1ea682f028e76ebe546b3dc892981bc10d2009a5
[lttng-tools.git] / src / common / consumer / consumer.c
1 /*
2 * Copyright (C) 2011 - Julien Desfossez <julien.desfossez@polymtl.ca>
3 * Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
4 * 2012 - David Goulet <dgoulet@efficios.com>
5 *
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License, version 2 only,
8 * as published by the Free Software Foundation.
9 *
10 * This program is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
13 * more details.
14 *
15 * You should have received a copy of the GNU General Public License along
16 * with this program; if not, write to the Free Software Foundation, Inc.,
17 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
18 */
19
20 #define _LGPL_SOURCE
21 #include <assert.h>
22 #include <poll.h>
23 #include <pthread.h>
24 #include <stdlib.h>
25 #include <string.h>
26 #include <sys/mman.h>
27 #include <sys/socket.h>
28 #include <sys/types.h>
29 #include <unistd.h>
30 #include <inttypes.h>
31 #include <signal.h>
32
33 #include <bin/lttng-consumerd/health-consumerd.h>
34 #include <common/common.h>
35 #include <common/utils.h>
36 #include <common/time.h>
37 #include <common/compat/poll.h>
38 #include <common/compat/endian.h>
39 #include <common/index/index.h>
40 #include <common/kernel-ctl/kernel-ctl.h>
41 #include <common/sessiond-comm/relayd.h>
42 #include <common/sessiond-comm/sessiond-comm.h>
43 #include <common/kernel-consumer/kernel-consumer.h>
44 #include <common/relayd/relayd.h>
45 #include <common/ust-consumer/ust-consumer.h>
46 #include <common/consumer/consumer-timer.h>
47 #include <common/consumer/consumer.h>
48 #include <common/consumer/consumer-stream.h>
49 #include <common/consumer/consumer-testpoint.h>
50 #include <common/align.h>
51 #include <common/consumer/consumer-metadata-cache.h>
52 #include <common/trace-chunk.h>
53 #include <common/trace-chunk-registry.h>
54 #include <common/string-utils/format.h>
55 #include <common/dynamic-array.h>
56
57 struct lttng_consumer_global_data consumer_data = {
58 .stream_count = 0,
59 .need_update = 1,
60 .type = LTTNG_CONSUMER_UNKNOWN,
61 };
62
63 enum consumer_channel_action {
64 CONSUMER_CHANNEL_ADD,
65 CONSUMER_CHANNEL_DEL,
66 CONSUMER_CHANNEL_QUIT,
67 };
68
69 struct consumer_channel_msg {
70 enum consumer_channel_action action;
71 struct lttng_consumer_channel *chan; /* add */
72 uint64_t key; /* del */
73 };
74
75 /* Flag used to temporarily pause data consumption from testpoints. */
76 int data_consumption_paused;
77
78 /*
79 * Flag to inform the polling thread to quit when all fd hung up. Updated by
80 * the consumer_thread_receive_fds when it notices that all fds has hung up.
81 * Also updated by the signal handler (consumer_should_exit()). Read by the
82 * polling threads.
83 */
84 int consumer_quit;
85
86 /*
87 * Global hash table containing respectively metadata and data streams. The
88 * stream element in this ht should only be updated by the metadata poll thread
89 * for the metadata and the data poll thread for the data.
90 */
91 static struct lttng_ht *metadata_ht;
92 static struct lttng_ht *data_ht;
93
94 /*
95 * Notify a thread lttng pipe to poll back again. This usually means that some
96 * global state has changed so we just send back the thread in a poll wait
97 * call.
98 */
99 static void notify_thread_lttng_pipe(struct lttng_pipe *pipe)
100 {
101 struct lttng_consumer_stream *null_stream = NULL;
102
103 assert(pipe);
104
105 (void) lttng_pipe_write(pipe, &null_stream, sizeof(null_stream));
106 }
107
108 static void notify_health_quit_pipe(int *pipe)
109 {
110 ssize_t ret;
111
112 ret = lttng_write(pipe[1], "4", 1);
113 if (ret < 1) {
114 PERROR("write consumer health quit");
115 }
116 }
117
118 static void notify_channel_pipe(struct lttng_consumer_local_data *ctx,
119 struct lttng_consumer_channel *chan,
120 uint64_t key,
121 enum consumer_channel_action action)
122 {
123 struct consumer_channel_msg msg;
124 ssize_t ret;
125
126 memset(&msg, 0, sizeof(msg));
127
128 msg.action = action;
129 msg.chan = chan;
130 msg.key = key;
131 ret = lttng_write(ctx->consumer_channel_pipe[1], &msg, sizeof(msg));
132 if (ret < sizeof(msg)) {
133 PERROR("notify_channel_pipe write error");
134 }
135 }
136
137 void notify_thread_del_channel(struct lttng_consumer_local_data *ctx,
138 uint64_t key)
139 {
140 notify_channel_pipe(ctx, NULL, key, CONSUMER_CHANNEL_DEL);
141 }
142
143 static int read_channel_pipe(struct lttng_consumer_local_data *ctx,
144 struct lttng_consumer_channel **chan,
145 uint64_t *key,
146 enum consumer_channel_action *action)
147 {
148 struct consumer_channel_msg msg;
149 ssize_t ret;
150
151 ret = lttng_read(ctx->consumer_channel_pipe[0], &msg, sizeof(msg));
152 if (ret < sizeof(msg)) {
153 ret = -1;
154 goto error;
155 }
156 *action = msg.action;
157 *chan = msg.chan;
158 *key = msg.key;
159 error:
160 return (int) ret;
161 }
162
163 /*
164 * Cleanup the stream list of a channel. Those streams are not yet globally
165 * visible
166 */
167 static void clean_channel_stream_list(struct lttng_consumer_channel *channel)
168 {
169 struct lttng_consumer_stream *stream, *stmp;
170
171 assert(channel);
172
173 /* Delete streams that might have been left in the stream list. */
174 cds_list_for_each_entry_safe(stream, stmp, &channel->streams.head,
175 send_node) {
176 cds_list_del(&stream->send_node);
177 /*
178 * Once a stream is added to this list, the buffers were created so we
179 * have a guarantee that this call will succeed. Setting the monitor
180 * mode to 0 so we don't lock nor try to delete the stream from the
181 * global hash table.
182 */
183 stream->monitor = 0;
184 consumer_stream_destroy(stream, NULL);
185 }
186 }
187
188 /*
189 * Find a stream. The consumer_data.lock must be locked during this
190 * call.
191 */
192 static struct lttng_consumer_stream *find_stream(uint64_t key,
193 struct lttng_ht *ht)
194 {
195 struct lttng_ht_iter iter;
196 struct lttng_ht_node_u64 *node;
197 struct lttng_consumer_stream *stream = NULL;
198
199 assert(ht);
200
201 /* -1ULL keys are lookup failures */
202 if (key == (uint64_t) -1ULL) {
203 return NULL;
204 }
205
206 rcu_read_lock();
207
208 lttng_ht_lookup(ht, &key, &iter);
209 node = lttng_ht_iter_get_node_u64(&iter);
210 if (node != NULL) {
211 stream = caa_container_of(node, struct lttng_consumer_stream, node);
212 }
213
214 rcu_read_unlock();
215
216 return stream;
217 }
218
219 static void steal_stream_key(uint64_t key, struct lttng_ht *ht)
220 {
221 struct lttng_consumer_stream *stream;
222
223 rcu_read_lock();
224 stream = find_stream(key, ht);
225 if (stream) {
226 stream->key = (uint64_t) -1ULL;
227 /*
228 * We don't want the lookup to match, but we still need
229 * to iterate on this stream when iterating over the hash table. Just
230 * change the node key.
231 */
232 stream->node.key = (uint64_t) -1ULL;
233 }
234 rcu_read_unlock();
235 }
236
237 /*
238 * Return a channel object for the given key.
239 *
240 * RCU read side lock MUST be acquired before calling this function and
241 * protects the channel ptr.
242 */
243 struct lttng_consumer_channel *consumer_find_channel(uint64_t key)
244 {
245 struct lttng_ht_iter iter;
246 struct lttng_ht_node_u64 *node;
247 struct lttng_consumer_channel *channel = NULL;
248
249 /* -1ULL keys are lookup failures */
250 if (key == (uint64_t) -1ULL) {
251 return NULL;
252 }
253
254 lttng_ht_lookup(consumer_data.channel_ht, &key, &iter);
255 node = lttng_ht_iter_get_node_u64(&iter);
256 if (node != NULL) {
257 channel = caa_container_of(node, struct lttng_consumer_channel, node);
258 }
259
260 return channel;
261 }
262
263 /*
264 * There is a possibility that the consumer does not have enough time between
265 * the close of the channel on the session daemon and the cleanup in here thus
266 * once we have a channel add with an existing key, we know for sure that this
267 * channel will eventually get cleaned up by all streams being closed.
268 *
269 * This function just nullifies the already existing channel key.
270 */
271 static void steal_channel_key(uint64_t key)
272 {
273 struct lttng_consumer_channel *channel;
274
275 rcu_read_lock();
276 channel = consumer_find_channel(key);
277 if (channel) {
278 channel->key = (uint64_t) -1ULL;
279 /*
280 * We don't want the lookup to match, but we still need to iterate on
281 * this channel when iterating over the hash table. Just change the
282 * node key.
283 */
284 channel->node.key = (uint64_t) -1ULL;
285 }
286 rcu_read_unlock();
287 }
288
289 static void free_channel_rcu(struct rcu_head *head)
290 {
291 struct lttng_ht_node_u64 *node =
292 caa_container_of(head, struct lttng_ht_node_u64, head);
293 struct lttng_consumer_channel *channel =
294 caa_container_of(node, struct lttng_consumer_channel, node);
295
296 switch (consumer_data.type) {
297 case LTTNG_CONSUMER_KERNEL:
298 break;
299 case LTTNG_CONSUMER32_UST:
300 case LTTNG_CONSUMER64_UST:
301 lttng_ustconsumer_free_channel(channel);
302 break;
303 default:
304 ERR("Unknown consumer_data type");
305 abort();
306 }
307 free(channel);
308 }
309
310 /*
311 * RCU protected relayd socket pair free.
312 */
313 static void free_relayd_rcu(struct rcu_head *head)
314 {
315 struct lttng_ht_node_u64 *node =
316 caa_container_of(head, struct lttng_ht_node_u64, head);
317 struct consumer_relayd_sock_pair *relayd =
318 caa_container_of(node, struct consumer_relayd_sock_pair, node);
319
320 /*
321 * Close all sockets. This is done in the call RCU since we don't want the
322 * socket fds to be reassigned thus potentially creating bad state of the
323 * relayd object.
324 *
325 * We do not have to lock the control socket mutex here since at this stage
326 * there is no one referencing to this relayd object.
327 */
328 (void) relayd_close(&relayd->control_sock);
329 (void) relayd_close(&relayd->data_sock);
330
331 pthread_mutex_destroy(&relayd->ctrl_sock_mutex);
332 free(relayd);
333 }
334
335 /*
336 * Destroy and free relayd socket pair object.
337 */
338 void consumer_destroy_relayd(struct consumer_relayd_sock_pair *relayd)
339 {
340 int ret;
341 struct lttng_ht_iter iter;
342
343 if (relayd == NULL) {
344 return;
345 }
346
347 DBG("Consumer destroy and close relayd socket pair");
348
349 iter.iter.node = &relayd->node.node;
350 ret = lttng_ht_del(consumer_data.relayd_ht, &iter);
351 if (ret != 0) {
352 /* We assume the relayd is being or is destroyed */
353 return;
354 }
355
356 /* RCU free() call */
357 call_rcu(&relayd->node.head, free_relayd_rcu);
358 }
359
360 /*
361 * Remove a channel from the global list protected by a mutex. This function is
362 * also responsible for freeing its data structures.
363 */
364 void consumer_del_channel(struct lttng_consumer_channel *channel)
365 {
366 struct lttng_ht_iter iter;
367
368 DBG("Consumer delete channel key %" PRIu64, channel->key);
369
370 pthread_mutex_lock(&consumer_data.lock);
371 pthread_mutex_lock(&channel->lock);
372
373 /* Destroy streams that might have been left in the stream list. */
374 clean_channel_stream_list(channel);
375
376 if (channel->live_timer_enabled == 1) {
377 consumer_timer_live_stop(channel);
378 }
379 if (channel->monitor_timer_enabled == 1) {
380 consumer_timer_monitor_stop(channel);
381 }
382
383 switch (consumer_data.type) {
384 case LTTNG_CONSUMER_KERNEL:
385 break;
386 case LTTNG_CONSUMER32_UST:
387 case LTTNG_CONSUMER64_UST:
388 lttng_ustconsumer_del_channel(channel);
389 break;
390 default:
391 ERR("Unknown consumer_data type");
392 assert(0);
393 goto end;
394 }
395
396 lttng_trace_chunk_put(channel->trace_chunk);
397 channel->trace_chunk = NULL;
398
399 if (channel->is_published) {
400 int ret;
401
402 rcu_read_lock();
403 iter.iter.node = &channel->node.node;
404 ret = lttng_ht_del(consumer_data.channel_ht, &iter);
405 assert(!ret);
406
407 iter.iter.node = &channel->channels_by_session_id_ht_node.node;
408 ret = lttng_ht_del(consumer_data.channels_by_session_id_ht,
409 &iter);
410 assert(!ret);
411 rcu_read_unlock();
412 }
413
414 channel->is_deleted = true;
415 call_rcu(&channel->node.head, free_channel_rcu);
416 end:
417 pthread_mutex_unlock(&channel->lock);
418 pthread_mutex_unlock(&consumer_data.lock);
419 }
420
421 /*
422 * Iterate over the relayd hash table and destroy each element. Finally,
423 * destroy the whole hash table.
424 */
425 static void cleanup_relayd_ht(void)
426 {
427 struct lttng_ht_iter iter;
428 struct consumer_relayd_sock_pair *relayd;
429
430 rcu_read_lock();
431
432 cds_lfht_for_each_entry(consumer_data.relayd_ht->ht, &iter.iter, relayd,
433 node.node) {
434 consumer_destroy_relayd(relayd);
435 }
436
437 rcu_read_unlock();
438
439 lttng_ht_destroy(consumer_data.relayd_ht);
440 }
441
442 /*
443 * Update the end point status of all streams having the given network sequence
444 * index (relayd index).
445 *
446 * It's atomically set without having the stream mutex locked which is fine
447 * because we handle the write/read race with a pipe wakeup for each thread.
448 */
449 static void update_endpoint_status_by_netidx(uint64_t net_seq_idx,
450 enum consumer_endpoint_status status)
451 {
452 struct lttng_ht_iter iter;
453 struct lttng_consumer_stream *stream;
454
455 DBG("Consumer set delete flag on stream by idx %" PRIu64, net_seq_idx);
456
457 rcu_read_lock();
458
459 /* Let's begin with metadata */
460 cds_lfht_for_each_entry(metadata_ht->ht, &iter.iter, stream, node.node) {
461 if (stream->net_seq_idx == net_seq_idx) {
462 uatomic_set(&stream->endpoint_status, status);
463 DBG("Delete flag set to metadata stream %d", stream->wait_fd);
464 }
465 }
466
467 /* Follow up by the data streams */
468 cds_lfht_for_each_entry(data_ht->ht, &iter.iter, stream, node.node) {
469 if (stream->net_seq_idx == net_seq_idx) {
470 uatomic_set(&stream->endpoint_status, status);
471 DBG("Delete flag set to data stream %d", stream->wait_fd);
472 }
473 }
474 rcu_read_unlock();
475 }
476
477 /*
478 * Cleanup a relayd object by flagging every associated streams for deletion,
479 * destroying the object meaning removing it from the relayd hash table,
480 * closing the sockets and freeing the memory in a RCU call.
481 *
482 * If a local data context is available, notify the threads that the streams'
483 * state have changed.
484 */
485 void lttng_consumer_cleanup_relayd(struct consumer_relayd_sock_pair *relayd)
486 {
487 uint64_t netidx;
488
489 assert(relayd);
490
491 DBG("Cleaning up relayd object ID %"PRIu64, relayd->net_seq_idx);
492
493 /* Save the net sequence index before destroying the object */
494 netidx = relayd->net_seq_idx;
495
496 /*
497 * Delete the relayd from the relayd hash table, close the sockets and free
498 * the object in a RCU call.
499 */
500 consumer_destroy_relayd(relayd);
501
502 /* Set inactive endpoint to all streams */
503 update_endpoint_status_by_netidx(netidx, CONSUMER_ENDPOINT_INACTIVE);
504
505 /*
506 * With a local data context, notify the threads that the streams' state
507 * have changed. The write() action on the pipe acts as an "implicit"
508 * memory barrier ordering the updates of the end point status from the
509 * read of this status which happens AFTER receiving this notify.
510 */
511 notify_thread_lttng_pipe(relayd->ctx->consumer_data_pipe);
512 notify_thread_lttng_pipe(relayd->ctx->consumer_metadata_pipe);
513 }
514
515 /*
516 * Flag a relayd socket pair for destruction. Destroy it if the refcount
517 * reaches zero.
518 *
519 * RCU read side lock MUST be aquired before calling this function.
520 */
521 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair *relayd)
522 {
523 assert(relayd);
524
525 /* Set destroy flag for this object */
526 uatomic_set(&relayd->destroy_flag, 1);
527
528 /* Destroy the relayd if refcount is 0 */
529 if (uatomic_read(&relayd->refcount) == 0) {
530 consumer_destroy_relayd(relayd);
531 }
532 }
533
534 /*
535 * Completly destroy stream from every visiable data structure and the given
536 * hash table if one.
537 *
538 * One this call returns, the stream object is not longer usable nor visible.
539 */
540 void consumer_del_stream(struct lttng_consumer_stream *stream,
541 struct lttng_ht *ht)
542 {
543 consumer_stream_destroy(stream, ht);
544 }
545
546 /*
547 * XXX naming of del vs destroy is all mixed up.
548 */
549 void consumer_del_stream_for_data(struct lttng_consumer_stream *stream)
550 {
551 consumer_stream_destroy(stream, data_ht);
552 }
553
554 void consumer_del_stream_for_metadata(struct lttng_consumer_stream *stream)
555 {
556 consumer_stream_destroy(stream, metadata_ht);
557 }
558
559 void consumer_stream_update_channel_attributes(
560 struct lttng_consumer_stream *stream,
561 struct lttng_consumer_channel *channel)
562 {
563 stream->channel_read_only_attributes.tracefile_size =
564 channel->tracefile_size;
565 }
566
567 struct lttng_consumer_stream *consumer_allocate_stream(uint64_t channel_key,
568 uint64_t stream_key,
569 const char *channel_name,
570 uint64_t relayd_id,
571 uint64_t session_id,
572 struct lttng_trace_chunk *trace_chunk,
573 int cpu,
574 int *alloc_ret,
575 enum consumer_channel_type type,
576 unsigned int monitor)
577 {
578 int ret;
579 struct lttng_consumer_stream *stream;
580
581 stream = zmalloc(sizeof(*stream));
582 if (stream == NULL) {
583 PERROR("malloc struct lttng_consumer_stream");
584 ret = -ENOMEM;
585 goto end;
586 }
587
588 if (trace_chunk && !lttng_trace_chunk_get(trace_chunk)) {
589 ERR("Failed to acquire trace chunk reference during the creation of a stream");
590 ret = -1;
591 goto error;
592 }
593
594 rcu_read_lock();
595 stream->key = stream_key;
596 stream->trace_chunk = trace_chunk;
597 stream->out_fd = -1;
598 stream->out_fd_offset = 0;
599 stream->output_written = 0;
600 stream->net_seq_idx = relayd_id;
601 stream->session_id = session_id;
602 stream->monitor = monitor;
603 stream->endpoint_status = CONSUMER_ENDPOINT_ACTIVE;
604 stream->index_file = NULL;
605 stream->last_sequence_number = -1ULL;
606 pthread_mutex_init(&stream->lock, NULL);
607 pthread_mutex_init(&stream->metadata_timer_lock, NULL);
608
609 /* If channel is the metadata, flag this stream as metadata. */
610 if (type == CONSUMER_CHANNEL_TYPE_METADATA) {
611 stream->metadata_flag = 1;
612 /* Metadata is flat out. */
613 strncpy(stream->name, DEFAULT_METADATA_NAME, sizeof(stream->name));
614 /* Live rendez-vous point. */
615 pthread_cond_init(&stream->metadata_rdv, NULL);
616 pthread_mutex_init(&stream->metadata_rdv_lock, NULL);
617 } else {
618 /* Format stream name to <channel_name>_<cpu_number> */
619 ret = snprintf(stream->name, sizeof(stream->name), "%s_%d",
620 channel_name, cpu);
621 if (ret < 0) {
622 PERROR("snprintf stream name");
623 goto error;
624 }
625 }
626
627 /* Key is always the wait_fd for streams. */
628 lttng_ht_node_init_u64(&stream->node, stream->key);
629
630 /* Init node per channel id key */
631 lttng_ht_node_init_u64(&stream->node_channel_id, channel_key);
632
633 /* Init session id node with the stream session id */
634 lttng_ht_node_init_u64(&stream->node_session_id, stream->session_id);
635
636 DBG3("Allocated stream %s (key %" PRIu64 ", chan_key %" PRIu64
637 " relayd_id %" PRIu64 ", session_id %" PRIu64,
638 stream->name, stream->key, channel_key,
639 stream->net_seq_idx, stream->session_id);
640
641 rcu_read_unlock();
642 return stream;
643
644 error:
645 rcu_read_unlock();
646 lttng_trace_chunk_put(stream->trace_chunk);
647 free(stream);
648 end:
649 if (alloc_ret) {
650 *alloc_ret = ret;
651 }
652 return NULL;
653 }
654
655 /*
656 * Add a stream to the global list protected by a mutex.
657 */
658 void consumer_add_data_stream(struct lttng_consumer_stream *stream)
659 {
660 struct lttng_ht *ht = data_ht;
661
662 assert(stream);
663 assert(ht);
664
665 DBG3("Adding consumer stream %" PRIu64, stream->key);
666
667 pthread_mutex_lock(&consumer_data.lock);
668 pthread_mutex_lock(&stream->chan->lock);
669 pthread_mutex_lock(&stream->chan->timer_lock);
670 pthread_mutex_lock(&stream->lock);
671 rcu_read_lock();
672
673 /* Steal stream identifier to avoid having streams with the same key */
674 steal_stream_key(stream->key, ht);
675
676 lttng_ht_add_unique_u64(ht, &stream->node);
677
678 lttng_ht_add_u64(consumer_data.stream_per_chan_id_ht,
679 &stream->node_channel_id);
680
681 /*
682 * Add stream to the stream_list_ht of the consumer data. No need to steal
683 * the key since the HT does not use it and we allow to add redundant keys
684 * into this table.
685 */
686 lttng_ht_add_u64(consumer_data.stream_list_ht, &stream->node_session_id);
687
688 /*
689 * When nb_init_stream_left reaches 0, we don't need to trigger any action
690 * in terms of destroying the associated channel, because the action that
691 * causes the count to become 0 also causes a stream to be added. The
692 * channel deletion will thus be triggered by the following removal of this
693 * stream.
694 */
695 if (uatomic_read(&stream->chan->nb_init_stream_left) > 0) {
696 /* Increment refcount before decrementing nb_init_stream_left */
697 cmm_smp_wmb();
698 uatomic_dec(&stream->chan->nb_init_stream_left);
699 }
700
701 /* Update consumer data once the node is inserted. */
702 consumer_data.stream_count++;
703 consumer_data.need_update = 1;
704
705 rcu_read_unlock();
706 pthread_mutex_unlock(&stream->lock);
707 pthread_mutex_unlock(&stream->chan->timer_lock);
708 pthread_mutex_unlock(&stream->chan->lock);
709 pthread_mutex_unlock(&consumer_data.lock);
710 }
711
712 void consumer_del_data_stream(struct lttng_consumer_stream *stream)
713 {
714 consumer_del_stream(stream, data_ht);
715 }
716
717 /*
718 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
719 * be acquired before calling this.
720 */
721 static int add_relayd(struct consumer_relayd_sock_pair *relayd)
722 {
723 int ret = 0;
724 struct lttng_ht_node_u64 *node;
725 struct lttng_ht_iter iter;
726
727 assert(relayd);
728
729 lttng_ht_lookup(consumer_data.relayd_ht,
730 &relayd->net_seq_idx, &iter);
731 node = lttng_ht_iter_get_node_u64(&iter);
732 if (node != NULL) {
733 goto end;
734 }
735 lttng_ht_add_unique_u64(consumer_data.relayd_ht, &relayd->node);
736
737 end:
738 return ret;
739 }
740
741 /*
742 * Allocate and return a consumer relayd socket.
743 */
744 static struct consumer_relayd_sock_pair *consumer_allocate_relayd_sock_pair(
745 uint64_t net_seq_idx)
746 {
747 struct consumer_relayd_sock_pair *obj = NULL;
748
749 /* net sequence index of -1 is a failure */
750 if (net_seq_idx == (uint64_t) -1ULL) {
751 goto error;
752 }
753
754 obj = zmalloc(sizeof(struct consumer_relayd_sock_pair));
755 if (obj == NULL) {
756 PERROR("zmalloc relayd sock");
757 goto error;
758 }
759
760 obj->net_seq_idx = net_seq_idx;
761 obj->refcount = 0;
762 obj->destroy_flag = 0;
763 obj->control_sock.sock.fd = -1;
764 obj->data_sock.sock.fd = -1;
765 lttng_ht_node_init_u64(&obj->node, obj->net_seq_idx);
766 pthread_mutex_init(&obj->ctrl_sock_mutex, NULL);
767
768 error:
769 return obj;
770 }
771
772 /*
773 * Find a relayd socket pair in the global consumer data.
774 *
775 * Return the object if found else NULL.
776 * RCU read-side lock must be held across this call and while using the
777 * returned object.
778 */
779 struct consumer_relayd_sock_pair *consumer_find_relayd(uint64_t key)
780 {
781 struct lttng_ht_iter iter;
782 struct lttng_ht_node_u64 *node;
783 struct consumer_relayd_sock_pair *relayd = NULL;
784
785 /* Negative keys are lookup failures */
786 if (key == (uint64_t) -1ULL) {
787 goto error;
788 }
789
790 lttng_ht_lookup(consumer_data.relayd_ht, &key,
791 &iter);
792 node = lttng_ht_iter_get_node_u64(&iter);
793 if (node != NULL) {
794 relayd = caa_container_of(node, struct consumer_relayd_sock_pair, node);
795 }
796
797 error:
798 return relayd;
799 }
800
801 /*
802 * Find a relayd and send the stream
803 *
804 * Returns 0 on success, < 0 on error
805 */
806 int consumer_send_relayd_stream(struct lttng_consumer_stream *stream,
807 char *path)
808 {
809 int ret = 0;
810 struct consumer_relayd_sock_pair *relayd;
811
812 assert(stream);
813 assert(stream->net_seq_idx != -1ULL);
814 assert(path);
815
816 /* The stream is not metadata. Get relayd reference if exists. */
817 rcu_read_lock();
818 relayd = consumer_find_relayd(stream->net_seq_idx);
819 if (relayd != NULL) {
820 /* Add stream on the relayd */
821 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
822 ret = relayd_add_stream(&relayd->control_sock, stream->name,
823 path, &stream->relayd_stream_id,
824 stream->chan->tracefile_size,
825 stream->chan->tracefile_count,
826 stream->trace_chunk);
827 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
828 if (ret < 0) {
829 ERR("Relayd add stream failed. Cleaning up relayd %" PRIu64".", relayd->net_seq_idx);
830 lttng_consumer_cleanup_relayd(relayd);
831 goto end;
832 }
833
834 uatomic_inc(&relayd->refcount);
835 stream->sent_to_relayd = 1;
836 } else {
837 ERR("Stream %" PRIu64 " relayd ID %" PRIu64 " unknown. Can't send it.",
838 stream->key, stream->net_seq_idx);
839 ret = -1;
840 goto end;
841 }
842
843 DBG("Stream %s with key %" PRIu64 " sent to relayd id %" PRIu64,
844 stream->name, stream->key, stream->net_seq_idx);
845
846 end:
847 rcu_read_unlock();
848 return ret;
849 }
850
851 /*
852 * Find a relayd and send the streams sent message
853 *
854 * Returns 0 on success, < 0 on error
855 */
856 int consumer_send_relayd_streams_sent(uint64_t net_seq_idx)
857 {
858 int ret = 0;
859 struct consumer_relayd_sock_pair *relayd;
860
861 assert(net_seq_idx != -1ULL);
862
863 /* The stream is not metadata. Get relayd reference if exists. */
864 rcu_read_lock();
865 relayd = consumer_find_relayd(net_seq_idx);
866 if (relayd != NULL) {
867 /* Add stream on the relayd */
868 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
869 ret = relayd_streams_sent(&relayd->control_sock);
870 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
871 if (ret < 0) {
872 ERR("Relayd streams sent failed. Cleaning up relayd %" PRIu64".", relayd->net_seq_idx);
873 lttng_consumer_cleanup_relayd(relayd);
874 goto end;
875 }
876 } else {
877 ERR("Relayd ID %" PRIu64 " unknown. Can't send streams_sent.",
878 net_seq_idx);
879 ret = -1;
880 goto end;
881 }
882
883 ret = 0;
884 DBG("All streams sent relayd id %" PRIu64, net_seq_idx);
885
886 end:
887 rcu_read_unlock();
888 return ret;
889 }
890
891 /*
892 * Find a relayd and close the stream
893 */
894 void close_relayd_stream(struct lttng_consumer_stream *stream)
895 {
896 struct consumer_relayd_sock_pair *relayd;
897
898 /* The stream is not metadata. Get relayd reference if exists. */
899 rcu_read_lock();
900 relayd = consumer_find_relayd(stream->net_seq_idx);
901 if (relayd) {
902 consumer_stream_relayd_close(stream, relayd);
903 }
904 rcu_read_unlock();
905 }
906
907 /*
908 * Handle stream for relayd transmission if the stream applies for network
909 * streaming where the net sequence index is set.
910 *
911 * Return destination file descriptor or negative value on error.
912 */
913 static int write_relayd_stream_header(struct lttng_consumer_stream *stream,
914 size_t data_size, unsigned long padding,
915 struct consumer_relayd_sock_pair *relayd)
916 {
917 int outfd = -1, ret;
918 struct lttcomm_relayd_data_hdr data_hdr;
919
920 /* Safety net */
921 assert(stream);
922 assert(relayd);
923
924 /* Reset data header */
925 memset(&data_hdr, 0, sizeof(data_hdr));
926
927 if (stream->metadata_flag) {
928 /* Caller MUST acquire the relayd control socket lock */
929 ret = relayd_send_metadata(&relayd->control_sock, data_size);
930 if (ret < 0) {
931 goto error;
932 }
933
934 /* Metadata are always sent on the control socket. */
935 outfd = relayd->control_sock.sock.fd;
936 } else {
937 /* Set header with stream information */
938 data_hdr.stream_id = htobe64(stream->relayd_stream_id);
939 data_hdr.data_size = htobe32(data_size);
940 data_hdr.padding_size = htobe32(padding);
941
942 /*
943 * Note that net_seq_num below is assigned with the *current* value of
944 * next_net_seq_num and only after that the next_net_seq_num will be
945 * increment. This is why when issuing a command on the relayd using
946 * this next value, 1 should always be substracted in order to compare
947 * the last seen sequence number on the relayd side to the last sent.
948 */
949 data_hdr.net_seq_num = htobe64(stream->next_net_seq_num);
950 /* Other fields are zeroed previously */
951
952 ret = relayd_send_data_hdr(&relayd->data_sock, &data_hdr,
953 sizeof(data_hdr));
954 if (ret < 0) {
955 goto error;
956 }
957
958 ++stream->next_net_seq_num;
959
960 /* Set to go on data socket */
961 outfd = relayd->data_sock.sock.fd;
962 }
963
964 error:
965 return outfd;
966 }
967
968 /*
969 * Trigger a dump of the metadata content. Following/during the succesful
970 * completion of this call, the metadata poll thread will start receiving
971 * metadata packets to consume.
972 *
973 * The caller must hold the channel and stream locks.
974 */
975 static
976 int consumer_metadata_stream_dump(struct lttng_consumer_stream *stream)
977 {
978 int ret;
979
980 ASSERT_LOCKED(stream->chan->lock);
981 ASSERT_LOCKED(stream->lock);
982 assert(stream->metadata_flag);
983 assert(stream->chan->trace_chunk);
984
985 switch (consumer_data.type) {
986 case LTTNG_CONSUMER_KERNEL:
987 /*
988 * Reset the position of what has been read from the
989 * metadata cache to 0 so we can dump it again.
990 */
991 ret = kernctl_metadata_cache_dump(stream->wait_fd);
992 break;
993 case LTTNG_CONSUMER32_UST:
994 case LTTNG_CONSUMER64_UST:
995 /*
996 * Reset the position pushed from the metadata cache so it
997 * will write from the beginning on the next push.
998 */
999 stream->ust_metadata_pushed = 0;
1000 ret = consumer_metadata_wakeup_pipe(stream->chan);
1001 break;
1002 default:
1003 ERR("Unknown consumer_data type");
1004 abort();
1005 }
1006 if (ret < 0) {
1007 ERR("Failed to dump the metadata cache");
1008 }
1009 return ret;
1010 }
1011
1012 static
1013 int lttng_consumer_channel_set_trace_chunk(
1014 struct lttng_consumer_channel *channel,
1015 struct lttng_trace_chunk *new_trace_chunk)
1016 {
1017 pthread_mutex_lock(&channel->lock);
1018 if (channel->is_deleted) {
1019 /*
1020 * The channel has been logically deleted and should no longer
1021 * be used. It has released its reference to its current trace
1022 * chunk and should not acquire a new one.
1023 *
1024 * Return success as there is nothing for the caller to do.
1025 */
1026 goto end;
1027 }
1028
1029 /*
1030 * The acquisition of the reference cannot fail (barring
1031 * a severe internal error) since a reference to the published
1032 * chunk is already held by the caller.
1033 */
1034 if (new_trace_chunk) {
1035 const bool acquired_reference = lttng_trace_chunk_get(
1036 new_trace_chunk);
1037
1038 assert(acquired_reference);
1039 }
1040
1041 lttng_trace_chunk_put(channel->trace_chunk);
1042 channel->trace_chunk = new_trace_chunk;
1043 end:
1044 pthread_mutex_unlock(&channel->lock);
1045 return 0;
1046 }
1047
1048 /*
1049 * Allocate and return a new lttng_consumer_channel object using the given key
1050 * to initialize the hash table node.
1051 *
1052 * On error, return NULL.
1053 */
1054 struct lttng_consumer_channel *consumer_allocate_channel(uint64_t key,
1055 uint64_t session_id,
1056 const uint64_t *chunk_id,
1057 const char *pathname,
1058 const char *name,
1059 uint64_t relayd_id,
1060 enum lttng_event_output output,
1061 uint64_t tracefile_size,
1062 uint64_t tracefile_count,
1063 uint64_t session_id_per_pid,
1064 unsigned int monitor,
1065 unsigned int live_timer_interval,
1066 const char *root_shm_path,
1067 const char *shm_path)
1068 {
1069 struct lttng_consumer_channel *channel = NULL;
1070 struct lttng_trace_chunk *trace_chunk = NULL;
1071
1072 if (chunk_id) {
1073 trace_chunk = lttng_trace_chunk_registry_find_chunk(
1074 consumer_data.chunk_registry, session_id,
1075 *chunk_id);
1076 if (!trace_chunk) {
1077 ERR("Failed to find trace chunk reference during creation of channel");
1078 goto end;
1079 }
1080 }
1081
1082 channel = zmalloc(sizeof(*channel));
1083 if (channel == NULL) {
1084 PERROR("malloc struct lttng_consumer_channel");
1085 goto end;
1086 }
1087
1088 channel->key = key;
1089 channel->refcount = 0;
1090 channel->session_id = session_id;
1091 channel->session_id_per_pid = session_id_per_pid;
1092 channel->relayd_id = relayd_id;
1093 channel->tracefile_size = tracefile_size;
1094 channel->tracefile_count = tracefile_count;
1095 channel->monitor = monitor;
1096 channel->live_timer_interval = live_timer_interval;
1097 pthread_mutex_init(&channel->lock, NULL);
1098 pthread_mutex_init(&channel->timer_lock, NULL);
1099
1100 switch (output) {
1101 case LTTNG_EVENT_SPLICE:
1102 channel->output = CONSUMER_CHANNEL_SPLICE;
1103 break;
1104 case LTTNG_EVENT_MMAP:
1105 channel->output = CONSUMER_CHANNEL_MMAP;
1106 break;
1107 default:
1108 assert(0);
1109 free(channel);
1110 channel = NULL;
1111 goto end;
1112 }
1113
1114 /*
1115 * In monitor mode, the streams associated with the channel will be put in
1116 * a special list ONLY owned by this channel. So, the refcount is set to 1
1117 * here meaning that the channel itself has streams that are referenced.
1118 *
1119 * On a channel deletion, once the channel is no longer visible, the
1120 * refcount is decremented and checked for a zero value to delete it. With
1121 * streams in no monitor mode, it will now be safe to destroy the channel.
1122 */
1123 if (!channel->monitor) {
1124 channel->refcount = 1;
1125 }
1126
1127 strncpy(channel->pathname, pathname, sizeof(channel->pathname));
1128 channel->pathname[sizeof(channel->pathname) - 1] = '\0';
1129
1130 strncpy(channel->name, name, sizeof(channel->name));
1131 channel->name[sizeof(channel->name) - 1] = '\0';
1132
1133 if (root_shm_path) {
1134 strncpy(channel->root_shm_path, root_shm_path, sizeof(channel->root_shm_path));
1135 channel->root_shm_path[sizeof(channel->root_shm_path) - 1] = '\0';
1136 }
1137 if (shm_path) {
1138 strncpy(channel->shm_path, shm_path, sizeof(channel->shm_path));
1139 channel->shm_path[sizeof(channel->shm_path) - 1] = '\0';
1140 }
1141
1142 lttng_ht_node_init_u64(&channel->node, channel->key);
1143 lttng_ht_node_init_u64(&channel->channels_by_session_id_ht_node,
1144 channel->session_id);
1145
1146 channel->wait_fd = -1;
1147 CDS_INIT_LIST_HEAD(&channel->streams.head);
1148
1149 if (trace_chunk) {
1150 int ret = lttng_consumer_channel_set_trace_chunk(channel,
1151 trace_chunk);
1152 if (ret) {
1153 goto error;
1154 }
1155 }
1156
1157 DBG("Allocated channel (key %" PRIu64 ")", channel->key);
1158
1159 end:
1160 lttng_trace_chunk_put(trace_chunk);
1161 return channel;
1162 error:
1163 consumer_del_channel(channel);
1164 channel = NULL;
1165 goto end;
1166 }
1167
1168 /*
1169 * Add a channel to the global list protected by a mutex.
1170 *
1171 * Always return 0 indicating success.
1172 */
1173 int consumer_add_channel(struct lttng_consumer_channel *channel,
1174 struct lttng_consumer_local_data *ctx)
1175 {
1176 pthread_mutex_lock(&consumer_data.lock);
1177 pthread_mutex_lock(&channel->lock);
1178 pthread_mutex_lock(&channel->timer_lock);
1179
1180 /*
1181 * This gives us a guarantee that the channel we are about to add to the
1182 * channel hash table will be unique. See this function comment on the why
1183 * we need to steel the channel key at this stage.
1184 */
1185 steal_channel_key(channel->key);
1186
1187 rcu_read_lock();
1188 lttng_ht_add_unique_u64(consumer_data.channel_ht, &channel->node);
1189 lttng_ht_add_u64(consumer_data.channels_by_session_id_ht,
1190 &channel->channels_by_session_id_ht_node);
1191 rcu_read_unlock();
1192 channel->is_published = true;
1193
1194 pthread_mutex_unlock(&channel->timer_lock);
1195 pthread_mutex_unlock(&channel->lock);
1196 pthread_mutex_unlock(&consumer_data.lock);
1197
1198 if (channel->wait_fd != -1 && channel->type == CONSUMER_CHANNEL_TYPE_DATA) {
1199 notify_channel_pipe(ctx, channel, -1, CONSUMER_CHANNEL_ADD);
1200 }
1201
1202 return 0;
1203 }
1204
1205 /*
1206 * Allocate the pollfd structure and the local view of the out fds to avoid
1207 * doing a lookup in the linked list and concurrency issues when writing is
1208 * needed. Called with consumer_data.lock held.
1209 *
1210 * Returns the number of fds in the structures.
1211 */
1212 static int update_poll_array(struct lttng_consumer_local_data *ctx,
1213 struct pollfd **pollfd, struct lttng_consumer_stream **local_stream,
1214 struct lttng_ht *ht, int *nb_inactive_fd)
1215 {
1216 int i = 0;
1217 struct lttng_ht_iter iter;
1218 struct lttng_consumer_stream *stream;
1219
1220 assert(ctx);
1221 assert(ht);
1222 assert(pollfd);
1223 assert(local_stream);
1224
1225 DBG("Updating poll fd array");
1226 *nb_inactive_fd = 0;
1227 rcu_read_lock();
1228 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
1229 /*
1230 * Only active streams with an active end point can be added to the
1231 * poll set and local stream storage of the thread.
1232 *
1233 * There is a potential race here for endpoint_status to be updated
1234 * just after the check. However, this is OK since the stream(s) will
1235 * be deleted once the thread is notified that the end point state has
1236 * changed where this function will be called back again.
1237 *
1238 * We track the number of inactive FDs because they still need to be
1239 * closed by the polling thread after a wakeup on the data_pipe or
1240 * metadata_pipe.
1241 */
1242 if (stream->endpoint_status == CONSUMER_ENDPOINT_INACTIVE) {
1243 (*nb_inactive_fd)++;
1244 continue;
1245 }
1246 /*
1247 * This clobbers way too much the debug output. Uncomment that if you
1248 * need it for debugging purposes.
1249 */
1250 (*pollfd)[i].fd = stream->wait_fd;
1251 (*pollfd)[i].events = POLLIN | POLLPRI;
1252 local_stream[i] = stream;
1253 i++;
1254 }
1255 rcu_read_unlock();
1256
1257 /*
1258 * Insert the consumer_data_pipe at the end of the array and don't
1259 * increment i so nb_fd is the number of real FD.
1260 */
1261 (*pollfd)[i].fd = lttng_pipe_get_readfd(ctx->consumer_data_pipe);
1262 (*pollfd)[i].events = POLLIN | POLLPRI;
1263
1264 (*pollfd)[i + 1].fd = lttng_pipe_get_readfd(ctx->consumer_wakeup_pipe);
1265 (*pollfd)[i + 1].events = POLLIN | POLLPRI;
1266 return i;
1267 }
1268
1269 /*
1270 * Poll on the should_quit pipe and the command socket return -1 on
1271 * error, 1 if should exit, 0 if data is available on the command socket
1272 */
1273 int lttng_consumer_poll_socket(struct pollfd *consumer_sockpoll)
1274 {
1275 int num_rdy;
1276
1277 restart:
1278 num_rdy = poll(consumer_sockpoll, 2, -1);
1279 if (num_rdy == -1) {
1280 /*
1281 * Restart interrupted system call.
1282 */
1283 if (errno == EINTR) {
1284 goto restart;
1285 }
1286 PERROR("Poll error");
1287 return -1;
1288 }
1289 if (consumer_sockpoll[0].revents & (POLLIN | POLLPRI)) {
1290 DBG("consumer_should_quit wake up");
1291 return 1;
1292 }
1293 return 0;
1294 }
1295
1296 /*
1297 * Set the error socket.
1298 */
1299 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data *ctx,
1300 int sock)
1301 {
1302 ctx->consumer_error_socket = sock;
1303 }
1304
1305 /*
1306 * Set the command socket path.
1307 */
1308 void lttng_consumer_set_command_sock_path(
1309 struct lttng_consumer_local_data *ctx, char *sock)
1310 {
1311 ctx->consumer_command_sock_path = sock;
1312 }
1313
1314 /*
1315 * Send return code to the session daemon.
1316 * If the socket is not defined, we return 0, it is not a fatal error
1317 */
1318 int lttng_consumer_send_error(struct lttng_consumer_local_data *ctx, int cmd)
1319 {
1320 if (ctx->consumer_error_socket > 0) {
1321 return lttcomm_send_unix_sock(ctx->consumer_error_socket, &cmd,
1322 sizeof(enum lttcomm_sessiond_command));
1323 }
1324
1325 return 0;
1326 }
1327
1328 /*
1329 * Close all the tracefiles and stream fds and MUST be called when all
1330 * instances are destroyed i.e. when all threads were joined and are ended.
1331 */
1332 void lttng_consumer_cleanup(void)
1333 {
1334 struct lttng_ht_iter iter;
1335 struct lttng_consumer_channel *channel;
1336 unsigned int trace_chunks_left;
1337
1338 rcu_read_lock();
1339
1340 cds_lfht_for_each_entry(consumer_data.channel_ht->ht, &iter.iter, channel,
1341 node.node) {
1342 consumer_del_channel(channel);
1343 }
1344
1345 rcu_read_unlock();
1346
1347 lttng_ht_destroy(consumer_data.channel_ht);
1348 lttng_ht_destroy(consumer_data.channels_by_session_id_ht);
1349
1350 cleanup_relayd_ht();
1351
1352 lttng_ht_destroy(consumer_data.stream_per_chan_id_ht);
1353
1354 /*
1355 * This HT contains streams that are freed by either the metadata thread or
1356 * the data thread so we do *nothing* on the hash table and simply destroy
1357 * it.
1358 */
1359 lttng_ht_destroy(consumer_data.stream_list_ht);
1360
1361 /*
1362 * Trace chunks in the registry may still exist if the session
1363 * daemon has encountered an internal error and could not
1364 * tear down its sessions and/or trace chunks properly.
1365 *
1366 * Release the session daemon's implicit reference to any remaining
1367 * trace chunk and print an error if any trace chunk was found. Note
1368 * that there are _no_ legitimate cases for trace chunks to be left,
1369 * it is a leak. However, it can happen following a crash of the
1370 * session daemon and not emptying the registry would cause an assertion
1371 * to hit.
1372 */
1373 trace_chunks_left = lttng_trace_chunk_registry_put_each_chunk(
1374 consumer_data.chunk_registry);
1375 if (trace_chunks_left) {
1376 ERR("%u trace chunks are leaked by lttng-consumerd. "
1377 "This can be caused by an internal error of the session daemon.",
1378 trace_chunks_left);
1379 }
1380 /* Run all callbacks freeing each chunk. */
1381 rcu_barrier();
1382 lttng_trace_chunk_registry_destroy(consumer_data.chunk_registry);
1383 }
1384
1385 /*
1386 * Called from signal handler.
1387 */
1388 void lttng_consumer_should_exit(struct lttng_consumer_local_data *ctx)
1389 {
1390 ssize_t ret;
1391
1392 CMM_STORE_SHARED(consumer_quit, 1);
1393 ret = lttng_write(ctx->consumer_should_quit[1], "4", 1);
1394 if (ret < 1) {
1395 PERROR("write consumer quit");
1396 }
1397
1398 DBG("Consumer flag that it should quit");
1399 }
1400
1401
1402 /*
1403 * Flush pending writes to trace output disk file.
1404 */
1405 static
1406 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream *stream,
1407 off_t orig_offset)
1408 {
1409 int ret;
1410 int outfd = stream->out_fd;
1411
1412 /*
1413 * This does a blocking write-and-wait on any page that belongs to the
1414 * subbuffer prior to the one we just wrote.
1415 * Don't care about error values, as these are just hints and ways to
1416 * limit the amount of page cache used.
1417 */
1418 if (orig_offset < stream->max_sb_size) {
1419 return;
1420 }
1421 lttng_sync_file_range(outfd, orig_offset - stream->max_sb_size,
1422 stream->max_sb_size,
1423 SYNC_FILE_RANGE_WAIT_BEFORE
1424 | SYNC_FILE_RANGE_WRITE
1425 | SYNC_FILE_RANGE_WAIT_AFTER);
1426 /*
1427 * Give hints to the kernel about how we access the file:
1428 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1429 * we write it.
1430 *
1431 * We need to call fadvise again after the file grows because the
1432 * kernel does not seem to apply fadvise to non-existing parts of the
1433 * file.
1434 *
1435 * Call fadvise _after_ having waited for the page writeback to
1436 * complete because the dirty page writeback semantic is not well
1437 * defined. So it can be expected to lead to lower throughput in
1438 * streaming.
1439 */
1440 ret = posix_fadvise(outfd, orig_offset - stream->max_sb_size,
1441 stream->max_sb_size, POSIX_FADV_DONTNEED);
1442 if (ret && ret != -ENOSYS) {
1443 errno = ret;
1444 PERROR("posix_fadvise on fd %i", outfd);
1445 }
1446 }
1447
1448 /*
1449 * Initialise the necessary environnement :
1450 * - create a new context
1451 * - create the poll_pipe
1452 * - create the should_quit pipe (for signal handler)
1453 * - create the thread pipe (for splice)
1454 *
1455 * Takes a function pointer as argument, this function is called when data is
1456 * available on a buffer. This function is responsible to do the
1457 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1458 * buffer configuration and then kernctl_put_next_subbuf at the end.
1459 *
1460 * Returns a pointer to the new context or NULL on error.
1461 */
1462 struct lttng_consumer_local_data *lttng_consumer_create(
1463 enum lttng_consumer_type type,
1464 ssize_t (*buffer_ready)(struct lttng_consumer_stream *stream,
1465 struct lttng_consumer_local_data *ctx),
1466 int (*recv_channel)(struct lttng_consumer_channel *channel),
1467 int (*recv_stream)(struct lttng_consumer_stream *stream),
1468 int (*update_stream)(uint64_t stream_key, uint32_t state))
1469 {
1470 int ret;
1471 struct lttng_consumer_local_data *ctx;
1472
1473 assert(consumer_data.type == LTTNG_CONSUMER_UNKNOWN ||
1474 consumer_data.type == type);
1475 consumer_data.type = type;
1476
1477 ctx = zmalloc(sizeof(struct lttng_consumer_local_data));
1478 if (ctx == NULL) {
1479 PERROR("allocating context");
1480 goto error;
1481 }
1482
1483 ctx->consumer_error_socket = -1;
1484 ctx->consumer_metadata_socket = -1;
1485 pthread_mutex_init(&ctx->metadata_socket_lock, NULL);
1486 /* assign the callbacks */
1487 ctx->on_buffer_ready = buffer_ready;
1488 ctx->on_recv_channel = recv_channel;
1489 ctx->on_recv_stream = recv_stream;
1490 ctx->on_update_stream = update_stream;
1491
1492 ctx->consumer_data_pipe = lttng_pipe_open(0);
1493 if (!ctx->consumer_data_pipe) {
1494 goto error_poll_pipe;
1495 }
1496
1497 ctx->consumer_wakeup_pipe = lttng_pipe_open(0);
1498 if (!ctx->consumer_wakeup_pipe) {
1499 goto error_wakeup_pipe;
1500 }
1501
1502 ret = pipe(ctx->consumer_should_quit);
1503 if (ret < 0) {
1504 PERROR("Error creating recv pipe");
1505 goto error_quit_pipe;
1506 }
1507
1508 ret = pipe(ctx->consumer_channel_pipe);
1509 if (ret < 0) {
1510 PERROR("Error creating channel pipe");
1511 goto error_channel_pipe;
1512 }
1513
1514 ctx->consumer_metadata_pipe = lttng_pipe_open(0);
1515 if (!ctx->consumer_metadata_pipe) {
1516 goto error_metadata_pipe;
1517 }
1518
1519 ctx->channel_monitor_pipe = -1;
1520
1521 return ctx;
1522
1523 error_metadata_pipe:
1524 utils_close_pipe(ctx->consumer_channel_pipe);
1525 error_channel_pipe:
1526 utils_close_pipe(ctx->consumer_should_quit);
1527 error_quit_pipe:
1528 lttng_pipe_destroy(ctx->consumer_wakeup_pipe);
1529 error_wakeup_pipe:
1530 lttng_pipe_destroy(ctx->consumer_data_pipe);
1531 error_poll_pipe:
1532 free(ctx);
1533 error:
1534 return NULL;
1535 }
1536
1537 /*
1538 * Iterate over all streams of the hashtable and free them properly.
1539 */
1540 static void destroy_data_stream_ht(struct lttng_ht *ht)
1541 {
1542 struct lttng_ht_iter iter;
1543 struct lttng_consumer_stream *stream;
1544
1545 if (ht == NULL) {
1546 return;
1547 }
1548
1549 rcu_read_lock();
1550 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
1551 /*
1552 * Ignore return value since we are currently cleaning up so any error
1553 * can't be handled.
1554 */
1555 (void) consumer_del_stream(stream, ht);
1556 }
1557 rcu_read_unlock();
1558
1559 lttng_ht_destroy(ht);
1560 }
1561
1562 /*
1563 * Iterate over all streams of the metadata hashtable and free them
1564 * properly.
1565 */
1566 static void destroy_metadata_stream_ht(struct lttng_ht *ht)
1567 {
1568 struct lttng_ht_iter iter;
1569 struct lttng_consumer_stream *stream;
1570
1571 if (ht == NULL) {
1572 return;
1573 }
1574
1575 rcu_read_lock();
1576 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
1577 /*
1578 * Ignore return value since we are currently cleaning up so any error
1579 * can't be handled.
1580 */
1581 (void) consumer_del_metadata_stream(stream, ht);
1582 }
1583 rcu_read_unlock();
1584
1585 lttng_ht_destroy(ht);
1586 }
1587
1588 /*
1589 * Close all fds associated with the instance and free the context.
1590 */
1591 void lttng_consumer_destroy(struct lttng_consumer_local_data *ctx)
1592 {
1593 int ret;
1594
1595 DBG("Consumer destroying it. Closing everything.");
1596
1597 if (!ctx) {
1598 return;
1599 }
1600
1601 destroy_data_stream_ht(data_ht);
1602 destroy_metadata_stream_ht(metadata_ht);
1603
1604 ret = close(ctx->consumer_error_socket);
1605 if (ret) {
1606 PERROR("close");
1607 }
1608 ret = close(ctx->consumer_metadata_socket);
1609 if (ret) {
1610 PERROR("close");
1611 }
1612 utils_close_pipe(ctx->consumer_channel_pipe);
1613 lttng_pipe_destroy(ctx->consumer_data_pipe);
1614 lttng_pipe_destroy(ctx->consumer_metadata_pipe);
1615 lttng_pipe_destroy(ctx->consumer_wakeup_pipe);
1616 utils_close_pipe(ctx->consumer_should_quit);
1617
1618 unlink(ctx->consumer_command_sock_path);
1619 free(ctx);
1620 }
1621
1622 /*
1623 * Write the metadata stream id on the specified file descriptor.
1624 */
1625 static int write_relayd_metadata_id(int fd,
1626 struct lttng_consumer_stream *stream,
1627 unsigned long padding)
1628 {
1629 ssize_t ret;
1630 struct lttcomm_relayd_metadata_payload hdr;
1631
1632 hdr.stream_id = htobe64(stream->relayd_stream_id);
1633 hdr.padding_size = htobe32(padding);
1634 ret = lttng_write(fd, (void *) &hdr, sizeof(hdr));
1635 if (ret < sizeof(hdr)) {
1636 /*
1637 * This error means that the fd's end is closed so ignore the PERROR
1638 * not to clubber the error output since this can happen in a normal
1639 * code path.
1640 */
1641 if (errno != EPIPE) {
1642 PERROR("write metadata stream id");
1643 }
1644 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno);
1645 /*
1646 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1647 * handle writting the missing part so report that as an error and
1648 * don't lie to the caller.
1649 */
1650 ret = -1;
1651 goto end;
1652 }
1653 DBG("Metadata stream id %" PRIu64 " with padding %lu written before data",
1654 stream->relayd_stream_id, padding);
1655
1656 end:
1657 return (int) ret;
1658 }
1659
1660 /*
1661 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1662 * core function for writing trace buffers to either the local filesystem or
1663 * the network.
1664 *
1665 * It must be called with the stream and the channel lock held.
1666 *
1667 * Careful review MUST be put if any changes occur!
1668 *
1669 * Returns the number of bytes written
1670 */
1671 ssize_t lttng_consumer_on_read_subbuffer_mmap(
1672 struct lttng_consumer_local_data *ctx,
1673 struct lttng_consumer_stream *stream, unsigned long len,
1674 unsigned long padding,
1675 struct ctf_packet_index *index)
1676 {
1677 unsigned long mmap_offset;
1678 void *mmap_base;
1679 ssize_t ret = 0;
1680 off_t orig_offset = stream->out_fd_offset;
1681 /* Default is on the disk */
1682 int outfd = stream->out_fd;
1683 struct consumer_relayd_sock_pair *relayd = NULL;
1684 unsigned int relayd_hang_up = 0;
1685
1686 /* RCU lock for the relayd pointer */
1687 rcu_read_lock();
1688 assert(stream->net_seq_idx != (uint64_t) -1ULL ||
1689 stream->trace_chunk);
1690
1691 /* Flag that the current stream if set for network streaming. */
1692 if (stream->net_seq_idx != (uint64_t) -1ULL) {
1693 relayd = consumer_find_relayd(stream->net_seq_idx);
1694 if (relayd == NULL) {
1695 ret = -EPIPE;
1696 goto end;
1697 }
1698 }
1699
1700 /* get the offset inside the fd to mmap */
1701 switch (consumer_data.type) {
1702 case LTTNG_CONSUMER_KERNEL:
1703 mmap_base = stream->mmap_base;
1704 ret = kernctl_get_mmap_read_offset(stream->wait_fd, &mmap_offset);
1705 if (ret < 0) {
1706 PERROR("tracer ctl get_mmap_read_offset");
1707 goto end;
1708 }
1709 break;
1710 case LTTNG_CONSUMER32_UST:
1711 case LTTNG_CONSUMER64_UST:
1712 mmap_base = lttng_ustctl_get_mmap_base(stream);
1713 if (!mmap_base) {
1714 ERR("read mmap get mmap base for stream %s", stream->name);
1715 ret = -EPERM;
1716 goto end;
1717 }
1718 ret = lttng_ustctl_get_mmap_read_offset(stream, &mmap_offset);
1719 if (ret != 0) {
1720 PERROR("tracer ctl get_mmap_read_offset");
1721 ret = -EINVAL;
1722 goto end;
1723 }
1724 break;
1725 default:
1726 ERR("Unknown consumer_data type");
1727 assert(0);
1728 }
1729
1730 /* Handle stream on the relayd if the output is on the network */
1731 if (relayd) {
1732 unsigned long netlen = len;
1733
1734 /*
1735 * Lock the control socket for the complete duration of the function
1736 * since from this point on we will use the socket.
1737 */
1738 if (stream->metadata_flag) {
1739 /* Metadata requires the control socket. */
1740 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
1741 if (stream->reset_metadata_flag) {
1742 ret = relayd_reset_metadata(&relayd->control_sock,
1743 stream->relayd_stream_id,
1744 stream->metadata_version);
1745 if (ret < 0) {
1746 relayd_hang_up = 1;
1747 goto write_error;
1748 }
1749 stream->reset_metadata_flag = 0;
1750 }
1751 netlen += sizeof(struct lttcomm_relayd_metadata_payload);
1752 }
1753
1754 ret = write_relayd_stream_header(stream, netlen, padding, relayd);
1755 if (ret < 0) {
1756 relayd_hang_up = 1;
1757 goto write_error;
1758 }
1759 /* Use the returned socket. */
1760 outfd = ret;
1761
1762 /* Write metadata stream id before payload */
1763 if (stream->metadata_flag) {
1764 ret = write_relayd_metadata_id(outfd, stream, padding);
1765 if (ret < 0) {
1766 relayd_hang_up = 1;
1767 goto write_error;
1768 }
1769 }
1770 } else {
1771 /* No streaming, we have to set the len with the full padding */
1772 len += padding;
1773
1774 if (stream->metadata_flag && stream->reset_metadata_flag) {
1775 ret = utils_truncate_stream_file(stream->out_fd, 0);
1776 if (ret < 0) {
1777 ERR("Reset metadata file");
1778 goto end;
1779 }
1780 stream->reset_metadata_flag = 0;
1781 }
1782
1783 /*
1784 * Check if we need to change the tracefile before writing the packet.
1785 */
1786 if (stream->chan->tracefile_size > 0 &&
1787 (stream->tracefile_size_current + len) >
1788 stream->chan->tracefile_size) {
1789 ret = consumer_stream_rotate_output_files(stream);
1790 if (ret) {
1791 goto end;
1792 }
1793 outfd = stream->out_fd;
1794 orig_offset = 0;
1795 }
1796 stream->tracefile_size_current += len;
1797 if (index) {
1798 index->offset = htobe64(stream->out_fd_offset);
1799 }
1800 }
1801
1802 /*
1803 * This call guarantee that len or less is returned. It's impossible to
1804 * receive a ret value that is bigger than len.
1805 */
1806 ret = lttng_write(outfd, mmap_base + mmap_offset, len);
1807 DBG("Consumer mmap write() ret %zd (len %lu)", ret, len);
1808 if (ret < 0 || ((size_t) ret != len)) {
1809 /*
1810 * Report error to caller if nothing was written else at least send the
1811 * amount written.
1812 */
1813 if (ret < 0) {
1814 ret = -errno;
1815 }
1816 relayd_hang_up = 1;
1817
1818 /* Socket operation failed. We consider the relayd dead */
1819 if (errno == EPIPE) {
1820 /*
1821 * This is possible if the fd is closed on the other side
1822 * (outfd) or any write problem. It can be verbose a bit for a
1823 * normal execution if for instance the relayd is stopped
1824 * abruptly. This can happen so set this to a DBG statement.
1825 */
1826 DBG("Consumer mmap write detected relayd hang up");
1827 } else {
1828 /* Unhandled error, print it and stop function right now. */
1829 PERROR("Error in write mmap (ret %zd != len %lu)", ret, len);
1830 }
1831 goto write_error;
1832 }
1833 stream->output_written += ret;
1834
1835 /* This call is useless on a socket so better save a syscall. */
1836 if (!relayd) {
1837 /* This won't block, but will start writeout asynchronously */
1838 lttng_sync_file_range(outfd, stream->out_fd_offset, len,
1839 SYNC_FILE_RANGE_WRITE);
1840 stream->out_fd_offset += len;
1841 lttng_consumer_sync_trace_file(stream, orig_offset);
1842 }
1843
1844 write_error:
1845 /*
1846 * This is a special case that the relayd has closed its socket. Let's
1847 * cleanup the relayd object and all associated streams.
1848 */
1849 if (relayd && relayd_hang_up) {
1850 ERR("Relayd hangup. Cleaning up relayd %" PRIu64".", relayd->net_seq_idx);
1851 lttng_consumer_cleanup_relayd(relayd);
1852 }
1853
1854 end:
1855 /* Unlock only if ctrl socket used */
1856 if (relayd && stream->metadata_flag) {
1857 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
1858 }
1859
1860 rcu_read_unlock();
1861 return ret;
1862 }
1863
1864 /*
1865 * Splice the data from the ring buffer to the tracefile.
1866 *
1867 * It must be called with the stream lock held.
1868 *
1869 * Returns the number of bytes spliced.
1870 */
1871 ssize_t lttng_consumer_on_read_subbuffer_splice(
1872 struct lttng_consumer_local_data *ctx,
1873 struct lttng_consumer_stream *stream, unsigned long len,
1874 unsigned long padding,
1875 struct ctf_packet_index *index)
1876 {
1877 ssize_t ret = 0, written = 0, ret_splice = 0;
1878 loff_t offset = 0;
1879 off_t orig_offset = stream->out_fd_offset;
1880 int fd = stream->wait_fd;
1881 /* Default is on the disk */
1882 int outfd = stream->out_fd;
1883 struct consumer_relayd_sock_pair *relayd = NULL;
1884 int *splice_pipe;
1885 unsigned int relayd_hang_up = 0;
1886
1887 switch (consumer_data.type) {
1888 case LTTNG_CONSUMER_KERNEL:
1889 break;
1890 case LTTNG_CONSUMER32_UST:
1891 case LTTNG_CONSUMER64_UST:
1892 /* Not supported for user space tracing */
1893 return -ENOSYS;
1894 default:
1895 ERR("Unknown consumer_data type");
1896 assert(0);
1897 }
1898
1899 /* RCU lock for the relayd pointer */
1900 rcu_read_lock();
1901
1902 /* Flag that the current stream if set for network streaming. */
1903 if (stream->net_seq_idx != (uint64_t) -1ULL) {
1904 relayd = consumer_find_relayd(stream->net_seq_idx);
1905 if (relayd == NULL) {
1906 written = -ret;
1907 goto end;
1908 }
1909 }
1910 splice_pipe = stream->splice_pipe;
1911
1912 /* Write metadata stream id before payload */
1913 if (relayd) {
1914 unsigned long total_len = len;
1915
1916 if (stream->metadata_flag) {
1917 /*
1918 * Lock the control socket for the complete duration of the function
1919 * since from this point on we will use the socket.
1920 */
1921 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
1922
1923 if (stream->reset_metadata_flag) {
1924 ret = relayd_reset_metadata(&relayd->control_sock,
1925 stream->relayd_stream_id,
1926 stream->metadata_version);
1927 if (ret < 0) {
1928 relayd_hang_up = 1;
1929 goto write_error;
1930 }
1931 stream->reset_metadata_flag = 0;
1932 }
1933 ret = write_relayd_metadata_id(splice_pipe[1], stream,
1934 padding);
1935 if (ret < 0) {
1936 written = ret;
1937 relayd_hang_up = 1;
1938 goto write_error;
1939 }
1940
1941 total_len += sizeof(struct lttcomm_relayd_metadata_payload);
1942 }
1943
1944 ret = write_relayd_stream_header(stream, total_len, padding, relayd);
1945 if (ret < 0) {
1946 written = ret;
1947 relayd_hang_up = 1;
1948 goto write_error;
1949 }
1950 /* Use the returned socket. */
1951 outfd = ret;
1952 } else {
1953 /* No streaming, we have to set the len with the full padding */
1954 len += padding;
1955
1956 if (stream->metadata_flag && stream->reset_metadata_flag) {
1957 ret = utils_truncate_stream_file(stream->out_fd, 0);
1958 if (ret < 0) {
1959 ERR("Reset metadata file");
1960 goto end;
1961 }
1962 stream->reset_metadata_flag = 0;
1963 }
1964 /*
1965 * Check if we need to change the tracefile before writing the packet.
1966 */
1967 if (stream->chan->tracefile_size > 0 &&
1968 (stream->tracefile_size_current + len) >
1969 stream->chan->tracefile_size) {
1970 ret = consumer_stream_rotate_output_files(stream);
1971 if (ret < 0) {
1972 written = ret;
1973 goto end;
1974 }
1975 outfd = stream->out_fd;
1976 orig_offset = 0;
1977 }
1978 stream->tracefile_size_current += len;
1979 index->offset = htobe64(stream->out_fd_offset);
1980 }
1981
1982 while (len > 0) {
1983 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1984 (unsigned long)offset, len, fd, splice_pipe[1]);
1985 ret_splice = splice(fd, &offset, splice_pipe[1], NULL, len,
1986 SPLICE_F_MOVE | SPLICE_F_MORE);
1987 DBG("splice chan to pipe, ret %zd", ret_splice);
1988 if (ret_splice < 0) {
1989 ret = errno;
1990 written = -ret;
1991 PERROR("Error in relay splice");
1992 goto splice_error;
1993 }
1994
1995 /* Handle stream on the relayd if the output is on the network */
1996 if (relayd && stream->metadata_flag) {
1997 size_t metadata_payload_size =
1998 sizeof(struct lttcomm_relayd_metadata_payload);
1999
2000 /* Update counter to fit the spliced data */
2001 ret_splice += metadata_payload_size;
2002 len += metadata_payload_size;
2003 /*
2004 * We do this so the return value can match the len passed as
2005 * argument to this function.
2006 */
2007 written -= metadata_payload_size;
2008 }
2009
2010 /* Splice data out */
2011 ret_splice = splice(splice_pipe[0], NULL, outfd, NULL,
2012 ret_splice, SPLICE_F_MOVE | SPLICE_F_MORE);
2013 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
2014 outfd, ret_splice);
2015 if (ret_splice < 0) {
2016 ret = errno;
2017 written = -ret;
2018 relayd_hang_up = 1;
2019 goto write_error;
2020 } else if (ret_splice > len) {
2021 /*
2022 * We don't expect this code path to be executed but you never know
2023 * so this is an extra protection agains a buggy splice().
2024 */
2025 ret = errno;
2026 written += ret_splice;
2027 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice,
2028 len);
2029 goto splice_error;
2030 } else {
2031 /* All good, update current len and continue. */
2032 len -= ret_splice;
2033 }
2034
2035 /* This call is useless on a socket so better save a syscall. */
2036 if (!relayd) {
2037 /* This won't block, but will start writeout asynchronously */
2038 lttng_sync_file_range(outfd, stream->out_fd_offset, ret_splice,
2039 SYNC_FILE_RANGE_WRITE);
2040 stream->out_fd_offset += ret_splice;
2041 }
2042 stream->output_written += ret_splice;
2043 written += ret_splice;
2044 }
2045 if (!relayd) {
2046 lttng_consumer_sync_trace_file(stream, orig_offset);
2047 }
2048 goto end;
2049
2050 write_error:
2051 /*
2052 * This is a special case that the relayd has closed its socket. Let's
2053 * cleanup the relayd object and all associated streams.
2054 */
2055 if (relayd && relayd_hang_up) {
2056 ERR("Relayd hangup. Cleaning up relayd %" PRIu64".", relayd->net_seq_idx);
2057 lttng_consumer_cleanup_relayd(relayd);
2058 /* Skip splice error so the consumer does not fail */
2059 goto end;
2060 }
2061
2062 splice_error:
2063 /* send the appropriate error description to sessiond */
2064 switch (ret) {
2065 case EINVAL:
2066 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_EINVAL);
2067 break;
2068 case ENOMEM:
2069 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_ENOMEM);
2070 break;
2071 case ESPIPE:
2072 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_ESPIPE);
2073 break;
2074 }
2075
2076 end:
2077 if (relayd && stream->metadata_flag) {
2078 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
2079 }
2080
2081 rcu_read_unlock();
2082 return written;
2083 }
2084
2085 /*
2086 * Sample the snapshot positions for a specific fd
2087 *
2088 * Returns 0 on success, < 0 on error
2089 */
2090 int lttng_consumer_sample_snapshot_positions(struct lttng_consumer_stream *stream)
2091 {
2092 switch (consumer_data.type) {
2093 case LTTNG_CONSUMER_KERNEL:
2094 return lttng_kconsumer_sample_snapshot_positions(stream);
2095 case LTTNG_CONSUMER32_UST:
2096 case LTTNG_CONSUMER64_UST:
2097 return lttng_ustconsumer_sample_snapshot_positions(stream);
2098 default:
2099 ERR("Unknown consumer_data type");
2100 assert(0);
2101 return -ENOSYS;
2102 }
2103 }
2104 /*
2105 * Take a snapshot for a specific fd
2106 *
2107 * Returns 0 on success, < 0 on error
2108 */
2109 int lttng_consumer_take_snapshot(struct lttng_consumer_stream *stream)
2110 {
2111 switch (consumer_data.type) {
2112 case LTTNG_CONSUMER_KERNEL:
2113 return lttng_kconsumer_take_snapshot(stream);
2114 case LTTNG_CONSUMER32_UST:
2115 case LTTNG_CONSUMER64_UST:
2116 return lttng_ustconsumer_take_snapshot(stream);
2117 default:
2118 ERR("Unknown consumer_data type");
2119 assert(0);
2120 return -ENOSYS;
2121 }
2122 }
2123
2124 /*
2125 * Get the produced position
2126 *
2127 * Returns 0 on success, < 0 on error
2128 */
2129 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream *stream,
2130 unsigned long *pos)
2131 {
2132 switch (consumer_data.type) {
2133 case LTTNG_CONSUMER_KERNEL:
2134 return lttng_kconsumer_get_produced_snapshot(stream, pos);
2135 case LTTNG_CONSUMER32_UST:
2136 case LTTNG_CONSUMER64_UST:
2137 return lttng_ustconsumer_get_produced_snapshot(stream, pos);
2138 default:
2139 ERR("Unknown consumer_data type");
2140 assert(0);
2141 return -ENOSYS;
2142 }
2143 }
2144
2145 /*
2146 * Get the consumed position (free-running counter position in bytes).
2147 *
2148 * Returns 0 on success, < 0 on error
2149 */
2150 int lttng_consumer_get_consumed_snapshot(struct lttng_consumer_stream *stream,
2151 unsigned long *pos)
2152 {
2153 switch (consumer_data.type) {
2154 case LTTNG_CONSUMER_KERNEL:
2155 return lttng_kconsumer_get_consumed_snapshot(stream, pos);
2156 case LTTNG_CONSUMER32_UST:
2157 case LTTNG_CONSUMER64_UST:
2158 return lttng_ustconsumer_get_consumed_snapshot(stream, pos);
2159 default:
2160 ERR("Unknown consumer_data type");
2161 assert(0);
2162 return -ENOSYS;
2163 }
2164 }
2165
2166 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data *ctx,
2167 int sock, struct pollfd *consumer_sockpoll)
2168 {
2169 switch (consumer_data.type) {
2170 case LTTNG_CONSUMER_KERNEL:
2171 return lttng_kconsumer_recv_cmd(ctx, sock, consumer_sockpoll);
2172 case LTTNG_CONSUMER32_UST:
2173 case LTTNG_CONSUMER64_UST:
2174 return lttng_ustconsumer_recv_cmd(ctx, sock, consumer_sockpoll);
2175 default:
2176 ERR("Unknown consumer_data type");
2177 assert(0);
2178 return -ENOSYS;
2179 }
2180 }
2181
2182 void lttng_consumer_close_all_metadata(void)
2183 {
2184 switch (consumer_data.type) {
2185 case LTTNG_CONSUMER_KERNEL:
2186 /*
2187 * The Kernel consumer has a different metadata scheme so we don't
2188 * close anything because the stream will be closed by the session
2189 * daemon.
2190 */
2191 break;
2192 case LTTNG_CONSUMER32_UST:
2193 case LTTNG_CONSUMER64_UST:
2194 /*
2195 * Close all metadata streams. The metadata hash table is passed and
2196 * this call iterates over it by closing all wakeup fd. This is safe
2197 * because at this point we are sure that the metadata producer is
2198 * either dead or blocked.
2199 */
2200 lttng_ustconsumer_close_all_metadata(metadata_ht);
2201 break;
2202 default:
2203 ERR("Unknown consumer_data type");
2204 assert(0);
2205 }
2206 }
2207
2208 /*
2209 * Clean up a metadata stream and free its memory.
2210 */
2211 void consumer_del_metadata_stream(struct lttng_consumer_stream *stream,
2212 struct lttng_ht *ht)
2213 {
2214 struct lttng_consumer_channel *channel = NULL;
2215 bool free_channel = false;
2216
2217 assert(stream);
2218 /*
2219 * This call should NEVER receive regular stream. It must always be
2220 * metadata stream and this is crucial for data structure synchronization.
2221 */
2222 assert(stream->metadata_flag);
2223
2224 DBG3("Consumer delete metadata stream %d", stream->wait_fd);
2225
2226 pthread_mutex_lock(&consumer_data.lock);
2227 /*
2228 * Note that this assumes that a stream's channel is never changed and
2229 * that the stream's lock doesn't need to be taken to sample its
2230 * channel.
2231 */
2232 channel = stream->chan;
2233 pthread_mutex_lock(&channel->lock);
2234 pthread_mutex_lock(&stream->lock);
2235 if (channel->metadata_cache) {
2236 /* Only applicable to userspace consumers. */
2237 pthread_mutex_lock(&channel->metadata_cache->lock);
2238 }
2239
2240 /* Remove any reference to that stream. */
2241 consumer_stream_delete(stream, ht);
2242
2243 /* Close down everything including the relayd if one. */
2244 consumer_stream_close(stream);
2245 /* Destroy tracer buffers of the stream. */
2246 consumer_stream_destroy_buffers(stream);
2247
2248 /* Atomically decrement channel refcount since other threads can use it. */
2249 if (!uatomic_sub_return(&channel->refcount, 1)
2250 && !uatomic_read(&channel->nb_init_stream_left)) {
2251 /* Go for channel deletion! */
2252 free_channel = true;
2253 }
2254 stream->chan = NULL;
2255
2256 /*
2257 * Nullify the stream reference so it is not used after deletion. The
2258 * channel lock MUST be acquired before being able to check for a NULL
2259 * pointer value.
2260 */
2261 channel->metadata_stream = NULL;
2262
2263 if (channel->metadata_cache) {
2264 pthread_mutex_unlock(&channel->metadata_cache->lock);
2265 }
2266 pthread_mutex_unlock(&stream->lock);
2267 pthread_mutex_unlock(&channel->lock);
2268 pthread_mutex_unlock(&consumer_data.lock);
2269
2270 if (free_channel) {
2271 consumer_del_channel(channel);
2272 }
2273
2274 lttng_trace_chunk_put(stream->trace_chunk);
2275 stream->trace_chunk = NULL;
2276 consumer_stream_free(stream);
2277 }
2278
2279 /*
2280 * Action done with the metadata stream when adding it to the consumer internal
2281 * data structures to handle it.
2282 */
2283 void consumer_add_metadata_stream(struct lttng_consumer_stream *stream)
2284 {
2285 struct lttng_ht *ht = metadata_ht;
2286 struct lttng_ht_iter iter;
2287 struct lttng_ht_node_u64 *node;
2288
2289 assert(stream);
2290 assert(ht);
2291
2292 DBG3("Adding metadata stream %" PRIu64 " to hash table", stream->key);
2293
2294 pthread_mutex_lock(&consumer_data.lock);
2295 pthread_mutex_lock(&stream->chan->lock);
2296 pthread_mutex_lock(&stream->chan->timer_lock);
2297 pthread_mutex_lock(&stream->lock);
2298
2299 /*
2300 * From here, refcounts are updated so be _careful_ when returning an error
2301 * after this point.
2302 */
2303
2304 rcu_read_lock();
2305
2306 /*
2307 * Lookup the stream just to make sure it does not exist in our internal
2308 * state. This should NEVER happen.
2309 */
2310 lttng_ht_lookup(ht, &stream->key, &iter);
2311 node = lttng_ht_iter_get_node_u64(&iter);
2312 assert(!node);
2313
2314 /*
2315 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2316 * in terms of destroying the associated channel, because the action that
2317 * causes the count to become 0 also causes a stream to be added. The
2318 * channel deletion will thus be triggered by the following removal of this
2319 * stream.
2320 */
2321 if (uatomic_read(&stream->chan->nb_init_stream_left) > 0) {
2322 /* Increment refcount before decrementing nb_init_stream_left */
2323 cmm_smp_wmb();
2324 uatomic_dec(&stream->chan->nb_init_stream_left);
2325 }
2326
2327 lttng_ht_add_unique_u64(ht, &stream->node);
2328
2329 lttng_ht_add_u64(consumer_data.stream_per_chan_id_ht,
2330 &stream->node_channel_id);
2331
2332 /*
2333 * Add stream to the stream_list_ht of the consumer data. No need to steal
2334 * the key since the HT does not use it and we allow to add redundant keys
2335 * into this table.
2336 */
2337 lttng_ht_add_u64(consumer_data.stream_list_ht, &stream->node_session_id);
2338
2339 rcu_read_unlock();
2340
2341 pthread_mutex_unlock(&stream->lock);
2342 pthread_mutex_unlock(&stream->chan->lock);
2343 pthread_mutex_unlock(&stream->chan->timer_lock);
2344 pthread_mutex_unlock(&consumer_data.lock);
2345 }
2346
2347 /*
2348 * Delete data stream that are flagged for deletion (endpoint_status).
2349 */
2350 static void validate_endpoint_status_data_stream(void)
2351 {
2352 struct lttng_ht_iter iter;
2353 struct lttng_consumer_stream *stream;
2354
2355 DBG("Consumer delete flagged data stream");
2356
2357 rcu_read_lock();
2358 cds_lfht_for_each_entry(data_ht->ht, &iter.iter, stream, node.node) {
2359 /* Validate delete flag of the stream */
2360 if (stream->endpoint_status == CONSUMER_ENDPOINT_ACTIVE) {
2361 continue;
2362 }
2363 /* Delete it right now */
2364 consumer_del_stream(stream, data_ht);
2365 }
2366 rcu_read_unlock();
2367 }
2368
2369 /*
2370 * Delete metadata stream that are flagged for deletion (endpoint_status).
2371 */
2372 static void validate_endpoint_status_metadata_stream(
2373 struct lttng_poll_event *pollset)
2374 {
2375 struct lttng_ht_iter iter;
2376 struct lttng_consumer_stream *stream;
2377
2378 DBG("Consumer delete flagged metadata stream");
2379
2380 assert(pollset);
2381
2382 rcu_read_lock();
2383 cds_lfht_for_each_entry(metadata_ht->ht, &iter.iter, stream, node.node) {
2384 /* Validate delete flag of the stream */
2385 if (stream->endpoint_status == CONSUMER_ENDPOINT_ACTIVE) {
2386 continue;
2387 }
2388 /*
2389 * Remove from pollset so the metadata thread can continue without
2390 * blocking on a deleted stream.
2391 */
2392 lttng_poll_del(pollset, stream->wait_fd);
2393
2394 /* Delete it right now */
2395 consumer_del_metadata_stream(stream, metadata_ht);
2396 }
2397 rcu_read_unlock();
2398 }
2399
2400 /*
2401 * Thread polls on metadata file descriptor and write them on disk or on the
2402 * network.
2403 */
2404 void *consumer_thread_metadata_poll(void *data)
2405 {
2406 int ret, i, pollfd, err = -1;
2407 uint32_t revents, nb_fd;
2408 struct lttng_consumer_stream *stream = NULL;
2409 struct lttng_ht_iter iter;
2410 struct lttng_ht_node_u64 *node;
2411 struct lttng_poll_event events;
2412 struct lttng_consumer_local_data *ctx = data;
2413 ssize_t len;
2414
2415 rcu_register_thread();
2416
2417 health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_METADATA);
2418
2419 if (testpoint(consumerd_thread_metadata)) {
2420 goto error_testpoint;
2421 }
2422
2423 health_code_update();
2424
2425 DBG("Thread metadata poll started");
2426
2427 /* Size is set to 1 for the consumer_metadata pipe */
2428 ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC);
2429 if (ret < 0) {
2430 ERR("Poll set creation failed");
2431 goto end_poll;
2432 }
2433
2434 ret = lttng_poll_add(&events,
2435 lttng_pipe_get_readfd(ctx->consumer_metadata_pipe), LPOLLIN);
2436 if (ret < 0) {
2437 goto end;
2438 }
2439
2440 /* Main loop */
2441 DBG("Metadata main loop started");
2442
2443 while (1) {
2444 restart:
2445 health_code_update();
2446 health_poll_entry();
2447 DBG("Metadata poll wait");
2448 ret = lttng_poll_wait(&events, -1);
2449 DBG("Metadata poll return from wait with %d fd(s)",
2450 LTTNG_POLL_GETNB(&events));
2451 health_poll_exit();
2452 DBG("Metadata event caught in thread");
2453 if (ret < 0) {
2454 if (errno == EINTR) {
2455 ERR("Poll EINTR caught");
2456 goto restart;
2457 }
2458 if (LTTNG_POLL_GETNB(&events) == 0) {
2459 err = 0; /* All is OK */
2460 }
2461 goto end;
2462 }
2463
2464 nb_fd = ret;
2465
2466 /* From here, the event is a metadata wait fd */
2467 for (i = 0; i < nb_fd; i++) {
2468 health_code_update();
2469
2470 revents = LTTNG_POLL_GETEV(&events, i);
2471 pollfd = LTTNG_POLL_GETFD(&events, i);
2472
2473 if (pollfd == lttng_pipe_get_readfd(ctx->consumer_metadata_pipe)) {
2474 if (revents & LPOLLIN) {
2475 ssize_t pipe_len;
2476
2477 pipe_len = lttng_pipe_read(ctx->consumer_metadata_pipe,
2478 &stream, sizeof(stream));
2479 if (pipe_len < sizeof(stream)) {
2480 if (pipe_len < 0) {
2481 PERROR("read metadata stream");
2482 }
2483 /*
2484 * Remove the pipe from the poll set and continue the loop
2485 * since their might be data to consume.
2486 */
2487 lttng_poll_del(&events,
2488 lttng_pipe_get_readfd(ctx->consumer_metadata_pipe));
2489 lttng_pipe_read_close(ctx->consumer_metadata_pipe);
2490 continue;
2491 }
2492
2493 /* A NULL stream means that the state has changed. */
2494 if (stream == NULL) {
2495 /* Check for deleted streams. */
2496 validate_endpoint_status_metadata_stream(&events);
2497 goto restart;
2498 }
2499
2500 DBG("Adding metadata stream %d to poll set",
2501 stream->wait_fd);
2502
2503 /* Add metadata stream to the global poll events list */
2504 lttng_poll_add(&events, stream->wait_fd,
2505 LPOLLIN | LPOLLPRI | LPOLLHUP);
2506 } else if (revents & (LPOLLERR | LPOLLHUP)) {
2507 DBG("Metadata thread pipe hung up");
2508 /*
2509 * Remove the pipe from the poll set and continue the loop
2510 * since their might be data to consume.
2511 */
2512 lttng_poll_del(&events,
2513 lttng_pipe_get_readfd(ctx->consumer_metadata_pipe));
2514 lttng_pipe_read_close(ctx->consumer_metadata_pipe);
2515 continue;
2516 } else {
2517 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
2518 goto end;
2519 }
2520
2521 /* Handle other stream */
2522 continue;
2523 }
2524
2525 rcu_read_lock();
2526 {
2527 uint64_t tmp_id = (uint64_t) pollfd;
2528
2529 lttng_ht_lookup(metadata_ht, &tmp_id, &iter);
2530 }
2531 node = lttng_ht_iter_get_node_u64(&iter);
2532 assert(node);
2533
2534 stream = caa_container_of(node, struct lttng_consumer_stream,
2535 node);
2536
2537 if (revents & (LPOLLIN | LPOLLPRI)) {
2538 /* Get the data out of the metadata file descriptor */
2539 DBG("Metadata available on fd %d", pollfd);
2540 assert(stream->wait_fd == pollfd);
2541
2542 do {
2543 health_code_update();
2544
2545 len = ctx->on_buffer_ready(stream, ctx);
2546 /*
2547 * We don't check the return value here since if we get
2548 * a negative len, it means an error occurred thus we
2549 * simply remove it from the poll set and free the
2550 * stream.
2551 */
2552 } while (len > 0);
2553
2554 /* It's ok to have an unavailable sub-buffer */
2555 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2556 /* Clean up stream from consumer and free it. */
2557 lttng_poll_del(&events, stream->wait_fd);
2558 consumer_del_metadata_stream(stream, metadata_ht);
2559 }
2560 } else if (revents & (LPOLLERR | LPOLLHUP)) {
2561 DBG("Metadata fd %d is hup|err.", pollfd);
2562 if (!stream->hangup_flush_done
2563 && (consumer_data.type == LTTNG_CONSUMER32_UST
2564 || consumer_data.type == LTTNG_CONSUMER64_UST)) {
2565 DBG("Attempting to flush and consume the UST buffers");
2566 lttng_ustconsumer_on_stream_hangup(stream);
2567
2568 /* We just flushed the stream now read it. */
2569 do {
2570 health_code_update();
2571
2572 len = ctx->on_buffer_ready(stream, ctx);
2573 /*
2574 * We don't check the return value here since if we get
2575 * a negative len, it means an error occurred thus we
2576 * simply remove it from the poll set and free the
2577 * stream.
2578 */
2579 } while (len > 0);
2580 }
2581
2582 lttng_poll_del(&events, stream->wait_fd);
2583 /*
2584 * This call update the channel states, closes file descriptors
2585 * and securely free the stream.
2586 */
2587 consumer_del_metadata_stream(stream, metadata_ht);
2588 } else {
2589 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
2590 rcu_read_unlock();
2591 goto end;
2592 }
2593 /* Release RCU lock for the stream looked up */
2594 rcu_read_unlock();
2595 }
2596 }
2597
2598 /* All is OK */
2599 err = 0;
2600 end:
2601 DBG("Metadata poll thread exiting");
2602
2603 lttng_poll_clean(&events);
2604 end_poll:
2605 error_testpoint:
2606 if (err) {
2607 health_error();
2608 ERR("Health error occurred in %s", __func__);
2609 }
2610 health_unregister(health_consumerd);
2611 rcu_unregister_thread();
2612 return NULL;
2613 }
2614
2615 /*
2616 * This thread polls the fds in the set to consume the data and write
2617 * it to tracefile if necessary.
2618 */
2619 void *consumer_thread_data_poll(void *data)
2620 {
2621 int num_rdy, num_hup, high_prio, ret, i, err = -1;
2622 struct pollfd *pollfd = NULL;
2623 /* local view of the streams */
2624 struct lttng_consumer_stream **local_stream = NULL, *new_stream = NULL;
2625 /* local view of consumer_data.fds_count */
2626 int nb_fd = 0;
2627 /* 2 for the consumer_data_pipe and wake up pipe */
2628 const int nb_pipes_fd = 2;
2629 /* Number of FDs with CONSUMER_ENDPOINT_INACTIVE but still open. */
2630 int nb_inactive_fd = 0;
2631 struct lttng_consumer_local_data *ctx = data;
2632 ssize_t len;
2633
2634 rcu_register_thread();
2635
2636 health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_DATA);
2637
2638 if (testpoint(consumerd_thread_data)) {
2639 goto error_testpoint;
2640 }
2641
2642 health_code_update();
2643
2644 local_stream = zmalloc(sizeof(struct lttng_consumer_stream *));
2645 if (local_stream == NULL) {
2646 PERROR("local_stream malloc");
2647 goto end;
2648 }
2649
2650 while (1) {
2651 health_code_update();
2652
2653 high_prio = 0;
2654 num_hup = 0;
2655
2656 /*
2657 * the fds set has been updated, we need to update our
2658 * local array as well
2659 */
2660 pthread_mutex_lock(&consumer_data.lock);
2661 if (consumer_data.need_update) {
2662 free(pollfd);
2663 pollfd = NULL;
2664
2665 free(local_stream);
2666 local_stream = NULL;
2667
2668 /* Allocate for all fds */
2669 pollfd = zmalloc((consumer_data.stream_count + nb_pipes_fd) * sizeof(struct pollfd));
2670 if (pollfd == NULL) {
2671 PERROR("pollfd malloc");
2672 pthread_mutex_unlock(&consumer_data.lock);
2673 goto end;
2674 }
2675
2676 local_stream = zmalloc((consumer_data.stream_count + nb_pipes_fd) *
2677 sizeof(struct lttng_consumer_stream *));
2678 if (local_stream == NULL) {
2679 PERROR("local_stream malloc");
2680 pthread_mutex_unlock(&consumer_data.lock);
2681 goto end;
2682 }
2683 ret = update_poll_array(ctx, &pollfd, local_stream,
2684 data_ht, &nb_inactive_fd);
2685 if (ret < 0) {
2686 ERR("Error in allocating pollfd or local_outfds");
2687 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
2688 pthread_mutex_unlock(&consumer_data.lock);
2689 goto end;
2690 }
2691 nb_fd = ret;
2692 consumer_data.need_update = 0;
2693 }
2694 pthread_mutex_unlock(&consumer_data.lock);
2695
2696 /* No FDs and consumer_quit, consumer_cleanup the thread */
2697 if (nb_fd == 0 && nb_inactive_fd == 0 &&
2698 CMM_LOAD_SHARED(consumer_quit) == 1) {
2699 err = 0; /* All is OK */
2700 goto end;
2701 }
2702 /* poll on the array of fds */
2703 restart:
2704 DBG("polling on %d fd", nb_fd + nb_pipes_fd);
2705 if (testpoint(consumerd_thread_data_poll)) {
2706 goto end;
2707 }
2708 health_poll_entry();
2709 num_rdy = poll(pollfd, nb_fd + nb_pipes_fd, -1);
2710 health_poll_exit();
2711 DBG("poll num_rdy : %d", num_rdy);
2712 if (num_rdy == -1) {
2713 /*
2714 * Restart interrupted system call.
2715 */
2716 if (errno == EINTR) {
2717 goto restart;
2718 }
2719 PERROR("Poll error");
2720 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
2721 goto end;
2722 } else if (num_rdy == 0) {
2723 DBG("Polling thread timed out");
2724 goto end;
2725 }
2726
2727 if (caa_unlikely(data_consumption_paused)) {
2728 DBG("Data consumption paused, sleeping...");
2729 sleep(1);
2730 goto restart;
2731 }
2732
2733 /*
2734 * If the consumer_data_pipe triggered poll go directly to the
2735 * beginning of the loop to update the array. We want to prioritize
2736 * array update over low-priority reads.
2737 */
2738 if (pollfd[nb_fd].revents & (POLLIN | POLLPRI)) {
2739 ssize_t pipe_readlen;
2740
2741 DBG("consumer_data_pipe wake up");
2742 pipe_readlen = lttng_pipe_read(ctx->consumer_data_pipe,
2743 &new_stream, sizeof(new_stream));
2744 if (pipe_readlen < sizeof(new_stream)) {
2745 PERROR("Consumer data pipe");
2746 /* Continue so we can at least handle the current stream(s). */
2747 continue;
2748 }
2749
2750 /*
2751 * If the stream is NULL, just ignore it. It's also possible that
2752 * the sessiond poll thread changed the consumer_quit state and is
2753 * waking us up to test it.
2754 */
2755 if (new_stream == NULL) {
2756 validate_endpoint_status_data_stream();
2757 continue;
2758 }
2759
2760 /* Continue to update the local streams and handle prio ones */
2761 continue;
2762 }
2763
2764 /* Handle wakeup pipe. */
2765 if (pollfd[nb_fd + 1].revents & (POLLIN | POLLPRI)) {
2766 char dummy;
2767 ssize_t pipe_readlen;
2768
2769 pipe_readlen = lttng_pipe_read(ctx->consumer_wakeup_pipe, &dummy,
2770 sizeof(dummy));
2771 if (pipe_readlen < 0) {
2772 PERROR("Consumer data wakeup pipe");
2773 }
2774 /* We've been awakened to handle stream(s). */
2775 ctx->has_wakeup = 0;
2776 }
2777
2778 /* Take care of high priority channels first. */
2779 for (i = 0; i < nb_fd; i++) {
2780 health_code_update();
2781
2782 if (local_stream[i] == NULL) {
2783 continue;
2784 }
2785 if (pollfd[i].revents & POLLPRI) {
2786 DBG("Urgent read on fd %d", pollfd[i].fd);
2787 high_prio = 1;
2788 len = ctx->on_buffer_ready(local_stream[i], ctx);
2789 /* it's ok to have an unavailable sub-buffer */
2790 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2791 /* Clean the stream and free it. */
2792 consumer_del_stream(local_stream[i], data_ht);
2793 local_stream[i] = NULL;
2794 } else if (len > 0) {
2795 local_stream[i]->data_read = 1;
2796 }
2797 }
2798 }
2799
2800 /*
2801 * If we read high prio channel in this loop, try again
2802 * for more high prio data.
2803 */
2804 if (high_prio) {
2805 continue;
2806 }
2807
2808 /* Take care of low priority channels. */
2809 for (i = 0; i < nb_fd; i++) {
2810 health_code_update();
2811
2812 if (local_stream[i] == NULL) {
2813 continue;
2814 }
2815 if ((pollfd[i].revents & POLLIN) ||
2816 local_stream[i]->hangup_flush_done ||
2817 local_stream[i]->has_data) {
2818 DBG("Normal read on fd %d", pollfd[i].fd);
2819 len = ctx->on_buffer_ready(local_stream[i], ctx);
2820 /* it's ok to have an unavailable sub-buffer */
2821 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2822 /* Clean the stream and free it. */
2823 consumer_del_stream(local_stream[i], data_ht);
2824 local_stream[i] = NULL;
2825 } else if (len > 0) {
2826 local_stream[i]->data_read = 1;
2827 }
2828 }
2829 }
2830
2831 /* Handle hangup and errors */
2832 for (i = 0; i < nb_fd; i++) {
2833 health_code_update();
2834
2835 if (local_stream[i] == NULL) {
2836 continue;
2837 }
2838 if (!local_stream[i]->hangup_flush_done
2839 && (pollfd[i].revents & (POLLHUP | POLLERR | POLLNVAL))
2840 && (consumer_data.type == LTTNG_CONSUMER32_UST
2841 || consumer_data.type == LTTNG_CONSUMER64_UST)) {
2842 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2843 pollfd[i].fd);
2844 lttng_ustconsumer_on_stream_hangup(local_stream[i]);
2845 /* Attempt read again, for the data we just flushed. */
2846 local_stream[i]->data_read = 1;
2847 }
2848 /*
2849 * If the poll flag is HUP/ERR/NVAL and we have
2850 * read no data in this pass, we can remove the
2851 * stream from its hash table.
2852 */
2853 if ((pollfd[i].revents & POLLHUP)) {
2854 DBG("Polling fd %d tells it has hung up.", pollfd[i].fd);
2855 if (!local_stream[i]->data_read) {
2856 consumer_del_stream(local_stream[i], data_ht);
2857 local_stream[i] = NULL;
2858 num_hup++;
2859 }
2860 } else if (pollfd[i].revents & POLLERR) {
2861 ERR("Error returned in polling fd %d.", pollfd[i].fd);
2862 if (!local_stream[i]->data_read) {
2863 consumer_del_stream(local_stream[i], data_ht);
2864 local_stream[i] = NULL;
2865 num_hup++;
2866 }
2867 } else if (pollfd[i].revents & POLLNVAL) {
2868 ERR("Polling fd %d tells fd is not open.", pollfd[i].fd);
2869 if (!local_stream[i]->data_read) {
2870 consumer_del_stream(local_stream[i], data_ht);
2871 local_stream[i] = NULL;
2872 num_hup++;
2873 }
2874 }
2875 if (local_stream[i] != NULL) {
2876 local_stream[i]->data_read = 0;
2877 }
2878 }
2879 }
2880 /* All is OK */
2881 err = 0;
2882 end:
2883 DBG("polling thread exiting");
2884 free(pollfd);
2885 free(local_stream);
2886
2887 /*
2888 * Close the write side of the pipe so epoll_wait() in
2889 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2890 * read side of the pipe. If we close them both, epoll_wait strangely does
2891 * not return and could create a endless wait period if the pipe is the
2892 * only tracked fd in the poll set. The thread will take care of closing
2893 * the read side.
2894 */
2895 (void) lttng_pipe_write_close(ctx->consumer_metadata_pipe);
2896
2897 error_testpoint:
2898 if (err) {
2899 health_error();
2900 ERR("Health error occurred in %s", __func__);
2901 }
2902 health_unregister(health_consumerd);
2903
2904 rcu_unregister_thread();
2905 return NULL;
2906 }
2907
2908 /*
2909 * Close wake-up end of each stream belonging to the channel. This will
2910 * allow the poll() on the stream read-side to detect when the
2911 * write-side (application) finally closes them.
2912 */
2913 static
2914 void consumer_close_channel_streams(struct lttng_consumer_channel *channel)
2915 {
2916 struct lttng_ht *ht;
2917 struct lttng_consumer_stream *stream;
2918 struct lttng_ht_iter iter;
2919
2920 ht = consumer_data.stream_per_chan_id_ht;
2921
2922 rcu_read_lock();
2923 cds_lfht_for_each_entry_duplicate(ht->ht,
2924 ht->hash_fct(&channel->key, lttng_ht_seed),
2925 ht->match_fct, &channel->key,
2926 &iter.iter, stream, node_channel_id.node) {
2927 /*
2928 * Protect against teardown with mutex.
2929 */
2930 pthread_mutex_lock(&stream->lock);
2931 if (cds_lfht_is_node_deleted(&stream->node.node)) {
2932 goto next;
2933 }
2934 switch (consumer_data.type) {
2935 case LTTNG_CONSUMER_KERNEL:
2936 break;
2937 case LTTNG_CONSUMER32_UST:
2938 case LTTNG_CONSUMER64_UST:
2939 if (stream->metadata_flag) {
2940 /* Safe and protected by the stream lock. */
2941 lttng_ustconsumer_close_metadata(stream->chan);
2942 } else {
2943 /*
2944 * Note: a mutex is taken internally within
2945 * liblttng-ust-ctl to protect timer wakeup_fd
2946 * use from concurrent close.
2947 */
2948 lttng_ustconsumer_close_stream_wakeup(stream);
2949 }
2950 break;
2951 default:
2952 ERR("Unknown consumer_data type");
2953 assert(0);
2954 }
2955 next:
2956 pthread_mutex_unlock(&stream->lock);
2957 }
2958 rcu_read_unlock();
2959 }
2960
2961 static void destroy_channel_ht(struct lttng_ht *ht)
2962 {
2963 struct lttng_ht_iter iter;
2964 struct lttng_consumer_channel *channel;
2965 int ret;
2966
2967 if (ht == NULL) {
2968 return;
2969 }
2970
2971 rcu_read_lock();
2972 cds_lfht_for_each_entry(ht->ht, &iter.iter, channel, wait_fd_node.node) {
2973 ret = lttng_ht_del(ht, &iter);
2974 assert(ret != 0);
2975 }
2976 rcu_read_unlock();
2977
2978 lttng_ht_destroy(ht);
2979 }
2980
2981 /*
2982 * This thread polls the channel fds to detect when they are being
2983 * closed. It closes all related streams if the channel is detected as
2984 * closed. It is currently only used as a shim layer for UST because the
2985 * consumerd needs to keep the per-stream wakeup end of pipes open for
2986 * periodical flush.
2987 */
2988 void *consumer_thread_channel_poll(void *data)
2989 {
2990 int ret, i, pollfd, err = -1;
2991 uint32_t revents, nb_fd;
2992 struct lttng_consumer_channel *chan = NULL;
2993 struct lttng_ht_iter iter;
2994 struct lttng_ht_node_u64 *node;
2995 struct lttng_poll_event events;
2996 struct lttng_consumer_local_data *ctx = data;
2997 struct lttng_ht *channel_ht;
2998
2999 rcu_register_thread();
3000
3001 health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_CHANNEL);
3002
3003 if (testpoint(consumerd_thread_channel)) {
3004 goto error_testpoint;
3005 }
3006
3007 health_code_update();
3008
3009 channel_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
3010 if (!channel_ht) {
3011 /* ENOMEM at this point. Better to bail out. */
3012 goto end_ht;
3013 }
3014
3015 DBG("Thread channel poll started");
3016
3017 /* Size is set to 1 for the consumer_channel pipe */
3018 ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC);
3019 if (ret < 0) {
3020 ERR("Poll set creation failed");
3021 goto end_poll;
3022 }
3023
3024 ret = lttng_poll_add(&events, ctx->consumer_channel_pipe[0], LPOLLIN);
3025 if (ret < 0) {
3026 goto end;
3027 }
3028
3029 /* Main loop */
3030 DBG("Channel main loop started");
3031
3032 while (1) {
3033 restart:
3034 health_code_update();
3035 DBG("Channel poll wait");
3036 health_poll_entry();
3037 ret = lttng_poll_wait(&events, -1);
3038 DBG("Channel poll return from wait with %d fd(s)",
3039 LTTNG_POLL_GETNB(&events));
3040 health_poll_exit();
3041 DBG("Channel event caught in thread");
3042 if (ret < 0) {
3043 if (errno == EINTR) {
3044 ERR("Poll EINTR caught");
3045 goto restart;
3046 }
3047 if (LTTNG_POLL_GETNB(&events) == 0) {
3048 err = 0; /* All is OK */
3049 }
3050 goto end;
3051 }
3052
3053 nb_fd = ret;
3054
3055 /* From here, the event is a channel wait fd */
3056 for (i = 0; i < nb_fd; i++) {
3057 health_code_update();
3058
3059 revents = LTTNG_POLL_GETEV(&events, i);
3060 pollfd = LTTNG_POLL_GETFD(&events, i);
3061
3062 if (pollfd == ctx->consumer_channel_pipe[0]) {
3063 if (revents & LPOLLIN) {
3064 enum consumer_channel_action action;
3065 uint64_t key;
3066
3067 ret = read_channel_pipe(ctx, &chan, &key, &action);
3068 if (ret <= 0) {
3069 if (ret < 0) {
3070 ERR("Error reading channel pipe");
3071 }
3072 lttng_poll_del(&events, ctx->consumer_channel_pipe[0]);
3073 continue;
3074 }
3075
3076 switch (action) {
3077 case CONSUMER_CHANNEL_ADD:
3078 DBG("Adding channel %d to poll set",
3079 chan->wait_fd);
3080
3081 lttng_ht_node_init_u64(&chan->wait_fd_node,
3082 chan->wait_fd);
3083 rcu_read_lock();
3084 lttng_ht_add_unique_u64(channel_ht,
3085 &chan->wait_fd_node);
3086 rcu_read_unlock();
3087 /* Add channel to the global poll events list */
3088 lttng_poll_add(&events, chan->wait_fd,
3089 LPOLLERR | LPOLLHUP);
3090 break;
3091 case CONSUMER_CHANNEL_DEL:
3092 {
3093 /*
3094 * This command should never be called if the channel
3095 * has streams monitored by either the data or metadata
3096 * thread. The consumer only notify this thread with a
3097 * channel del. command if it receives a destroy
3098 * channel command from the session daemon that send it
3099 * if a command prior to the GET_CHANNEL failed.
3100 */
3101
3102 rcu_read_lock();
3103 chan = consumer_find_channel(key);
3104 if (!chan) {
3105 rcu_read_unlock();
3106 ERR("UST consumer get channel key %" PRIu64 " not found for del channel", key);
3107 break;
3108 }
3109 lttng_poll_del(&events, chan->wait_fd);
3110 iter.iter.node = &chan->wait_fd_node.node;
3111 ret = lttng_ht_del(channel_ht, &iter);
3112 assert(ret == 0);
3113
3114 switch (consumer_data.type) {
3115 case LTTNG_CONSUMER_KERNEL:
3116 break;
3117 case LTTNG_CONSUMER32_UST:
3118 case LTTNG_CONSUMER64_UST:
3119 health_code_update();
3120 /* Destroy streams that might have been left in the stream list. */
3121 clean_channel_stream_list(chan);
3122 break;
3123 default:
3124 ERR("Unknown consumer_data type");
3125 assert(0);
3126 }
3127
3128 /*
3129 * Release our own refcount. Force channel deletion even if
3130 * streams were not initialized.
3131 */
3132 if (!uatomic_sub_return(&chan->refcount, 1)) {
3133 consumer_del_channel(chan);
3134 }
3135 rcu_read_unlock();
3136 goto restart;
3137 }
3138 case CONSUMER_CHANNEL_QUIT:
3139 /*
3140 * Remove the pipe from the poll set and continue the loop
3141 * since their might be data to consume.
3142 */
3143 lttng_poll_del(&events, ctx->consumer_channel_pipe[0]);
3144 continue;
3145 default:
3146 ERR("Unknown action");
3147 break;
3148 }
3149 } else if (revents & (LPOLLERR | LPOLLHUP)) {
3150 DBG("Channel thread pipe hung up");
3151 /*
3152 * Remove the pipe from the poll set and continue the loop
3153 * since their might be data to consume.
3154 */
3155 lttng_poll_del(&events, ctx->consumer_channel_pipe[0]);
3156 continue;
3157 } else {
3158 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
3159 goto end;
3160 }
3161
3162 /* Handle other stream */
3163 continue;
3164 }
3165
3166 rcu_read_lock();
3167 {
3168 uint64_t tmp_id = (uint64_t) pollfd;
3169
3170 lttng_ht_lookup(channel_ht, &tmp_id, &iter);
3171 }
3172 node = lttng_ht_iter_get_node_u64(&iter);
3173 assert(node);
3174
3175 chan = caa_container_of(node, struct lttng_consumer_channel,
3176 wait_fd_node);
3177
3178 /* Check for error event */
3179 if (revents & (LPOLLERR | LPOLLHUP)) {
3180 DBG("Channel fd %d is hup|err.", pollfd);
3181
3182 lttng_poll_del(&events, chan->wait_fd);
3183 ret = lttng_ht_del(channel_ht, &iter);
3184 assert(ret == 0);
3185
3186 /*
3187 * This will close the wait fd for each stream associated to
3188 * this channel AND monitored by the data/metadata thread thus
3189 * will be clean by the right thread.
3190 */
3191 consumer_close_channel_streams(chan);
3192
3193 /* Release our own refcount */
3194 if (!uatomic_sub_return(&chan->refcount, 1)
3195 && !uatomic_read(&chan->nb_init_stream_left)) {
3196 consumer_del_channel(chan);
3197 }
3198 } else {
3199 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
3200 rcu_read_unlock();
3201 goto end;
3202 }
3203
3204 /* Release RCU lock for the channel looked up */
3205 rcu_read_unlock();
3206 }
3207 }
3208
3209 /* All is OK */
3210 err = 0;
3211 end:
3212 lttng_poll_clean(&events);
3213 end_poll:
3214 destroy_channel_ht(channel_ht);
3215 end_ht:
3216 error_testpoint:
3217 DBG("Channel poll thread exiting");
3218 if (err) {
3219 health_error();
3220 ERR("Health error occurred in %s", __func__);
3221 }
3222 health_unregister(health_consumerd);
3223 rcu_unregister_thread();
3224 return NULL;
3225 }
3226
3227 static int set_metadata_socket(struct lttng_consumer_local_data *ctx,
3228 struct pollfd *sockpoll, int client_socket)
3229 {
3230 int ret;
3231
3232 assert(ctx);
3233 assert(sockpoll);
3234
3235 ret = lttng_consumer_poll_socket(sockpoll);
3236 if (ret) {
3237 goto error;
3238 }
3239 DBG("Metadata connection on client_socket");
3240
3241 /* Blocking call, waiting for transmission */
3242 ctx->consumer_metadata_socket = lttcomm_accept_unix_sock(client_socket);
3243 if (ctx->consumer_metadata_socket < 0) {
3244 WARN("On accept metadata");
3245 ret = -1;
3246 goto error;
3247 }
3248 ret = 0;
3249
3250 error:
3251 return ret;
3252 }
3253
3254 /*
3255 * This thread listens on the consumerd socket and receives the file
3256 * descriptors from the session daemon.
3257 */
3258 void *consumer_thread_sessiond_poll(void *data)
3259 {
3260 int sock = -1, client_socket, ret, err = -1;
3261 /*
3262 * structure to poll for incoming data on communication socket avoids
3263 * making blocking sockets.
3264 */
3265 struct pollfd consumer_sockpoll[2];
3266 struct lttng_consumer_local_data *ctx = data;
3267
3268 rcu_register_thread();
3269
3270 health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_SESSIOND);
3271
3272 if (testpoint(consumerd_thread_sessiond)) {
3273 goto error_testpoint;
3274 }
3275
3276 health_code_update();
3277
3278 DBG("Creating command socket %s", ctx->consumer_command_sock_path);
3279 unlink(ctx->consumer_command_sock_path);
3280 client_socket = lttcomm_create_unix_sock(ctx->consumer_command_sock_path);
3281 if (client_socket < 0) {
3282 ERR("Cannot create command socket");
3283 goto end;
3284 }
3285
3286 ret = lttcomm_listen_unix_sock(client_socket);
3287 if (ret < 0) {
3288 goto end;
3289 }
3290
3291 DBG("Sending ready command to lttng-sessiond");
3292 ret = lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY);
3293 /* return < 0 on error, but == 0 is not fatal */
3294 if (ret < 0) {
3295 ERR("Error sending ready command to lttng-sessiond");
3296 goto end;
3297 }
3298
3299 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3300 consumer_sockpoll[0].fd = ctx->consumer_should_quit[0];
3301 consumer_sockpoll[0].events = POLLIN | POLLPRI;
3302 consumer_sockpoll[1].fd = client_socket;
3303 consumer_sockpoll[1].events = POLLIN | POLLPRI;
3304
3305 ret = lttng_consumer_poll_socket(consumer_sockpoll);
3306 if (ret) {
3307 if (ret > 0) {
3308 /* should exit */
3309 err = 0;
3310 }
3311 goto end;
3312 }
3313 DBG("Connection on client_socket");
3314
3315 /* Blocking call, waiting for transmission */
3316 sock = lttcomm_accept_unix_sock(client_socket);
3317 if (sock < 0) {
3318 WARN("On accept");
3319 goto end;
3320 }
3321
3322 /*
3323 * Setup metadata socket which is the second socket connection on the
3324 * command unix socket.
3325 */
3326 ret = set_metadata_socket(ctx, consumer_sockpoll, client_socket);
3327 if (ret) {
3328 if (ret > 0) {
3329 /* should exit */
3330 err = 0;
3331 }
3332 goto end;
3333 }
3334
3335 /* This socket is not useful anymore. */
3336 ret = close(client_socket);
3337 if (ret < 0) {
3338 PERROR("close client_socket");
3339 }
3340 client_socket = -1;
3341
3342 /* update the polling structure to poll on the established socket */
3343 consumer_sockpoll[1].fd = sock;
3344 consumer_sockpoll[1].events = POLLIN | POLLPRI;
3345
3346 while (1) {
3347 health_code_update();
3348
3349 health_poll_entry();
3350 ret = lttng_consumer_poll_socket(consumer_sockpoll);
3351 health_poll_exit();
3352 if (ret) {
3353 if (ret > 0) {
3354 /* should exit */
3355 err = 0;
3356 }
3357 goto end;
3358 }
3359 DBG("Incoming command on sock");
3360 ret = lttng_consumer_recv_cmd(ctx, sock, consumer_sockpoll);
3361 if (ret <= 0) {
3362 /*
3363 * This could simply be a session daemon quitting. Don't output
3364 * ERR() here.
3365 */
3366 DBG("Communication interrupted on command socket");
3367 err = 0;
3368 goto end;
3369 }
3370 if (CMM_LOAD_SHARED(consumer_quit)) {
3371 DBG("consumer_thread_receive_fds received quit from signal");
3372 err = 0; /* All is OK */
3373 goto end;
3374 }
3375 DBG("received command on sock");
3376 }
3377 /* All is OK */
3378 err = 0;
3379
3380 end:
3381 DBG("Consumer thread sessiond poll exiting");
3382
3383 /*
3384 * Close metadata streams since the producer is the session daemon which
3385 * just died.
3386 *
3387 * NOTE: for now, this only applies to the UST tracer.
3388 */
3389 lttng_consumer_close_all_metadata();
3390
3391 /*
3392 * when all fds have hung up, the polling thread
3393 * can exit cleanly
3394 */
3395 CMM_STORE_SHARED(consumer_quit, 1);
3396
3397 /*
3398 * Notify the data poll thread to poll back again and test the
3399 * consumer_quit state that we just set so to quit gracefully.
3400 */
3401 notify_thread_lttng_pipe(ctx->consumer_data_pipe);
3402
3403 notify_channel_pipe(ctx, NULL, -1, CONSUMER_CHANNEL_QUIT);
3404
3405 notify_health_quit_pipe(health_quit_pipe);
3406
3407 /* Cleaning up possibly open sockets. */
3408 if (sock >= 0) {
3409 ret = close(sock);
3410 if (ret < 0) {
3411 PERROR("close sock sessiond poll");
3412 }
3413 }
3414 if (client_socket >= 0) {
3415 ret = close(client_socket);
3416 if (ret < 0) {
3417 PERROR("close client_socket sessiond poll");
3418 }
3419 }
3420
3421 error_testpoint:
3422 if (err) {
3423 health_error();
3424 ERR("Health error occurred in %s", __func__);
3425 }
3426 health_unregister(health_consumerd);
3427
3428 rcu_unregister_thread();
3429 return NULL;
3430 }
3431
3432 ssize_t lttng_consumer_read_subbuffer(struct lttng_consumer_stream *stream,
3433 struct lttng_consumer_local_data *ctx)
3434 {
3435 ssize_t ret;
3436
3437 pthread_mutex_lock(&stream->chan->lock);
3438 pthread_mutex_lock(&stream->lock);
3439 if (stream->metadata_flag) {
3440 pthread_mutex_lock(&stream->metadata_rdv_lock);
3441 }
3442
3443 switch (consumer_data.type) {
3444 case LTTNG_CONSUMER_KERNEL:
3445 ret = lttng_kconsumer_read_subbuffer(stream, ctx);
3446 break;
3447 case LTTNG_CONSUMER32_UST:
3448 case LTTNG_CONSUMER64_UST:
3449 ret = lttng_ustconsumer_read_subbuffer(stream, ctx);
3450 break;
3451 default:
3452 ERR("Unknown consumer_data type");
3453 assert(0);
3454 ret = -ENOSYS;
3455 break;
3456 }
3457
3458 if (stream->metadata_flag) {
3459 pthread_cond_broadcast(&stream->metadata_rdv);
3460 pthread_mutex_unlock(&stream->metadata_rdv_lock);
3461 }
3462 pthread_mutex_unlock(&stream->lock);
3463 pthread_mutex_unlock(&stream->chan->lock);
3464
3465 return ret;
3466 }
3467
3468 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream *stream)
3469 {
3470 switch (consumer_data.type) {
3471 case LTTNG_CONSUMER_KERNEL:
3472 return lttng_kconsumer_on_recv_stream(stream);
3473 case LTTNG_CONSUMER32_UST:
3474 case LTTNG_CONSUMER64_UST:
3475 return lttng_ustconsumer_on_recv_stream(stream);
3476 default:
3477 ERR("Unknown consumer_data type");
3478 assert(0);
3479 return -ENOSYS;
3480 }
3481 }
3482
3483 /*
3484 * Allocate and set consumer data hash tables.
3485 */
3486 int lttng_consumer_init(void)
3487 {
3488 consumer_data.channel_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
3489 if (!consumer_data.channel_ht) {
3490 goto error;
3491 }
3492
3493 consumer_data.channels_by_session_id_ht =
3494 lttng_ht_new(0, LTTNG_HT_TYPE_U64);
3495 if (!consumer_data.channels_by_session_id_ht) {
3496 goto error;
3497 }
3498
3499 consumer_data.relayd_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
3500 if (!consumer_data.relayd_ht) {
3501 goto error;
3502 }
3503
3504 consumer_data.stream_list_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
3505 if (!consumer_data.stream_list_ht) {
3506 goto error;
3507 }
3508
3509 consumer_data.stream_per_chan_id_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
3510 if (!consumer_data.stream_per_chan_id_ht) {
3511 goto error;
3512 }
3513
3514 data_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
3515 if (!data_ht) {
3516 goto error;
3517 }
3518
3519 metadata_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
3520 if (!metadata_ht) {
3521 goto error;
3522 }
3523
3524 consumer_data.chunk_registry = lttng_trace_chunk_registry_create();
3525 if (!consumer_data.chunk_registry) {
3526 goto error;
3527 }
3528
3529 return 0;
3530
3531 error:
3532 return -1;
3533 }
3534
3535 /*
3536 * Process the ADD_RELAYD command receive by a consumer.
3537 *
3538 * This will create a relayd socket pair and add it to the relayd hash table.
3539 * The caller MUST acquire a RCU read side lock before calling it.
3540 */
3541 void consumer_add_relayd_socket(uint64_t net_seq_idx, int sock_type,
3542 struct lttng_consumer_local_data *ctx, int sock,
3543 struct pollfd *consumer_sockpoll,
3544 struct lttcomm_relayd_sock *relayd_sock, uint64_t sessiond_id,
3545 uint64_t relayd_session_id)
3546 {
3547 int fd = -1, ret = -1, relayd_created = 0;
3548 enum lttcomm_return_code ret_code = LTTCOMM_CONSUMERD_SUCCESS;
3549 struct consumer_relayd_sock_pair *relayd = NULL;
3550
3551 assert(ctx);
3552 assert(relayd_sock);
3553
3554 DBG("Consumer adding relayd socket (idx: %" PRIu64 ")", net_seq_idx);
3555
3556 /* Get relayd reference if exists. */
3557 relayd = consumer_find_relayd(net_seq_idx);
3558 if (relayd == NULL) {
3559 assert(sock_type == LTTNG_STREAM_CONTROL);
3560 /* Not found. Allocate one. */
3561 relayd = consumer_allocate_relayd_sock_pair(net_seq_idx);
3562 if (relayd == NULL) {
3563 ret_code = LTTCOMM_CONSUMERD_ENOMEM;
3564 goto error;
3565 } else {
3566 relayd->sessiond_session_id = sessiond_id;
3567 relayd_created = 1;
3568 }
3569
3570 /*
3571 * This code path MUST continue to the consumer send status message to
3572 * we can notify the session daemon and continue our work without
3573 * killing everything.
3574 */
3575 } else {
3576 /*
3577 * relayd key should never be found for control socket.
3578 */
3579 assert(sock_type != LTTNG_STREAM_CONTROL);
3580 }
3581
3582 /* First send a status message before receiving the fds. */
3583 ret = consumer_send_status_msg(sock, LTTCOMM_CONSUMERD_SUCCESS);
3584 if (ret < 0) {
3585 /* Somehow, the session daemon is not responding anymore. */
3586 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_FATAL);
3587 goto error_nosignal;
3588 }
3589
3590 /* Poll on consumer socket. */
3591 ret = lttng_consumer_poll_socket(consumer_sockpoll);
3592 if (ret) {
3593 /* Needing to exit in the middle of a command: error. */
3594 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
3595 goto error_nosignal;
3596 }
3597
3598 /* Get relayd socket from session daemon */
3599 ret = lttcomm_recv_fds_unix_sock(sock, &fd, 1);
3600 if (ret != sizeof(fd)) {
3601 fd = -1; /* Just in case it gets set with an invalid value. */
3602
3603 /*
3604 * Failing to receive FDs might indicate a major problem such as
3605 * reaching a fd limit during the receive where the kernel returns a
3606 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3607 * don't take any chances and stop everything.
3608 *
3609 * XXX: Feature request #558 will fix that and avoid this possible
3610 * issue when reaching the fd limit.
3611 */
3612 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_ERROR_RECV_FD);
3613 ret_code = LTTCOMM_CONSUMERD_ERROR_RECV_FD;
3614 goto error;
3615 }
3616
3617 /* Copy socket information and received FD */
3618 switch (sock_type) {
3619 case LTTNG_STREAM_CONTROL:
3620 /* Copy received lttcomm socket */
3621 lttcomm_copy_sock(&relayd->control_sock.sock, &relayd_sock->sock);
3622 ret = lttcomm_create_sock(&relayd->control_sock.sock);
3623 /* Handle create_sock error. */
3624 if (ret < 0) {
3625 ret_code = LTTCOMM_CONSUMERD_ENOMEM;
3626 goto error;
3627 }
3628 /*
3629 * Close the socket created internally by
3630 * lttcomm_create_sock, so we can replace it by the one
3631 * received from sessiond.
3632 */
3633 if (close(relayd->control_sock.sock.fd)) {
3634 PERROR("close");
3635 }
3636
3637 /* Assign new file descriptor */
3638 relayd->control_sock.sock.fd = fd;
3639 /* Assign version values. */
3640 relayd->control_sock.major = relayd_sock->major;
3641 relayd->control_sock.minor = relayd_sock->minor;
3642
3643 relayd->relayd_session_id = relayd_session_id;
3644
3645 break;
3646 case LTTNG_STREAM_DATA:
3647 /* Copy received lttcomm socket */
3648 lttcomm_copy_sock(&relayd->data_sock.sock, &relayd_sock->sock);
3649 ret = lttcomm_create_sock(&relayd->data_sock.sock);
3650 /* Handle create_sock error. */
3651 if (ret < 0) {
3652 ret_code = LTTCOMM_CONSUMERD_ENOMEM;
3653 goto error;
3654 }
3655 /*
3656 * Close the socket created internally by
3657 * lttcomm_create_sock, so we can replace it by the one
3658 * received from sessiond.
3659 */
3660 if (close(relayd->data_sock.sock.fd)) {
3661 PERROR("close");
3662 }
3663
3664 /* Assign new file descriptor */
3665 relayd->data_sock.sock.fd = fd;
3666 /* Assign version values. */
3667 relayd->data_sock.major = relayd_sock->major;
3668 relayd->data_sock.minor = relayd_sock->minor;
3669 break;
3670 default:
3671 ERR("Unknown relayd socket type (%d)", sock_type);
3672 ret_code = LTTCOMM_CONSUMERD_FATAL;
3673 goto error;
3674 }
3675
3676 DBG("Consumer %s socket created successfully with net idx %" PRIu64 " (fd: %d)",
3677 sock_type == LTTNG_STREAM_CONTROL ? "control" : "data",
3678 relayd->net_seq_idx, fd);
3679 /*
3680 * We gave the ownership of the fd to the relayd structure. Set the
3681 * fd to -1 so we don't call close() on it in the error path below.
3682 */
3683 fd = -1;
3684
3685 /* We successfully added the socket. Send status back. */
3686 ret = consumer_send_status_msg(sock, ret_code);
3687 if (ret < 0) {
3688 /* Somehow, the session daemon is not responding anymore. */
3689 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_FATAL);
3690 goto error_nosignal;
3691 }
3692
3693 /*
3694 * Add relayd socket pair to consumer data hashtable. If object already
3695 * exists or on error, the function gracefully returns.
3696 */
3697 relayd->ctx = ctx;
3698 add_relayd(relayd);
3699
3700 /* All good! */
3701 return;
3702
3703 error:
3704 if (consumer_send_status_msg(sock, ret_code) < 0) {
3705 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_FATAL);
3706 }
3707
3708 error_nosignal:
3709 /* Close received socket if valid. */
3710 if (fd >= 0) {
3711 if (close(fd)) {
3712 PERROR("close received socket");
3713 }
3714 }
3715
3716 if (relayd_created) {
3717 free(relayd);
3718 }
3719 }
3720
3721 /*
3722 * Search for a relayd associated to the session id and return the reference.
3723 *
3724 * A rcu read side lock MUST be acquire before calling this function and locked
3725 * until the relayd object is no longer necessary.
3726 */
3727 static struct consumer_relayd_sock_pair *find_relayd_by_session_id(uint64_t id)
3728 {
3729 struct lttng_ht_iter iter;
3730 struct consumer_relayd_sock_pair *relayd = NULL;
3731
3732 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3733 cds_lfht_for_each_entry(consumer_data.relayd_ht->ht, &iter.iter, relayd,
3734 node.node) {
3735 /*
3736 * Check by sessiond id which is unique here where the relayd session
3737 * id might not be when having multiple relayd.
3738 */
3739 if (relayd->sessiond_session_id == id) {
3740 /* Found the relayd. There can be only one per id. */
3741 goto found;
3742 }
3743 }
3744
3745 return NULL;
3746
3747 found:
3748 return relayd;
3749 }
3750
3751 /*
3752 * Check if for a given session id there is still data needed to be extract
3753 * from the buffers.