to all streams */
486 update_endpoint_status_by_netidx(netidx, CONSUMER_ENDPOINT_INACTIVE);
487
488 /*
489 * With a local data context, notify the threads that the streams' state
490 * have changed. The write() action on the pipe acts as an "implicit"
491 * memory barrier ordering the updates of the end point status from the
492 * read of this status which happens AFTER receiving this notify.
493 */
494 notify_thread_lttng_pipe(relayd->ctx->consumer_data_pipe);
495 notify_thread_lttng_pipe(relayd->ctx->consumer_metadata_pipe);
496 }
497
498 /*
499 * Flag a relayd socket pair for destruction. Destroy it if the refcount
500 * reaches zero.
501 *
502 * RCU read side lock MUST be aquired before calling this function.
503 */
504 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair *relayd)
505 {
506 assert(relayd);
507
508 /* Set destroy flag for this object */
509 uatomic_set(&relayd->destroy_flag, 1);
510
511 /* Destroy the relayd if refcount is 0 */
512 if (uatomic_read(&relayd->refcount) == 0) {
513 consumer_destroy_relayd(relayd);
514 }
515 }
516
517 /*
518 * Completly destroy stream from every visiable data structure and the given
519 * hash table if one.
520 *
521 * One this call returns, the stream object is not longer usable nor visible.
522 */
523 void consumer_del_stream(struct lttng_consumer_stream *stream,
524 struct lttng_ht *ht)
525 {
526 consumer_stream_destroy(stream, ht);
527 }
528
529 /*
530 * XXX naming of del vs destroy is all mixed up.
531 */
532 void consumer_del_stream_for_data(struct lttng_consumer_stream *stream)
533 {
534 consumer_stream_destroy(stream, data_ht);
535 }
536
537 void consumer_del_stream_for_metadata(struct lttng_consumer_stream *stream)
538 {
539 consumer_stream_destroy(stream, metadata_ht);
540 }
541
542 void consumer_stream_update_channel_attributes(
543 struct lttng_consumer_stream *stream,
544 struct lttng_consumer_channel *channel)
545 {
546 stream->channel_read_only_attributes.tracefile_size =
547 channel->tracefile_size;
548 memcpy(stream->channel_read_only_attributes.path, channel->pathname,
549 sizeof(stream->channel_read_only_attributes.path));
550 }
551
552 struct lttng_consumer_stream *consumer_allocate_stream(uint64_t channel_key,
553 uint64_t stream_key,
554 enum lttng_consumer_stream_state state,
555 const char *channel_name,
556 uid_t uid,
557 gid_t gid,
558 uint64_t relayd_id,
559 uint64_t session_id,
560 int cpu,
561 int *alloc_ret,
562 enum consumer_channel_type type,
563 unsigned int monitor,
564 uint64_t trace_archive_id)
565 {
566 int ret;
567 struct lttng_consumer_stream *stream;
568
569 stream = zmalloc(sizeof(*stream));
570 if (stream == NULL) {
571 PERROR("malloc struct lttng_consumer_stream");
572 ret = -ENOMEM;
573 goto end;
574 }
575
576 rcu_read_lock();
577
578 stream->key = stream_key;
579 stream->out_fd = -1;
580 stream->out_fd_offset = 0;
581 stream->output_written = 0;
582 stream->state = state;
583 stream->uid = uid;
584 stream->gid = gid;
585 stream->net_seq_idx = relayd_id;
586 stream->session_id = session_id;
587 stream->monitor = monitor;
588 stream->endpoint_status = CONSUMER_ENDPOINT_ACTIVE;
589 stream->index_file = NULL;
590 stream->last_sequence_number = -1ULL;
591 stream->trace_archive_id = trace_archive_id;
592 pthread_mutex_init(&stream->lock, NULL);
593 pthread_mutex_init(&stream->metadata_timer_lock, NULL);
594
595 /* If channel is the metadata, flag this stream as metadata. */
596 if (type == CONSUMER_CHANNEL_TYPE_METADATA) {
597 stream->metadata_flag = 1;
598 /* Metadata is flat out. */
599 strncpy(stream->name, DEFAULT_METADATA_NAME, sizeof(stream->name));
600 /* Live rendez-vous point. */
601 pthread_cond_init(&stream->metadata_rdv, NULL);
602 pthread_mutex_init(&stream->metadata_rdv_lock, NULL);
603 } else {
604 /* Format stream name to <channel_name>_<cpu_number> */
605 ret = snprintf(stream->name, sizeof(stream->name), "%s_%d",
606 channel_name, cpu);
607 if (ret < 0) {
608 PERROR("snprintf stream name");
609 goto error;
610 }
611 }
612
613 /* Key is always the wait_fd for streams. */
614 lttng_ht_node_init_u64(&stream->node, stream->key);
615
616 /* Init node per channel id key */
617 lttng_ht_node_init_u64(&stream->node_channel_id, channel_key);
618
619 /* Init session id node with the stream session id */
620 lttng_ht_node_init_u64(&stream->node_session_id, stream->session_id);
621
622 DBG3("Allocated stream %s (key %" PRIu64 ", chan_key %" PRIu64
623 " relayd_id %" PRIu64 ", session_id %" PRIu64,
624 stream->name, stream->key, channel_key,
625 stream->net_seq_idx, stream->session_id);
626
627 rcu_read_unlock();
628 return stream;
629
630 error:
631 rcu_read_unlock();
632 free(stream);
633 end:
634 if (alloc_ret) {
635 *alloc_ret = ret;
636 }
637 return NULL;
638 }
639
640 /*
641 * Add a stream to the global list protected by a mutex.
642 */
643 void consumer_add_data_stream(struct lttng_consumer_stream *stream)
644 {
645 struct lttng_ht *ht = data_ht;
646
647 assert(stream);
648 assert(ht);
649
650 DBG3("Adding consumer stream %" PRIu64, stream->key);
651
652 pthread_mutex_lock(&consumer_data.lock);
653 pthread_mutex_lock(&stream->chan->lock);
654 pthread_mutex_lock(&stream->chan->timer_lock);
655 pthread_mutex_lock(&stream->lock);
656 rcu_read_lock();
657
658 /* Steal stream identifier to avoid having streams with the same key */
659 steal_stream_key(stream->key, ht);
660
661 lttng_ht_add_unique_u64(ht, &stream->node);
662
663 lttng_ht_add_u64(consumer_data.stream_per_chan_id_ht,
664 &stream->node_channel_id);
665
666 /*
667 * Add stream to the stream_list_ht of the consumer data. No need to steal
668 * the key since the HT does not use it and we allow to add redundant keys
669 * into this table.
670 */
671 lttng_ht_add_u64(consumer_data.stream_list_ht, &stream->node_session_id);
672
673 /*
674 * When nb_init_stream_left reaches 0, we don't need to trigger any action
675 * in terms of destroying the associated channel, because the action that
676 * causes the count to become 0 also causes a stream to be added. The
677 * channel deletion will thus be triggered by the following removal of this
678 * stream.
679 */
680 if (uatomic_read(&stream->chan->nb_init_stream_left) > 0) {
681 /* Increment refcount before decrementing nb_init_stream_left */
682 cmm_smp_wmb();
683 uatomic_dec(&stream->chan->nb_init_stream_left);
684 }
685
686 /* Update consumer data once the node is inserted. */
687 consumer_data.stream_count++;
688 consumer_data.need_update = 1;
689
690 rcu_read_unlock();
691 pthread_mutex_unlock(&stream->lock);
692 pthread_mutex_unlock(&stream->chan->timer_lock);
693 pthread_mutex_unlock(&stream->chan->lock);
694 pthread_mutex_unlock(&consumer_data.lock);
695 }
696
697 void consumer_del_data_stream(struct lttng_consumer_stream *stream)
698 {
699 consumer_del_stream(stream, data_ht);
700 }
701
702 /*
703 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
704 * be acquired before calling this.
705 */
706 static int add_relayd(struct consumer_relayd_sock_pair *relayd)
707 {
708 int ret = 0;
709 struct lttng_ht_node_u64 *node;
710 struct lttng_ht_iter iter;
711
712 assert(relayd);
713
714 lttng_ht_lookup(consumer_data.relayd_ht,
715 &relayd->net_seq_idx, &iter);
716 node = lttng_ht_iter_get_node_u64(&iter);
717 if (node != NULL) {
718 goto end;
719 }
720 lttng_ht_add_unique_u64(consumer_data.relayd_ht, &relayd->node);
721
722 end:
723 return ret;
724 }
725
726 /*
727 * Allocate and return a consumer relayd socket.
728 */
729 static struct consumer_relayd_sock_pair *consumer_allocate_relayd_sock_pair(
730 uint64_t net_seq_idx)
731 {
732 struct consumer_relayd_sock_pair *obj = NULL;
733
734 /* net sequence index of -1 is a failure */
735 if (net_seq_idx == (uint64_t) -1ULL) {
736 goto error;
737 }
738
739 obj = zmalloc(sizeof(struct consumer_relayd_sock_pair));
740 if (obj == NULL) {
741 PERROR("zmalloc relayd sock");
742 goto error;
743 }
744
745 obj->net_seq_idx = net_seq_idx;
746 obj->refcount = 0;
747 obj->destroy_flag = 0;
748 obj->control_sock.sock.fd = -1;
749 obj->data_sock.sock.fd = -1;
750 lttng_ht_node_init_u64(&obj->node, obj->net_seq_idx);
751 pthread_mutex_init(&obj->ctrl_sock_mutex, NULL);
752
753 error:
754 return obj;
755 }
756
757 /*
758 * Find a relayd socket pair in the global consumer data.
759 *
760 * Return the object if found else NULL.
761 * RCU read-side lock must be held across this call and while using the
762 * returned object.
763 */
764 struct consumer_relayd_sock_pair *consumer_find_relayd(uint64_t key)
765 {
766 struct lttng_ht_iter iter;
767 struct lttng_ht_node_u64 *node;
768 struct consumer_relayd_sock_pair *relayd = NULL;
769
770 /* Negative keys are lookup failures */
771 if (key == (uint64_t) -1ULL) {
772 goto error;
773 }
774
775 lttng_ht_lookup(consumer_data.relayd_ht, &key,
776 &iter);
777 node = lttng_ht_iter_get_node_u64(&iter);
778 if (node != NULL) {
779 relayd = caa_container_of(node, struct consumer_relayd_sock_pair, node);
780 }
781
782 error:
783 return relayd;
784 }
785
786 /*
787 * Find a relayd and send the stream
788 *
789 * Returns 0 on success, < 0 on error
790 */
791 int consumer_send_relayd_stream(struct lttng_consumer_stream *stream,
792 char *path)
793 {
794 int ret = 0;
795 struct consumer_relayd_sock_pair *relayd;
796
797 assert(stream);
798 assert(stream->net_seq_idx != -1ULL);
799 assert(path);
800
801 /* The stream is not metadata. Get relayd reference if exists. */
802 rcu_read_lock();
803 relayd = consumer_find_relayd(stream->net_seq_idx);
804 if (relayd != NULL) {
805 /* Add stream on the relayd */
806 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
807 ret = relayd_add_stream(&relayd->control_sock, stream->name,
808 path, &stream->relayd_stream_id,
809 stream->chan->tracefile_size, stream->chan->tracefile_count,
810 stream->trace_archive_id);
811 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
812 if (ret < 0) {
813 ERR("Relayd add stream failed. Cleaning up relayd %" PRIu64".", relayd->net_seq_idx);
814 lttng_consumer_cleanup_relayd(relayd);
815 goto end;
816 }
817
818 uatomic_inc(&relayd->refcount);
819 stream->sent_to_relayd = 1;
820 } else {
821 ERR("Stream %" PRIu64 " relayd ID %" PRIu64 " unknown. Can't send it.",
822 stream->key, stream->net_seq_idx);
823 ret = -1;
824 goto end;
825 }
826
827 DBG("Stream %s with key %" PRIu64 " sent to relayd id %" PRIu64,
828 stream->name, stream->key, stream->net_seq_idx);
829
830 end:
831 rcu_read_unlock();
832 return ret;
833 }
834
835 /*
836 * Find a relayd and send the streams sent message
837 *
838 * Returns 0 on success, < 0 on error
839 */
840 int consumer_send_relayd_streams_sent(uint64_t net_seq_idx)
841 {
842 int ret = 0;
843 struct consumer_relayd_sock_pair *relayd;
844
845 assert(net_seq_idx != -1ULL);
846
847 /* The stream is not metadata. Get relayd reference if exists. */
848 rcu_read_lock();
849 relayd = consumer_find_relayd(net_seq_idx);
850 if (relayd != NULL) {
851 /* Add stream on the relayd */
852 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
853 ret = relayd_streams_sent(&relayd->control_sock);
854 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
855 if (ret < 0) {
856 ERR("Relayd streams sent failed. Cleaning up relayd %" PRIu64".", relayd->net_seq_idx);
857 lttng_consumer_cleanup_relayd(relayd);
858 goto end;
859 }
860 } else {
861 ERR("Relayd ID %" PRIu64 " unknown. Can't send streams_sent.",
862 net_seq_idx);
863 ret = -1;
864 goto end;
865 }
866
867 ret = 0;
868 DBG("All streams sent relayd id %" PRIu64, net_seq_idx);
869
870 end:
871 rcu_read_unlock();
872 return ret;
873 }
874
875 /*
876 * Find a relayd and close the stream
877 */
878 void close_relayd_stream(struct lttng_consumer_stream *stream)
879 {
880 struct consumer_relayd_sock_pair *relayd;
881
882 /* The stream is not metadata. Get relayd reference if exists. */
883 rcu_read_lock();
884 relayd = consumer_find_relayd(stream->net_seq_idx);
885 if (relayd) {
886 consumer_stream_relayd_close(stream, relayd);
887 }
888 rcu_read_unlock();
889 }
890
891 /*
892 * Handle stream for relayd transmission if the stream applies for network
893 * streaming where the net sequence index is set.
894 *
895 * Return destination file descriptor or negative value on error.
896 */
897 static int write_relayd_stream_header(struct lttng_consumer_stream *stream,
898 size_t data_size, unsigned long padding,
899 struct consumer_relayd_sock_pair *relayd)
900 {
901 int outfd = -1, ret;
902 struct lttcomm_relayd_data_hdr data_hdr;
903
904 /* Safety net */
905 assert(stream);
906 assert(relayd);
907
908 /* Reset data header */
909 memset(&data_hdr, 0, sizeof(data_hdr));
910
911 if (stream->metadata_flag) {
912 /* Caller MUST acquire the relayd control socket lock */
913 ret = relayd_send_metadata(&relayd->control_sock, data_size);
914 if (ret < 0) {
915 goto error;
916 }
917
918 /* Metadata are always sent on the control socket. */
919 outfd = relayd->control_sock.sock.fd;
920 } else {
921 /* Set header with stream information */
922 data_hdr.stream_id = htobe64(stream->relayd_stream_id);
923 data_hdr.data_size = htobe32(data_size);
924 data_hdr.padding_size = htobe32(padding);
925 /*
926 * Note that net_seq_num below is assigned with the *current* value of
927 * next_net_seq_num and only after that the next_net_seq_num will be
928 * increment. This is why when issuing a command on the relayd using
929 * this next value, 1 should always be substracted in order to compare
930 * the last seen sequence number on the relayd side to the last sent.
931 */
932 data_hdr.net_seq_num = htobe64(stream->next_net_seq_num);
933 /* Other fields are zeroed previously */
934
935 ret = relayd_send_data_hdr(&relayd->data_sock, &data_hdr,
936 sizeof(data_hdr));
937 if (ret < 0) {
938 goto error;
939 }
940
941 ++stream->next_net_seq_num;
942
943 /* Set to go on data socket */
944 outfd = relayd->data_sock.sock.fd;
945 }
946
947 error:
948 return outfd;
949 }
950
951 /*
952 * Allocate and return a new lttng_consumer_channel object using the given key
953 * to initialize the hash table node.
954 *
955 * On error, return NULL.
956 */
957 struct lttng_consumer_channel *consumer_allocate_channel(uint64_t key,
958 uint64_t session_id,
959 const char *pathname,
960 const char *name,
961 uid_t uid,
962 gid_t gid,
963 uint64_t relayd_id,
964 enum lttng_event_output output,
965 uint64_t tracefile_size,
966 uint64_t tracefile_count,
967 uint64_t session_id_per_pid,
968 unsigned int monitor,
969 unsigned int live_timer_interval,
970 const char *root_shm_path,
971 const char *shm_path)
972 {
973 struct lttng_consumer_channel *channel;
974
975 channel = zmalloc(sizeof(*channel));
976 if (channel == NULL) {
977 PERROR("malloc struct lttng_consumer_channel");
978 goto end;
979 }
980
981 channel->key = key;
982 channel->refcount = 0;
983 channel->session_id = session_id;
984 channel->session_id_per_pid = session_id_per_pid;
985 channel->uid = uid;
986 channel->gid = gid;
987 channel->relayd_id = relayd_id;
988 channel->tracefile_size = tracefile_size;
989 channel->tracefile_count = tracefile_count;
990 channel->monitor = monitor;
991 channel->live_timer_interval = live_timer_interval;
992 pthread_mutex_init(&channel->lock, NULL);
993 pthread_mutex_init(&channel->timer_lock, NULL);
994
995 switch (output) {
996 case LTTNG_EVENT_SPLICE:
997 channel->output = CONSUMER_CHANNEL_SPLICE;
998 break;
999 case LTTNG_EVENT_MMAP:
1000 channel->output = CONSUMER_CHANNEL_MMAP;
1001 break;
1002 default:
1003 assert(0);
1004 free(channel);
1005 channel = NULL;
1006 goto end;
1007 }
1008
1009 /*
1010 * In monitor mode, the streams associated with the channel will be put in
1011 * a special list ONLY owned by this channel. So, the refcount is set to 1
1012 * here meaning that the channel itself has streams that are referenced.
1013 *
1014 * On a channel deletion, once the channel is no longer visible, the
1015 * refcount is decremented and checked for a zero value to delete it. With
1016 * streams in no monitor mode, it will now be safe to destroy the channel.
1017 */
1018 if (!channel->monitor) {
1019 channel->refcount = 1;
1020 }
1021
1022 strncpy(channel->pathname, pathname, sizeof(channel->pathname));
1023 channel->pathname[sizeof(channel->pathname) - 1] = '\0';
1024
1025 strncpy(channel->name, name, sizeof(channel->name));
1026 channel->name[sizeof(channel->name) - 1] = '\0';
1027
1028 if (root_shm_path) {
1029 strncpy(channel->root_shm_path, root_shm_path, sizeof(channel->root_shm_path));
1030 channel->root_shm_path[sizeof(channel->root_shm_path) - 1] = '\0';
1031 }
1032 if (shm_path) {
1033 strncpy(channel->shm_path, shm_path, sizeof(channel->shm_path));
1034 channel->shm_path[sizeof(channel->shm_path) - 1] = '\0';
1035 }
1036
1037 lttng_ht_node_init_u64(&channel->node, channel->key);
1038
1039 channel->wait_fd = -1;
1040
1041 CDS_INIT_LIST_HEAD(&channel->streams.head);
1042
1043 DBG("Allocated channel (key %" PRIu64 ")", channel->key);
1044
1045 end:
1046 return channel;
1047 }
1048
1049 /*
1050 * Add a channel to the global list protected by a mutex.
1051 *
1052 * Always return 0 indicating success.
1053 */
1054 int consumer_add_channel(struct lttng_consumer_channel *channel,
1055 struct lttng_consumer_local_data *ctx)
1056 {
1057 pthread_mutex_lock(&consumer_data.lock);
1058 pthread_mutex_lock(&channel->lock);
1059 pthread_mutex_lock(&channel->timer_lock);
1060
1061 /*
1062 * This gives us a guarantee that the channel we are about to add to the
1063 * channel hash table will be unique. See this function comment on the why
1064 * we need to steel the channel key at this stage.
1065 */
1066 steal_channel_key(channel->key);
1067
1068 rcu_read_lock();
1069 lttng_ht_add_unique_u64(consumer_data.channel_ht, &channel->node);
1070 rcu_read_unlock();
1071
1072 pthread_mutex_unlock(&channel->timer_lock);
1073 pthread_mutex_unlock(&channel->lock);
1074 pthread_mutex_unlock(&consumer_data.lock);
1075
1076 if (channel->wait_fd != -1 && channel->type == CONSUMER_CHANNEL_TYPE_DATA) {
1077 notify_channel_pipe(ctx, channel, -1, CONSUMER_CHANNEL_ADD);
1078 }
1079
1080 return 0;
1081 }
1082
1083 /*
1084 * Allocate the pollfd structure and the local view of the out fds to avoid
1085 * doing a lookup in the linked list and concurrency issues when writing is
1086 * needed. Called with consumer_data.lock held.
1087 *
1088 * Returns the number of fds in the structures.
1089 */
1090 static int update_poll_array(struct lttng_consumer_local_data *ctx,
1091 struct pollfd **pollfd, struct lttng_consumer_stream **local_stream,
1092 struct lttng_ht *ht, int *nb_inactive_fd)
1093 {
1094 int i = 0;
1095 struct lttng_ht_iter iter;
1096 struct lttng_consumer_stream *stream;
1097
1098 assert(ctx);
1099 assert(ht);
1100 assert(pollfd);
1101 assert(local_stream);
1102
1103 DBG("Updating poll fd array");
1104 *nb_inactive_fd = 0;
1105 rcu_read_lock();
1106 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
1107 /*
1108 * Only active streams with an active end point can be added to the
1109 * poll set and local stream storage of the thread.
1110 *
1111 * There is a potential race here for endpoint_status to be updated
1112 * just after the check. However, this is OK since the stream(s) will
1113 * be deleted once the thread is notified that the end point state has
1114 * changed where this function will be called back again.
1115 *
1116 * We track the number of inactive FDs because they still need to be
1117 * closed by the polling thread after a wakeup on the data_pipe or
1118 * metadata_pipe.
1119 */
1120 if (stream->state != LTTNG_CONSUMER_ACTIVE_STREAM ||
1121 stream->endpoint_status == CONSUMER_ENDPOINT_INACTIVE) {
1122 (*nb_inactive_fd)++;
1123 continue;
1124 }
1125 /*
1126 * This clobbers way too much the debug output. Uncomment that if you
1127 * need it for debugging purposes.
1128 *
1129 * DBG("Active FD %d", stream->wait_fd);
1130 */
1131 (*pollfd)[i].fd = stream->wait_fd;
1132 (*pollfd)[i].events = POLLIN | POLLPRI;
1133 local_stream[i] = stream;
1134 i++;
1135 }
1136 rcu_read_unlock();
1137
1138 /*
1139 * Insert the consumer_data_pipe at the end of the array and don't
1140 * increment i so nb_fd is the number of real FD.
1141 */
1142 (*pollfd)[i].fd = lttng_pipe_get_readfd(ctx->consumer_data_pipe);
1143 (*pollfd)[i].events = POLLIN | POLLPRI;
1144
1145 (*pollfd)[i + 1].fd = lttng_pipe_get_readfd(ctx->consumer_wakeup_pipe);
1146 (*pollfd)[i + 1].events = POLLIN | POLLPRI;
1147 return i;
1148 }
1149
1150 /*
1151 * Poll on the should_quit pipe and the command socket return -1 on
1152 * error, 1 if should exit, 0 if data is available on the command socket
1153 */
1154 int lttng_consumer_poll_socket(struct pollfd *consumer_sockpoll)
1155 {
1156 int num_rdy;
1157
1158 restart:
1159 num_rdy = poll(consumer_sockpoll, 2, -1);
1160 if (num_rdy == -1) {
1161 /*
1162 * Restart interrupted system call.
1163 */
1164 if (errno == EINTR) {
1165 goto restart;
1166 }
1167 PERROR("Poll error");
1168 return -1;
1169 }
1170 if (consumer_sockpoll[0].revents & (POLLIN | POLLPRI)) {
1171 DBG("consumer_should_quit wake up");
1172 return 1;
1173 }
1174 return 0;
1175 }
1176
1177 /*
1178 * Set the error socket.
1179 */
1180 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data *ctx,
1181 int sock)
1182 {
1183 ctx->consumer_error_socket = sock;
1184 }
1185
1186 /*
1187 * Set the command socket path.
1188 */
1189 void lttng_consumer_set_command_sock_path(
1190 struct lttng_consumer_local_data *ctx, char *sock)
1191 {
1192 ctx->consumer_command_sock_path = sock;
1193 }
1194
1195 /*
1196 * Send return code to the session daemon.
1197 * If the socket is not defined, we return 0, it is not a fatal error
1198 */
1199 int lttng_consumer_send_error(struct lttng_consumer_local_data *ctx, int cmd)
1200 {
1201 if (ctx->consumer_error_socket > 0) {
1202 return lttcomm_send_unix_sock(ctx->consumer_error_socket, &cmd,
1203 sizeof(enum lttcomm_sessiond_command));
1204 }
1205
1206 return 0;
1207 }
1208
1209 /*
1210 * Close all the tracefiles and stream fds and MUST be called when all
1211 * instances are destroyed i.e. when all threads were joined and are ended.
1212 */
1213 void lttng_consumer_cleanup(void)
1214 {
1215 struct lttng_ht_iter iter;
1216 struct lttng_consumer_channel *channel;
1217
1218 rcu_read_lock();
1219
1220 cds_lfht_for_each_entry(consumer_data.channel_ht->ht, &iter.iter, channel,
1221 node.node) {
1222 consumer_del_channel(channel);
1223 }
1224
1225 rcu_read_unlock();
1226
1227 lttng_ht_destroy(consumer_data.channel_ht);
1228
1229 cleanup_relayd_ht();
1230
1231 lttng_ht_destroy(consumer_data.stream_per_chan_id_ht);
1232
1233 /*
1234 * This HT contains streams that are freed by either the metadata thread or
1235 * the data thread so we do *nothing* on the hash table and simply destroy
1236 * it.
1237 */
1238 lttng_ht_destroy(consumer_data.stream_list_ht);
1239 }
1240
1241 /*
1242 * Called from signal handler.
1243 */
1244 void lttng_consumer_should_exit(struct lttng_consumer_local_data *ctx)
1245 {
1246 ssize_t ret;
1247
1248 CMM_STORE_SHARED(consumer_quit, 1);
1249 ret = lttng_write(ctx->consumer_should_quit[1], "4", 1);
1250 if (ret < 1) {
1251 PERROR("write consumer quit");
1252 }
1253
1254 DBG("Consumer flag that it should quit");
1255 }
1256
1257
1258 /*
1259 * Flush pending writes to trace output disk file.
1260 */
1261 static
1262 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream *stream,
1263 off_t orig_offset)
1264 {
1265 int ret;
1266 int outfd = stream->out_fd;
1267
1268 /*
1269 * This does a blocking write-and-wait on any page that belongs to the
1270 * subbuffer prior to the one we just wrote.
1271 * Don't care about error values, as these are just hints and ways to
1272 * limit the amount of page cache used.
1273 */
1274 if (orig_offset < stream->max_sb_size) {
1275 return;
1276 }
1277 lttng_sync_file_range(outfd, orig_offset - stream->max_sb_size,
1278 stream->max_sb_size,
1279 SYNC_FILE_RANGE_WAIT_BEFORE
1280 | SYNC_FILE_RANGE_WRITE
1281 | SYNC_FILE_RANGE_WAIT_AFTER);
1282 /*
1283 * Give hints to the kernel about how we access the file:
1284 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1285 * we write it.
1286 *
1287 * We need to call fadvise again after the file grows because the
1288 * kernel does not seem to apply fadvise to non-existing parts of the
1289 * file.
1290 *
1291 * Call fadvise _after_ having waited for the page writeback to
1292 * complete because the dirty page writeback semantic is not well
1293 * defined. So it can be expected to lead to lower throughput in
1294 * streaming.
1295 */
1296 ret = posix_fadvise(outfd, orig_offset - stream->max_sb_size,
1297 stream->max_sb_size, POSIX_FADV_DONTNEED);
1298 if (ret && ret != -ENOSYS) {
1299 errno = ret;
1300 PERROR("posix_fadvise on fd %i", outfd);
1301 }
1302 }
1303
1304 /*
1305 * Initialise the necessary environnement :
1306 * - create a new context
1307 * - create the poll_pipe
1308 * - create the should_quit pipe (for signal handler)
1309 * - create the thread pipe (for splice)
1310 *
1311 * Takes a function pointer as argument, this function is called when data is
1312 * available on a buffer. This function is responsible to do the
1313 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1314 * buffer configuration and then kernctl_put_next_subbuf at the end.
1315 *
1316 * Returns a pointer to the new context or NULL on error.
1317 */
1318 struct lttng_consumer_local_data *lttng_consumer_create(
1319 enum lttng_consumer_type type,
1320 ssize_t (*buffer_ready)(struct lttng_consumer_stream *stream,
1321 struct lttng_consumer_local_data *ctx),
1322 int (*recv_channel)(struct lttng_consumer_channel *channel),
1323 int (*recv_stream)(struct lttng_consumer_stream *stream),
1324 int (*update_stream)(uint64_t stream_key, uint32_t state))
1325 {
1326 int ret;
1327 struct lttng_consumer_local_data *ctx;
1328
1329 assert(consumer_data.type == LTTNG_CONSUMER_UNKNOWN ||
1330 consumer_data.type == type);
1331 consumer_data.type = type;
1332
1333 ctx = zmalloc(sizeof(struct lttng_consumer_local_data));
1334 if (ctx == NULL) {
1335 PERROR("allocating context");
1336 goto error;
1337 }
1338
1339 ctx->consumer_error_socket = -1;
1340 ctx->consumer_metadata_socket = -1;
1341 pthread_mutex_init(&ctx->metadata_socket_lock, NULL);
1342 /* assign the callbacks */
1343 ctx->on_buffer_ready = buffer_ready;
1344 ctx->on_recv_channel = recv_channel;
1345 ctx->on_recv_stream = recv_stream;
1346 ctx->on_update_stream = update_stream;
1347
1348 ctx->consumer_data_pipe = lttng_pipe_open(0);
1349 if (!ctx->consumer_data_pipe) {
1350 goto error_poll_pipe;
1351 }
1352
1353 ctx->consumer_wakeup_pipe = lttng_pipe_open(0);
1354 if (!ctx->consumer_wakeup_pipe) {
1355 goto error_wakeup_pipe;
1356 }
1357
1358 ret = pipe(ctx->consumer_should_quit);
1359 if (ret < 0) {
1360 PERROR("Error creating recv pipe");
1361 goto error_quit_pipe;
1362 }
1363
1364 ret = pipe(ctx->consumer_channel_pipe);
1365 if (ret < 0) {
1366 PERROR("Error creating channel pipe");
1367 goto error_channel_pipe;
1368 }
1369
1370 ctx->consumer_metadata_pipe = lttng_pipe_open(0);
1371 if (!ctx->consumer_metadata_pipe) {
1372 goto error_metadata_pipe;
1373 }
1374
1375 ctx->channel_monitor_pipe = -1;
1376
1377 return ctx;
1378
1379 error_metadata_pipe:
1380 utils_close_pipe(ctx->consumer_channel_pipe);
1381 error_channel_pipe:
1382 utils_close_pipe(ctx->consumer_should_quit);
1383 error_quit_pipe:
1384 lttng_pipe_destroy(ctx->consumer_wakeup_pipe);
1385 error_wakeup_pipe:
1386 lttng_pipe_destroy(ctx->consumer_data_pipe);
1387 error_poll_pipe:
1388 free(ctx);
1389 error:
1390 return NULL;
1391 }
1392
1393 /*
1394 * Iterate over all streams of the hashtable and free them properly.
1395 */
1396 static void destroy_data_stream_ht(struct lttng_ht *ht)
1397 {
1398 struct lttng_ht_iter iter;
1399 struct lttng_consumer_stream *stream;
1400
1401 if (ht == NULL) {
1402 return;
1403 }
1404
1405 rcu_read_lock();
1406 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
1407 /*
1408 * Ignore return value since we are currently cleaning up so any error
1409 * can't be handled.
1410 */
1411 (void) consumer_del_stream(stream, ht);
1412 }
1413 rcu_read_unlock();
1414
1415 lttng_ht_destroy(ht);
1416 }
1417
1418 /*
1419 * Iterate over all streams of the metadata hashtable and free them
1420 * properly.
1421 */
1422 static void destroy_metadata_stream_ht(struct lttng_ht *ht)
1423 {
1424 struct lttng_ht_iter iter;
1425 struct lttng_consumer_stream *stream;
1426
1427 if (ht == NULL) {
1428 return;
1429 }
1430
1431 rcu_read_lock();
1432 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
1433 /*
1434 * Ignore return value since we are currently cleaning up so any error
1435 * can't be handled.
1436 */
1437 (void) consumer_del_metadata_stream(stream, ht);
1438 }
1439 rcu_read_unlock();
1440
1441 lttng_ht_destroy(ht);
1442 }
1443
1444 /*
1445 * Close all fds associated with the instance and free the context.
1446 */
1447 void lttng_consumer_destroy(struct lttng_consumer_local_data *ctx)
1448 {
1449 int ret;
1450
1451 DBG("Consumer destroying it. Closing everything.");
1452
1453 if (!ctx) {
1454 return;
1455 }
1456
1457 destroy_data_stream_ht(data_ht);
1458 destroy_metadata_stream_ht(metadata_ht);
1459
1460 ret = close(ctx->consumer_error_socket);
1461 if (ret) {
1462 PERROR("close");
1463 }
1464 ret = close(ctx->consumer_metadata_socket);
1465 if (ret) {
1466 PERROR("close");
1467 }
1468 utils_close_pipe(ctx->consumer_channel_pipe);
1469 lttng_pipe_destroy(ctx->consumer_data_pipe);
1470 lttng_pipe_destroy(ctx->consumer_metadata_pipe);
1471 lttng_pipe_destroy(ctx->consumer_wakeup_pipe);
1472 utils_close_pipe(ctx->consumer_should_quit);
1473
1474 unlink(ctx->consumer_command_sock_path);
1475 free(ctx);
1476 }
1477
1478 /*
1479 * Write the metadata stream id on the specified file descriptor.
1480 */
1481 static int write_relayd_metadata_id(int fd,
1482 struct lttng_consumer_stream *stream,
1483 unsigned long padding)
1484 {
1485 ssize_t ret;
1486 struct lttcomm_relayd_metadata_payload hdr;
1487
1488 hdr.stream_id = htobe64(stream->relayd_stream_id);
1489 hdr.padding_size = htobe32(padding);
1490 ret = lttng_write(fd, (void *) &hdr, sizeof(hdr));
1491 if (ret < sizeof(hdr)) {
1492 /*
1493 * This error means that the fd's end is closed so ignore the PERROR
1494 * not to clubber the error output since this can happen in a normal
1495 * code path.
1496 */
1497 if (errno != EPIPE) {
1498 PERROR("write metadata stream id");
1499 }
1500 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno);
1501 /*
1502 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1503 * handle writting the missing part so report that as an error and
1504 * don't lie to the caller.
1505 */
1506 ret = -1;
1507 goto end;
1508 }
1509 DBG("Metadata stream id %" PRIu64 " with padding %lu written before data",
1510 stream->relayd_stream_id, padding);
1511
1512 end:
1513 return (int) ret;
1514 }
1515
1516 /*
1517 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1518 * core function for writing trace buffers to either the local filesystem or
1519 * the network.
1520 *
1521 * It must be called with the stream lock held.
1522 *
1523 * Careful review MUST be put if any changes occur!
1524 *
1525 * Returns the number of bytes written
1526 */
1527 ssize_t lttng_consumer_on_read_subbuffer_mmap(
1528 struct lttng_consumer_local_data *ctx,
1529 struct lttng_consumer_stream *stream, unsigned long len,
1530 unsigned long padding,
1531 struct ctf_packet_index *index)
1532 {
1533 unsigned long mmap_offset;
1534 void *mmap_base;
1535 ssize_t ret = 0;
1536 off_t orig_offset = stream->out_fd_offset;
1537 /* Default is on the disk */
1538 int outfd = stream->out_fd;
1539 struct consumer_relayd_sock_pair *relayd = NULL;
1540 unsigned int relayd_hang_up = 0;
1541
1542 /* RCU lock for the relayd pointer */
1543 rcu_read_lock();
1544
1545 /* Flag that the current stream if set for network streaming. */
1546 if (stream->net_seq_idx != (uint64_t) -1ULL) {
1547 relayd = consumer_find_relayd(stream->net_seq_idx);
1548 if (relayd == NULL) {
1549 ret = -EPIPE;
1550 goto end;
1551 }
1552 }
1553
1554 /* get the offset inside the fd to mmap */
1555 switch (consumer_data.type) {
1556 case LTTNG_CONSUMER_KERNEL:
1557 mmap_base = stream->mmap_base;
1558 ret = kernctl_get_mmap_read_offset(stream->wait_fd, &mmap_offset);
1559 if (ret < 0) {
1560 PERROR("tracer ctl get_mmap_read_offset");
1561 goto end;
1562 }
1563 break;
1564 case LTTNG_CONSUMER32_UST:
1565 case LTTNG_CONSUMER64_UST:
1566 mmap_base = lttng_ustctl_get_mmap_base(stream);
1567 if (!mmap_base) {
1568 ERR("read mmap get mmap base for stream %s", stream->name);
1569 ret = -EPERM;
1570 goto end;
1571 }
1572 ret = lttng_ustctl_get_mmap_read_offset(stream, &mmap_offset);
1573 if (ret != 0) {
1574 PERROR("tracer ctl get_mmap_read_offset");
1575 ret = -EINVAL;
1576 goto end;
1577 }
1578 break;
1579 default:
1580 ERR("Unknown consumer_data type");
1581 assert(0);
1582 }
1583
1584 /* Handle stream on the relayd if the output is on the network */
1585 if (relayd) {
1586 unsigned long netlen = len;
1587
1588 /*
1589 * Lock the control socket for the complete duration of the function
1590 * since from this point on we will use the socket.
1591 */
1592 if (stream->metadata_flag) {
1593 /* Metadata requires the control socket. */
1594 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
1595 if (stream->reset_metadata_flag) {
1596 ret = relayd_reset_metadata(&relayd->control_sock,
1597 stream->relayd_stream_id,
1598 stream->metadata_version);
1599 if (ret < 0) {
1600 relayd_hang_up = 1;
1601 goto write_error;
1602 }
1603 stream->reset_metadata_flag = 0;
1604 }
1605 netlen += sizeof(struct lttcomm_relayd_metadata_payload);
1606 }
1607
1608 ret = write_relayd_stream_header(stream, netlen, padding, relayd);
1609 if (ret < 0) {
1610 relayd_hang_up = 1;
1611 goto write_error;
1612 }
1613 /* Use the returned socket. */
1614 outfd = ret;
1615
1616 /* Write metadata stream id before payload */
1617 if (stream->metadata_flag) {
1618 ret = write_relayd_metadata_id(outfd, stream, padding);
1619 if (ret < 0) {
1620 relayd_hang_up = 1;
1621 goto write_error;
1622 }
1623 }
1624 } else {
1625 /* No streaming, we have to set the len with the full padding */
1626 len += padding;
1627
1628 if (stream->metadata_flag && stream->reset_metadata_flag) {
1629 ret = utils_truncate_stream_file(stream->out_fd, 0);
1630 if (ret < 0) {
1631 ERR("Reset metadata file");
1632 goto end;
1633 }
1634 stream->reset_metadata_flag = 0;
1635 }
1636
1637 /*
1638 * Check if we need to change the tracefile before writing the packet.
1639 */
1640 if (stream->chan->tracefile_size > 0 &&
1641 (stream->tracefile_size_current + len) >
1642 stream->chan->tracefile_size) {
1643 ret = utils_rotate_stream_file(stream->chan->pathname,
1644 stream->name, stream->chan->tracefile_size,
1645 stream->chan->tracefile_count, stream->uid, stream->gid,
1646 stream->out_fd, &(stream->tracefile_count_current),
1647 &stream->out_fd);
1648 if (ret < 0) {
1649 ERR("Rotating output file");
1650 goto end;
1651 }
1652 outfd = stream->out_fd;
1653
1654 if (stream->index_file) {
1655 lttng_index_file_put(stream->index_file);
1656 stream->index_file = lttng_index_file_create(stream->chan->pathname,
1657 stream->name, stream->uid, stream->gid,
1658 stream->chan->tracefile_size,
1659 stream->tracefile_count_current,
1660 CTF_INDEX_MAJOR, CTF_INDEX_MINOR);
1661 if (!stream->index_file) {
1662 goto end;
1663 }
1664 }
1665
1666 /* Reset current size because we just perform a rotation. */
1667 stream->tracefile_size_current = 0;
1668 stream->out_fd_offset = 0;
1669 orig_offset = 0;
1670 }
1671 stream->tracefile_size_current += len;
1672 if (index) {
1673 index->offset = htobe64(stream->out_fd_offset);
1674 }
1675 }
1676
1677 /*
1678 * This call guarantee that len or less is returned. It's impossible to
1679 * receive a ret value that is bigger than len.
1680 */
1681 ret = lttng_write(outfd, mmap_base + mmap_offset, len);
1682 DBG("Consumer mmap write() ret %zd (len %lu)", ret, len);
1683 if (ret < 0 || ((size_t) ret != len)) {
1684 /*
1685 * Report error to caller if nothing was written else at least send the
1686 * amount written.
1687 */
1688 if (ret < 0) {
1689 ret = -errno;
1690 }
1691 relayd_hang_up = 1;
1692
1693 /* Socket operation failed. We consider the relayd dead */
1694 if (errno == EPIPE || errno == EINVAL || errno == EBADF) {
1695 /*
1696 * This is possible if the fd is closed on the other side
1697 * (outfd) or any write problem. It can be verbose a bit for a
1698 * normal execution if for instance the relayd is stopped
1699 * abruptly. This can happen so set this to a DBG statement.
1700 */
1701 DBG("Consumer mmap write detected relayd hang up");
1702 } else {
1703 /* Unhandled error, print it and stop function right now. */
1704 PERROR("Error in write mmap (ret %zd != len %lu)", ret, len);
1705 }
1706 goto write_error;
1707 }
1708 stream->output_written += ret;
1709
1710 /* This call is useless on a socket so better save a syscall. */
1711 if (!relayd) {
1712 /* This won't block, but will start writeout asynchronously */
1713 lttng_sync_file_range(outfd, stream->out_fd_offset, len,
1714 SYNC_FILE_RANGE_WRITE);
1715 stream->out_fd_offset += len;
1716 lttng_consumer_sync_trace_file(stream, orig_offset);
1717 }
1718
1719 write_error:
1720 /*
1721 * This is a special case that the relayd has closed its socket. Let's
1722 * cleanup the relayd object and all associated streams.
1723 */
1724 if (relayd && relayd_hang_up) {
1725 ERR("Relayd hangup. Cleaning up relayd %" PRIu64".", relayd->net_seq_idx);
1726 lttng_consumer_cleanup_relayd(relayd);
1727 }
1728
1729 end:
1730 /* Unlock only if ctrl socket used */
1731 if (relayd && stream->metadata_flag) {
1732 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
1733 }
1734
1735 rcu_read_unlock();
1736 return ret;
1737 }
1738
1739 /*
1740 * Splice the data from the ring buffer to the tracefile.
1741 *
1742 * It must be called with the stream lock held.
1743 *
1744 * Returns the number of bytes spliced.
1745 */
1746 ssize_t lttng_consumer_on_read_subbuffer_splice(
1747 struct lttng_consumer_local_data *ctx,
1748 struct lttng_consumer_stream *stream, unsigned long len,
1749 unsigned long padding,
1750 struct ctf_packet_index *index)
1751 {
1752 ssize_t ret = 0, written = 0, ret_splice = 0;
1753 loff_t offset = 0;
1754 off_t orig_offset = stream->out_fd_offset;
1755 int fd = stream->wait_fd;
1756 /* Default is on the disk */
1757 int outfd = stream->out_fd;
1758 struct consumer_relayd_sock_pair *relayd = NULL;
1759 int *splice_pipe;
1760 unsigned int relayd_hang_up = 0;
1761
1762 switch (consumer_data.type) {
1763 case LTTNG_CONSUMER_KERNEL:
1764 break;
1765 case LTTNG_CONSUMER32_UST:
1766 case LTTNG_CONSUMER64_UST:
1767 /* Not supported for user space tracing */
1768 return -ENOSYS;
1769 default:
1770 ERR("Unknown consumer_data type");
1771 assert(0);
1772 }
1773
1774 /* RCU lock for the relayd pointer */
1775 rcu_read_lock();
1776
1777 /* Flag that the current stream if set for network streaming. */
1778 if (stream->net_seq_idx != (uint64_t) -1ULL) {
1779 relayd = consumer_find_relayd(stream->net_seq_idx);
1780 if (relayd == NULL) {
1781 written = -ret;
1782 goto end;
1783 }
1784 }
1785 splice_pipe = stream->splice_pipe;
1786
1787 /* Write metadata stream id before payload */
1788 if (relayd) {
1789 unsigned long total_len = len;
1790
1791 if (stream->metadata_flag) {
1792 /*
1793 * Lock the control socket for the complete duration of the function
1794 * since from this point on we will use the socket.
1795 */
1796 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
1797
1798 if (stream->reset_metadata_flag) {
1799 ret = relayd_reset_metadata(&relayd->control_sock,
1800 stream->relayd_stream_id,
1801 stream->metadata_version);
1802 if (ret < 0) {
1803 relayd_hang_up = 1;
1804 goto write_error;
1805 }
1806 stream->reset_metadata_flag = 0;
1807 }
1808 ret = write_relayd_metadata_id(splice_pipe[1], stream,
1809 padding);
1810 if (ret < 0) {
1811 written = ret;
1812 relayd_hang_up = 1;
1813 goto write_error;
1814 }
1815
1816 total_len += sizeof(struct lttcomm_relayd_metadata_payload);
1817 }
1818
1819 ret = write_relayd_stream_header(stream, total_len, padding, relayd);
1820 if (ret < 0) {
1821 written = ret;
1822 relayd_hang_up = 1;
1823 goto write_error;
1824 }
1825 /* Use the returned socket. */
1826 outfd = ret;
1827 } else {
1828 /* No streaming, we have to set the len with the full padding */
1829 len += padding;
1830
1831 if (stream->metadata_flag && stream->reset_metadata_flag) {
1832 ret = utils_truncate_stream_file(stream->out_fd, 0);
1833 if (ret < 0) {
1834 ERR("Reset metadata file");
1835 goto end;
1836 }
1837 stream->reset_metadata_flag = 0;
1838 }
1839 /*
1840 * Check if we need to change the tracefile before writing the packet.
1841 */
1842 if (stream->chan->tracefile_size > 0 &&
1843 (stream->tracefile_size_current + len) >
1844 stream->chan->tracefile_size) {
1845 ret = utils_rotate_stream_file(stream->chan->pathname,
1846 stream->name, stream->chan->tracefile_size,
1847 stream->chan->tracefile_count, stream->uid, stream->gid,
1848 stream->out_fd, &(stream->tracefile_count_current),
1849 &stream->out_fd);
1850 if (ret < 0) {
1851 written = ret;
1852 ERR("Rotating output file");
1853 goto end;
1854 }
1855 outfd = stream->out_fd;
1856
1857 if (stream->index_file) {
1858 lttng_index_file_put(stream->index_file);
1859 stream->index_file = lttng_index_file_create(stream->chan->pathname,
1860 stream->name, stream->uid, stream->gid,
1861 stream->chan->tracefile_size,
1862 stream->tracefile_count_current,
1863 CTF_INDEX_MAJOR, CTF_INDEX_MINOR);
1864 if (!stream->index_file) {
1865 goto end;
1866 }
1867 }
1868
1869 /* Reset current size because we just perform a rotation. */
1870 stream->tracefile_size_current = 0;
1871 stream->out_fd_offset = 0;
1872 orig_offset = 0;
1873 }
1874 stream->tracefile_size_current += len;
1875 index->offset = htobe64(stream->out_fd_offset);
1876 }
1877
1878 while (len > 0) {
1879 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1880 (unsigned long)offset, len, fd, splice_pipe[1]);
1881 ret_splice = splice(fd, &offset, splice_pipe[1], NULL, len,
1882 SPLICE_F_MOVE | SPLICE_F_MORE);
1883 DBG("splice chan to pipe, ret %zd", ret_splice);
1884 if (ret_splice < 0) {
1885 ret = errno;
1886 written = -ret;
1887 PERROR("Error in relay splice");
1888 goto splice_error;
1889 }
1890
1891 /* Handle stream on the relayd if the output is on the network */
1892 if (relayd && stream->metadata_flag) {
1893 size_t metadata_payload_size =
1894 sizeof(struct lttcomm_relayd_metadata_payload);
1895
1896 /* Update counter to fit the spliced data */
1897 ret_splice += metadata_payload_size;
1898 len += metadata_payload_size;
1899 /*
1900 * We do this so the return value can match the len passed as
1901 * argument to this function.
1902 */
1903 written -= metadata_payload_size;
1904 }
1905
1906 /* Splice data out */
1907 ret_splice = splice(splice_pipe[0], NULL, outfd, NULL,
1908 ret_splice, SPLICE_F_MOVE | SPLICE_F_MORE);
1909 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
1910 outfd, ret_splice);
1911 if (ret_splice < 0) {
1912 ret = errno;
1913 written = -ret;
1914 relayd_hang_up = 1;
1915 goto write_error;
1916 } else if (ret_splice > len) {
1917 /*
1918 * We don't expect this code path to be executed but you never know
1919 * so this is an extra protection agains a buggy splice().
1920 */
1921 ret = errno;
1922 written += ret_splice;
1923 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice,
1924 len);
1925 goto splice_error;
1926 } else {
1927 /* All good, update current len and continue. */
1928 len -= ret_splice;
1929 }
1930
1931 /* This call is useless on a socket so better save a syscall. */
1932 if (!relayd) {
1933 /* This won't block, but will start writeout asynchronously */
1934 lttng_sync_file_range(outfd, stream->out_fd_offset, ret_splice,
1935 SYNC_FILE_RANGE_WRITE);
1936 stream->out_fd_offset += ret_splice;
1937 }
1938 stream->output_written += ret_splice;
1939 written += ret_splice;
1940 }
1941 if (!relayd) {
1942 lttng_consumer_sync_trace_file(stream, orig_offset);
1943 }
1944 goto end;
1945
1946 write_error:
1947 /*
1948 * This is a special case that the relayd has closed its socket. Let's
1949 * cleanup the relayd object and all associated streams.
1950 */
1951 if (relayd && relayd_hang_up) {
1952 ERR("Relayd hangup. Cleaning up relayd %" PRIu64".", relayd->net_seq_idx);
1953 lttng_consumer_cleanup_relayd(relayd);
1954 /* Skip splice error so the consumer does not fail */
1955 goto end;
1956 }
1957
1958 splice_error:
1959 /* send the appropriate error description to sessiond */
1960 switch (ret) {
1961 case EINVAL:
1962 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_EINVAL);
1963 break;
1964 case ENOMEM:
1965 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_ENOMEM);
1966 break;
1967 case ESPIPE:
1968 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_ESPIPE);
1969 break;
1970 }
1971
1972 end:
1973 if (relayd && stream->metadata_flag) {
1974 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
1975 }
1976
1977 rcu_read_unlock();
1978 return written;
1979 }
1980
1981 /*
1982 * Sample the snapshot positions for a specific fd
1983 *
1984 * Returns 0 on success, < 0 on error
1985 */
1986 int lttng_consumer_sample_snapshot_positions(struct lttng_consumer_stream *stream)
1987 {
1988 switch (consumer_data.type) {
1989 case LTTNG_CONSUMER_KERNEL:
1990 return lttng_kconsumer_sample_snapshot_positions(stream);
1991 case LTTNG_CONSUMER32_UST:
1992 case LTTNG_CONSUMER64_UST:
1993 return lttng_ustconsumer_sample_snapshot_positions(stream);
1994 default:
1995 ERR("Unknown consumer_data type");
1996 assert(0);
1997 return -ENOSYS;
1998 }
1999 }
2000 /*
2001 * Take a snapshot for a specific fd
2002 *
2003 * Returns 0 on success, < 0 on error
2004 */
2005 int lttng_consumer_take_snapshot(struct lttng_consumer_stream *stream)
2006 {
2007 switch (consumer_data.type) {
2008 case LTTNG_CONSUMER_KERNEL:
2009 return lttng_kconsumer_take_snapshot(stream);
2010 case LTTNG_CONSUMER32_UST:
2011 case LTTNG_CONSUMER64_UST:
2012 return lttng_ustconsumer_take_snapshot(stream);
2013 default:
2014 ERR("Unknown consumer_data type");
2015 assert(0);
2016 return -ENOSYS;
2017 }
2018 }
2019
2020 /*
2021 * Get the produced position
2022 *
2023 * Returns 0 on success, < 0 on error
2024 */
2025 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream *stream,
2026 unsigned long *pos)
2027 {
2028 switch (consumer_data.type) {
2029 case LTTNG_CONSUMER_KERNEL:
2030 return lttng_kconsumer_get_produced_snapshot(stream, pos);
2031 case LTTNG_CONSUMER32_UST:
2032 case LTTNG_CONSUMER64_UST:
2033 return lttng_ustconsumer_get_produced_snapshot(stream, pos);
2034 default:
2035 ERR("Unknown consumer_data type");
2036 assert(0);
2037 return -ENOSYS;
2038 }
2039 }
2040
2041 /*
2042 * Get the consumed position (free-running counter position in bytes).
2043 *
2044 * Returns 0 on success, < 0 on error
2045 */
2046 int lttng_consumer_get_consumed_snapshot(struct lttng_consumer_stream *stream,
2047 unsigned long *pos)
2048 {
2049 switch (consumer_data.type) {
2050 case LTTNG_CONSUMER_KERNEL:
2051 return lttng_kconsumer_get_consumed_snapshot(stream, pos);
2052 case LTTNG_CONSUMER32_UST:
2053 case LTTNG_CONSUMER64_UST:
2054 return lttng_ustconsumer_get_consumed_snapshot(stream, pos);
2055 default:
2056 ERR("Unknown consumer_data type");
2057 assert(0);
2058 return -ENOSYS;
2059 }
2060 }
2061
2062 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data *ctx,
2063 int sock, struct pollfd *consumer_sockpoll)
2064 {
2065 switch (consumer_data.type) {
2066 case LTTNG_CONSUMER_KERNEL:
2067 return lttng_kconsumer_recv_cmd(ctx, sock, consumer_sockpoll);
2068 case LTTNG_CONSUMER32_UST:
2069 case LTTNG_CONSUMER64_UST:
2070 return lttng_ustconsumer_recv_cmd(ctx, sock, consumer_sockpoll);
2071 default:
2072 ERR("Unknown consumer_data type");
2073 assert(0);
2074 return -ENOSYS;
2075 }
2076 }
2077
2078 void lttng_consumer_close_all_metadata(void)
2079 {
2080 switch (consumer_data.type) {
2081 case LTTNG_CONSUMER_KERNEL:
2082 /*
2083 * The Kernel consumer has a different metadata scheme so we don't
2084 * close anything because the stream will be closed by the session
2085 * daemon.
2086 */
2087 break;
2088 case LTTNG_CONSUMER32_UST:
2089 case LTTNG_CONSUMER64_UST:
2090 /*
2091 * Close all metadata streams. The metadata hash table is passed and
2092 * this call iterates over it by closing all wakeup fd. This is safe
2093 * because at this point we are sure that the metadata producer is
2094 * either dead or blocked.
2095 */
2096 lttng_ustconsumer_close_all_metadata(metadata_ht);
2097 break;
2098 default:
2099 ERR("Unknown consumer_data type");
2100 assert(0);
2101 }
2102 }
2103
2104 /*
2105 * Clean up a metadata stream and free its memory.
2106 */
2107 void consumer_del_metadata_stream(struct lttng_consumer_stream *stream,
2108 struct lttng_ht *ht)
2109 {
2110 struct lttng_consumer_channel *free_chan = NULL;
2111
2112 assert(stream);
2113 /*
2114 * This call should NEVER receive regular stream. It must always be
2115 * metadata stream and this is crucial for data structure synchronization.
2116 */
2117 assert(stream->metadata_flag);
2118
2119 DBG3("Consumer delete metadata stream %d", stream->wait_fd);
2120
2121 pthread_mutex_lock(&consumer_data.lock);
2122 pthread_mutex_lock(&stream->chan->lock);
2123 pthread_mutex_lock(&stream->lock);
2124 if (stream->chan->metadata_cache) {
2125 /* Only applicable to userspace consumers. */
2126 pthread_mutex_lock(&stream->chan->metadata_cache->lock);
2127 }
2128
2129 /* Remove any reference to that stream. */
2130 consumer_stream_delete(stream, ht);
2131
2132 /* Close down everything including the relayd if one. */
2133 consumer_stream_close(stream);
2134 /* Destroy tracer buffers of the stream. */
2135 consumer_stream_destroy_buffers(stream);
2136
2137 /* Atomically decrement channel refcount since other threads can use it. */
2138 if (!uatomic_sub_return(&stream->chan->refcount, 1)
2139 && !uatomic_read(&stream->chan->nb_init_stream_left)) {
2140 /* Go for channel deletion! */
2141 free_chan = stream->chan;
2142 }
2143
2144 /*
2145 * Nullify the stream reference so it is not used after deletion. The
2146 * channel lock MUST be acquired before being able to check for a NULL
2147 * pointer value.
2148 */
2149 stream->chan->metadata_stream = NULL;
2150
2151 if (stream->chan->metadata_cache) {
2152 pthread_mutex_unlock(&stream->chan->metadata_cache->lock);
2153 }
2154 pthread_mutex_unlock(&stream->lock);
2155 pthread_mutex_unlock(&stream->chan->lock);
2156 pthread_mutex_unlock(&consumer_data.lock);
2157
2158 if (free_chan) {
2159 consumer_del_channel(free_chan);
2160 }
2161
2162 consumer_stream_free(stream);
2163 }
2164
2165 /*
2166 * Action done with the metadata stream when adding it to the consumer internal
2167 * data structures to handle it.
2168 */
2169 void consumer_add_metadata_stream(struct lttng_consumer_stream *stream)
2170 {
2171 struct lttng_ht *ht = metadata_ht;
2172 struct lttng_ht_iter iter;
2173 struct lttng_ht_node_u64 *node;
2174
2175 assert(stream);
2176 assert(ht);
2177
2178 DBG3("Adding metadata stream %" PRIu64 " to hash table", stream->key);
2179
2180 pthread_mutex_lock(&consumer_data.lock);
2181 pthread_mutex_lock(&stream->chan->lock);
2182 pthread_mutex_lock(&stream->chan->timer_lock);
2183 pthread_mutex_lock(&stream->lock);
2184
2185 /*
2186 * From here, refcounts are updated so be _careful_ when returning an error
2187 * after this point.
2188 */
2189
2190 rcu_read_lock();
2191
2192 /*
2193 * Lookup the stream just to make sure it does not exist in our internal
2194 * state. This should NEVER happen.
2195 */
2196 lttng_ht_lookup(ht, &stream->key, &iter);
2197 node = lttng_ht_iter_get_node_u64(&iter);
2198 assert(!node);
2199
2200 /*
2201 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2202 * in terms of destroying the associated channel, because the action that
2203 * causes the count to become 0 also causes a stream to be added. The
2204 * channel deletion will thus be triggered by the following removal of this
2205 * stream.
2206 */
2207 if (uatomic_read(&stream->chan->nb_init_stream_left) > 0) {
2208 /* Increment refcount before decrementing nb_init_stream_left */
2209 cmm_smp_wmb();
2210 uatomic_dec(&stream->chan->nb_init_stream_left);
2211 }
2212
2213 lttng_ht_add_unique_u64(ht, &stream->node);
2214
2215 lttng_ht_add_u64(consumer_data.stream_per_chan_id_ht,
2216 &stream->node_channel_id);
2217
2218 /*
2219 * Add stream to the stream_list_ht of the consumer data. No need to steal
2220 * the key since the HT does not use it and we allow to add redundant keys
2221 * into this table.
2222 */
2223 lttng_ht_add_u64(consumer_data.stream_list_ht, &stream->node_session_id);
2224
2225 rcu_read_unlock();
2226
2227 pthread_mutex_unlock(&stream->lock);
2228 pthread_mutex_unlock(&stream->chan->lock);
2229 pthread_mutex_unlock(&stream->chan->timer_lock);
2230 pthread_mutex_unlock(&consumer_data.lock);
2231 }
2232
2233 /*
2234 * Delete data stream that are flagged for deletion (endpoint_status).
2235 */
2236 static void validate_endpoint_status_data_stream(void)
2237 {
2238 struct lttng_ht_iter iter;
2239 struct lttng_consumer_stream *stream;
2240
2241 DBG("Consumer delete flagged data stream");
2242
2243 rcu_read_lock();
2244 cds_lfht_for_each_entry(data_ht->ht, &iter.iter, stream, node.node) {
2245 /* Validate delete flag of the stream */
2246 if (stream->endpoint_status == CONSUMER_ENDPOINT_ACTIVE) {
2247 continue;
2248 }
2249 /* Delete it right now */
2250 consumer_del_stream(stream, data_ht);
2251 }
2252 rcu_read_unlock();
2253 }
2254
2255 /*
2256 * Delete metadata stream that are flagged for deletion (endpoint_status).
2257 */
2258 static void validate_endpoint_status_metadata_stream(
2259 struct lttng_poll_event *pollset)
2260 {
2261 struct lttng_ht_iter iter;
2262 struct lttng_consumer_stream *stream;
2263
2264 DBG("Consumer delete flagged metadata stream");
2265
2266 assert(pollset);
2267
2268 rcu_read_lock();
2269 cds_lfht_for_each_entry(metadata_ht->ht, &iter.iter, stream, node.node) {
2270 /* Validate delete flag of the stream */
2271 if (stream->endpoint_status == CONSUMER_ENDPOINT_ACTIVE) {
2272 continue;
2273 }
2274 /*
2275 * Remove from pollset so the metadata thread can continue without
2276 * blocking on a deleted stream.
2277 */
2278 lttng_poll_del(pollset, stream->wait_fd);
2279
2280 /* Delete it right now */
2281 consumer_del_metadata_stream(stream, metadata_ht);
2282 }
2283 rcu_read_unlock();
2284 }
2285
2286 static
2287 int rotate_notify_sessiond(struct lttng_consumer_local_data *ctx,
2288 uint64_t key)
2289 {
2290 ssize_t ret;
2291
2292 do {
2293 ret = write(ctx->channel_rotate_pipe, &key, sizeof(key));
2294 } while (ret == -1 && errno == EINTR);
2295 if (ret == -1) {
2296 PERROR("Failed to write to the channel rotation pipe");
2297 } else {
2298 DBG("Sent channel rotation notification for channel key %"
2299 PRIu64, key);
2300 ret = 0;
2301 }
2302
2303 return (int) ret;
2304 }
2305
2306 /*
2307 * Perform operations that need to be done after a stream has
2308 * rotated and released the stream lock.
2309 *
2310 * Multiple rotations cannot occur simultaneously, so we know the state of the
2311 * "rotated" stream flag cannot change.
2312 *
2313 * This MUST be called WITHOUT the stream lock held.
2314 */
2315 static
2316 int consumer_post_rotation(struct lttng_consumer_stream *stream,
2317 struct lttng_consumer_local_data *ctx)
2318 {
2319 int ret = 0;
2320
2321 pthread_mutex_lock(&stream->chan->lock);
2322
2323 switch (consumer_data.type) {
2324 case LTTNG_CONSUMER_KERNEL:
2325 break;
2326 case LTTNG_CONSUMER32_UST:
2327 case LTTNG_CONSUMER64_UST:
2328 /*
2329 * The ust_metadata_pushed counter has been reset to 0, so now
2330 * we can wakeup the metadata thread so it dumps the metadata
2331 * cache to the new file.
2332 */
2333 if (stream->metadata_flag) {
2334 consumer_metadata_wakeup_pipe(stream->chan);
2335 }
2336 break;
2337 default:
2338 ERR("Unknown consumer_data type");
2339 abort();
2340 }
2341
2342 if (--stream->chan->nr_stream_rotate_pending == 0) {
2343 DBG("Rotation of channel \"%s\" completed, notifying the session daemon",
2344 stream->chan->name);
2345 ret = rotate_notify_sessiond(ctx, stream->chan->key);
2346 }
2347 pthread_mutex_unlock(&stream->chan->lock);
2348
2349 return ret;
2350 }
2351
2352 /*
2353 * Thread polls on metadata file descriptor and write them on disk or on the
2354 * network.
2355 */
2356 void *consumer_thread_metadata_poll(void *data)
2357 {
2358 int ret, i, pollfd, err = -1;
2359 uint32_t revents, nb_fd;
2360 struct lttng_consumer_stream *stream = NULL;
2361 struct lttng_ht_iter iter;
2362 struct lttng_ht_node_u64 *node;
2363 struct lttng_poll_event events;
2364 struct lttng_consumer_local_data *ctx = data;
2365 ssize_t len;
2366
2367 rcu_register_thread();
2368
2369 health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_METADATA);
2370
2371 if (testpoint(consumerd_thread_metadata)) {
2372 goto error_testpoint;
2373 }
2374
2375 health_code_update();
2376
2377 DBG("Thread metadata poll started");
2378
2379 /* Size is set to 1 for the consumer_metadata pipe */
2380 ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC);
2381 if (ret < 0) {
2382 ERR("Poll set creation failed");
2383 goto end_poll;
2384 }
2385
2386 ret = lttng_poll_add(&events,
2387 lttng_pipe_get_readfd(ctx->consumer_metadata_pipe), LPOLLIN);
2388 if (ret < 0) {
2389 goto end;
2390 }
2391
2392 /* Main loop */
2393 DBG("Metadata main loop started");
2394
2395 while (1) {
2396 restart:
2397 health_code_update();
2398 health_poll_entry();
2399 DBG("Metadata poll wait");
2400 ret = lttng_poll_wait(&events, -1);
2401 DBG("Metadata poll return from wait with %d fd(s)",
2402 LTTNG_POLL_GETNB(&events));
2403 health_poll_exit();
2404 DBG("Metadata event caught in thread");
2405 if (ret < 0) {
2406 if (errno == EINTR) {
2407 ERR("Poll EINTR caught");
2408 goto restart;
2409 }
2410 if (LTTNG_POLL_GETNB(&events) == 0) {
2411 err = 0; /* All is OK */
2412 }
2413 goto end;
2414 }
2415
2416 nb_fd = ret;
2417
2418 /* From here, the event is a metadata wait fd */
2419 for (i = 0; i < nb_fd; i++) {
2420 health_code_update();
2421
2422 revents = LTTNG_POLL_GETEV(&events, i);
2423 pollfd = LTTNG_POLL_GETFD(&events, i);
2424
2425 if (!revents) {
2426 /* No activity for this FD (poll implementation). */
2427 continue;
2428 }
2429
2430 if (pollfd == lttng_pipe_get_readfd(ctx->consumer_metadata_pipe)) {
2431 if (revents & LPOLLIN) {
2432 ssize_t pipe_len;
2433
2434 pipe_len = lttng_pipe_read(ctx->consumer_metadata_pipe,
2435 &stream, sizeof(stream));
2436 if (pipe_len < sizeof(stream)) {
2437 if (pipe_len < 0) {
2438 PERROR("read metadata stream");
2439 }
2440 /*
2441 * Remove the pipe from the poll set and continue the loop
2442 * since their might be data to consume.
2443 */
2444 lttng_poll_del(&events,
2445 lttng_pipe_get_readfd(ctx->consumer_metadata_pipe));
2446 lttng_pipe_read_close(ctx->consumer_metadata_pipe);
2447 continue;
2448 }
2449
2450 /* A NULL stream means that the state has changed. */
2451 if (stream == NULL) {
2452 /* Check for deleted streams. */
2453 validate_endpoint_status_metadata_stream(&events);
2454 goto restart;
2455 }
2456
2457 DBG("Adding metadata stream %d to poll set",
2458 stream->wait_fd);
2459
2460 /* Add metadata stream to the global poll events list */
2461 lttng_poll_add(&events, stream->wait_fd,
2462 LPOLLIN | LPOLLPRI | LPOLLHUP);
2463 } else if (revents & (LPOLLERR | LPOLLHUP)) {
2464 DBG("Metadata thread pipe hung up");
2465 /*
2466 * Remove the pipe from the poll set and continue the loop
2467 * since their might be data to consume.
2468 */
2469 lttng_poll_del(&events,
2470 lttng_pipe_get_readfd(ctx->consumer_metadata_pipe));
2471 lttng_pipe_read_close(ctx->consumer_metadata_pipe);
2472 continue;
2473 } else {
2474 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
2475 goto end;
2476 }
2477
2478 /* Handle other stream */
2479 continue;
2480 }
2481
2482 rcu_read_lock();
2483 {
2484 uint64_t tmp_id = (uint64_t) pollfd;
2485
2486 lttng_ht_lookup(metadata_ht, &tmp_id, &iter);
2487 }
2488 node = lttng_ht_iter_get_node_u64(&iter);
2489 assert(node);
2490
2491 stream = caa_container_of(node, struct lttng_consumer_stream,
2492 node);
2493
2494 if (revents & (LPOLLIN | LPOLLPRI)) {
2495 /* Get the data out of the metadata file descriptor */
2496 DBG("Metadata available on fd %d", pollfd);
2497 assert(stream->wait_fd == pollfd);
2498
2499 do {
2500 health_code_update();
2501
2502 len = ctx->on_buffer_ready(stream, ctx);
2503 /*
2504 * We don't check the return value here since if we get
2505 * a negative len, it means an error occurred thus we
2506 * simply remove it from the poll set and free the
2507 * stream.
2508 */
2509 } while (len > 0);
2510
2511 /* It's ok to have an unavailable sub-buffer */
2512 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2513 /* Clean up stream from consumer and free it. */
2514 lttng_poll_del(&events, stream->wait_fd);
2515 consumer_del_metadata_stream(stream, metadata_ht);
2516 }
2517 } else if (revents & (LPOLLERR | LPOLLHUP)) {
2518 DBG("Metadata fd %d is hup|err.", pollfd);
2519 if (!stream->hangup_flush_done
2520 && (consumer_data.type == LTTNG_CONSUMER32_UST
2521 || consumer_data.type == LTTNG_CONSUMER64_UST)) {
2522 DBG("Attempting to flush and consume the UST buffers");
2523 lttng_ustconsumer_on_stream_hangup(stream);
2524
2525 /* We just flushed the stream now read it. */
2526 do {
2527 health_code_update();
2528
2529 len = ctx->on_buffer_ready(stream, ctx);
2530 /*
2531 * We don't check the return value here since if we get
2532 * a negative len, it means an error occurred thus we
2533 * simply remove it from the poll set and free the
2534 * stream.
2535 */
2536 } while (len > 0);
2537 }
2538
2539 lttng_poll_del(&events, stream->wait_fd);
2540 /*
2541 * This call update the channel states, closes file descriptors
2542 * and securely free the stream.
2543 */
2544 consumer_del_metadata_stream(stream, metadata_ht);
2545 } else {
2546 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
2547 rcu_read_unlock();
2548 goto end;
2549 }
2550 /* Release RCU lock for the stream looked up */
2551 rcu_read_unlock();
2552 }
2553 }
2554
2555 /* All is OK */
2556 err = 0;
2557 end:
2558 DBG("Metadata poll thread exiting");
2559
2560 lttng_poll_clean(&events);
2561 end_poll:
2562 error_testpoint:
2563 if (err) {
2564 health_error();
2565 ERR("Health error occurred in %s", __func__);
2566 }
2567 health_unregister(health_consumerd);
2568 rcu_unregister_thread();
2569 return NULL;
2570 }
2571
2572 /*
2573 * This thread polls the fds in the set to consume the data and write
2574 * it to tracefile if necessary.
2575 */
2576 void *consumer_thread_data_poll(void *data)
2577 {
2578 int num_rdy, num_hup, high_prio, ret, i, err = -1;
2579 struct pollfd *pollfd = NULL;
2580 /* local view of the streams */
2581 struct lttng_consumer_stream **local_stream = NULL, *new_stream = NULL;
2582 /* local view of consumer_data.fds_count */
2583 int nb_fd = 0;
2584 /* 2 for the consumer_data_pipe and wake up pipe */
2585 const int nb_pipes_fd = 2;
2586 /* Number of FDs with CONSUMER_ENDPOINT_INACTIVE but still open. */
2587 int nb_inactive_fd = 0;
2588 struct lttng_consumer_local_data *ctx = data;
2589 ssize_t len;
2590
2591 rcu_register_thread();
2592
2593 health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_DATA);
2594
2595 if (testpoint(consumerd_thread_data)) {
2596 goto error_testpoint;
2597 }
2598
2599 health_code_update();
2600
2601 local_stream = zmalloc(sizeof(struct lttng_consumer_stream *));
2602 if (local_stream == NULL) {
2603 PERROR("local_stream malloc");
2604 goto end;
2605 }
2606
2607 while (1) {
2608 health_code_update();
2609
2610 high_prio = 0;
2611 num_hup = 0;
2612
2613 /*
2614 * the fds set has been updated, we need to update our
2615 * local array as well
2616 */
2617 pthread_mutex_lock(&consumer_data.lock);
2618 if (consumer_data.need_update) {
2619 free(pollfd);
2620 pollfd = NULL;
2621
2622 free(local_stream);
2623 local_stream = NULL;
2624
2625 /* Allocate for all fds */
2626 pollfd = zmalloc((consumer_data.stream_count + nb_pipes_fd) * sizeof(struct pollfd));
2627 if (pollfd == NULL) {
2628 PERROR("pollfd malloc");
2629 pthread_mutex_unlock(&consumer_data.lock);
2630 goto end;
2631 }
2632
2633 local_stream = zmalloc((consumer_data.stream_count + nb_pipes_fd) *
2634 sizeof(struct lttng_consumer_stream *));
2635 if (local_stream == NULL) {
2636 PERROR("local_stream malloc");
2637 pthread_mutex_unlock(&consumer_data.lock);
2638 goto end;
2639 }
2640 ret = update_poll_array(ctx, &pollfd, local_stream,
2641 data_ht, &nb_inactive_fd);
2642 if (ret < 0) {
2643 ERR("Error in allocating pollfd or local_outfds");
2644 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
2645 pthread_mutex_unlock(&consumer_data.lock);
2646 goto end;
2647 }
2648 nb_fd = ret;
2649 consumer_data.need_update = 0;
2650 }
2651 pthread_mutex_unlock(&consumer_data.lock);
2652
2653 /* No FDs and consumer_quit, consumer_cleanup the thread */
2654 if (nb_fd == 0 && nb_inactive_fd == 0 &&
2655 CMM_LOAD_SHARED(consumer_quit) == 1) {
2656 err = 0; /* All is OK */
2657 goto end;
2658 }
2659 /* poll on the array of fds */
2660 restart:
2661 DBG("polling on %d fd", nb_fd + nb_pipes_fd);
2662 if (testpoint(consumerd_thread_data_poll)) {
2663 goto end;
2664 }
2665 health_poll_entry();
2666 num_rdy = poll(pollfd, nb_fd + nb_pipes_fd, -1);
2667 health_poll_exit();
2668 DBG("poll num_rdy : %d", num_rdy);
2669 if (num_rdy == -1) {
2670 /*
2671 * Restart interrupted system call.
2672 */
2673 if (errno == EINTR) {
2674 goto restart;
2675 }
2676 PERROR("Poll error");
2677 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
2678 goto end;
2679 } else if (num_rdy == 0) {
2680 DBG("Polling thread timed out");
2681 goto end;
2682 }
2683
2684 if (caa_unlikely(data_consumption_paused)) {
2685 DBG("Data consumption paused, sleeping...");
2686 sleep(1);
2687 goto restart;
2688 }
2689
2690 /*
2691 * If the consumer_data_pipe triggered poll go directly to the
2692 * beginning of the loop to update the array. We want to prioritize
2693 * array update over low-priority reads.
2694 */
2695 if (pollfd[nb_fd].revents & (POLLIN | POLLPRI)) {
2696 ssize_t pipe_readlen;
2697
2698 DBG("consumer_data_pipe wake up");
2699 pipe_readlen = lttng_pipe_read(ctx->consumer_data_pipe,
2700 &new_stream, sizeof(new_stream));
2701 if (pipe_readlen < sizeof(new_stream)) {
2702 PERROR("Consumer data pipe");
2703 /* Continue so we can at least handle the current stream(s). */
2704 continue;
2705 }
2706
2707 /*
2708 * If the stream is NULL, just ignore it. It's also possible that
2709 * the sessiond poll thread changed the consumer_quit state and is
2710 * waking us up to test it.
2711 */
2712 if (new_stream == NULL) {
2713 validate_endpoint_status_data_stream();
2714 continue;
2715 }
2716
2717 /* Continue to update the local streams and handle prio ones */
2718 continue;
2719 }
2720
2721 /* Handle wakeup pipe. */
2722 if (pollfd[nb_fd + 1].revents & (POLLIN | POLLPRI)) {
2723 char dummy;
2724 ssize_t pipe_readlen;
2725
2726 pipe_readlen = lttng_pipe_read(ctx->consumer_wakeup_pipe, &dummy,
2727 sizeof(dummy));
2728 if (pipe_readlen < 0) {
2729 PERROR("Consumer data wakeup pipe");
2730 }
2731 /* We've been awakened to handle stream(s). */
2732 ctx->has_wakeup = 0;
2733 }
2734
2735 /* Take care of high priority channels first. */
2736 for (i = 0; i < nb_fd; i++) {
2737 health_code_update();
2738
2739 if (local_stream[i] == NULL) {
2740 continue;
2741 }
2742 if (pollfd[i].revents & POLLPRI) {
2743 DBG("Urgent read on fd %d", pollfd[i].fd);
2744 high_prio = 1;
2745 len = ctx->on_buffer_ready(local_stream[i], ctx);
2746 /* it's ok to have an unavailable sub-buffer */
2747 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2748 /* Clean the stream and free it. */
2749 consumer_del_stream(local_stream[i], data_ht);
2750 local_stream[i] = NULL;
2751 } else if (len > 0) {
2752 local_stream[i]->data_read = 1;
2753 }
2754 }
2755 }
2756
2757 /*
2758 * If we read high prio channel in this loop, try again
2759 * for more high prio data.
2760 */
2761 if (high_prio) {
2762 continue;
2763 }
2764
2765 /* Take care of low priority channels. */
2766 for (i = 0; i < nb_fd; i++) {
2767 health_code_update();
2768
2769 if (local_stream[i] == NULL) {
2770 continue;
2771 }
2772 if ((pollfd[i].revents & POLLIN) ||
2773 local_stream[i]->hangup_flush_done ||
2774 local_stream[i]->has_data) {
2775 DBG("Normal read on fd %d", pollfd[i].fd);
2776 len = ctx->on_buffer_ready(local_stream[i], ctx);
2777 /* it's ok to have an unavailable sub-buffer */
2778 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2779 /* Clean the stream and free it. */
2780 consumer_del_stream(local_stream[i], data_ht);
2781 local_stream[i] = NULL;
2782 } else if (len > 0) {
2783 local_stream[i]->data_read = 1;
2784 }
2785 }
2786 }
2787
2788 /* Handle hangup and errors */
2789 for (i = 0; i < nb_fd; i++) {
2790 health_code_update();
2791
2792 if (local_stream[i] == NULL) {
2793 continue;
2794 }
2795 if (!local_stream[i]->hangup_flush_done
2796 && (pollfd[i].revents & (POLLHUP | POLLERR | POLLNVAL))
2797 && (consumer_data.type == LTTNG_CONSUMER32_UST
2798 || consumer_data.type == LTTNG_CONSUMER64_UST)) {
2799 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2800 pollfd[i].fd);
2801 lttng_ustconsumer_on_stream_hangup(local_stream[i]);
2802 /* Attempt read again, for the data we just flushed. */
2803 local_stream[i]->data_read = 1;
2804 }
2805 /*
2806 * If the poll flag is HUP/ERR/NVAL and we have
2807 * read no data in this pass, we can remove the
2808 * stream from its hash table.
2809 */
2810 if ((pollfd[i].revents & POLLHUP)) {
2811 DBG("Polling fd %d tells it has hung up.", pollfd[i].fd);
2812 if (!local_stream[i]->data_read) {
2813 consumer_del_stream(local_stream[i], data_ht);
2814 local_stream[i] = NULL;
2815 num_hup++;
2816 }
2817 } else if (pollfd[i].revents & POLLERR) {
2818 ERR("Error returned in polling fd %d.", pollfd[i].fd);
2819 if (!local_stream[i]->data_read) {
2820 consumer_del_stream(local_stream[i], data_ht);
2821 local_stream[i] = NULL;
2822 num_hup++;
2823 }
2824 } else if (pollfd[i].revents & POLLNVAL) {
2825 ERR("Polling fd %d tells fd is not open.", pollfd[i].fd);
2826 if (!local_stream[i]->data_read) {
2827 consumer_del_stream(local_stream[i], data_ht);
2828 local_stream[i] = NULL;
2829 num_hup++;
2830 }
2831 }
2832 if (local_stream[i] != NULL) {
2833 local_stream[i]->data_read = 0;
2834 }
2835 }
2836 }
2837 /* All is OK */
2838 err = 0;
2839 end:
2840 DBG("polling thread exiting");
2841 free(pollfd);
2842 free(local_stream);
2843
2844 /*
2845 * Close the write side of the pipe so epoll_wait() in
2846 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2847 * read side of the pipe. If we close them both, epoll_wait strangely does
2848 * not return and could create a endless wait period if the pipe is the
2849 * only tracked fd in the poll set. The thread will take care of closing
2850 * the read side.
2851 */
2852 (void) lttng_pipe_write_close(ctx->consumer_metadata_pipe);
2853
2854 error_testpoint:
2855 if (err) {
2856 health_error();
2857 ERR("Health error occurred in %s", __func__);
2858 }
2859 health_unregister(health_consumerd);
2860
2861 rcu_unregister_thread();
2862 return NULL;
2863 }
2864
2865 /*
2866 * Close wake-up end of each stream belonging to the channel. This will
2867 * allow the poll() on the stream read-side to detect when the
2868 * write-side (application) finally closes them.
2869 */
2870 static
2871 void consumer_close_channel_streams(struct lttng_consumer_channel *channel)
2872 {
2873 struct lttng_ht *ht;
2874 struct lttng_consumer_stream *stream;
2875 struct lttng_ht_iter iter;
2876
2877 ht = consumer_data.stream_per_chan_id_ht;
2878
2879 rcu_read_lock();
2880 cds_lfht_for_each_entry_duplicate(ht->ht,
2881 ht->hash_fct(&channel->key, lttng_ht_seed),
2882 ht->match_fct, &channel->key,
2883 &iter.iter, stream, node_channel_id.node) {
2884 /*
2885 * Protect against teardown with mutex.
2886 */
2887 pthread_mutex_lock(&stream->lock);
2888 if (cds_lfht_is_node_deleted(&stream->node.node)) {
2889 goto next;
2890 }
2891 switch (consumer_data.type) {
2892 case LTTNG_CONSUMER_KERNEL:
2893 break;
2894 case LTTNG_CONSUMER32_UST:
2895 case LTTNG_CONSUMER64_UST:
2896 if (stream->metadata_flag) {
2897 /* Safe and protected by the stream lock. */
2898 lttng_ustconsumer_close_metadata(stream->chan);
2899 } else {
2900 /*
2901 * Note: a mutex is taken internally within
2902 * liblttng-ust-ctl to protect timer wakeup_fd
2903 * use from concurrent close.
2904 */
2905 lttng_ustconsumer_close_stream_wakeup(stream);
2906 }
2907 break;
2908 default:
2909 ERR("Unknown consumer_data type");
2910 assert(0);
2911 }
2912 next:
2913 pthread_mutex_unlock(&stream->lock);
2914 }
2915 rcu_read_unlock();
2916 }
2917
2918 static void destroy_channel_ht(struct lttng_ht *ht)
2919 {
2920 struct lttng_ht_iter iter;
2921 struct lttng_consumer_channel *channel;
2922 int ret;
2923
2924 if (ht == NULL) {
2925 return;
2926 }
2927
2928 rcu_read_lock();
2929 cds_lfht_for_each_entry(ht->ht, &iter.iter, channel, wait_fd_node.node) {
2930 ret = lttng_ht_del(ht, &iter);
2931 assert(ret != 0);
2932 }
2933 rcu_read_unlock();
2934
2935 lttng_ht_destroy(ht);
2936 }
2937
2938 /*
2939 * This thread polls the channel fds to detect when they are being
2940 * closed. It closes all related streams if the channel is detected as
2941 * closed. It is currently only used as a shim layer for UST because the
2942 * consumerd needs to keep the per-stream wakeup end of pipes open for
2943 * periodical flush.
2944 */
2945 void *consumer_thread_channel_poll(void *data)
2946 {
2947 int ret, i, pollfd, err = -1;
2948 uint32_t revents, nb_fd;
2949 struct lttng_consumer_channel *chan = NULL;
2950 struct lttng_ht_iter iter;
2951 struct lttng_ht_node_u64 *node;
2952 struct lttng_poll_event events;
2953 struct lttng_consumer_local_data *ctx = data;
2954 struct lttng_ht *channel_ht;
2955
2956 rcu_register_thread();
2957
2958 health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_CHANNEL);
2959
2960 if (testpoint(consumerd_thread_channel)) {
2961 goto error_testpoint;
2962 }
2963
2964 health_code_update();
2965
2966 channel_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
2967 if (!channel_ht) {
2968 /* ENOMEM at this point. Better to bail out. */
2969 goto end_ht;
2970 }
2971
2972 DBG("Thread channel poll started");
2973
2974 /* Size is set to 1 for the consumer_channel pipe */
2975 ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC);
2976 if (ret < 0) {
2977 ERR("Poll set creation failed");
2978 goto end_poll;
2979 }
2980
2981 ret = lttng_poll_add(&events, ctx->consumer_channel_pipe[0], LPOLLIN);
2982 if (ret < 0) {
2983 goto end;
2984 }
2985
2986 /* Main loop */
2987 DBG("Channel main loop started");
2988
2989 while (1) {
2990 restart:
2991 health_code_update();
2992 DBG("Channel poll wait");
2993 health_poll_entry();
2994 ret = lttng_poll_wait(&events, -1);
2995 DBG("Channel poll return from wait with %d fd(s)",
2996 LTTNG_POLL_GETNB(&events));
2997 health_poll_exit();
2998 DBG("Channel event caught in thread");
2999 if (ret < 0) {
3000 if (errno == EINTR) {
3001 ERR("Poll EINTR caught");
3002 goto restart;
3003 }
3004 if (LTTNG_POLL_GETNB(&events) == 0) {
3005 err = 0; /* All is OK */
3006 }
3007 goto end;
3008 }
3009
3010 nb_fd = ret;
3011
3012 /* From here, the event is a channel wait fd */
3013 for (i = 0; i < nb_fd; i++) {
3014 health_code_update();
3015
3016 revents = LTTNG_POLL_GETEV(&events, i);
3017 pollfd = LTTNG_POLL_GETFD(&events, i);
3018
3019 if (!revents) {
3020 /* No activity for this FD (poll implementation). */
3021 continue;
3022 }
3023
3024 if (pollfd == ctx->consumer_channel_pipe[0]) {
3025 if (revents & LPOLLIN) {
3026 enum consumer_channel_action action;
3027 uint64_t key;
3028
3029 ret = read_channel_pipe(ctx, &chan, &key, &action);
3030 if (ret <= 0) {
3031 if (ret < 0) {
3032 ERR("Error reading channel pipe");
3033 }
3034 lttng_poll_del(&events, ctx->consumer_channel_pipe[0]);
3035 continue;
3036 }
3037
3038 switch (action) {
3039 case CONSUMER_CHANNEL_ADD:
3040 DBG("Adding channel %d to poll set",
3041 chan->wait_fd);
3042
3043 lttng_ht_node_init_u64(&chan->wait_fd_node,
3044 chan->wait_fd);
3045 rcu_read_lock();
3046 lttng_ht_add_unique_u64(channel_ht,
3047 &chan->wait_fd_node);
3048 rcu_read_unlock();
3049 /* Add channel to the global poll events list */
3050 lttng_poll_add(&events, chan->wait_fd,
3051 LPOLLERR | LPOLLHUP);
3052 break;
3053 case CONSUMER_CHANNEL_DEL:
3054 {
3055 /*
3056 * This command should never be called if the channel
3057 * has streams monitored by either the data or metadata
3058 * thread. The consumer only notify this thread with a
3059 * channel del. command if it receives a destroy
3060 * channel command from the session daemon that send it
3061 * if a command prior to the GET_CHANNEL failed.
3062 */
3063
3064 rcu_read_lock();
3065 chan = consumer_find_channel(key);
3066 if (!chan) {
3067 rcu_read_unlock();
3068 ERR("UST consumer get channel key %" PRIu64 " not found for del channel", key);
3069 break;
3070 }
3071 lttng_poll_del(&events, chan->wait_fd);
3072 iter.iter.node = &chan->wait_fd_node.node;
3073 ret = lttng_ht_del(channel_ht, &iter);
3074 assert(ret == 0);
3075
3076 switch (consumer_data.type) {
3077 case LTTNG_CONSUMER_KERNEL:
3078 break;
3079 case LTTNG_CONSUMER32_UST:
3080 case LTTNG_CONSUMER64_UST:
3081 health_code_update();
3082 /* Destroy streams that might have been left in the stream list. */
3083 clean_channel_stream_list(chan);
3084 break;
3085 default:
3086 ERR("Unknown consumer_data type");
3087 assert(0);
3088 }
3089
3090 /*
3091 * Release our own refcount. Force channel deletion even if
3092 * streams were not initialized.
3093 */
3094 if (!uatomic_sub_return(&chan->refcount, 1)) {
3095 consumer_del_channel(chan);
3096 }
3097 rcu_read_unlock();
3098 goto restart;
3099 }
3100 case CONSUMER_CHANNEL_QUIT:
3101 /*
3102 * Remove the pipe from the poll set and continue the loop
3103 * since their might be data to consume.
3104 */
3105 lttng_poll_del(&events, ctx->consumer_channel_pipe[0]);
3106 continue;
3107 default:
3108 ERR("Unknown action");
3109 break;
3110 }
3111 } else if (revents & (LPOLLERR | LPOLLHUP)) {
3112 DBG("Channel thread pipe hung up");
3113 /*
3114 * Remove the pipe from the poll set and continue the loop
3115 * since their might be data to consume.
3116 */
3117 lttng_poll_del(&events, ctx->consumer_channel_pipe[0]);
3118 continue;
3119 } else {
3120 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
3121 goto end;
3122 }
3123
3124 /* Handle other stream */
3125 continue;
3126 }
3127
3128 rcu_read_lock();
3129 {
3130 uint64_t tmp_id = (uint64_t) pollfd;
3131
3132 lttng_ht_lookup(channel_ht, &tmp_id, &iter);
3133 }
3134 node = lttng_ht_iter_get_node_u64(&iter);
3135 assert(node);
3136
3137 chan = caa_container_of(node, struct lttng_consumer_channel,
3138 wait_fd_node);
3139
3140 /* Check for error event */
3141 if (revents & (LPOLLERR | LPOLLHUP)) {
3142 DBG("Channel fd %d is hup|err.", pollfd);
3143
3144 lttng_poll_del(&events, chan->wait_fd);
3145 ret = lttng_ht_del(channel_ht, &iter);
3146 assert(ret == 0);
3147
3148 /*
3149 * This will close the wait fd for each stream associated to
3150 * this channel AND monitored by the data/metadata thread thus
3151 * will be clean by the right thread.
3152 */
3153 consumer_close_channel_streams(chan);
3154
3155 /* Release our own refcount */
3156 if (!uatomic_sub_return(&chan->refcount, 1)
3157 && !uatomic_read(&chan->nb_init_stream_left)) {
3158 consumer_del_channel(chan);
3159 }
3160 } else {
3161 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
3162 rcu_read_unlock();
3163 goto end;
3164 }
3165
3166 /* Release RCU lock for the channel looked up */
3167 rcu_read_unlock();
3168 }
3169 }
3170
3171 /* All is OK */
3172 err = 0;
3173 end:
3174 lttng_poll_clean(&events);
3175 end_poll:
3176 destroy_channel_ht(channel_ht);
3177 end_ht:
3178 error_testpoint:
3179 DBG("Channel poll thread exiting");
3180 if (err) {
3181 health_error();
3182 ERR("Health error occurred in %s", __func__);
3183 }
3184 health_unregister(health_consumerd);
3185 rcu_unregister_thread();
3186 return NULL;
3187 }
3188
3189 static int set_metadata_socket(struct lttng_consumer_local_data *ctx,
3190 struct pollfd *sockpoll, int client_socket)
3191 {
3192 int ret;
3193
3194 assert(ctx);
3195 assert(sockpoll);
3196
3197 ret = lttng_consumer_poll_socket(sockpoll);
3198 if (ret) {
3199 goto error;
3200 }
3201 DBG("Metadata connection on client_socket");
3202
3203 /* Blocking call, waiting for transmission */
3204 ctx->consumer_metadata_socket = lttcomm_accept_unix_sock(client_socket);
3205 if (ctx->consumer_metadata_socket < 0) {
3206 WARN("On accept metadata");
3207 ret = -1;
3208 goto error;
3209 }
3210 ret = 0;
3211
3212 error:
3213 return ret;
3214 }
3215
3216 /*
3217 * This thread listens on the consumerd socket and receives the file
3218 * descriptors from the session daemon.
3219 */
3220 void *consumer_thread_sessiond_poll(void *data)
3221 {
3222 int sock = -1, client_socket