cb05a1eb411c9f522b37efaf8600ecdca930352f
[lttng-tools.git] / src / common / consumer / consumer.c
1 /*
2 * Copyright (C) 2011 - Julien Desfossez <julien.desfossez@polymtl.ca>
3 * Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
4 * 2012 - David Goulet <dgoulet@efficios.com>
5 *
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License, version 2 only,
8 * as published by the Free Software Foundation.
9 *
10 * This program is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
13 * more details.
14 *
15 * You should have received a copy of the GNU General Public License along
16 * with this program; if not, write to the Free Software Foundation, Inc.,
17 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
18 */
19
20 #define _LGPL_SOURCE
21 #include <assert.h>
22 #include <poll.h>
23 #include <pthread.h>
24 #include <stdlib.h>
25 #include <string.h>
26 #include <sys/mman.h>
27 #include <sys/socket.h>
28 #include <sys/types.h>
29 #include <unistd.h>
30 #include <inttypes.h>
31 #include <signal.h>
32
33 #include <bin/lttng-consumerd/health-consumerd.h>
34 #include <common/common.h>
35 #include <common/utils.h>
36 #include <common/compat/poll.h>
37 #include <common/compat/endian.h>
38 #include <common/index/index.h>
39 #include <common/kernel-ctl/kernel-ctl.h>
40 #include <common/sessiond-comm/relayd.h>
41 #include <common/sessiond-comm/sessiond-comm.h>
42 #include <common/kernel-consumer/kernel-consumer.h>
43 #include <common/relayd/relayd.h>
44 #include <common/ust-consumer/ust-consumer.h>
45 #include <common/consumer/consumer-timer.h>
46 #include <common/consumer/consumer.h>
47 #include <common/consumer/consumer-stream.h>
48 #include <common/consumer/consumer-testpoint.h>
49 #include <common/align.h>
50
51 struct lttng_consumer_global_data consumer_data = {
52 .stream_count = 0,
53 .need_update = 1,
54 .type = LTTNG_CONSUMER_UNKNOWN,
55 };
56
57 enum consumer_channel_action {
58 CONSUMER_CHANNEL_ADD,
59 CONSUMER_CHANNEL_DEL,
60 CONSUMER_CHANNEL_QUIT,
61 };
62
63 struct consumer_channel_msg {
64 enum consumer_channel_action action;
65 struct lttng_consumer_channel *chan; /* add */
66 uint64_t key; /* del */
67 };
68
69 /*
70 * Flag to inform the polling thread to quit when all fd hung up. Updated by
71 * the consumer_thread_receive_fds when it notices that all fds has hung up.
72 * Also updated by the signal handler (consumer_should_exit()). Read by the
73 * polling threads.
74 */
75 volatile int consumer_quit;
76
77 /*
78 * Global hash table containing respectively metadata and data streams. The
79 * stream element in this ht should only be updated by the metadata poll thread
80 * for the metadata and the data poll thread for the data.
81 */
82 static struct lttng_ht *metadata_ht;
83 static struct lttng_ht *data_ht;
84
85 /*
86 * Notify a thread lttng pipe to poll back again. This usually means that some
87 * global state has changed so we just send back the thread in a poll wait
88 * call.
89 */
90 static void notify_thread_lttng_pipe(struct lttng_pipe *pipe)
91 {
92 struct lttng_consumer_stream *null_stream = NULL;
93
94 assert(pipe);
95
96 (void) lttng_pipe_write(pipe, &null_stream, sizeof(null_stream));
97 }
98
99 static void notify_health_quit_pipe(int *pipe)
100 {
101 ssize_t ret;
102
103 ret = lttng_write(pipe[1], "4", 1);
104 if (ret < 1) {
105 PERROR("write consumer health quit");
106 }
107 }
108
109 static void notify_channel_pipe(struct lttng_consumer_local_data *ctx,
110 struct lttng_consumer_channel *chan,
111 uint64_t key,
112 enum consumer_channel_action action)
113 {
114 struct consumer_channel_msg msg;
115 ssize_t ret;
116
117 memset(&msg, 0, sizeof(msg));
118
119 msg.action = action;
120 msg.chan = chan;
121 msg.key = key;
122 ret = lttng_write(ctx->consumer_channel_pipe[1], &msg, sizeof(msg));
123 if (ret < sizeof(msg)) {
124 PERROR("notify_channel_pipe write error");
125 }
126 }
127
128 void notify_thread_del_channel(struct lttng_consumer_local_data *ctx,
129 uint64_t key)
130 {
131 notify_channel_pipe(ctx, NULL, key, CONSUMER_CHANNEL_DEL);
132 }
133
134 static int read_channel_pipe(struct lttng_consumer_local_data *ctx,
135 struct lttng_consumer_channel **chan,
136 uint64_t *key,
137 enum consumer_channel_action *action)
138 {
139 struct consumer_channel_msg msg;
140 ssize_t ret;
141
142 ret = lttng_read(ctx->consumer_channel_pipe[0], &msg, sizeof(msg));
143 if (ret < sizeof(msg)) {
144 ret = -1;
145 goto error;
146 }
147 *action = msg.action;
148 *chan = msg.chan;
149 *key = msg.key;
150 error:
151 return (int) ret;
152 }
153
154 /*
155 * Cleanup the stream list of a channel. Those streams are not yet globally
156 * visible
157 */
158 static void clean_channel_stream_list(struct lttng_consumer_channel *channel)
159 {
160 struct lttng_consumer_stream *stream, *stmp;
161
162 assert(channel);
163
164 /* Delete streams that might have been left in the stream list. */
165 cds_list_for_each_entry_safe(stream, stmp, &channel->streams.head,
166 send_node) {
167 cds_list_del(&stream->send_node);
168 /*
169 * Once a stream is added to this list, the buffers were created so we
170 * have a guarantee that this call will succeed. Setting the monitor
171 * mode to 0 so we don't lock nor try to delete the stream from the
172 * global hash table.
173 */
174 stream->monitor = 0;
175 consumer_stream_destroy(stream, NULL);
176 }
177 }
178
179 /*
180 * Find a stream. The consumer_data.lock must be locked during this
181 * call.
182 */
183 static struct lttng_consumer_stream *find_stream(uint64_t key,
184 struct lttng_ht *ht)
185 {
186 struct lttng_ht_iter iter;
187 struct lttng_ht_node_u64 *node;
188 struct lttng_consumer_stream *stream = NULL;
189
190 assert(ht);
191
192 /* -1ULL keys are lookup failures */
193 if (key == (uint64_t) -1ULL) {
194 return NULL;
195 }
196
197 rcu_read_lock();
198
199 lttng_ht_lookup(ht, &key, &iter);
200 node = lttng_ht_iter_get_node_u64(&iter);
201 if (node != NULL) {
202 stream = caa_container_of(node, struct lttng_consumer_stream, node);
203 }
204
205 rcu_read_unlock();
206
207 return stream;
208 }
209
210 static void steal_stream_key(uint64_t key, struct lttng_ht *ht)
211 {
212 struct lttng_consumer_stream *stream;
213
214 rcu_read_lock();
215 stream = find_stream(key, ht);
216 if (stream) {
217 stream->key = (uint64_t) -1ULL;
218 /*
219 * We don't want the lookup to match, but we still need
220 * to iterate on this stream when iterating over the hash table. Just
221 * change the node key.
222 */
223 stream->node.key = (uint64_t) -1ULL;
224 }
225 rcu_read_unlock();
226 }
227
228 /*
229 * Return a channel object for the given key.
230 *
231 * RCU read side lock MUST be acquired before calling this function and
232 * protects the channel ptr.
233 */
234 struct lttng_consumer_channel *consumer_find_channel(uint64_t key)
235 {
236 struct lttng_ht_iter iter;
237 struct lttng_ht_node_u64 *node;
238 struct lttng_consumer_channel *channel = NULL;
239
240 /* -1ULL keys are lookup failures */
241 if (key == (uint64_t) -1ULL) {
242 return NULL;
243 }
244
245 lttng_ht_lookup(consumer_data.channel_ht, &key, &iter);
246 node = lttng_ht_iter_get_node_u64(&iter);
247 if (node != NULL) {
248 channel = caa_container_of(node, struct lttng_consumer_channel, node);
249 }
250
251 return channel;
252 }
253
254 /*
255 * There is a possibility that the consumer does not have enough time between
256 * the close of the channel on the session daemon and the cleanup in here thus
257 * once we have a channel add with an existing key, we know for sure that this
258 * channel will eventually get cleaned up by all streams being closed.
259 *
260 * This function just nullifies the already existing channel key.
261 */
262 static void steal_channel_key(uint64_t key)
263 {
264 struct lttng_consumer_channel *channel;
265
266 rcu_read_lock();
267 channel = consumer_find_channel(key);
268 if (channel) {
269 channel->key = (uint64_t) -1ULL;
270 /*
271 * We don't want the lookup to match, but we still need to iterate on
272 * this channel when iterating over the hash table. Just change the
273 * node key.
274 */
275 channel->node.key = (uint64_t) -1ULL;
276 }
277 rcu_read_unlock();
278 }
279
280 static void free_channel_rcu(struct rcu_head *head)
281 {
282 struct lttng_ht_node_u64 *node =
283 caa_container_of(head, struct lttng_ht_node_u64, head);
284 struct lttng_consumer_channel *channel =
285 caa_container_of(node, struct lttng_consumer_channel, node);
286
287 switch (consumer_data.type) {
288 case LTTNG_CONSUMER_KERNEL:
289 break;
290 case LTTNG_CONSUMER32_UST:
291 case LTTNG_CONSUMER64_UST:
292 lttng_ustconsumer_free_channel(channel);
293 break;
294 default:
295 ERR("Unknown consumer_data type");
296 abort();
297 }
298 free(channel);
299 }
300
301 /*
302 * RCU protected relayd socket pair free.
303 */
304 static void free_relayd_rcu(struct rcu_head *head)
305 {
306 struct lttng_ht_node_u64 *node =
307 caa_container_of(head, struct lttng_ht_node_u64, head);
308 struct consumer_relayd_sock_pair *relayd =
309 caa_container_of(node, struct consumer_relayd_sock_pair, node);
310
311 /*
312 * Close all sockets. This is done in the call RCU since we don't want the
313 * socket fds to be reassigned thus potentially creating bad state of the
314 * relayd object.
315 *
316 * We do not have to lock the control socket mutex here since at this stage
317 * there is no one referencing to this relayd object.
318 */
319 (void) relayd_close(&relayd->control_sock);
320 (void) relayd_close(&relayd->data_sock);
321
322 free(relayd);
323 }
324
325 /*
326 * Destroy and free relayd socket pair object.
327 */
328 void consumer_destroy_relayd(struct consumer_relayd_sock_pair *relayd)
329 {
330 int ret;
331 struct lttng_ht_iter iter;
332
333 if (relayd == NULL) {
334 return;
335 }
336
337 DBG("Consumer destroy and close relayd socket pair");
338
339 iter.iter.node = &relayd->node.node;
340 ret = lttng_ht_del(consumer_data.relayd_ht, &iter);
341 if (ret != 0) {
342 /* We assume the relayd is being or is destroyed */
343 return;
344 }
345
346 /* RCU free() call */
347 call_rcu(&relayd->node.head, free_relayd_rcu);
348 }
349
350 /*
351 * Remove a channel from the global list protected by a mutex. This function is
352 * also responsible for freeing its data structures.
353 */
354 void consumer_del_channel(struct lttng_consumer_channel *channel)
355 {
356 int ret;
357 struct lttng_ht_iter iter;
358
359 DBG("Consumer delete channel key %" PRIu64, channel->key);
360
361 pthread_mutex_lock(&consumer_data.lock);
362 pthread_mutex_lock(&channel->lock);
363
364 /* Destroy streams that might have been left in the stream list. */
365 clean_channel_stream_list(channel);
366
367 if (channel->live_timer_enabled == 1) {
368 consumer_timer_live_stop(channel);
369 }
370
371 switch (consumer_data.type) {
372 case LTTNG_CONSUMER_KERNEL:
373 break;
374 case LTTNG_CONSUMER32_UST:
375 case LTTNG_CONSUMER64_UST:
376 lttng_ustconsumer_del_channel(channel);
377 break;
378 default:
379 ERR("Unknown consumer_data type");
380 assert(0);
381 goto end;
382 }
383
384 rcu_read_lock();
385 iter.iter.node = &channel->node.node;
386 ret = lttng_ht_del(consumer_data.channel_ht, &iter);
387 assert(!ret);
388 rcu_read_unlock();
389
390 call_rcu(&channel->node.head, free_channel_rcu);
391 end:
392 pthread_mutex_unlock(&channel->lock);
393 pthread_mutex_unlock(&consumer_data.lock);
394 }
395
396 /*
397 * Iterate over the relayd hash table and destroy each element. Finally,
398 * destroy the whole hash table.
399 */
400 static void cleanup_relayd_ht(void)
401 {
402 struct lttng_ht_iter iter;
403 struct consumer_relayd_sock_pair *relayd;
404
405 rcu_read_lock();
406
407 cds_lfht_for_each_entry(consumer_data.relayd_ht->ht, &iter.iter, relayd,
408 node.node) {
409 consumer_destroy_relayd(relayd);
410 }
411
412 rcu_read_unlock();
413
414 lttng_ht_destroy(consumer_data.relayd_ht);
415 }
416
417 /*
418 * Update the end point status of all streams having the given network sequence
419 * index (relayd index).
420 *
421 * It's atomically set without having the stream mutex locked which is fine
422 * because we handle the write/read race with a pipe wakeup for each thread.
423 */
424 static void update_endpoint_status_by_netidx(uint64_t net_seq_idx,
425 enum consumer_endpoint_status status)
426 {
427 struct lttng_ht_iter iter;
428 struct lttng_consumer_stream *stream;
429
430 DBG("Consumer set delete flag on stream by idx %" PRIu64, net_seq_idx);
431
432 rcu_read_lock();
433
434 /* Let's begin with metadata */
435 cds_lfht_for_each_entry(metadata_ht->ht, &iter.iter, stream, node.node) {
436 if (stream->net_seq_idx == net_seq_idx) {
437 uatomic_set(&stream->endpoint_status, status);
438 DBG("Delete flag set to metadata stream %d", stream->wait_fd);
439 }
440 }
441
442 /* Follow up by the data streams */
443 cds_lfht_for_each_entry(data_ht->ht, &iter.iter, stream, node.node) {
444 if (stream->net_seq_idx == net_seq_idx) {
445 uatomic_set(&stream->endpoint_status, status);
446 DBG("Delete flag set to data stream %d", stream->wait_fd);
447 }
448 }
449 rcu_read_unlock();
450 }
451
452 /*
453 * Cleanup a relayd object by flagging every associated streams for deletion,
454 * destroying the object meaning removing it from the relayd hash table,
455 * closing the sockets and freeing the memory in a RCU call.
456 *
457 * If a local data context is available, notify the threads that the streams'
458 * state have changed.
459 */
460 static void cleanup_relayd(struct consumer_relayd_sock_pair *relayd,
461 struct lttng_consumer_local_data *ctx)
462 {
463 uint64_t netidx;
464
465 assert(relayd);
466
467 DBG("Cleaning up relayd sockets");
468
469 /* Save the net sequence index before destroying the object */
470 netidx = relayd->net_seq_idx;
471
472 /*
473 * Delete the relayd from the relayd hash table, close the sockets and free
474 * the object in a RCU call.
475 */
476 consumer_destroy_relayd(relayd);
477
478 /* Set inactive endpoint to all streams */
479 update_endpoint_status_by_netidx(netidx, CONSUMER_ENDPOINT_INACTIVE);
480
481 /*
482 * With a local data context, notify the threads that the streams' state
483 * have changed. The write() action on the pipe acts as an "implicit"
484 * memory barrier ordering the updates of the end point status from the
485 * read of this status which happens AFTER receiving this notify.
486 */
487 if (ctx) {
488 notify_thread_lttng_pipe(ctx->consumer_data_pipe);
489 notify_thread_lttng_pipe(ctx->consumer_metadata_pipe);
490 }
491 }
492
493 /*
494 * Flag a relayd socket pair for destruction. Destroy it if the refcount
495 * reaches zero.
496 *
497 * RCU read side lock MUST be aquired before calling this function.
498 */
499 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair *relayd)
500 {
501 assert(relayd);
502
503 /* Set destroy flag for this object */
504 uatomic_set(&relayd->destroy_flag, 1);
505
506 /* Destroy the relayd if refcount is 0 */
507 if (uatomic_read(&relayd->refcount) == 0) {
508 consumer_destroy_relayd(relayd);
509 }
510 }
511
512 /*
513 * Completly destroy stream from every visiable data structure and the given
514 * hash table if one.
515 *
516 * One this call returns, the stream object is not longer usable nor visible.
517 */
518 void consumer_del_stream(struct lttng_consumer_stream *stream,
519 struct lttng_ht *ht)
520 {
521 consumer_stream_destroy(stream, ht);
522 }
523
524 /*
525 * XXX naming of del vs destroy is all mixed up.
526 */
527 void consumer_del_stream_for_data(struct lttng_consumer_stream *stream)
528 {
529 consumer_stream_destroy(stream, data_ht);
530 }
531
532 void consumer_del_stream_for_metadata(struct lttng_consumer_stream *stream)
533 {
534 consumer_stream_destroy(stream, metadata_ht);
535 }
536
537 struct lttng_consumer_stream *consumer_allocate_stream(uint64_t channel_key,
538 uint64_t stream_key,
539 enum lttng_consumer_stream_state state,
540 const char *channel_name,
541 uid_t uid,
542 gid_t gid,
543 uint64_t relayd_id,
544 uint64_t session_id,
545 int cpu,
546 int *alloc_ret,
547 enum consumer_channel_type type,
548 unsigned int monitor)
549 {
550 int ret;
551 struct lttng_consumer_stream *stream;
552
553 stream = zmalloc(sizeof(*stream));
554 if (stream == NULL) {
555 PERROR("malloc struct lttng_consumer_stream");
556 ret = -ENOMEM;
557 goto end;
558 }
559
560 rcu_read_lock();
561
562 stream->key = stream_key;
563 stream->out_fd = -1;
564 stream->out_fd_offset = 0;
565 stream->output_written = 0;
566 stream->state = state;
567 stream->uid = uid;
568 stream->gid = gid;
569 stream->net_seq_idx = relayd_id;
570 stream->session_id = session_id;
571 stream->monitor = monitor;
572 stream->endpoint_status = CONSUMER_ENDPOINT_ACTIVE;
573 stream->index_fd = -1;
574 stream->last_sequence_number = -1ULL;
575 pthread_mutex_init(&stream->lock, NULL);
576 pthread_mutex_init(&stream->metadata_timer_lock, NULL);
577
578 /* If channel is the metadata, flag this stream as metadata. */
579 if (type == CONSUMER_CHANNEL_TYPE_METADATA) {
580 stream->metadata_flag = 1;
581 /* Metadata is flat out. */
582 strncpy(stream->name, DEFAULT_METADATA_NAME, sizeof(stream->name));
583 /* Live rendez-vous point. */
584 pthread_cond_init(&stream->metadata_rdv, NULL);
585 pthread_mutex_init(&stream->metadata_rdv_lock, NULL);
586 } else {
587 /* Format stream name to <channel_name>_<cpu_number> */
588 ret = snprintf(stream->name, sizeof(stream->name), "%s_%d",
589 channel_name, cpu);
590 if (ret < 0) {
591 PERROR("snprintf stream name");
592 goto error;
593 }
594 }
595
596 /* Key is always the wait_fd for streams. */
597 lttng_ht_node_init_u64(&stream->node, stream->key);
598
599 /* Init node per channel id key */
600 lttng_ht_node_init_u64(&stream->node_channel_id, channel_key);
601
602 /* Init session id node with the stream session id */
603 lttng_ht_node_init_u64(&stream->node_session_id, stream->session_id);
604
605 DBG3("Allocated stream %s (key %" PRIu64 ", chan_key %" PRIu64
606 " relayd_id %" PRIu64 ", session_id %" PRIu64,
607 stream->name, stream->key, channel_key,
608 stream->net_seq_idx, stream->session_id);
609
610 rcu_read_unlock();
611 return stream;
612
613 error:
614 rcu_read_unlock();
615 free(stream);
616 end:
617 if (alloc_ret) {
618 *alloc_ret = ret;
619 }
620 return NULL;
621 }
622
623 /*
624 * Add a stream to the global list protected by a mutex.
625 */
626 int consumer_add_data_stream(struct lttng_consumer_stream *stream)
627 {
628 struct lttng_ht *ht = data_ht;
629 int ret = 0;
630
631 assert(stream);
632 assert(ht);
633
634 DBG3("Adding consumer stream %" PRIu64, stream->key);
635
636 pthread_mutex_lock(&consumer_data.lock);
637 pthread_mutex_lock(&stream->chan->lock);
638 pthread_mutex_lock(&stream->chan->timer_lock);
639 pthread_mutex_lock(&stream->lock);
640 rcu_read_lock();
641
642 /* Steal stream identifier to avoid having streams with the same key */
643 steal_stream_key(stream->key, ht);
644
645 lttng_ht_add_unique_u64(ht, &stream->node);
646
647 lttng_ht_add_u64(consumer_data.stream_per_chan_id_ht,
648 &stream->node_channel_id);
649
650 /*
651 * Add stream to the stream_list_ht of the consumer data. No need to steal
652 * the key since the HT does not use it and we allow to add redundant keys
653 * into this table.
654 */
655 lttng_ht_add_u64(consumer_data.stream_list_ht, &stream->node_session_id);
656
657 /*
658 * When nb_init_stream_left reaches 0, we don't need to trigger any action
659 * in terms of destroying the associated channel, because the action that
660 * causes the count to become 0 also causes a stream to be added. The
661 * channel deletion will thus be triggered by the following removal of this
662 * stream.
663 */
664 if (uatomic_read(&stream->chan->nb_init_stream_left) > 0) {
665 /* Increment refcount before decrementing nb_init_stream_left */
666 cmm_smp_wmb();
667 uatomic_dec(&stream->chan->nb_init_stream_left);
668 }
669
670 /* Update consumer data once the node is inserted. */
671 consumer_data.stream_count++;
672 consumer_data.need_update = 1;
673
674 rcu_read_unlock();
675 pthread_mutex_unlock(&stream->lock);
676 pthread_mutex_unlock(&stream->chan->timer_lock);
677 pthread_mutex_unlock(&stream->chan->lock);
678 pthread_mutex_unlock(&consumer_data.lock);
679
680 return ret;
681 }
682
683 void consumer_del_data_stream(struct lttng_consumer_stream *stream)
684 {
685 consumer_del_stream(stream, data_ht);
686 }
687
688 /*
689 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
690 * be acquired before calling this.
691 */
692 static int add_relayd(struct consumer_relayd_sock_pair *relayd)
693 {
694 int ret = 0;
695 struct lttng_ht_node_u64 *node;
696 struct lttng_ht_iter iter;
697
698 assert(relayd);
699
700 lttng_ht_lookup(consumer_data.relayd_ht,
701 &relayd->net_seq_idx, &iter);
702 node = lttng_ht_iter_get_node_u64(&iter);
703 if (node != NULL) {
704 goto end;
705 }
706 lttng_ht_add_unique_u64(consumer_data.relayd_ht, &relayd->node);
707
708 end:
709 return ret;
710 }
711
712 /*
713 * Allocate and return a consumer relayd socket.
714 */
715 static struct consumer_relayd_sock_pair *consumer_allocate_relayd_sock_pair(
716 uint64_t net_seq_idx)
717 {
718 struct consumer_relayd_sock_pair *obj = NULL;
719
720 /* net sequence index of -1 is a failure */
721 if (net_seq_idx == (uint64_t) -1ULL) {
722 goto error;
723 }
724
725 obj = zmalloc(sizeof(struct consumer_relayd_sock_pair));
726 if (obj == NULL) {
727 PERROR("zmalloc relayd sock");
728 goto error;
729 }
730
731 obj->net_seq_idx = net_seq_idx;
732 obj->refcount = 0;
733 obj->destroy_flag = 0;
734 obj->control_sock.sock.fd = -1;
735 obj->data_sock.sock.fd = -1;
736 lttng_ht_node_init_u64(&obj->node, obj->net_seq_idx);
737 pthread_mutex_init(&obj->ctrl_sock_mutex, NULL);
738
739 error:
740 return obj;
741 }
742
743 /*
744 * Find a relayd socket pair in the global consumer data.
745 *
746 * Return the object if found else NULL.
747 * RCU read-side lock must be held across this call and while using the
748 * returned object.
749 */
750 struct consumer_relayd_sock_pair *consumer_find_relayd(uint64_t key)
751 {
752 struct lttng_ht_iter iter;
753 struct lttng_ht_node_u64 *node;
754 struct consumer_relayd_sock_pair *relayd = NULL;
755
756 /* Negative keys are lookup failures */
757 if (key == (uint64_t) -1ULL) {
758 goto error;
759 }
760
761 lttng_ht_lookup(consumer_data.relayd_ht, &key,
762 &iter);
763 node = lttng_ht_iter_get_node_u64(&iter);
764 if (node != NULL) {
765 relayd = caa_container_of(node, struct consumer_relayd_sock_pair, node);
766 }
767
768 error:
769 return relayd;
770 }
771
772 /*
773 * Find a relayd and send the stream
774 *
775 * Returns 0 on success, < 0 on error
776 */
777 int consumer_send_relayd_stream(struct lttng_consumer_stream *stream,
778 char *path)
779 {
780 int ret = 0;
781 struct consumer_relayd_sock_pair *relayd;
782
783 assert(stream);
784 assert(stream->net_seq_idx != -1ULL);
785 assert(path);
786
787 /* The stream is not metadata. Get relayd reference if exists. */
788 rcu_read_lock();
789 relayd = consumer_find_relayd(stream->net_seq_idx);
790 if (relayd != NULL) {
791 /* Add stream on the relayd */
792 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
793 ret = relayd_add_stream(&relayd->control_sock, stream->name,
794 path, &stream->relayd_stream_id,
795 stream->chan->tracefile_size, stream->chan->tracefile_count);
796 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
797 if (ret < 0) {
798 goto end;
799 }
800
801 uatomic_inc(&relayd->refcount);
802 stream->sent_to_relayd = 1;
803 } else {
804 ERR("Stream %" PRIu64 " relayd ID %" PRIu64 " unknown. Can't send it.",
805 stream->key, stream->net_seq_idx);
806 ret = -1;
807 goto end;
808 }
809
810 DBG("Stream %s with key %" PRIu64 " sent to relayd id %" PRIu64,
811 stream->name, stream->key, stream->net_seq_idx);
812
813 end:
814 rcu_read_unlock();
815 return ret;
816 }
817
818 /*
819 * Find a relayd and send the streams sent message
820 *
821 * Returns 0 on success, < 0 on error
822 */
823 int consumer_send_relayd_streams_sent(uint64_t net_seq_idx)
824 {
825 int ret = 0;
826 struct consumer_relayd_sock_pair *relayd;
827
828 assert(net_seq_idx != -1ULL);
829
830 /* The stream is not metadata. Get relayd reference if exists. */
831 rcu_read_lock();
832 relayd = consumer_find_relayd(net_seq_idx);
833 if (relayd != NULL) {
834 /* Add stream on the relayd */
835 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
836 ret = relayd_streams_sent(&relayd->control_sock);
837 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
838 if (ret < 0) {
839 goto end;
840 }
841 } else {
842 ERR("Relayd ID %" PRIu64 " unknown. Can't send streams_sent.",
843 net_seq_idx);
844 ret = -1;
845 goto end;
846 }
847
848 ret = 0;
849 DBG("All streams sent relayd id %" PRIu64, net_seq_idx);
850
851 end:
852 rcu_read_unlock();
853 return ret;
854 }
855
856 /*
857 * Find a relayd and close the stream
858 */
859 void close_relayd_stream(struct lttng_consumer_stream *stream)
860 {
861 struct consumer_relayd_sock_pair *relayd;
862
863 /* The stream is not metadata. Get relayd reference if exists. */
864 rcu_read_lock();
865 relayd = consumer_find_relayd(stream->net_seq_idx);
866 if (relayd) {
867 consumer_stream_relayd_close(stream, relayd);
868 }
869 rcu_read_unlock();
870 }
871
872 /*
873 * Handle stream for relayd transmission if the stream applies for network
874 * streaming where the net sequence index is set.
875 *
876 * Return destination file descriptor or negative value on error.
877 */
878 static int write_relayd_stream_header(struct lttng_consumer_stream *stream,
879 size_t data_size, unsigned long padding,
880 struct consumer_relayd_sock_pair *relayd)
881 {
882 int outfd = -1, ret;
883 struct lttcomm_relayd_data_hdr data_hdr;
884
885 /* Safety net */
886 assert(stream);
887 assert(relayd);
888
889 /* Reset data header */
890 memset(&data_hdr, 0, sizeof(data_hdr));
891
892 if (stream->metadata_flag) {
893 /* Caller MUST acquire the relayd control socket lock */
894 ret = relayd_send_metadata(&relayd->control_sock, data_size);
895 if (ret < 0) {
896 goto error;
897 }
898
899 /* Metadata are always sent on the control socket. */
900 outfd = relayd->control_sock.sock.fd;
901 } else {
902 /* Set header with stream information */
903 data_hdr.stream_id = htobe64(stream->relayd_stream_id);
904 data_hdr.data_size = htobe32(data_size);
905 data_hdr.padding_size = htobe32(padding);
906 /*
907 * Note that net_seq_num below is assigned with the *current* value of
908 * next_net_seq_num and only after that the next_net_seq_num will be
909 * increment. This is why when issuing a command on the relayd using
910 * this next value, 1 should always be substracted in order to compare
911 * the last seen sequence number on the relayd side to the last sent.
912 */
913 data_hdr.net_seq_num = htobe64(stream->next_net_seq_num);
914 /* Other fields are zeroed previously */
915
916 ret = relayd_send_data_hdr(&relayd->data_sock, &data_hdr,
917 sizeof(data_hdr));
918 if (ret < 0) {
919 goto error;
920 }
921
922 ++stream->next_net_seq_num;
923
924 /* Set to go on data socket */
925 outfd = relayd->data_sock.sock.fd;
926 }
927
928 error:
929 return outfd;
930 }
931
932 /*
933 * Allocate and return a new lttng_consumer_channel object using the given key
934 * to initialize the hash table node.
935 *
936 * On error, return NULL.
937 */
938 struct lttng_consumer_channel *consumer_allocate_channel(uint64_t key,
939 uint64_t session_id,
940 const char *pathname,
941 const char *name,
942 uid_t uid,
943 gid_t gid,
944 uint64_t relayd_id,
945 enum lttng_event_output output,
946 uint64_t tracefile_size,
947 uint64_t tracefile_count,
948 uint64_t session_id_per_pid,
949 unsigned int monitor,
950 unsigned int live_timer_interval,
951 const char *root_shm_path,
952 const char *shm_path)
953 {
954 struct lttng_consumer_channel *channel;
955
956 channel = zmalloc(sizeof(*channel));
957 if (channel == NULL) {
958 PERROR("malloc struct lttng_consumer_channel");
959 goto end;
960 }
961
962 channel->key = key;
963 channel->refcount = 0;
964 channel->session_id = session_id;
965 channel->session_id_per_pid = session_id_per_pid;
966 channel->uid = uid;
967 channel->gid = gid;
968 channel->relayd_id = relayd_id;
969 channel->tracefile_size = tracefile_size;
970 channel->tracefile_count = tracefile_count;
971 channel->monitor = monitor;
972 channel->live_timer_interval = live_timer_interval;
973 pthread_mutex_init(&channel->lock, NULL);
974 pthread_mutex_init(&channel->timer_lock, NULL);
975
976 switch (output) {
977 case LTTNG_EVENT_SPLICE:
978 channel->output = CONSUMER_CHANNEL_SPLICE;
979 break;
980 case LTTNG_EVENT_MMAP:
981 channel->output = CONSUMER_CHANNEL_MMAP;
982 break;
983 default:
984 assert(0);
985 free(channel);
986 channel = NULL;
987 goto end;
988 }
989
990 /*
991 * In monitor mode, the streams associated with the channel will be put in
992 * a special list ONLY owned by this channel. So, the refcount is set to 1
993 * here meaning that the channel itself has streams that are referenced.
994 *
995 * On a channel deletion, once the channel is no longer visible, the
996 * refcount is decremented and checked for a zero value to delete it. With
997 * streams in no monitor mode, it will now be safe to destroy the channel.
998 */
999 if (!channel->monitor) {
1000 channel->refcount = 1;
1001 }
1002
1003 strncpy(channel->pathname, pathname, sizeof(channel->pathname));
1004 channel->pathname[sizeof(channel->pathname) - 1] = '\0';
1005
1006 strncpy(channel->name, name, sizeof(channel->name));
1007 channel->name[sizeof(channel->name) - 1] = '\0';
1008
1009 if (root_shm_path) {
1010 strncpy(channel->root_shm_path, root_shm_path, sizeof(channel->root_shm_path));
1011 channel->root_shm_path[sizeof(channel->root_shm_path) - 1] = '\0';
1012 }
1013 if (shm_path) {
1014 strncpy(channel->shm_path, shm_path, sizeof(channel->shm_path));
1015 channel->shm_path[sizeof(channel->shm_path) - 1] = '\0';
1016 }
1017
1018 lttng_ht_node_init_u64(&channel->node, channel->key);
1019
1020 channel->wait_fd = -1;
1021
1022 CDS_INIT_LIST_HEAD(&channel->streams.head);
1023
1024 DBG("Allocated channel (key %" PRIu64 ")", channel->key)
1025
1026 end:
1027 return channel;
1028 }
1029
1030 /*
1031 * Add a channel to the global list protected by a mutex.
1032 *
1033 * Always return 0 indicating success.
1034 */
1035 int consumer_add_channel(struct lttng_consumer_channel *channel,
1036 struct lttng_consumer_local_data *ctx)
1037 {
1038 pthread_mutex_lock(&consumer_data.lock);
1039 pthread_mutex_lock(&channel->lock);
1040 pthread_mutex_lock(&channel->timer_lock);
1041
1042 /*
1043 * This gives us a guarantee that the channel we are about to add to the
1044 * channel hash table will be unique. See this function comment on the why
1045 * we need to steel the channel key at this stage.
1046 */
1047 steal_channel_key(channel->key);
1048
1049 rcu_read_lock();
1050 lttng_ht_add_unique_u64(consumer_data.channel_ht, &channel->node);
1051 rcu_read_unlock();
1052
1053 pthread_mutex_unlock(&channel->timer_lock);
1054 pthread_mutex_unlock(&channel->lock);
1055 pthread_mutex_unlock(&consumer_data.lock);
1056
1057 if (channel->wait_fd != -1 && channel->type == CONSUMER_CHANNEL_TYPE_DATA) {
1058 notify_channel_pipe(ctx, channel, -1, CONSUMER_CHANNEL_ADD);
1059 }
1060
1061 return 0;
1062 }
1063
1064 /*
1065 * Allocate the pollfd structure and the local view of the out fds to avoid
1066 * doing a lookup in the linked list and concurrency issues when writing is
1067 * needed. Called with consumer_data.lock held.
1068 *
1069 * Returns the number of fds in the structures.
1070 */
1071 static int update_poll_array(struct lttng_consumer_local_data *ctx,
1072 struct pollfd **pollfd, struct lttng_consumer_stream **local_stream,
1073 struct lttng_ht *ht)
1074 {
1075 int i = 0;
1076 struct lttng_ht_iter iter;
1077 struct lttng_consumer_stream *stream;
1078
1079 assert(ctx);
1080 assert(ht);
1081 assert(pollfd);
1082 assert(local_stream);
1083
1084 DBG("Updating poll fd array");
1085 rcu_read_lock();
1086 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
1087 /*
1088 * Only active streams with an active end point can be added to the
1089 * poll set and local stream storage of the thread.
1090 *
1091 * There is a potential race here for endpoint_status to be updated
1092 * just after the check. However, this is OK since the stream(s) will
1093 * be deleted once the thread is notified that the end point state has
1094 * changed where this function will be called back again.
1095 */
1096 if (stream->state != LTTNG_CONSUMER_ACTIVE_STREAM ||
1097 stream->endpoint_status == CONSUMER_ENDPOINT_INACTIVE) {
1098 continue;
1099 }
1100 /*
1101 * This clobbers way too much the debug output. Uncomment that if you
1102 * need it for debugging purposes.
1103 *
1104 * DBG("Active FD %d", stream->wait_fd);
1105 */
1106 (*pollfd)[i].fd = stream->wait_fd;
1107 (*pollfd)[i].events = POLLIN | POLLPRI;
1108 local_stream[i] = stream;
1109 i++;
1110 }
1111 rcu_read_unlock();
1112
1113 /*
1114 * Insert the consumer_data_pipe at the end of the array and don't
1115 * increment i so nb_fd is the number of real FD.
1116 */
1117 (*pollfd)[i].fd = lttng_pipe_get_readfd(ctx->consumer_data_pipe);
1118 (*pollfd)[i].events = POLLIN | POLLPRI;
1119
1120 (*pollfd)[i + 1].fd = lttng_pipe_get_readfd(ctx->consumer_wakeup_pipe);
1121 (*pollfd)[i + 1].events = POLLIN | POLLPRI;
1122 return i;
1123 }
1124
1125 /*
1126 * Poll on the should_quit pipe and the command socket return -1 on
1127 * error, 1 if should exit, 0 if data is available on the command socket
1128 */
1129 int lttng_consumer_poll_socket(struct pollfd *consumer_sockpoll)
1130 {
1131 int num_rdy;
1132
1133 restart:
1134 num_rdy = poll(consumer_sockpoll, 2, -1);
1135 if (num_rdy == -1) {
1136 /*
1137 * Restart interrupted system call.
1138 */
1139 if (errno == EINTR) {
1140 goto restart;
1141 }
1142 PERROR("Poll error");
1143 return -1;
1144 }
1145 if (consumer_sockpoll[0].revents & (POLLIN | POLLPRI)) {
1146 DBG("consumer_should_quit wake up");
1147 return 1;
1148 }
1149 return 0;
1150 }
1151
1152 /*
1153 * Set the error socket.
1154 */
1155 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data *ctx,
1156 int sock)
1157 {
1158 ctx->consumer_error_socket = sock;
1159 }
1160
1161 /*
1162 * Set the command socket path.
1163 */
1164 void lttng_consumer_set_command_sock_path(
1165 struct lttng_consumer_local_data *ctx, char *sock)
1166 {
1167 ctx->consumer_command_sock_path = sock;
1168 }
1169
1170 /*
1171 * Send return code to the session daemon.
1172 * If the socket is not defined, we return 0, it is not a fatal error
1173 */
1174 int lttng_consumer_send_error(struct lttng_consumer_local_data *ctx, int cmd)
1175 {
1176 if (ctx->consumer_error_socket > 0) {
1177 return lttcomm_send_unix_sock(ctx->consumer_error_socket, &cmd,
1178 sizeof(enum lttcomm_sessiond_command));
1179 }
1180
1181 return 0;
1182 }
1183
1184 /*
1185 * Close all the tracefiles and stream fds and MUST be called when all
1186 * instances are destroyed i.e. when all threads were joined and are ended.
1187 */
1188 void lttng_consumer_cleanup(void)
1189 {
1190 struct lttng_ht_iter iter;
1191 struct lttng_consumer_channel *channel;
1192
1193 rcu_read_lock();
1194
1195 cds_lfht_for_each_entry(consumer_data.channel_ht->ht, &iter.iter, channel,
1196 node.node) {
1197 consumer_del_channel(channel);
1198 }
1199
1200 rcu_read_unlock();
1201
1202 lttng_ht_destroy(consumer_data.channel_ht);
1203
1204 cleanup_relayd_ht();
1205
1206 lttng_ht_destroy(consumer_data.stream_per_chan_id_ht);
1207
1208 /*
1209 * This HT contains streams that are freed by either the metadata thread or
1210 * the data thread so we do *nothing* on the hash table and simply destroy
1211 * it.
1212 */
1213 lttng_ht_destroy(consumer_data.stream_list_ht);
1214 }
1215
1216 /*
1217 * Called from signal handler.
1218 */
1219 void lttng_consumer_should_exit(struct lttng_consumer_local_data *ctx)
1220 {
1221 ssize_t ret;
1222
1223 consumer_quit = 1;
1224 ret = lttng_write(ctx->consumer_should_quit[1], "4", 1);
1225 if (ret < 1) {
1226 PERROR("write consumer quit");
1227 }
1228
1229 DBG("Consumer flag that it should quit");
1230 }
1231
1232 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream *stream,
1233 off_t orig_offset)
1234 {
1235 int outfd = stream->out_fd;
1236
1237 /*
1238 * This does a blocking write-and-wait on any page that belongs to the
1239 * subbuffer prior to the one we just wrote.
1240 * Don't care about error values, as these are just hints and ways to
1241 * limit the amount of page cache used.
1242 */
1243 if (orig_offset < stream->max_sb_size) {
1244 return;
1245 }
1246 lttng_sync_file_range(outfd, orig_offset - stream->max_sb_size,
1247 stream->max_sb_size,
1248 SYNC_FILE_RANGE_WAIT_BEFORE
1249 | SYNC_FILE_RANGE_WRITE
1250 | SYNC_FILE_RANGE_WAIT_AFTER);
1251 /*
1252 * Give hints to the kernel about how we access the file:
1253 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1254 * we write it.
1255 *
1256 * We need to call fadvise again after the file grows because the
1257 * kernel does not seem to apply fadvise to non-existing parts of the
1258 * file.
1259 *
1260 * Call fadvise _after_ having waited for the page writeback to
1261 * complete because the dirty page writeback semantic is not well
1262 * defined. So it can be expected to lead to lower throughput in
1263 * streaming.
1264 */
1265 posix_fadvise(outfd, orig_offset - stream->max_sb_size,
1266 stream->max_sb_size, POSIX_FADV_DONTNEED);
1267 }
1268
1269 /*
1270 * Initialise the necessary environnement :
1271 * - create a new context
1272 * - create the poll_pipe
1273 * - create the should_quit pipe (for signal handler)
1274 * - create the thread pipe (for splice)
1275 *
1276 * Takes a function pointer as argument, this function is called when data is
1277 * available on a buffer. This function is responsible to do the
1278 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1279 * buffer configuration and then kernctl_put_next_subbuf at the end.
1280 *
1281 * Returns a pointer to the new context or NULL on error.
1282 */
1283 struct lttng_consumer_local_data *lttng_consumer_create(
1284 enum lttng_consumer_type type,
1285 ssize_t (*buffer_ready)(struct lttng_consumer_stream *stream,
1286 struct lttng_consumer_local_data *ctx),
1287 int (*recv_channel)(struct lttng_consumer_channel *channel),
1288 int (*recv_stream)(struct lttng_consumer_stream *stream),
1289 int (*update_stream)(uint64_t stream_key, uint32_t state))
1290 {
1291 int ret;
1292 struct lttng_consumer_local_data *ctx;
1293
1294 assert(consumer_data.type == LTTNG_CONSUMER_UNKNOWN ||
1295 consumer_data.type == type);
1296 consumer_data.type = type;
1297
1298 ctx = zmalloc(sizeof(struct lttng_consumer_local_data));
1299 if (ctx == NULL) {
1300 PERROR("allocating context");
1301 goto error;
1302 }
1303
1304 ctx->consumer_error_socket = -1;
1305 ctx->consumer_metadata_socket = -1;
1306 pthread_mutex_init(&ctx->metadata_socket_lock, NULL);
1307 /* assign the callbacks */
1308 ctx->on_buffer_ready = buffer_ready;
1309 ctx->on_recv_channel = recv_channel;
1310 ctx->on_recv_stream = recv_stream;
1311 ctx->on_update_stream = update_stream;
1312
1313 ctx->consumer_data_pipe = lttng_pipe_open(0);
1314 if (!ctx->consumer_data_pipe) {
1315 goto error_poll_pipe;
1316 }
1317
1318 ctx->consumer_wakeup_pipe = lttng_pipe_open(0);
1319 if (!ctx->consumer_wakeup_pipe) {
1320 goto error_wakeup_pipe;
1321 }
1322
1323 ret = pipe(ctx->consumer_should_quit);
1324 if (ret < 0) {
1325 PERROR("Error creating recv pipe");
1326 goto error_quit_pipe;
1327 }
1328
1329 ret = pipe(ctx->consumer_channel_pipe);
1330 if (ret < 0) {
1331 PERROR("Error creating channel pipe");
1332 goto error_channel_pipe;
1333 }
1334
1335 ctx->consumer_metadata_pipe = lttng_pipe_open(0);
1336 if (!ctx->consumer_metadata_pipe) {
1337 goto error_metadata_pipe;
1338 }
1339
1340 return ctx;
1341
1342 error_metadata_pipe:
1343 utils_close_pipe(ctx->consumer_channel_pipe);
1344 error_channel_pipe:
1345 utils_close_pipe(ctx->consumer_should_quit);
1346 error_quit_pipe:
1347 lttng_pipe_destroy(ctx->consumer_wakeup_pipe);
1348 error_wakeup_pipe:
1349 lttng_pipe_destroy(ctx->consumer_data_pipe);
1350 error_poll_pipe:
1351 free(ctx);
1352 error:
1353 return NULL;
1354 }
1355
1356 /*
1357 * Iterate over all streams of the hashtable and free them properly.
1358 */
1359 static void destroy_data_stream_ht(struct lttng_ht *ht)
1360 {
1361 struct lttng_ht_iter iter;
1362 struct lttng_consumer_stream *stream;
1363
1364 if (ht == NULL) {
1365 return;
1366 }
1367
1368 rcu_read_lock();
1369 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
1370 /*
1371 * Ignore return value since we are currently cleaning up so any error
1372 * can't be handled.
1373 */
1374 (void) consumer_del_stream(stream, ht);
1375 }
1376 rcu_read_unlock();
1377
1378 lttng_ht_destroy(ht);
1379 }
1380
1381 /*
1382 * Iterate over all streams of the metadata hashtable and free them
1383 * properly.
1384 */
1385 static void destroy_metadata_stream_ht(struct lttng_ht *ht)
1386 {
1387 struct lttng_ht_iter iter;
1388 struct lttng_consumer_stream *stream;
1389
1390 if (ht == NULL) {
1391 return;
1392 }
1393
1394 rcu_read_lock();
1395 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
1396 /*
1397 * Ignore return value since we are currently cleaning up so any error
1398 * can't be handled.
1399 */
1400 (void) consumer_del_metadata_stream(stream, ht);
1401 }
1402 rcu_read_unlock();
1403
1404 lttng_ht_destroy(ht);
1405 }
1406
1407 /*
1408 * Close all fds associated with the instance and free the context.
1409 */
1410 void lttng_consumer_destroy(struct lttng_consumer_local_data *ctx)
1411 {
1412 int ret;
1413
1414 DBG("Consumer destroying it. Closing everything.");
1415
1416 if (!ctx) {
1417 return;
1418 }
1419
1420 destroy_data_stream_ht(data_ht);
1421 destroy_metadata_stream_ht(metadata_ht);
1422
1423 ret = close(ctx->consumer_error_socket);
1424 if (ret) {
1425 PERROR("close");
1426 }
1427 ret = close(ctx->consumer_metadata_socket);
1428 if (ret) {
1429 PERROR("close");
1430 }
1431 utils_close_pipe(ctx->consumer_channel_pipe);
1432 lttng_pipe_destroy(ctx->consumer_data_pipe);
1433 lttng_pipe_destroy(ctx->consumer_metadata_pipe);
1434 lttng_pipe_destroy(ctx->consumer_wakeup_pipe);
1435 utils_close_pipe(ctx->consumer_should_quit);
1436
1437 unlink(ctx->consumer_command_sock_path);
1438 free(ctx);
1439 }
1440
1441 /*
1442 * Write the metadata stream id on the specified file descriptor.
1443 */
1444 static int write_relayd_metadata_id(int fd,
1445 struct lttng_consumer_stream *stream,
1446 struct consumer_relayd_sock_pair *relayd, unsigned long padding)
1447 {
1448 ssize_t ret;
1449 struct lttcomm_relayd_metadata_payload hdr;
1450
1451 hdr.stream_id = htobe64(stream->relayd_stream_id);
1452 hdr.padding_size = htobe32(padding);
1453 ret = lttng_write(fd, (void *) &hdr, sizeof(hdr));
1454 if (ret < sizeof(hdr)) {
1455 /*
1456 * This error means that the fd's end is closed so ignore the PERROR
1457 * not to clubber the error output since this can happen in a normal
1458 * code path.
1459 */
1460 if (errno != EPIPE) {
1461 PERROR("write metadata stream id");
1462 }
1463 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno);
1464 /*
1465 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1466 * handle writting the missing part so report that as an error and
1467 * don't lie to the caller.
1468 */
1469 ret = -1;
1470 goto end;
1471 }
1472 DBG("Metadata stream id %" PRIu64 " with padding %lu written before data",
1473 stream->relayd_stream_id, padding);
1474
1475 end:
1476 return (int) ret;
1477 }
1478
1479 /*
1480 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1481 * core function for writing trace buffers to either the local filesystem or
1482 * the network.
1483 *
1484 * It must be called with the stream lock held.
1485 *
1486 * Careful review MUST be put if any changes occur!
1487 *
1488 * Returns the number of bytes written
1489 */
1490 ssize_t lttng_consumer_on_read_subbuffer_mmap(
1491 struct lttng_consumer_local_data *ctx,
1492 struct lttng_consumer_stream *stream, unsigned long len,
1493 unsigned long padding,
1494 struct ctf_packet_index *index)
1495 {
1496 unsigned long mmap_offset;
1497 void *mmap_base;
1498 ssize_t ret = 0;
1499 off_t orig_offset = stream->out_fd_offset;
1500 /* Default is on the disk */
1501 int outfd = stream->out_fd;
1502 struct consumer_relayd_sock_pair *relayd = NULL;
1503 unsigned int relayd_hang_up = 0;
1504
1505 /* RCU lock for the relayd pointer */
1506 rcu_read_lock();
1507
1508 /* Flag that the current stream if set for network streaming. */
1509 if (stream->net_seq_idx != (uint64_t) -1ULL) {
1510 relayd = consumer_find_relayd(stream->net_seq_idx);
1511 if (relayd == NULL) {
1512 ret = -EPIPE;
1513 goto end;
1514 }
1515 }
1516
1517 /* get the offset inside the fd to mmap */
1518 switch (consumer_data.type) {
1519 case LTTNG_CONSUMER_KERNEL:
1520 mmap_base = stream->mmap_base;
1521 ret = kernctl_get_mmap_read_offset(stream->wait_fd, &mmap_offset);
1522 if (ret < 0) {
1523 ret = -errno;
1524 PERROR("tracer ctl get_mmap_read_offset");
1525 goto end;
1526 }
1527 break;
1528 case LTTNG_CONSUMER32_UST:
1529 case LTTNG_CONSUMER64_UST:
1530 mmap_base = lttng_ustctl_get_mmap_base(stream);
1531 if (!mmap_base) {
1532 ERR("read mmap get mmap base for stream %s", stream->name);
1533 ret = -EPERM;
1534 goto end;
1535 }
1536 ret = lttng_ustctl_get_mmap_read_offset(stream, &mmap_offset);
1537 if (ret != 0) {
1538 PERROR("tracer ctl get_mmap_read_offset");
1539 ret = -EINVAL;
1540 goto end;
1541 }
1542 break;
1543 default:
1544 ERR("Unknown consumer_data type");
1545 assert(0);
1546 }
1547
1548 /* Handle stream on the relayd if the output is on the network */
1549 if (relayd) {
1550 unsigned long netlen = len;
1551
1552 /*
1553 * Lock the control socket for the complete duration of the function
1554 * since from this point on we will use the socket.
1555 */
1556 if (stream->metadata_flag) {
1557 /* Metadata requires the control socket. */
1558 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
1559 if (stream->reset_metadata_flag) {
1560 ret = relayd_reset_metadata(&relayd->control_sock,
1561 stream->relayd_stream_id,
1562 stream->metadata_version);
1563 if (ret < 0) {
1564 relayd_hang_up = 1;
1565 goto write_error;
1566 }
1567 stream->reset_metadata_flag = 0;
1568 }
1569 netlen += sizeof(struct lttcomm_relayd_metadata_payload);
1570 }
1571
1572 ret = write_relayd_stream_header(stream, netlen, padding, relayd);
1573 if (ret < 0) {
1574 relayd_hang_up = 1;
1575 goto write_error;
1576 }
1577 /* Use the returned socket. */
1578 outfd = ret;
1579
1580 /* Write metadata stream id before payload */
1581 if (stream->metadata_flag) {
1582 ret = write_relayd_metadata_id(outfd, stream, relayd, padding);
1583 if (ret < 0) {
1584 relayd_hang_up = 1;
1585 goto write_error;
1586 }
1587 }
1588 } else {
1589 /* No streaming, we have to set the len with the full padding */
1590 len += padding;
1591
1592 if (stream->metadata_flag && stream->reset_metadata_flag) {
1593 ret = utils_truncate_stream_file(stream->out_fd, 0);
1594 if (ret < 0) {
1595 ERR("Reset metadata file");
1596 goto end;
1597 }
1598 stream->reset_metadata_flag = 0;
1599 }
1600
1601 /*
1602 * Check if we need to change the tracefile before writing the packet.
1603 */
1604 if (stream->chan->tracefile_size > 0 &&
1605 (stream->tracefile_size_current + len) >
1606 stream->chan->tracefile_size) {
1607 ret = utils_rotate_stream_file(stream->chan->pathname,
1608 stream->name, stream->chan->tracefile_size,
1609 stream->chan->tracefile_count, stream->uid, stream->gid,
1610 stream->out_fd, &(stream->tracefile_count_current),
1611 &stream->out_fd);
1612 if (ret < 0) {
1613 ERR("Rotating output file");
1614 goto end;
1615 }
1616 outfd = stream->out_fd;
1617
1618 if (stream->index_fd >= 0) {
1619 ret = close(stream->index_fd);
1620 if (ret < 0) {
1621 PERROR("Closing index");
1622 goto end;
1623 }
1624 stream->index_fd = -1;
1625 ret = index_create_file(stream->chan->pathname,
1626 stream->name, stream->uid, stream->gid,
1627 stream->chan->tracefile_size,
1628 stream->tracefile_count_current);
1629 if (ret < 0) {
1630 goto end;
1631 }
1632 stream->index_fd = ret;
1633 }
1634
1635 /* Reset current size because we just perform a rotation. */
1636 stream->tracefile_size_current = 0;
1637 stream->out_fd_offset = 0;
1638 orig_offset = 0;
1639 }
1640 stream->tracefile_size_current += len;
1641 if (index) {
1642 index->offset = htobe64(stream->out_fd_offset);
1643 }
1644 }
1645
1646 /*
1647 * This call guarantee that len or less is returned. It's impossible to
1648 * receive a ret value that is bigger than len.
1649 */
1650 ret = lttng_write(outfd, mmap_base + mmap_offset, len);
1651 DBG("Consumer mmap write() ret %zd (len %lu)", ret, len);
1652 if (ret < 0 || ((size_t) ret != len)) {
1653 /*
1654 * Report error to caller if nothing was written else at least send the
1655 * amount written.
1656 */
1657 if (ret < 0) {
1658 ret = -errno;
1659 }
1660 relayd_hang_up = 1;
1661
1662 /* Socket operation failed. We consider the relayd dead */
1663 if (errno == EPIPE || errno == EINVAL || errno == EBADF) {
1664 /*
1665 * This is possible if the fd is closed on the other side
1666 * (outfd) or any write problem. It can be verbose a bit for a
1667 * normal execution if for instance the relayd is stopped
1668 * abruptly. This can happen so set this to a DBG statement.
1669 */
1670 DBG("Consumer mmap write detected relayd hang up");
1671 } else {
1672 /* Unhandled error, print it and stop function right now. */
1673 PERROR("Error in write mmap (ret %zd != len %lu)", ret, len);
1674 }
1675 goto write_error;
1676 }
1677 stream->output_written += ret;
1678
1679 /* This call is useless on a socket so better save a syscall. */
1680 if (!relayd) {
1681 /* This won't block, but will start writeout asynchronously */
1682 lttng_sync_file_range(outfd, stream->out_fd_offset, len,
1683 SYNC_FILE_RANGE_WRITE);
1684 stream->out_fd_offset += len;
1685 }
1686 lttng_consumer_sync_trace_file(stream, orig_offset);
1687
1688 write_error:
1689 /*
1690 * This is a special case that the relayd has closed its socket. Let's
1691 * cleanup the relayd object and all associated streams.
1692 */
1693 if (relayd && relayd_hang_up) {
1694 cleanup_relayd(relayd, ctx);
1695 }
1696
1697 end:
1698 /* Unlock only if ctrl socket used */
1699 if (relayd && stream->metadata_flag) {
1700 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
1701 }
1702
1703 rcu_read_unlock();
1704 return ret;
1705 }
1706
1707 /*
1708 * Splice the data from the ring buffer to the tracefile.
1709 *
1710 * It must be called with the stream lock held.
1711 *
1712 * Returns the number of bytes spliced.
1713 */
1714 ssize_t lttng_consumer_on_read_subbuffer_splice(
1715 struct lttng_consumer_local_data *ctx,
1716 struct lttng_consumer_stream *stream, unsigned long len,
1717 unsigned long padding,
1718 struct ctf_packet_index *index)
1719 {
1720 ssize_t ret = 0, written = 0, ret_splice = 0;
1721 loff_t offset = 0;
1722 off_t orig_offset = stream->out_fd_offset;
1723 int fd = stream->wait_fd;
1724 /* Default is on the disk */
1725 int outfd = stream->out_fd;
1726 struct consumer_relayd_sock_pair *relayd = NULL;
1727 int *splice_pipe;
1728 unsigned int relayd_hang_up = 0;
1729
1730 switch (consumer_data.type) {
1731 case LTTNG_CONSUMER_KERNEL:
1732 break;
1733 case LTTNG_CONSUMER32_UST:
1734 case LTTNG_CONSUMER64_UST:
1735 /* Not supported for user space tracing */
1736 return -ENOSYS;
1737 default:
1738 ERR("Unknown consumer_data type");
1739 assert(0);
1740 }
1741
1742 /* RCU lock for the relayd pointer */
1743 rcu_read_lock();
1744
1745 /* Flag that the current stream if set for network streaming. */
1746 if (stream->net_seq_idx != (uint64_t) -1ULL) {
1747 relayd = consumer_find_relayd(stream->net_seq_idx);
1748 if (relayd == NULL) {
1749 written = -ret;
1750 goto end;
1751 }
1752 }
1753 splice_pipe = stream->splice_pipe;
1754
1755 /* Write metadata stream id before payload */
1756 if (relayd) {
1757 unsigned long total_len = len;
1758
1759 if (stream->metadata_flag) {
1760 /*
1761 * Lock the control socket for the complete duration of the function
1762 * since from this point on we will use the socket.
1763 */
1764 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
1765
1766 if (stream->reset_metadata_flag) {
1767 ret = relayd_reset_metadata(&relayd->control_sock,
1768 stream->relayd_stream_id,
1769 stream->metadata_version);
1770 if (ret < 0) {
1771 relayd_hang_up = 1;
1772 goto write_error;
1773 }
1774 stream->reset_metadata_flag = 0;
1775 }
1776 ret = write_relayd_metadata_id(splice_pipe[1], stream, relayd,
1777 padding);
1778 if (ret < 0) {
1779 written = ret;
1780 relayd_hang_up = 1;
1781 goto write_error;
1782 }
1783
1784 total_len += sizeof(struct lttcomm_relayd_metadata_payload);
1785 }
1786
1787 ret = write_relayd_stream_header(stream, total_len, padding, relayd);
1788 if (ret < 0) {
1789 written = ret;
1790 relayd_hang_up = 1;
1791 goto write_error;
1792 }
1793 /* Use the returned socket. */
1794 outfd = ret;
1795 } else {
1796 /* No streaming, we have to set the len with the full padding */
1797 len += padding;
1798
1799 if (stream->metadata_flag && stream->reset_metadata_flag) {
1800 ret = utils_truncate_stream_file(stream->out_fd, 0);
1801 if (ret < 0) {
1802 ERR("Reset metadata file");
1803 goto end;
1804 }
1805 stream->reset_metadata_flag = 0;
1806 }
1807 /*
1808 * Check if we need to change the tracefile before writing the packet.
1809 */
1810 if (stream->chan->tracefile_size > 0 &&
1811 (stream->tracefile_size_current + len) >
1812 stream->chan->tracefile_size) {
1813 ret = utils_rotate_stream_file(stream->chan->pathname,
1814 stream->name, stream->chan->tracefile_size,
1815 stream->chan->tracefile_count, stream->uid, stream->gid,
1816 stream->out_fd, &(stream->tracefile_count_current),
1817 &stream->out_fd);
1818 if (ret < 0) {
1819 written = ret;
1820 ERR("Rotating output file");
1821 goto end;
1822 }
1823 outfd = stream->out_fd;
1824
1825 if (stream->index_fd >= 0) {
1826 ret = close(stream->index_fd);
1827 if (ret < 0) {
1828 PERROR("Closing index");
1829 goto end;
1830 }
1831 stream->index_fd = -1;
1832 ret = index_create_file(stream->chan->pathname,
1833 stream->name, stream->uid, stream->gid,
1834 stream->chan->tracefile_size,
1835 stream->tracefile_count_current);
1836 if (ret < 0) {
1837 written = ret;
1838 goto end;
1839 }
1840 stream->index_fd = ret;
1841 }
1842
1843 /* Reset current size because we just perform a rotation. */
1844 stream->tracefile_size_current = 0;
1845 stream->out_fd_offset = 0;
1846 orig_offset = 0;
1847 }
1848 stream->tracefile_size_current += len;
1849 index->offset = htobe64(stream->out_fd_offset);
1850 }
1851
1852 while (len > 0) {
1853 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1854 (unsigned long)offset, len, fd, splice_pipe[1]);
1855 ret_splice = splice(fd, &offset, splice_pipe[1], NULL, len,
1856 SPLICE_F_MOVE | SPLICE_F_MORE);
1857 DBG("splice chan to pipe, ret %zd", ret_splice);
1858 if (ret_splice < 0) {
1859 ret = errno;
1860 written = -ret;
1861 PERROR("Error in relay splice");
1862 goto splice_error;
1863 }
1864
1865 /* Handle stream on the relayd if the output is on the network */
1866 if (relayd && stream->metadata_flag) {
1867 size_t metadata_payload_size =
1868 sizeof(struct lttcomm_relayd_metadata_payload);
1869
1870 /* Update counter to fit the spliced data */
1871 ret_splice += metadata_payload_size;
1872 len += metadata_payload_size;
1873 /*
1874 * We do this so the return value can match the len passed as
1875 * argument to this function.
1876 */
1877 written -= metadata_payload_size;
1878 }
1879
1880 /* Splice data out */
1881 ret_splice = splice(splice_pipe[0], NULL, outfd, NULL,
1882 ret_splice, SPLICE_F_MOVE | SPLICE_F_MORE);
1883 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
1884 outfd, ret_splice);
1885 if (ret_splice < 0) {
1886 ret = errno;
1887 written = -ret;
1888 relayd_hang_up = 1;
1889 goto write_error;
1890 } else if (ret_splice > len) {
1891 /*
1892 * We don't expect this code path to be executed but you never know
1893 * so this is an extra protection agains a buggy splice().
1894 */
1895 ret = errno;
1896 written += ret_splice;
1897 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice,
1898 len);
1899 goto splice_error;
1900 } else {
1901 /* All good, update current len and continue. */
1902 len -= ret_splice;
1903 }
1904
1905 /* This call is useless on a socket so better save a syscall. */
1906 if (!relayd) {
1907 /* This won't block, but will start writeout asynchronously */
1908 lttng_sync_file_range(outfd, stream->out_fd_offset, ret_splice,
1909 SYNC_FILE_RANGE_WRITE);
1910 stream->out_fd_offset += ret_splice;
1911 }
1912 stream->output_written += ret_splice;
1913 written += ret_splice;
1914 }
1915 lttng_consumer_sync_trace_file(stream, orig_offset);
1916 goto end;
1917
1918 write_error:
1919 /*
1920 * This is a special case that the relayd has closed its socket. Let's
1921 * cleanup the relayd object and all associated streams.
1922 */
1923 if (relayd && relayd_hang_up) {
1924 cleanup_relayd(relayd, ctx);
1925 /* Skip splice error so the consumer does not fail */
1926 goto end;
1927 }
1928
1929 splice_error:
1930 /* send the appropriate error description to sessiond */
1931 switch (ret) {
1932 case EINVAL:
1933 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_EINVAL);
1934 break;
1935 case ENOMEM:
1936 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_ENOMEM);
1937 break;
1938 case ESPIPE:
1939 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_ESPIPE);
1940 break;
1941 }
1942
1943 end:
1944 if (relayd && stream->metadata_flag) {
1945 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
1946 }
1947
1948 rcu_read_unlock();
1949 return written;
1950 }
1951
1952 /*
1953 * Take a snapshot for a specific fd
1954 *
1955 * Returns 0 on success, < 0 on error
1956 */
1957 int lttng_consumer_take_snapshot(struct lttng_consumer_stream *stream)
1958 {
1959 switch (consumer_data.type) {
1960 case LTTNG_CONSUMER_KERNEL:
1961 return lttng_kconsumer_take_snapshot(stream);
1962 case LTTNG_CONSUMER32_UST:
1963 case LTTNG_CONSUMER64_UST:
1964 return lttng_ustconsumer_take_snapshot(stream);
1965 default:
1966 ERR("Unknown consumer_data type");
1967 assert(0);
1968 return -ENOSYS;
1969 }
1970 }
1971
1972 /*
1973 * Get the produced position
1974 *
1975 * Returns 0 on success, < 0 on error
1976 */
1977 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream *stream,
1978 unsigned long *pos)
1979 {
1980 switch (consumer_data.type) {
1981 case LTTNG_CONSUMER_KERNEL:
1982 return lttng_kconsumer_get_produced_snapshot(stream, pos);
1983 case LTTNG_CONSUMER32_UST:
1984 case LTTNG_CONSUMER64_UST:
1985 return lttng_ustconsumer_get_produced_snapshot(stream, pos);
1986 default:
1987 ERR("Unknown consumer_data type");
1988 assert(0);
1989 return -ENOSYS;
1990 }
1991 }
1992
1993 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data *ctx,
1994 int sock, struct pollfd *consumer_sockpoll)
1995 {
1996 switch (consumer_data.type) {
1997 case LTTNG_CONSUMER_KERNEL:
1998 return lttng_kconsumer_recv_cmd(ctx, sock, consumer_sockpoll);
1999 case LTTNG_CONSUMER32_UST:
2000 case LTTNG_CONSUMER64_UST:
2001 return lttng_ustconsumer_recv_cmd(ctx, sock, consumer_sockpoll);
2002 default:
2003 ERR("Unknown consumer_data type");
2004 assert(0);
2005 return -ENOSYS;
2006 }
2007 }
2008
2009 void lttng_consumer_close_all_metadata(void)
2010 {
2011 switch (consumer_data.type) {
2012 case LTTNG_CONSUMER_KERNEL:
2013 /*
2014 * The Kernel consumer has a different metadata scheme so we don't
2015 * close anything because the stream will be closed by the session
2016 * daemon.
2017 */
2018 break;
2019 case LTTNG_CONSUMER32_UST:
2020 case LTTNG_CONSUMER64_UST:
2021 /*
2022 * Close all metadata streams. The metadata hash table is passed and
2023 * this call iterates over it by closing all wakeup fd. This is safe
2024 * because at this point we are sure that the metadata producer is
2025 * either dead or blocked.
2026 */
2027 lttng_ustconsumer_close_all_metadata(metadata_ht);
2028 break;
2029 default:
2030 ERR("Unknown consumer_data type");
2031 assert(0);
2032 }
2033 }
2034
2035 /*
2036 * Clean up a metadata stream and free its memory.
2037 */
2038 void consumer_del_metadata_stream(struct lttng_consumer_stream *stream,
2039 struct lttng_ht *ht)
2040 {
2041 struct lttng_consumer_channel *free_chan = NULL;
2042
2043 assert(stream);
2044 /*
2045 * This call should NEVER receive regular stream. It must always be
2046 * metadata stream and this is crucial for data structure synchronization.
2047 */
2048 assert(stream->metadata_flag);
2049
2050 DBG3("Consumer delete metadata stream %d", stream->wait_fd);
2051
2052 pthread_mutex_lock(&consumer_data.lock);
2053 pthread_mutex_lock(&stream->chan->lock);
2054 pthread_mutex_lock(&stream->lock);
2055
2056 /* Remove any reference to that stream. */
2057 consumer_stream_delete(stream, ht);
2058
2059 /* Close down everything including the relayd if one. */
2060 consumer_stream_close(stream);
2061 /* Destroy tracer buffers of the stream. */
2062 consumer_stream_destroy_buffers(stream);
2063
2064 /* Atomically decrement channel refcount since other threads can use it. */
2065 if (!uatomic_sub_return(&stream->chan->refcount, 1)
2066 && !uatomic_read(&stream->chan->nb_init_stream_left)) {
2067 /* Go for channel deletion! */
2068 free_chan = stream->chan;
2069 }
2070
2071 /*
2072 * Nullify the stream reference so it is not used after deletion. The
2073 * channel lock MUST be acquired before being able to check for a NULL
2074 * pointer value.
2075 */
2076 stream->chan->metadata_stream = NULL;
2077
2078 pthread_mutex_unlock(&stream->lock);
2079 pthread_mutex_unlock(&stream->chan->lock);
2080 pthread_mutex_unlock(&consumer_data.lock);
2081
2082 if (free_chan) {
2083 consumer_del_channel(free_chan);
2084 }
2085
2086 consumer_stream_free(stream);
2087 }
2088
2089 /*
2090 * Action done with the metadata stream when adding it to the consumer internal
2091 * data structures to handle it.
2092 */
2093 int consumer_add_metadata_stream(struct lttng_consumer_stream *stream)
2094 {
2095 struct lttng_ht *ht = metadata_ht;
2096 int ret = 0;
2097 struct lttng_ht_iter iter;
2098 struct lttng_ht_node_u64 *node;
2099
2100 assert(stream);
2101 assert(ht);
2102
2103 DBG3("Adding metadata stream %" PRIu64 " to hash table", stream->key);
2104
2105 pthread_mutex_lock(&consumer_data.lock);
2106 pthread_mutex_lock(&stream->chan->lock);
2107 pthread_mutex_lock(&stream->chan->timer_lock);
2108 pthread_mutex_lock(&stream->lock);
2109
2110 /*
2111 * From here, refcounts are updated so be _careful_ when returning an error
2112 * after this point.
2113 */
2114
2115 rcu_read_lock();
2116
2117 /*
2118 * Lookup the stream just to make sure it does not exist in our internal
2119 * state. This should NEVER happen.
2120 */
2121 lttng_ht_lookup(ht, &stream->key, &iter);
2122 node = lttng_ht_iter_get_node_u64(&iter);
2123 assert(!node);
2124
2125 /*
2126 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2127 * in terms of destroying the associated channel, because the action that
2128 * causes the count to become 0 also causes a stream to be added. The
2129 * channel deletion will thus be triggered by the following removal of this
2130 * stream.
2131 */
2132 if (uatomic_read(&stream->chan->nb_init_stream_left) > 0) {
2133 /* Increment refcount before decrementing nb_init_stream_left */
2134 cmm_smp_wmb();
2135 uatomic_dec(&stream->chan->nb_init_stream_left);
2136 }
2137
2138 lttng_ht_add_unique_u64(ht, &stream->node);
2139
2140 lttng_ht_add_unique_u64(consumer_data.stream_per_chan_id_ht,
2141 &stream->node_channel_id);
2142
2143 /*
2144 * Add stream to the stream_list_ht of the consumer data. No need to steal
2145 * the key since the HT does not use it and we allow to add redundant keys
2146 * into this table.
2147 */
2148 lttng_ht_add_u64(consumer_data.stream_list_ht, &stream->node_session_id);
2149
2150 rcu_read_unlock();
2151
2152 pthread_mutex_unlock(&stream->lock);
2153 pthread_mutex_unlock(&stream->chan->lock);
2154 pthread_mutex_unlock(&stream->chan->timer_lock);
2155 pthread_mutex_unlock(&consumer_data.lock);
2156 return ret;
2157 }
2158
2159 /*
2160 * Delete data stream that are flagged for deletion (endpoint_status).
2161 */
2162 static void validate_endpoint_status_data_stream(void)
2163 {
2164 struct lttng_ht_iter iter;
2165 struct lttng_consumer_stream *stream;
2166
2167 DBG("Consumer delete flagged data stream");
2168
2169 rcu_read_lock();
2170 cds_lfht_for_each_entry(data_ht->ht, &iter.iter, stream, node.node) {
2171 /* Validate delete flag of the stream */
2172 if (stream->endpoint_status == CONSUMER_ENDPOINT_ACTIVE) {
2173 continue;
2174 }
2175 /* Delete it right now */
2176 consumer_del_stream(stream, data_ht);
2177 }
2178 rcu_read_unlock();
2179 }
2180
2181 /*
2182 * Delete metadata stream that are flagged for deletion (endpoint_status).
2183 */
2184 static void validate_endpoint_status_metadata_stream(
2185 struct lttng_poll_event *pollset)
2186 {
2187 struct lttng_ht_iter iter;
2188 struct lttng_consumer_stream *stream;
2189
2190 DBG("Consumer delete flagged metadata stream");
2191
2192 assert(pollset);
2193
2194 rcu_read_lock();
2195 cds_lfht_for_each_entry(metadata_ht->ht, &iter.iter, stream, node.node) {
2196 /* Validate delete flag of the stream */
2197 if (stream->endpoint_status == CONSUMER_ENDPOINT_ACTIVE) {
2198 continue;
2199 }
2200 /*
2201 * Remove from pollset so the metadata thread can continue without
2202 * blocking on a deleted stream.
2203 */
2204 lttng_poll_del(pollset, stream->wait_fd);
2205
2206 /* Delete it right now */
2207 consumer_del_metadata_stream(stream, metadata_ht);
2208 }
2209 rcu_read_unlock();
2210 }
2211
2212 /*
2213 * Thread polls on metadata file descriptor and write them on disk or on the
2214 * network.
2215 */
2216 void *consumer_thread_metadata_poll(void *data)
2217 {
2218 int ret, i, pollfd, err = -1;
2219 uint32_t revents, nb_fd;
2220 struct lttng_consumer_stream *stream = NULL;
2221 struct lttng_ht_iter iter;
2222 struct lttng_ht_node_u64 *node;
2223 struct lttng_poll_event events;
2224 struct lttng_consumer_local_data *ctx = data;
2225 ssize_t len;
2226
2227 rcu_register_thread();
2228
2229 health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_METADATA);
2230
2231 if (testpoint(consumerd_thread_metadata)) {
2232 goto error_testpoint;
2233 }
2234
2235 health_code_update();
2236
2237 DBG("Thread metadata poll started");
2238
2239 /* Size is set to 1 for the consumer_metadata pipe */
2240 ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC);
2241 if (ret < 0) {
2242 ERR("Poll set creation failed");
2243 goto end_poll;
2244 }
2245
2246 ret = lttng_poll_add(&events,
2247 lttng_pipe_get_readfd(ctx->consumer_metadata_pipe), LPOLLIN);
2248 if (ret < 0) {
2249 goto end;
2250 }
2251
2252 /* Main loop */
2253 DBG("Metadata main loop started");
2254
2255 while (1) {
2256 restart:
2257 health_code_update();
2258 health_poll_entry();
2259 DBG("Metadata poll wait");
2260 ret = lttng_poll_wait(&events, -1);
2261 DBG("Metadata poll return from wait with %d fd(s)",
2262 LTTNG_POLL_GETNB(&events));
2263 health_poll_exit();
2264 DBG("Metadata event caught in thread");
2265 if (ret < 0) {
2266 if (errno == EINTR) {
2267 ERR("Poll EINTR caught");
2268 goto restart;
2269 }
2270 if (LTTNG_POLL_GETNB(&events) == 0) {
2271 err = 0; /* All is OK */
2272 }
2273 goto end;
2274 }
2275
2276 nb_fd = ret;
2277
2278 /* From here, the event is a metadata wait fd */
2279 for (i = 0; i < nb_fd; i++) {
2280 health_code_update();
2281
2282 revents = LTTNG_POLL_GETEV(&events, i);
2283 pollfd = LTTNG_POLL_GETFD(&events, i);
2284
2285 if (!revents) {
2286 /* No activity for this FD (poll implementation). */
2287 continue;
2288 }
2289
2290 if (pollfd == lttng_pipe_get_readfd(ctx->consumer_metadata_pipe)) {
2291 if (revents & LPOLLIN) {
2292 ssize_t pipe_len;
2293
2294 pipe_len = lttng_pipe_read(ctx->consumer_metadata_pipe,
2295 &stream, sizeof(stream));
2296 if (pipe_len < sizeof(stream)) {
2297 if (pipe_len < 0) {
2298 PERROR("read metadata stream");
2299 }
2300 /*
2301 * Remove the pipe from the poll set and continue the loop
2302 * since their might be data to consume.
2303 */
2304 lttng_poll_del(&events,
2305 lttng_pipe_get_readfd(ctx->consumer_metadata_pipe));
2306 lttng_pipe_read_close(ctx->consumer_metadata_pipe);
2307 continue;
2308 }
2309
2310 /* A NULL stream means that the state has changed. */
2311 if (stream == NULL) {
2312 /* Check for deleted streams. */
2313 validate_endpoint_status_metadata_stream(&events);
2314 goto restart;
2315 }
2316
2317 DBG("Adding metadata stream %d to poll set",
2318 stream->wait_fd);
2319
2320 /* Add metadata stream to the global poll events list */
2321 lttng_poll_add(&events, stream->wait_fd,
2322 LPOLLIN | LPOLLPRI | LPOLLHUP);
2323 } else if (revents & (LPOLLERR | LPOLLHUP)) {
2324 DBG("Metadata thread pipe hung up");
2325 /*
2326 * Remove the pipe from the poll set and continue the loop
2327 * since their might be data to consume.
2328 */
2329 lttng_poll_del(&events,
2330 lttng_pipe_get_readfd(ctx->consumer_metadata_pipe));
2331 lttng_pipe_read_close(ctx->consumer_metadata_pipe);
2332 continue;
2333 } else {
2334 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
2335 goto end;
2336 }
2337
2338 /* Handle other stream */
2339 continue;
2340 }
2341
2342 rcu_read_lock();
2343 {
2344 uint64_t tmp_id = (uint64_t) pollfd;
2345
2346 lttng_ht_lookup(metadata_ht, &tmp_id, &iter);
2347 }
2348 node = lttng_ht_iter_get_node_u64(&iter);
2349 assert(node);
2350
2351 stream = caa_container_of(node, struct lttng_consumer_stream,
2352 node);
2353
2354 if (revents & (LPOLLIN | LPOLLPRI)) {
2355 /* Get the data out of the metadata file descriptor */
2356 DBG("Metadata available on fd %d", pollfd);
2357 assert(stream->wait_fd == pollfd);
2358
2359 do {
2360 health_code_update();
2361
2362 len = ctx->on_buffer_ready(stream, ctx);
2363 /*
2364 * We don't check the return value here since if we get
2365 * a negative len, it means an error occured thus we
2366 * simply remove it from the poll set and free the
2367 * stream.
2368 */
2369 } while (len > 0);
2370
2371 /* It's ok to have an unavailable sub-buffer */
2372 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2373 /* Clean up stream from consumer and free it. */
2374 lttng_poll_del(&events, stream->wait_fd);
2375 consumer_del_metadata_stream(stream, metadata_ht);
2376 }
2377 } else if (revents & (LPOLLERR | LPOLLHUP)) {
2378 DBG("Metadata fd %d is hup|err.", pollfd);
2379 if (!stream->hangup_flush_done
2380 && (consumer_data.type == LTTNG_CONSUMER32_UST
2381 || consumer_data.type == LTTNG_CONSUMER64_UST)) {
2382 DBG("Attempting to flush and consume the UST buffers");
2383 lttng_ustconsumer_on_stream_hangup(stream);
2384
2385 /* We just flushed the stream now read it. */
2386 do {
2387 health_code_update();
2388
2389 len = ctx->on_buffer_ready(stream, ctx);
2390 /*
2391 * We don't check the return value here since if we get
2392 * a negative len, it means an error occured thus we
2393 * simply remove it from the poll set and free the
2394 * stream.
2395 */
2396 } while (len > 0);
2397 }
2398
2399 lttng_poll_del(&events, stream->wait_fd);
2400 /*
2401 * This call update the channel states, closes file descriptors
2402 * and securely free the stream.
2403 */
2404 consumer_del_metadata_stream(stream, metadata_ht);
2405 } else {
2406 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
2407 rcu_read_unlock();
2408 goto end;
2409 }
2410 /* Release RCU lock for the stream looked up */
2411 rcu_read_unlock();
2412 }
2413 }
2414
2415 /* All is OK */
2416 err = 0;
2417 end:
2418 DBG("Metadata poll thread exiting");
2419
2420 lttng_poll_clean(&events);
2421 end_poll:
2422 error_testpoint:
2423 if (err) {
2424 health_error();
2425 ERR("Health error occurred in %s", __func__);
2426 }
2427 health_unregister(health_consumerd);
2428 rcu_unregister_thread();
2429 return NULL;
2430 }
2431
2432 /*
2433 * This thread polls the fds in the set to consume the data and write
2434 * it to tracefile if necessary.
2435 */
2436 void *consumer_thread_data_poll(void *data)
2437 {
2438 int num_rdy, num_hup, high_prio, ret, i, err = -1;
2439 struct pollfd *pollfd = NULL;
2440 /* local view of the streams */
2441 struct lttng_consumer_stream **local_stream = NULL, *new_stream = NULL;
2442 /* local view of consumer_data.fds_count */
2443 int nb_fd = 0;
2444 struct lttng_consumer_local_data *ctx = data;
2445 ssize_t len;
2446
2447 rcu_register_thread();
2448
2449 health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_DATA);
2450
2451 if (testpoint(consumerd_thread_data)) {
2452 goto error_testpoint;
2453 }
2454
2455 health_code_update();
2456
2457 local_stream = zmalloc(sizeof(struct lttng_consumer_stream *));
2458 if (local_stream == NULL) {
2459 PERROR("local_stream malloc");
2460 goto end;
2461 }
2462
2463 while (1) {
2464 health_code_update();
2465
2466 high_prio = 0;
2467 num_hup = 0;
2468
2469 /*
2470 * the fds set has been updated, we need to update our
2471 * local array as well
2472 */
2473 pthread_mutex_lock(&consumer_data.lock);
2474 if (consumer_data.need_update) {
2475 free(pollfd);
2476 pollfd = NULL;
2477
2478 free(local_stream);
2479 local_stream = NULL;
2480
2481 /*
2482 * Allocate for all fds +1 for the consumer_data_pipe and +1 for
2483 * wake up pipe.
2484 */
2485 pollfd = zmalloc((consumer_data.stream_count + 2) * sizeof(struct pollfd));
2486 if (pollfd == NULL) {
2487 PERROR("pollfd malloc");
2488 pthread_mutex_unlock(&consumer_data.lock);
2489 goto end;
2490 }
2491
2492 local_stream = zmalloc((consumer_data.stream_count + 2) *
2493 sizeof(struct lttng_consumer_stream *));
2494 if (local_stream == NULL) {
2495 PERROR("local_stream malloc");
2496 pthread_mutex_unlock(&consumer_data.lock);
2497 goto end;
2498 }
2499 ret = update_poll_array(ctx, &pollfd, local_stream,
2500 data_ht);
2501 if (ret < 0) {
2502 ERR("Error in allocating pollfd or local_outfds");
2503 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
2504 pthread_mutex_unlock(&consumer_data.lock);
2505 goto end;
2506 }
2507 nb_fd = ret;
2508 consumer_data.need_update = 0;
2509 }
2510 pthread_mutex_unlock(&consumer_data.lock);
2511
2512 /* No FDs and consumer_quit, consumer_cleanup the thread */
2513 if (nb_fd == 0 && consumer_quit == 1) {
2514 err = 0; /* All is OK */
2515 goto end;
2516 }
2517 /* poll on the array of fds */
2518 restart:
2519 DBG("polling on %d fd", nb_fd + 2);
2520 health_poll_entry();
2521 num_rdy = poll(pollfd, nb_fd + 2, -1);
2522 health_poll_exit();
2523 DBG("poll num_rdy : %d", num_rdy);
2524 if (num_rdy == -1) {
2525 /*
2526 * Restart interrupted system call.
2527 */
2528 if (errno == EINTR) {
2529 goto restart;
2530 }
2531 PERROR("Poll error");
2532 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
2533 goto end;
2534 } else if (num_rdy == 0) {
2535 DBG("Polling thread timed out");
2536 goto end;
2537 }
2538
2539 /*
2540 * If the consumer_data_pipe triggered poll go directly to the
2541 * beginning of the loop to update the array. We want to prioritize
2542 * array update over low-priority reads.
2543 */
2544 if (pollfd[nb_fd].revents & (POLLIN | POLLPRI)) {
2545 ssize_t pipe_readlen;
2546
2547 DBG("consumer_data_pipe wake up");
2548 pipe_readlen = lttng_pipe_read(ctx->consumer_data_pipe,
2549 &new_stream, sizeof(new_stream));
2550 if (pipe_readlen < sizeof(new_stream)) {
2551 PERROR("Consumer data pipe");
2552 /* Continue so we can at least handle the current stream(s). */
2553 continue;
2554 }
2555
2556 /*
2557 * If the stream is NULL, just ignore it. It's also possible that
2558 * the sessiond poll thread changed the consumer_quit state and is
2559 * waking us up to test it.
2560 */
2561 if (new_stream == NULL) {
2562 validate_endpoint_status_data_stream();
2563 continue;
2564 }
2565
2566 /* Continue to update the local streams and handle prio ones */
2567 continue;
2568 }
2569
2570 /* Handle wakeup pipe. */
2571 if (pollfd[nb_fd + 1].revents & (POLLIN | POLLPRI)) {
2572 char dummy;
2573 ssize_t pipe_readlen;
2574
2575 pipe_readlen = lttng_pipe_read(ctx->consumer_wakeup_pipe, &dummy,
2576 sizeof(dummy));
2577 if (pipe_readlen < 0) {
2578 PERROR("Consumer data wakeup pipe");
2579 }
2580 /* We've been awakened to handle stream(s). */
2581 ctx->has_wakeup = 0;
2582 }
2583
2584 /* Take care of high priority channels first. */
2585 for (i = 0; i < nb_fd; i++) {
2586 health_code_update();
2587
2588 if (local_stream[i] == NULL) {
2589 continue;
2590 }
2591 if (pollfd[i].revents & POLLPRI) {
2592 DBG("Urgent read on fd %d", pollfd[i].fd);
2593 high_prio = 1;
2594 len = ctx->on_buffer_ready(local_stream[i], ctx);
2595 /* it's ok to have an unavailable sub-buffer */
2596 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2597 /* Clean the stream and free it. */
2598 consumer_del_stream(local_stream[i], data_ht);
2599 local_stream[i] = NULL;
2600 } else if (len > 0) {
2601 local_stream[i]->data_read = 1;
2602 }
2603 }
2604 }
2605
2606 /*
2607 * If we read high prio channel in this loop, try again
2608 * for more high prio data.
2609 */
2610 if (high_prio) {
2611 continue;
2612 }
2613
2614 /* Take care of low priority channels. */
2615 for (i = 0; i < nb_fd; i++) {
2616 health_code_update();
2617
2618 if (local_stream[i] == NULL) {
2619 continue;
2620 }
2621 if ((pollfd[i].revents & POLLIN) ||
2622 local_stream[i]->hangup_flush_done ||
2623 local_stream[i]->has_data) {
2624 DBG("Normal read on fd %d", pollfd[i].fd);
2625 len = ctx->on_buffer_ready(local_stream[i], ctx);
2626 /* it's ok to have an unavailable sub-buffer */
2627 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2628 /* Clean the stream and free it. */
2629 consumer_del_stream(local_stream[i], data_ht);
2630 local_stream[i] = NULL;
2631 } else if (len > 0) {
2632 local_stream[i]->data_read = 1;
2633 }
2634 }
2635 }
2636
2637 /* Handle hangup and errors */
2638 for (i = 0; i < nb_fd; i++) {
2639 health_code_update();
2640
2641 if (local_stream[i] == NULL) {
2642 continue;
2643 }
2644 if (!local_stream[i]->hangup_flush_done
2645 && (pollfd[i].revents & (POLLHUP | POLLERR | POLLNVAL))
2646 && (consumer_data.type == LTTNG_CONSUMER32_UST
2647 || consumer_data.type == LTTNG_CONSUMER64_UST)) {
2648 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2649 pollfd[i].fd);
2650 lttng_ustconsumer_on_stream_hangup(local_stream[i]);
2651 /* Attempt read again, for the data we just flushed. */
2652 local_stream[i]->data_read = 1;
2653 }
2654 /*
2655 * If the poll flag is HUP/ERR/NVAL and we have
2656 * read no data in this pass, we can remove the
2657 * stream from its hash table.
2658 */
2659 if ((pollfd[i].revents & POLLHUP)) {
2660 DBG("Polling fd %d tells it has hung up.", pollfd[i].fd);
2661 if (!local_stream[i]->data_read) {
2662 consumer_del_stream(local_stream[i], data_ht);
2663 local_stream[i] = NULL;
2664 num_hup++;
2665 }
2666 } else if (pollfd[i].revents & POLLERR) {
2667 ERR("Error returned in polling fd %d.", pollfd[i].fd);
2668 if (!local_stream[i]->data_read) {
2669 consumer_del_stream(local_stream[i], data_ht);
2670 local_stream[i] = NULL;
2671 num_hup++;
2672 }
2673 } else if (pollfd[i].revents & POLLNVAL) {
2674 ERR("Polling fd %d tells fd is not open.", pollfd[i].fd);
2675 if (!local_stream[i]->data_read) {
2676 consumer_del_stream(local_stream[i], data_ht);
2677 local_stream[i] = NULL;
2678 num_hup++;
2679 }
2680 }
2681 if (local_stream[i] != NULL) {
2682 local_stream[i]->data_read = 0;
2683 }
2684 }
2685 }
2686 /* All is OK */
2687 err = 0;
2688 end:
2689 DBG("polling thread exiting");
2690 free(pollfd);
2691 free(local_stream);
2692
2693 /*
2694 * Close the write side of the pipe so epoll_wait() in
2695 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2696 * read side of the pipe. If we close them both, epoll_wait strangely does
2697 * not return and could create a endless wait period if the pipe is the
2698 * only tracked fd in the poll set. The thread will take care of closing
2699 * the read side.
2700 */
2701 (void) lttng_pipe_write_close(ctx->consumer_metadata_pipe);
2702
2703 error_testpoint:
2704 if (err) {
2705 health_error();
2706 ERR("Health error occurred in %s", __func__);
2707 }
2708 health_unregister(health_consumerd);
2709
2710 rcu_unregister_thread();
2711 return NULL;
2712 }
2713
2714 /*
2715 * Close wake-up end of each stream belonging to the channel. This will
2716 * allow the poll() on the stream read-side to detect when the
2717 * write-side (application) finally closes them.
2718 */
2719 static
2720 void consumer_close_channel_streams(struct lttng_consumer_channel *channel)
2721 {
2722 struct lttng_ht *ht;
2723 struct lttng_consumer_stream *stream;
2724 struct lttng_ht_iter iter;
2725
2726 ht = consumer_data.stream_per_chan_id_ht;
2727
2728 rcu_read_lock();
2729 cds_lfht_for_each_entry_duplicate(ht->ht,
2730 ht->hash_fct(&channel->key, lttng_ht_seed),
2731 ht->match_fct, &channel->key,
2732 &iter.iter, stream, node_channel_id.node) {
2733 /*
2734 * Protect against teardown with mutex.
2735 */
2736 pthread_mutex_lock(&stream->lock);
2737 if (cds_lfht_is_node_deleted(&stream->node.node)) {
2738 goto next;
2739 }
2740 switch (consumer_data.type) {
2741 case LTTNG_CONSUMER_KERNEL:
2742 break;
2743 case LTTNG_CONSUMER32_UST:
2744 case LTTNG_CONSUMER64_UST:
2745 if (stream->metadata_flag) {
2746 /* Safe and protected by the stream lock. */
2747 lttng_ustconsumer_close_metadata(stream->chan);
2748 } else {
2749 /*
2750 * Note: a mutex is taken internally within
2751 * liblttng-ust-ctl to protect timer wakeup_fd
2752 * use from concurrent close.
2753 */
2754 lttng_ustconsumer_close_stream_wakeup(stream);
2755 }
2756 break;
2757 default:
2758 ERR("Unknown consumer_data type");
2759 assert(0);
2760 }
2761 next:
2762 pthread_mutex_unlock(&stream->lock);
2763 }
2764 rcu_read_unlock();
2765 }
2766
2767 static void destroy_channel_ht(struct lttng_ht *ht)
2768 {
2769 struct lttng_ht_iter iter;
2770 struct lttng_consumer_channel *channel;
2771 int ret;
2772
2773 if (ht == NULL) {
2774 return;
2775 }
2776
2777 rcu_read_lock();
2778 cds_lfht_for_each_entry(ht->ht, &iter.iter, channel, wait_fd_node.node) {
2779 ret = lttng_ht_del(ht, &iter);
2780 assert(ret != 0);
2781 }
2782 rcu_read_unlock();
2783
2784 lttng_ht_destroy(ht);
2785 }
2786
2787 /*
2788 * This thread polls the channel fds to detect when they are being
2789 * closed. It closes all related streams if the channel is detected as
2790 * closed. It is currently only used as a shim layer for UST because the
2791 * consumerd needs to keep the per-stream wakeup end of pipes open for
2792 * periodical flush.
2793 */
2794 void *consumer_thread_channel_poll(void *data)
2795 {
2796 int ret, i, pollfd, err = -1;
2797 uint32_t revents, nb_fd;
2798 struct lttng_consumer_channel *chan = NULL;
2799 struct lttng_ht_iter iter;
2800 struct lttng_ht_node_u64 *node;
2801 struct lttng_poll_event events;
2802 struct lttng_consumer_local_data *ctx = data;
2803 struct lttng_ht *channel_ht;
2804
2805 rcu_register_thread();
2806
2807 health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_CHANNEL);
2808
2809 if (testpoint(consumerd_thread_channel)) {
2810 goto error_testpoint;
2811 }
2812
2813 health_code_update();
2814
2815 channel_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
2816 if (!channel_ht) {
2817 /* ENOMEM at this point. Better to bail out. */
2818 goto end_ht;
2819 }
2820
2821 DBG("Thread channel poll started");
2822
2823 /* Size is set to 1 for the consumer_channel pipe */
2824 ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC);
2825 if (ret < 0) {
2826 ERR("Poll set creation failed");
2827 goto end_poll;
2828 }
2829
2830 ret = lttng_poll_add(&events, ctx->consumer_channel_pipe[0], LPOLLIN);
2831 if (ret < 0) {
2832 goto end;
2833 }
2834
2835 /* Main loop */
2836 DBG("Channel main loop started");
2837
2838 while (1) {
2839 restart:
2840 health_code_update();
2841 DBG("Channel poll wait");
2842 health_poll_entry();
2843 ret = lttng_poll_wait(&events, -1);
2844 DBG("Channel poll return from wait with %d fd(s)",
2845 LTTNG_POLL_GETNB(&events));
2846 health_poll_exit();
2847 DBG("Channel event caught in thread");
2848 if (ret < 0) {
2849 if (errno == EINTR) {
2850 ERR("Poll EINTR caught");
2851 goto restart;
2852 }
2853 if (LTTNG_POLL_GETNB(&events) == 0) {
2854 err = 0; /* All is OK */
2855 }
2856 goto end;
2857 }
2858
2859 nb_fd = ret;
2860
2861 /* From here, the event is a channel wait fd */
2862 for (i = 0; i < nb_fd; i++) {
2863 health_code_update();
2864
2865 revents = LTTNG_POLL_GETEV(&events, i);
2866 pollfd = LTTNG_POLL_GETFD(&events, i);
2867
2868 if (!revents) {
2869 /* No activity for this FD (poll implementation). */
2870 continue;
2871 }
2872
2873 if (pollfd == ctx->consumer_channel_pipe[0]) {
2874 if (revents & LPOLLIN) {
2875 enum consumer_channel_action action;
2876 uint64_t key;
2877
2878 ret = read_channel_pipe(ctx, &chan, &key, &action);
2879 if (ret <= 0) {
2880 if (ret < 0) {
2881 ERR("Error reading channel pipe");
2882 }
2883 lttng_poll_del(&events, ctx->consumer_channel_pipe[0]);
2884 continue;
2885 }
2886
2887 switch (action) {
2888 case CONSUMER_CHANNEL_ADD:
2889 DBG("Adding channel %d to poll set",
2890 chan->wait_fd);
2891
2892 lttng_ht_node_init_u64(&chan->wait_fd_node,
2893 chan->wait_fd);
2894 rcu_read_lock();
2895 lttng_ht_add_unique_u64(channel_ht,
2896 &chan->wait_fd_node);
2897 rcu_read_unlock();
2898 /* Add channel to the global poll events list */
2899 lttng_poll_add(&events, chan->wait_fd,
2900 LPOLLERR | LPOLLHUP);
2901 break;
2902 case CONSUMER_CHANNEL_DEL:
2903 {
2904 /*
2905 * This command should never be called if the channel
2906 * has streams monitored by either the data or metadata
2907 * thread. The consumer only notify this thread with a
2908 * channel del. command if it receives a destroy
2909 * channel command from the session daemon that send it
2910 * if a command prior to the GET_CHANNEL failed.
2911 */
2912
2913 rcu_read_lock();
2914 chan = consumer_find_channel(key);
2915 if (!chan) {
2916 rcu_read_unlock();
2917 ERR("UST consumer get channel key %" PRIu64 " not found for del channel", key);
2918 break;
2919 }
2920 lttng_poll_del(&events, chan->wait_fd);
2921 iter.iter.node = &chan->wait_fd_node.node;
2922 ret = lttng_ht_del(channel_ht, &iter);
2923 assert(ret == 0);
2924
2925 switch (consumer_data.type) {
2926 case LTTNG_CONSUMER_KERNEL:
2927 break;
2928 case LTTNG_CONSUMER32_UST:
2929 case LTTNG_CONSUMER64_UST:
2930 health_code_update();
2931 /* Destroy streams that might have been left in the stream list. */
2932 clean_channel_stream_list(chan);
2933 break;
2934 default:
2935 ERR("Unknown consumer_data type");
2936 assert(0);
2937 }
2938
2939 /*
2940 * Release our own refcount. Force channel deletion even if
2941 * streams were not initialized.
2942 */
2943 if (!uatomic_sub_return(&chan->refcount, 1)) {
2944 consumer_del_channel(chan);
2945 }
2946 rcu_read_unlock();
2947 goto restart;
2948 }
2949 case CONSUMER_CHANNEL_QUIT:
2950 /*
2951 * Remove the pipe from the poll set and continue the loop
2952 * since their might be data to consume.
2953 */
2954 lttng_poll_del(&events, ctx->consumer_channel_pipe[0]);
2955 continue;
2956 default:
2957 ERR("Unknown action");
2958 break;
2959 }
2960 } else if (revents & (LPOLLERR | LPOLLHUP)) {
2961 DBG("Channel thread pipe hung up");
2962 /*
2963 * Remove the pipe from the poll set and continue the loop
2964 * since their might be data to consume.
2965 */
2966 lttng_poll_del(&events, ctx->consumer_channel_pipe[0]);
2967 continue;
2968 } else {
2969 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
2970 goto end;
2971 }
2972
2973 /* Handle other stream */
2974 continue;
2975 }
2976
2977 rcu_read_lock();
2978 {
2979 uint64_t tmp_id = (uint64_t) pollfd;
2980
2981 lttng_ht_lookup(channel_ht, &tmp_id, &iter);
2982 }
2983 node = lttng_ht_iter_get_node_u64(&iter);
2984 assert(node);
2985
2986 chan = caa_container_of(node, struct lttng_consumer_channel,
2987 wait_fd_node);
2988
2989 /* Check for error event */
2990 if (revents & (LPOLLERR | LPOLLHUP)) {
2991 DBG("Channel fd %d is hup|err.", pollfd);
2992
2993 lttng_poll_del(&events, chan->wait_fd);
2994 ret = lttng_ht_del(channel_ht, &iter);
2995 assert(ret == 0);
2996
2997 /*
2998 * This will close the wait fd for each stream associated to
2999 * this channel AND monitored by the data/metadata thread thus
3000 * will be clean by the right thread.
3001 */
3002 consumer_close_channel_streams(chan);
3003
3004 /* Release our own refcount */
3005 if (!uatomic_sub_return(&chan->refcount, 1)
3006 && !uatomic_read(&chan->nb_init_stream_left)) {
3007 consumer_del_channel(chan);
3008 }
3009 } else {
3010 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
3011 rcu_read_unlock();
3012 goto end;
3013 }
3014
3015 /* Release RCU lock for the channel looked up */
3016 rcu_read_unlock();
3017 }
3018 }
3019
3020 /* All is OK */
3021 err = 0;
3022 end:
3023 lttng_poll_clean(&events);
3024 end_poll:
3025 destroy_channel_ht(channel_ht);
3026 end_ht:
3027 error_testpoint:
3028 DBG("Channel poll thread exiting");
3029 if (err) {
3030 health_error();
3031 ERR("Health error occurred in %s", __func__);
3032 }
3033 health_unregister(health_consumerd);
3034 rcu_unregister_thread();
3035 return NULL;
3036 }
3037
3038 static int set_metadata_socket(struct lttng_consumer_local_data *ctx,
3039 struct pollfd *sockpoll, int client_socket)
3040 {
3041 int ret;
3042
3043 assert(ctx);
3044 assert(sockpoll);
3045
3046 ret = lttng_consumer_poll_socket(sockpoll);
3047 if (ret) {
3048 goto error;
3049 }
3050 DBG("Metadata connection on client_socket");
3051
3052 /* Blocking call, waiting for transmission */
3053 ctx->consumer_metadata_socket = lttcomm_accept_unix_sock(client_socket);
3054 if (ctx->consumer_metadata_socket < 0) {
3055 WARN("On accept metadata");
3056 ret = -1;
3057 goto error;
3058 }
3059 ret = 0;
3060
3061 error:
3062 return ret;
3063 }
3064
3065 /*
3066 * This thread listens on the consumerd socket and receives the file
3067 * descriptors from the session daemon.
3068 */
3069 void *consumer_thread_sessiond_poll(void *data)
3070 {
3071 int sock = -1, client_socket, ret, err = -1;
3072 /*
3073 * structure to poll for incoming data on communication socket avoids
3074 * making blocking sockets.
3075 */
3076 struct pollfd consumer_sockpoll[2];
3077 struct lttng_consumer_local_data *ctx = data;
3078
3079 rcu_register_thread();
3080
3081 health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_SESSIOND);
3082
3083 if (testpoint(consumerd_thread_sessiond)) {
3084 goto error_testpoint;
3085 }
3086
3087 health_code_update();
3088
3089 DBG("Creating command socket %s", ctx->consumer_command_sock_path);
3090 unlink(ctx->consumer_command_sock_path);
3091 client_socket = lttcomm_create_unix_sock(ctx->consumer_command_sock_path);
3092 if (client_socket < 0) {
3093 ERR("Cannot create command socket");
3094 goto end;
3095 }
3096
3097 ret = lttcomm_listen_unix_sock(client_socket);
3098 if (ret < 0) {
3099 goto end;
3100 }
3101
3102 DBG("Sending ready command to lttng-sessiond");
3103 ret = lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY);
3104 /* return < 0 on error, but == 0 is not fatal */
3105 if (ret < 0) {
3106 ERR("Error sending ready command to lttng-sessiond");
3107 goto end;
3108 }
3109
3110 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3111 consumer_sockpoll[0].fd = ctx->consumer_should_quit[0];
3112 consumer_sockpoll[0].events = POLLIN | POLLPRI;
3113 consumer_sockpoll[1].fd = client_socket;
3114 consumer_sockpoll[1].events = POLLIN | POLLPRI;
3115
3116 ret = lttng_consumer_poll_socket(consumer_sockpoll);
3117 if (ret) {
3118 if (ret > 0) {
3119 /* should exit */
3120 err = 0;
3121 }
3122 goto end;
3123 }
3124 DBG("Connection on client_socket");
3125
3126 /* Blocking call, waiting for transmission */
3127 sock = lttcomm_accept_unix_sock(client_socket);
3128 if (sock < 0) {
3129 WARN("On accept");
3130 goto end;
3131 }
3132
3133 /*
3134 * Setup metadata socket which is the second socket connection on the
3135 * command unix socket.
3136 */
3137 ret = set_metadata_socket(ctx, consumer_sockpoll, client_socket);
3138 if (ret) {
3139 if (ret > 0) {
3140 /* should exit */
3141 err = 0;
3142 }
3143 goto end;
3144 }
3145
3146 /* This socket is not useful anymore. */
3147 ret = close(client_socket);
3148 if (ret < 0) {
3149 PERROR("close client_socket");
3150 }
3151 client_socket = -1;
3152
3153 /* update the polling structure to poll on the established socket */
3154 consumer_sockpoll[1].fd = sock;
3155 consumer_sockpoll[1].events = POLLIN | POLLPRI;
3156
3157 while (1) {
3158 health_code_update();
3159
3160 health_poll_entry();
3161 ret = lttng_consumer_poll_socket(consumer_sockpoll);
3162 health_poll_exit();
3163 if (ret) {
3164 if (ret > 0) {
3165 /* should exit */
3166 err = 0;
3167 }
3168 goto end;
3169 }
3170 DBG("Incoming command on sock");
3171 ret = lttng_consumer_recv_cmd(ctx, sock, consumer_sockpoll);
3172 if (ret <= 0) {
3173 /*
3174 * This could simply be a session daemon quitting. Don't output
3175 * ERR() here.
3176 */
3177 DBG("Communication interrupted on command socket");
3178 err = 0;
3179 goto end;
3180 }
3181 if (consumer_quit) {
3182 DBG("consumer_thread_receive_fds received quit from signal");
3183 err = 0; /* All is OK */
3184 goto end;
3185 }
3186 DBG("received command on sock");
3187 }
3188 /* All is OK */
3189 err = 0;
3190
3191 end:
3192 DBG("Consumer thread sessiond poll exiting");
3193
3194 /*
3195 * Close metadata streams since the producer is the session daemon which
3196 * just died.
3197 *
3198 * NOTE: for now, this only applies to the UST tracer.
3199 */
3200 lttng_consumer_close_all_metadata();
3201
3202 /*
3203 * when all fds have hung up, the polling thread
3204 * can exit cleanly
3205 */
3206 consumer_quit = 1;
3207
3208 /*
3209 * Notify the data poll thread to poll back again and test the
3210 * consumer_quit state that we just set so to quit gracefully.
3211 */
3212 notify_thread_lttng_pipe(ctx->consumer_data_pipe);
3213
3214 notify_channel_pipe(ctx, NULL, -1, CONSUMER_CHANNEL_QUIT);
3215
3216 notify_health_quit_pipe(health_quit_pipe);
3217
3218 /* Cleaning up possibly open sockets. */
3219 if (sock >= 0) {
3220 ret = close(sock);
3221 if (ret < 0) {
3222 PERROR("close sock sessiond poll");
3223 }
3224 }
3225 if (client_socket >= 0) {
3226 ret = close(client_socket);
3227 if (ret < 0) {
3228 PERROR("close client_socket sessiond poll");
3229 }
3230 }
3231
3232 error_testpoint:
3233 if (err) {
3234 health_error();
3235 ERR("Health error occurred in %s", __func__);
3236 }
3237 health_unregister(health_consumerd);
3238
3239 rcu_unregister_thread();
3240 return NULL;
3241 }
3242
3243 ssize_t lttng_consumer_read_subbuffer(struct lttng_consumer_stream *stream,
3244 struct lttng_consumer_local_data *ctx)
3245 {
3246 ssize_t ret;
3247
3248 pthread_mutex_lock(&stream->lock);
3249 if (stream->metadata_flag) {
3250 pthread_mutex_lock(&stream->metadata_rdv_lock);
3251 }
3252
3253 switch (consumer_data.type) {
3254 case LTTNG_CONSUMER_KERNEL:
3255 ret = lttng_kconsumer_read_subbuffer(stream, ctx);
3256 break;
3257 case LTTNG_CONSUMER32_UST:
3258 case LTTNG_CONSUMER64_UST:
3259 ret = lttng_ustconsumer_read_subbuffer(stream, ctx);
3260 break;
3261 default:
3262 ERR("Unknown consumer_data type");
3263 assert(0);
3264 ret = -ENOSYS;
3265 break;
3266 }
3267
3268 if (stream->metadata_flag) {
3269 pthread_cond_broadcast(&stream->metadata_rdv);
3270 pthread_mutex_unlock(&stream->metadata_rdv_lock);
3271 }
3272 pthread_mutex_unlock(&stream->lock);
3273 return ret;
3274 }
3275
3276 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream *stream)
3277 {
3278 switch (consumer_data.type) {
3279 case LTTNG_CONSUMER_KERNEL:
3280 return lttng_kconsumer_on_recv_stream(stream);
3281 case LTTNG_CONSUMER32_UST:
3282 case LTTNG_CONSUMER64_UST:
3283 return lttng_ustconsumer_on_recv_stream(stream);
3284 default:
3285 ERR("Unknown consumer_data type");
3286 assert(0);
3287 return -ENOSYS;
3288 }
3289 }
3290
3291 /*
3292 * Allocate and set consumer data hash tables.
3293 */
3294 int lttng_consumer_init(void)
3295 {
3296 consumer_data.channel_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
3297 if (!consumer_data.channel_ht) {
3298 goto error;
3299 }
3300
3301 consumer_data.relayd_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
3302 if (!consumer_data.relayd_ht) {
3303 goto error;
3304 }
3305
3306 consumer_data.stream_list_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
3307 if (!consumer_data.stream_list_ht) {
3308 goto error;
3309 }
3310
3311 consumer_data.stream_per_chan_id_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
3312 if (!consumer_data.stream_per_chan_id_ht) {
3313 goto error;
3314 }
3315
3316 data_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
3317 if (!data_ht) {
3318 goto error;
3319 }
3320
3321 metadata_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
3322 if (!metadata_ht) {
3323 goto error;
3324 }
3325
3326 return 0;
3327
3328 error:
3329 return -1;
3330 }
3331
3332 /*
3333 * Process the ADD_RELAYD command receive by a consumer.
3334 *
3335 * This will create a relayd socket pair and add it to the relayd hash table.
3336 * The caller MUST acquire a RCU read side lock before calling it.
3337 */
3338 int consumer_add_relayd_socket(uint64_t net_seq_idx, int sock_type,
3339 struct lttng_consumer_local_data *ctx, int sock,
3340 struct pollfd *consumer_sockpoll,
3341 struct lttcomm_relayd_sock *relayd_sock, uint64_t sessiond_id,
3342 uint64_t relayd_session_id)
3343 {
3344 int fd = -1, ret = -1, relayd_created = 0;
3345 enum lttcomm_return_code ret_code = LTTCOMM_CONSUMERD_SUCCESS;
3346 struct consumer_relayd_sock_pair *relayd = NULL;
3347
3348 assert(ctx);
3349 assert(relayd_sock);
3350
3351 DBG("Consumer adding relayd socket (idx: %" PRIu64 ")", net_seq_idx);
3352
3353 /* Get relayd reference if exists. */
3354 relayd = consumer_find_relayd(net_seq_idx);
3355 if (relayd == NULL) {
3356 assert(sock_type == LTTNG_STREAM_CONTROL);
3357 /* Not found. Allocate one. */
3358 relayd = consumer_allocate_relayd_sock_pair(net_seq_idx);
3359 if (relayd == NULL) {
3360 ret = -ENOMEM;
3361 ret_code = LTTCOMM_CONSUMERD_ENOMEM;
3362 goto error;
3363 } else {
3364 relayd->sessiond_session_id = sessiond_id;
3365 relayd_created = 1;
3366 }
3367
3368 /*
3369 * This code path MUST continue to the consumer send status message to
3370 * we can notify the session daemon and continue our work without
3371 * killing everything.
3372 */
3373 } else {
3374 /*
3375 * relayd key should never be found for control socket.
3376 */
3377 assert(sock_type != LTTNG_STREAM_CONTROL);
3378 }
3379
3380 /* First send a status message before receiving the fds. */
3381 ret = consumer_send_status_msg(sock, LTTCOMM_CONSUMERD_SUCCESS);
3382 if (ret < 0) {
3383 /* Somehow, the session daemon is not responding anymore. */
3384 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_FATAL);
3385 goto error_nosignal;
3386 }
3387
3388 /* Poll on consumer socket. */
3389 ret = lttng_consumer_poll_socket(consumer_sockpoll);
3390 if (ret) {
3391 /* Needing to exit in the middle of a command: error. */
3392 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
3393 ret = -EINTR;
3394 goto error_nosignal;
3395 }
3396
3397 /* Get relayd socket from session daemon */
3398 ret = lttcomm_recv_fds_unix_sock(sock, &fd, 1);
3399 if (ret != sizeof(fd)) {
3400 ret = -1;
3401 fd = -1; /* Just in case it gets set with an invalid value. */
3402
3403 /*
3404 * Failing to receive FDs might indicate a major problem such as
3405 * reaching a fd limit during the receive where the kernel returns a
3406 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3407 * don't take any chances and stop everything.
3408 *
3409 * XXX: Feature request #558 will fix that and avoid this possible
3410 * issue when reaching the fd limit.
3411 */
3412 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_ERROR_RECV_FD);
3413 ret_code = LTTCOMM_CONSUMERD_ERROR_RECV_FD;
3414 goto error;
3415 }
3416
3417 /* Copy socket information and received FD */
3418 switch (sock_type) {
3419 case LTTNG_STREAM_CONTROL:
3420 /* Copy received lttcomm socket */
3421 lttcomm_copy_sock(&relayd->control_sock.sock, &relayd_sock->sock);
3422 ret = lttcomm_create_sock(&relayd->control_sock.sock);
3423 /* Handle create_sock error. */
3424 if (ret < 0) {
3425 ret_code = LTTCOMM_CONSUMERD_ENOMEM;
3426 goto error;
3427 }
3428 /*
3429 * Close the socket created internally by
3430 * lttcomm_create_sock, so we can replace it by the one
3431 * received from sessiond.
3432 */
3433 if (close(relayd->control_sock.sock.fd)) {
3434 PERROR("close");
3435 }
3436
3437 /* Assign new file descriptor */
3438 relayd->control_sock.sock.fd = fd;
3439 fd = -1; /* For error path */
3440 /* Assign version values. */
3441 relayd->control_sock.major = relayd_sock->major;
3442 relayd->control_sock.minor = relayd_sock->minor;
3443
3444 relayd->relayd_session_id = relayd_session_id;
3445
3446 break;
3447 case LTTNG_STREAM_DATA:
3448 /* Copy received lttcomm socket */
3449 lttcomm_copy_sock(&relayd->data_sock.sock, &relayd_sock->sock);
3450 ret = lttcomm_create_sock(&relayd->data_sock.sock);
3451 /* Handle create_sock error. */
3452 if (ret < 0) {
3453 ret_code = LTTCOMM_CONSUMERD_ENOMEM;
3454 goto error;
3455 }
3456 /*
3457 * Close the socket created internally by
3458 * lttcomm_create_sock, so we can replace it by the one
3459 * received from sessiond.
3460 */
3461 if (close(relayd->data_sock.sock.fd)) {
3462 PERROR("close");
3463 }
3464
3465 /* Assign new file descriptor */
3466 relayd->data_sock.sock.fd = fd;
3467 fd = -1; /* for eventual error paths */
3468 /* Assign version values. */
3469 relayd->data_sock.major = relayd_sock->major;
3470 relayd->data_sock.minor = relayd_sock->minor;
3471 break;
3472 default:
3473 ERR("Unknown relayd socket type (%d)", sock_type);
3474 ret = -1;
3475 ret_code = LTTCOMM_CONSUMERD_FATAL;
3476 goto error;
3477 }
3478
3479 DBG("Consumer %s socket created successfully with net idx %" PRIu64 " (fd: %d)",
3480 sock_type == LTTNG_STREAM_CONTROL ? "control" : "data",
3481 relayd->net_seq_idx, fd);
3482
3483 /* We successfully added the socket. Send status back. */
3484 ret = consumer_send_status_msg(sock, ret_code);
3485 if (ret < 0) {
3486 /* Somehow, the session daemon is not responding anymore. */
3487 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_FATAL);
3488 goto error_nosignal;
3489 }
3490
3491 /*
3492 * Add relayd socket pair to consumer data hashtable. If object already
3493 * exists or on error, the function gracefully returns.
3494 */
3495 add_relayd(relayd);
3496
3497 /* All good! */
3498 return 0;
3499
3500 error:
3501 if (consumer_send_status_msg(sock, ret_code) < 0) {
3502 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_FATAL);
3503 }
3504
3505 error_nosignal:
3506 /* Close received socket if valid. */
3507 if (fd >= 0) {
3508 if (close(fd)) {
3509 PERROR("close received socket");
3510 }
3511 }
3512
3513 if (relayd_created) {
3514 free(relayd);
3515 }
3516
3517 return ret;
3518 }
3519
3520 /*
3521 * Try to lock the stream mutex.
3522 *
3523 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
3524 */
3525 static int stream_try_lock(struct lttng_consumer_stream *stream)
3526 {
3527 int ret;
3528
3529 assert(stream);
3530
3531 /*
3532 * Try to lock the stream mutex. On failure, we know that the stream is
3533 * being used else where hence there is data still being extracted.
3534 */
3535 ret = pthread_mutex_trylock(&stream->lock);
3536 if (ret) {
3537 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
3538 ret = 0;
3539 goto end;
3540 }
3541
3542 ret = 1;
3543
3544 end:
3545 return ret;
3546 }
3547
3548 /*
3549 * Search for a relayd associated to the session id and return the reference.
3550 *
3551 * A rcu read side lock MUST be acquire before calling this function and locked
3552 * until the relayd object is no longer necessary.
3553 */
3554 static struct consumer_relayd_sock_pair *find_relayd_by_session_id(uint64_t id)
3555 {
3556 struct lttng_ht_iter iter;
3557 struct consumer_relayd_sock_pair *relayd = NULL;
3558
3559 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3560 cds_lfht_for_each_entry(consumer_data.relayd_ht->ht, &iter.iter, relayd,
3561 node.node) {
3562 /*
3563 * Check by sessiond id which is unique here where the relayd session
3564 * id might not be when having multiple relayd.
3565 */
3566 if (relayd->sessiond_session_id == id) {
3567 /* Found the relayd. There can be only one per id. */
3568 goto found;
3569 }
3570 }
3571
3572 return NULL;
3573
3574 found:
3575 return relayd;
3576 }
3577
3578 /*
3579 * Check if for a given session id there is still data needed to be extract
3580 * from the buffers.
3581 *
3582 * Return 1 if data is pending or else 0 meaning ready to be read.
3583 */
3584 int consumer_data_pending(uint64_t id)
3585 {
3586 int ret;
3587 struct lttng_ht_iter iter;
3588 struct lttng_ht *ht;
3589 struct lttng_consumer_stream *stream;
3590 struct consumer_relayd_sock_pair *relayd = NULL;
3591 int (*data_pending)(struct lttng_consumer_stream *);
3592
3593 DBG("Consumer data pending command on session id %" PRIu64, id);
3594
3595 rcu_read_lock();
3596 pthread_mutex_lock(&consumer_data.lock);
3597
3598 switch (consumer_data.type) {
3599 case LTTNG_CONSUMER_KERNEL:
3600 data_pending = lttng_kconsumer_data_pending;
3601 break;
3602 case LTTNG_CONSUMER32_UST:
3603 case LTTNG_CONSUMER64_UST:
3604 data_pending = lttng_ustconsumer_data_pending;
3605 break;
3606 default:
3607 ERR("Unknown consumer data type");
3608 assert(0);
3609 }
3610
3611 /* Ease our life a bit */
3612 ht = consumer_data.stream_list_ht;
3613
3614 relayd = find_relayd_by_session_id(id);
3615 if (relayd) {
3616 /* Send init command for data pending. */
3617 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
3618 ret = relayd_begin_data_pending(&relayd->control_sock,
3619 relayd->relayd_session_id);
3620 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
3621 if (ret < 0) {
3622 /* Communication error thus the relayd so no data pending. */
3623 goto data_not_pending;
3624 }
3625 }
3626
3627 cds_lfht_for_each_entry_duplicate(ht->ht,
3628 ht->hash_fct(&id, lttng_ht_seed),
3629 ht->match_fct, &id,
3630 &iter.iter, stream, node_session_id.node) {
3631 /* If this call fails, the stream is being used hence data pending. */
3632 ret = stream_try_lock(stream);
3633 if (!ret) {
3634 goto data_pending;
3635 }
3636
3637 /*
3638 * A removed node from the hash table indicates that the stream has
3639 * been deleted thus having a guarantee that the buffers are closed
3640 * on the consumer side. However, data can still be transmitted
3641 * over the network so don't skip the relayd check.
3642 */
3643 ret = cds_lfht_is_node_deleted(&stream->node.node);
3644 if (!ret) {
3645 /* Check the stream if there is data in the buffers. */
3646 ret = data_pending(stream);
3647 if (ret == 1) {
3648 pthread_mutex_unlock(&stream->lock);
3649 goto data_pending;
3650 }
3651 }
3652
3653 /* Relayd check */
3654 if (relayd) {
3655 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
3656 if (stream->metadata_flag) {
3657 ret = relayd_quiescent_control(&relayd->control_sock,
3658 stream->relayd_stream_id);
3659 } else {
3660 ret = relayd_data_pending(&relayd->control_sock,
3661 stream->relayd_stream_id,
3662 stream->next_net_seq_num - 1);
3663 }
3664 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
3665 if (ret == 1) {
3666 pthread_mutex_unlock(&stream->lock);
3667 goto data_pending;
3668 }
3669 }
3670 pthread_mutex_unlock(&stream->lock);
3671 }
3672
3673 if (relayd) {
3674 unsigned int is_data_inflight = 0;
3675
3676 /* Send init command for data pending. */
3677 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
3678 ret = relayd_end_data_pending(&relayd->control_sock,
3679 relayd->relayd_session_id, &is_data_inflight);
3680 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
3681 if (ret < 0) {
3682 goto data_not_pending;
3683 }
3684 if (is_data_inflight) {
3685 goto data_pending;
3686 }
3687 }
3688
3689 /*
3690 * Finding _no_ node in the hash table and no inflight data means that the
3691 * stream(s) have been removed thus data is guaranteed to be available for
3692 * analysis from the trace files.
3693 */
3694
3695 data_not_pending:
3696 /* Data is available to be read by a viewer. */
3697 pthread_mutex_unlock(&consumer_data.lock);
3698 rcu_read_unlock();
3699 return 0;
3700
3701 data_pending:
3702 /* Data is still being extracted from buffers. */
3703 pthread_mutex_unlock(&consumer_data.lock);
3704 rcu_read_unlock();
3705 return 1;
3706 }
3707
3708 /*
3709 * Send a ret code status message to the sessiond daemon.
3710 *
3711 * Return the sendmsg() return value.
3712 */
3713 int consumer_send_status_msg(int sock, int ret_code)
3714 {
3715 struct lttcomm_consumer_status_msg msg;
3716
3717 memset(&msg, 0, sizeof(msg));
3718 msg.ret_code = ret_code;
3719
3720 return lttcomm_send_unix_sock(sock, &msg, sizeof(msg));
3721 }
3722
3723 /*
3724 * Send a channel status message to the sessiond daemon.
3725 *
3726 * Return the sendmsg() return value.
3727 */
3728 int consumer_send_status_channel(int sock,
3729 struct lttng_consumer_channel *channel)
3730 {
3731 struct lttcomm_consumer_status_channel msg;
3732
3733 assert(sock >= 0);
3734
3735 memset(&msg, 0, sizeof(msg));
3736 if (!channel) {
3737 msg.ret_code = LTTCOMM_CONSUMERD_CHANNEL_FAIL;
3738 } else {
3739 msg.ret_code = LTTCOMM_CONSUMERD_SUCCESS;
3740 msg.key = channel->key;
3741 msg.stream_count = channel->streams.count;
3742 }
3743
3744 return lttcomm_send_unix_sock(sock, &msg, sizeof(msg));
3745 }
3746
3747 unsigned long consumer_get_consume_start_pos(unsigned long consumed_pos,
3748 unsigned long produced_pos,