Fix: add missing semicolons after MSG, DBG, ERR print macros
[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 ret;
1236 int outfd = stream->out_fd;
1237
1238 /*
1239 * This does a blocking write-and-wait on any page that belongs to the
1240 * subbuffer prior to the one we just wrote.
1241 * Don't care about error values, as these are just hints and ways to
1242 * limit the amount of page cache used.
1243 */
1244 if (orig_offset < stream->max_sb_size) {
1245 return;
1246 }
1247 lttng_sync_file_range(outfd, orig_offset - stream->max_sb_size,
1248 stream->max_sb_size,
1249 SYNC_FILE_RANGE_WAIT_BEFORE
1250 | SYNC_FILE_RANGE_WRITE
1251 | SYNC_FILE_RANGE_WAIT_AFTER);
1252 /*
1253 * Give hints to the kernel about how we access the file:
1254 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1255 * we write it.
1256 *
1257 * We need to call fadvise again after the file grows because the
1258 * kernel does not seem to apply fadvise to non-existing parts of the
1259 * file.
1260 *
1261 * Call fadvise _after_ having waited for the page writeback to
1262 * complete because the dirty page writeback semantic is not well
1263 * defined. So it can be expected to lead to lower throughput in
1264 * streaming.
1265 */
1266 ret = posix_fadvise(outfd, orig_offset - stream->max_sb_size,
1267 stream->max_sb_size, POSIX_FADV_DONTNEED);
1268 if (ret && ret != -ENOSYS) {
1269 errno = -ret;
1270 PERROR("posix_fadvise");
1271 }
1272 }
1273
1274 /*
1275 * Initialise the necessary environnement :
1276 * - create a new context
1277 * - create the poll_pipe
1278 * - create the should_quit pipe (for signal handler)
1279 * - create the thread pipe (for splice)
1280 *
1281 * Takes a function pointer as argument, this function is called when data is
1282 * available on a buffer. This function is responsible to do the
1283 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1284 * buffer configuration and then kernctl_put_next_subbuf at the end.
1285 *
1286 * Returns a pointer to the new context or NULL on error.
1287 */
1288 struct lttng_consumer_local_data *lttng_consumer_create(
1289 enum lttng_consumer_type type,
1290 ssize_t (*buffer_ready)(struct lttng_consumer_stream *stream,
1291 struct lttng_consumer_local_data *ctx),
1292 int (*recv_channel)(struct lttng_consumer_channel *channel),
1293 int (*recv_stream)(struct lttng_consumer_stream *stream),
1294 int (*update_stream)(uint64_t stream_key, uint32_t state))
1295 {
1296 int ret;
1297 struct lttng_consumer_local_data *ctx;
1298
1299 assert(consumer_data.type == LTTNG_CONSUMER_UNKNOWN ||
1300 consumer_data.type == type);
1301 consumer_data.type = type;
1302
1303 ctx = zmalloc(sizeof(struct lttng_consumer_local_data));
1304 if (ctx == NULL) {
1305 PERROR("allocating context");
1306 goto error;
1307 }
1308
1309 ctx->consumer_error_socket = -1;
1310 ctx->consumer_metadata_socket = -1;
1311 pthread_mutex_init(&ctx->metadata_socket_lock, NULL);
1312 /* assign the callbacks */
1313 ctx->on_buffer_ready = buffer_ready;
1314 ctx->on_recv_channel = recv_channel;
1315 ctx->on_recv_stream = recv_stream;
1316 ctx->on_update_stream = update_stream;
1317
1318 ctx->consumer_data_pipe = lttng_pipe_open(0);
1319 if (!ctx->consumer_data_pipe) {
1320 goto error_poll_pipe;
1321 }
1322
1323 ctx->consumer_wakeup_pipe = lttng_pipe_open(0);
1324 if (!ctx->consumer_wakeup_pipe) {
1325 goto error_wakeup_pipe;
1326 }
1327
1328 ret = pipe(ctx->consumer_should_quit);
1329 if (ret < 0) {
1330 PERROR("Error creating recv pipe");
1331 goto error_quit_pipe;
1332 }
1333
1334 ret = pipe(ctx->consumer_channel_pipe);
1335 if (ret < 0) {
1336 PERROR("Error creating channel pipe");
1337 goto error_channel_pipe;
1338 }
1339
1340 ctx->consumer_metadata_pipe = lttng_pipe_open(0);
1341 if (!ctx->consumer_metadata_pipe) {
1342 goto error_metadata_pipe;
1343 }
1344
1345 return ctx;
1346
1347 error_metadata_pipe:
1348 utils_close_pipe(ctx->consumer_channel_pipe);
1349 error_channel_pipe:
1350 utils_close_pipe(ctx->consumer_should_quit);
1351 error_quit_pipe:
1352 lttng_pipe_destroy(ctx->consumer_wakeup_pipe);
1353 error_wakeup_pipe:
1354 lttng_pipe_destroy(ctx->consumer_data_pipe);
1355 error_poll_pipe:
1356 free(ctx);
1357 error:
1358 return NULL;
1359 }
1360
1361 /*
1362 * Iterate over all streams of the hashtable and free them properly.
1363 */
1364 static void destroy_data_stream_ht(struct lttng_ht *ht)
1365 {
1366 struct lttng_ht_iter iter;
1367 struct lttng_consumer_stream *stream;
1368
1369 if (ht == NULL) {
1370 return;
1371 }
1372
1373 rcu_read_lock();
1374 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
1375 /*
1376 * Ignore return value since we are currently cleaning up so any error
1377 * can't be handled.
1378 */
1379 (void) consumer_del_stream(stream, ht);
1380 }
1381 rcu_read_unlock();
1382
1383 lttng_ht_destroy(ht);
1384 }
1385
1386 /*
1387 * Iterate over all streams of the metadata hashtable and free them
1388 * properly.
1389 */
1390 static void destroy_metadata_stream_ht(struct lttng_ht *ht)
1391 {
1392 struct lttng_ht_iter iter;
1393 struct lttng_consumer_stream *stream;
1394
1395 if (ht == NULL) {
1396 return;
1397 }
1398
1399 rcu_read_lock();
1400 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
1401 /*
1402 * Ignore return value since we are currently cleaning up so any error
1403 * can't be handled.
1404 */
1405 (void) consumer_del_metadata_stream(stream, ht);
1406 }
1407 rcu_read_unlock();
1408
1409 lttng_ht_destroy(ht);
1410 }
1411
1412 /*
1413 * Close all fds associated with the instance and free the context.
1414 */
1415 void lttng_consumer_destroy(struct lttng_consumer_local_data *ctx)
1416 {
1417 int ret;
1418
1419 DBG("Consumer destroying it. Closing everything.");
1420
1421 if (!ctx) {
1422 return;
1423 }
1424
1425 destroy_data_stream_ht(data_ht);
1426 destroy_metadata_stream_ht(metadata_ht);
1427
1428 ret = close(ctx->consumer_error_socket);
1429 if (ret) {
1430 PERROR("close");
1431 }
1432 ret = close(ctx->consumer_metadata_socket);
1433 if (ret) {
1434 PERROR("close");
1435 }
1436 utils_close_pipe(ctx->consumer_channel_pipe);
1437 lttng_pipe_destroy(ctx->consumer_data_pipe);
1438 lttng_pipe_destroy(ctx->consumer_metadata_pipe);
1439 lttng_pipe_destroy(ctx->consumer_wakeup_pipe);
1440 utils_close_pipe(ctx->consumer_should_quit);
1441
1442 unlink(ctx->consumer_command_sock_path);
1443 free(ctx);
1444 }
1445
1446 /*
1447 * Write the metadata stream id on the specified file descriptor.
1448 */
1449 static int write_relayd_metadata_id(int fd,
1450 struct lttng_consumer_stream *stream,
1451 struct consumer_relayd_sock_pair *relayd, unsigned long padding)
1452 {
1453 ssize_t ret;
1454 struct lttcomm_relayd_metadata_payload hdr;
1455
1456 hdr.stream_id = htobe64(stream->relayd_stream_id);
1457 hdr.padding_size = htobe32(padding);
1458 ret = lttng_write(fd, (void *) &hdr, sizeof(hdr));
1459 if (ret < sizeof(hdr)) {
1460 /*
1461 * This error means that the fd's end is closed so ignore the PERROR
1462 * not to clubber the error output since this can happen in a normal
1463 * code path.
1464 */
1465 if (errno != EPIPE) {
1466 PERROR("write metadata stream id");
1467 }
1468 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno);
1469 /*
1470 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1471 * handle writting the missing part so report that as an error and
1472 * don't lie to the caller.
1473 */
1474 ret = -1;
1475 goto end;
1476 }
1477 DBG("Metadata stream id %" PRIu64 " with padding %lu written before data",
1478 stream->relayd_stream_id, padding);
1479
1480 end:
1481 return (int) ret;
1482 }
1483
1484 /*
1485 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1486 * core function for writing trace buffers to either the local filesystem or
1487 * the network.
1488 *
1489 * It must be called with the stream lock held.
1490 *
1491 * Careful review MUST be put if any changes occur!
1492 *
1493 * Returns the number of bytes written
1494 */
1495 ssize_t lttng_consumer_on_read_subbuffer_mmap(
1496 struct lttng_consumer_local_data *ctx,
1497 struct lttng_consumer_stream *stream, unsigned long len,
1498 unsigned long padding,
1499 struct ctf_packet_index *index)
1500 {
1501 unsigned long mmap_offset;
1502 void *mmap_base;
1503 ssize_t ret = 0;
1504 off_t orig_offset = stream->out_fd_offset;
1505 /* Default is on the disk */
1506 int outfd = stream->out_fd;
1507 struct consumer_relayd_sock_pair *relayd = NULL;
1508 unsigned int relayd_hang_up = 0;
1509
1510 /* RCU lock for the relayd pointer */
1511 rcu_read_lock();
1512
1513 /* Flag that the current stream if set for network streaming. */
1514 if (stream->net_seq_idx != (uint64_t) -1ULL) {
1515 relayd = consumer_find_relayd(stream->net_seq_idx);
1516 if (relayd == NULL) {
1517 ret = -EPIPE;
1518 goto end;
1519 }
1520 }
1521
1522 /* get the offset inside the fd to mmap */
1523 switch (consumer_data.type) {
1524 case LTTNG_CONSUMER_KERNEL:
1525 mmap_base = stream->mmap_base;
1526 ret = kernctl_get_mmap_read_offset(stream->wait_fd, &mmap_offset);
1527 if (ret < 0) {
1528 ret = -errno;
1529 PERROR("tracer ctl get_mmap_read_offset");
1530 goto end;
1531 }
1532 break;
1533 case LTTNG_CONSUMER32_UST:
1534 case LTTNG_CONSUMER64_UST:
1535 mmap_base = lttng_ustctl_get_mmap_base(stream);
1536 if (!mmap_base) {
1537 ERR("read mmap get mmap base for stream %s", stream->name);
1538 ret = -EPERM;
1539 goto end;
1540 }
1541 ret = lttng_ustctl_get_mmap_read_offset(stream, &mmap_offset);
1542 if (ret != 0) {
1543 PERROR("tracer ctl get_mmap_read_offset");
1544 ret = -EINVAL;
1545 goto end;
1546 }
1547 break;
1548 default:
1549 ERR("Unknown consumer_data type");
1550 assert(0);
1551 }
1552
1553 /* Handle stream on the relayd if the output is on the network */
1554 if (relayd) {
1555 unsigned long netlen = len;
1556
1557 /*
1558 * Lock the control socket for the complete duration of the function
1559 * since from this point on we will use the socket.
1560 */
1561 if (stream->metadata_flag) {
1562 /* Metadata requires the control socket. */
1563 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
1564 if (stream->reset_metadata_flag) {
1565 ret = relayd_reset_metadata(&relayd->control_sock,
1566 stream->relayd_stream_id,
1567 stream->metadata_version);
1568 if (ret < 0) {
1569 relayd_hang_up = 1;
1570 goto write_error;
1571 }
1572 stream->reset_metadata_flag = 0;
1573 }
1574 netlen += sizeof(struct lttcomm_relayd_metadata_payload);
1575 }
1576
1577 ret = write_relayd_stream_header(stream, netlen, padding, relayd);
1578 if (ret < 0) {
1579 relayd_hang_up = 1;
1580 goto write_error;
1581 }
1582 /* Use the returned socket. */
1583 outfd = ret;
1584
1585 /* Write metadata stream id before payload */
1586 if (stream->metadata_flag) {
1587 ret = write_relayd_metadata_id(outfd, stream, relayd, padding);
1588 if (ret < 0) {
1589 relayd_hang_up = 1;
1590 goto write_error;
1591 }
1592 }
1593 } else {
1594 /* No streaming, we have to set the len with the full padding */
1595 len += padding;
1596
1597 if (stream->metadata_flag && stream->reset_metadata_flag) {
1598 ret = utils_truncate_stream_file(stream->out_fd, 0);
1599 if (ret < 0) {
1600 ERR("Reset metadata file");
1601 goto end;
1602 }
1603 stream->reset_metadata_flag = 0;
1604 }
1605
1606 /*
1607 * Check if we need to change the tracefile before writing the packet.
1608 */
1609 if (stream->chan->tracefile_size > 0 &&
1610 (stream->tracefile_size_current + len) >
1611 stream->chan->tracefile_size) {
1612 ret = utils_rotate_stream_file(stream->chan->pathname,
1613 stream->name, stream->chan->tracefile_size,
1614 stream->chan->tracefile_count, stream->uid, stream->gid,
1615 stream->out_fd, &(stream->tracefile_count_current),
1616 &stream->out_fd);
1617 if (ret < 0) {
1618 ERR("Rotating output file");
1619 goto end;
1620 }
1621 outfd = stream->out_fd;
1622
1623 if (stream->index_fd >= 0) {
1624 ret = close(stream->index_fd);
1625 if (ret < 0) {
1626 PERROR("Closing index");
1627 goto end;
1628 }
1629 stream->index_fd = -1;
1630 ret = index_create_file(stream->chan->pathname,
1631 stream->name, stream->uid, stream->gid,
1632 stream->chan->tracefile_size,
1633 stream->tracefile_count_current);
1634 if (ret < 0) {
1635 goto end;
1636 }
1637 stream->index_fd = ret;
1638 }
1639
1640 /* Reset current size because we just perform a rotation. */
1641 stream->tracefile_size_current = 0;
1642 stream->out_fd_offset = 0;
1643 orig_offset = 0;
1644 }
1645 stream->tracefile_size_current += len;
1646 if (index) {
1647 index->offset = htobe64(stream->out_fd_offset);
1648 }
1649 }
1650
1651 /*
1652 * This call guarantee that len or less is returned. It's impossible to
1653 * receive a ret value that is bigger than len.
1654 */
1655 ret = lttng_write(outfd, mmap_base + mmap_offset, len);
1656 DBG("Consumer mmap write() ret %zd (len %lu)", ret, len);
1657 if (ret < 0 || ((size_t) ret != len)) {
1658 /*
1659 * Report error to caller if nothing was written else at least send the
1660 * amount written.
1661 */
1662 if (ret < 0) {
1663 ret = -errno;
1664 }
1665 relayd_hang_up = 1;
1666
1667 /* Socket operation failed. We consider the relayd dead */
1668 if (errno == EPIPE || errno == EINVAL || errno == EBADF) {
1669 /*
1670 * This is possible if the fd is closed on the other side
1671 * (outfd) or any write problem. It can be verbose a bit for a
1672 * normal execution if for instance the relayd is stopped
1673 * abruptly. This can happen so set this to a DBG statement.
1674 */
1675 DBG("Consumer mmap write detected relayd hang up");
1676 } else {
1677 /* Unhandled error, print it and stop function right now. */
1678 PERROR("Error in write mmap (ret %zd != len %lu)", ret, len);
1679 }
1680 goto write_error;
1681 }
1682 stream->output_written += ret;
1683
1684 /* This call is useless on a socket so better save a syscall. */
1685 if (!relayd) {
1686 /* This won't block, but will start writeout asynchronously */
1687 lttng_sync_file_range(outfd, stream->out_fd_offset, len,
1688 SYNC_FILE_RANGE_WRITE);
1689 stream->out_fd_offset += len;
1690 }
1691 lttng_consumer_sync_trace_file(stream, orig_offset);
1692
1693 write_error:
1694 /*
1695 * This is a special case that the relayd has closed its socket. Let's
1696 * cleanup the relayd object and all associated streams.
1697 */
1698 if (relayd && relayd_hang_up) {
1699 cleanup_relayd(relayd, ctx);
1700 }
1701
1702 end:
1703 /* Unlock only if ctrl socket used */
1704 if (relayd && stream->metadata_flag) {
1705 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
1706 }
1707
1708 rcu_read_unlock();
1709 return ret;
1710 }
1711
1712 /*
1713 * Splice the data from the ring buffer to the tracefile.
1714 *
1715 * It must be called with the stream lock held.
1716 *
1717 * Returns the number of bytes spliced.
1718 */
1719 ssize_t lttng_consumer_on_read_subbuffer_splice(
1720 struct lttng_consumer_local_data *ctx,
1721 struct lttng_consumer_stream *stream, unsigned long len,
1722 unsigned long padding,
1723 struct ctf_packet_index *index)
1724 {
1725 ssize_t ret = 0, written = 0, ret_splice = 0;
1726 loff_t offset = 0;
1727 off_t orig_offset = stream->out_fd_offset;
1728 int fd = stream->wait_fd;
1729 /* Default is on the disk */
1730 int outfd = stream->out_fd;
1731 struct consumer_relayd_sock_pair *relayd = NULL;
1732 int *splice_pipe;
1733 unsigned int relayd_hang_up = 0;
1734
1735 switch (consumer_data.type) {
1736 case LTTNG_CONSUMER_KERNEL:
1737 break;
1738 case LTTNG_CONSUMER32_UST:
1739 case LTTNG_CONSUMER64_UST:
1740 /* Not supported for user space tracing */
1741 return -ENOSYS;
1742 default:
1743 ERR("Unknown consumer_data type");
1744 assert(0);
1745 }
1746
1747 /* RCU lock for the relayd pointer */
1748 rcu_read_lock();
1749
1750 /* Flag that the current stream if set for network streaming. */
1751 if (stream->net_seq_idx != (uint64_t) -1ULL) {
1752 relayd = consumer_find_relayd(stream->net_seq_idx);
1753 if (relayd == NULL) {
1754 written = -ret;
1755 goto end;
1756 }
1757 }
1758 splice_pipe = stream->splice_pipe;
1759
1760 /* Write metadata stream id before payload */
1761 if (relayd) {
1762 unsigned long total_len = len;
1763
1764 if (stream->metadata_flag) {
1765 /*
1766 * Lock the control socket for the complete duration of the function
1767 * since from this point on we will use the socket.
1768 */
1769 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
1770
1771 if (stream->reset_metadata_flag) {
1772 ret = relayd_reset_metadata(&relayd->control_sock,
1773 stream->relayd_stream_id,
1774 stream->metadata_version);
1775 if (ret < 0) {
1776 relayd_hang_up = 1;
1777 goto write_error;
1778 }
1779 stream->reset_metadata_flag = 0;
1780 }
1781 ret = write_relayd_metadata_id(splice_pipe[1], stream, relayd,
1782 padding);
1783 if (ret < 0) {
1784 written = ret;
1785 relayd_hang_up = 1;
1786 goto write_error;
1787 }
1788
1789 total_len += sizeof(struct lttcomm_relayd_metadata_payload);
1790 }
1791
1792 ret = write_relayd_stream_header(stream, total_len, padding, relayd);
1793 if (ret < 0) {
1794 written = ret;
1795 relayd_hang_up = 1;
1796 goto write_error;
1797 }
1798 /* Use the returned socket. */
1799 outfd = ret;
1800 } else {
1801 /* No streaming, we have to set the len with the full padding */
1802 len += padding;
1803
1804 if (stream->metadata_flag && stream->reset_metadata_flag) {
1805 ret = utils_truncate_stream_file(stream->out_fd, 0);
1806 if (ret < 0) {
1807 ERR("Reset metadata file");
1808 goto end;
1809 }
1810 stream->reset_metadata_flag = 0;
1811 }
1812 /*
1813 * Check if we need to change the tracefile before writing the packet.
1814 */
1815 if (stream->chan->tracefile_size > 0 &&
1816 (stream->tracefile_size_current + len) >
1817 stream->chan->tracefile_size) {
1818 ret = utils_rotate_stream_file(stream->chan->pathname,
1819 stream->name, stream->chan->tracefile_size,
1820 stream->chan->tracefile_count, stream->uid, stream->gid,
1821 stream->out_fd, &(stream->tracefile_count_current),
1822 &stream->out_fd);
1823 if (ret < 0) {
1824 written = ret;
1825 ERR("Rotating output file");
1826 goto end;
1827 }
1828 outfd = stream->out_fd;
1829
1830 if (stream->index_fd >= 0) {
1831 ret = close(stream->index_fd);
1832 if (ret < 0) {
1833 PERROR("Closing index");
1834 goto end;
1835 }
1836 stream->index_fd = -1;
1837 ret = index_create_file(stream->chan->pathname,
1838 stream->name, stream->uid, stream->gid,
1839 stream->chan->tracefile_size,
1840 stream->tracefile_count_current);
1841 if (ret < 0) {
1842 written = ret;
1843 goto end;
1844 }
1845 stream->index_fd = ret;
1846 }
1847
1848 /* Reset current size because we just perform a rotation. */
1849 stream->tracefile_size_current = 0;
1850 stream->out_fd_offset = 0;
1851 orig_offset = 0;
1852 }
1853 stream->tracefile_size_current += len;
1854 index->offset = htobe64(stream->out_fd_offset);
1855 }
1856
1857 while (len > 0) {
1858 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1859 (unsigned long)offset, len, fd, splice_pipe[1]);
1860 ret_splice = splice(fd, &offset, splice_pipe[1], NULL, len,
1861 SPLICE_F_MOVE | SPLICE_F_MORE);
1862 DBG("splice chan to pipe, ret %zd", ret_splice);
1863 if (ret_splice < 0) {
1864 ret = errno;
1865 written = -ret;
1866 PERROR("Error in relay splice");
1867 goto splice_error;
1868 }
1869
1870 /* Handle stream on the relayd if the output is on the network */
1871 if (relayd && stream->metadata_flag) {
1872 size_t metadata_payload_size =
1873 sizeof(struct lttcomm_relayd_metadata_payload);
1874
1875 /* Update counter to fit the spliced data */
1876 ret_splice += metadata_payload_size;
1877 len += metadata_payload_size;
1878 /*
1879 * We do this so the return value can match the len passed as
1880 * argument to this function.
1881 */
1882 written -= metadata_payload_size;
1883 }
1884
1885 /* Splice data out */
1886 ret_splice = splice(splice_pipe[0], NULL, outfd, NULL,
1887 ret_splice, SPLICE_F_MOVE | SPLICE_F_MORE);
1888 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
1889 outfd, ret_splice);
1890 if (ret_splice < 0) {
1891 ret = errno;
1892 written = -ret;
1893 relayd_hang_up = 1;
1894 goto write_error;
1895 } else if (ret_splice > len) {
1896 /*
1897 * We don't expect this code path to be executed but you never know
1898 * so this is an extra protection agains a buggy splice().
1899 */
1900 ret = errno;
1901 written += ret_splice;
1902 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice,
1903 len);
1904 goto splice_error;
1905 } else {
1906 /* All good, update current len and continue. */
1907 len -= ret_splice;
1908 }
1909
1910 /* This call is useless on a socket so better save a syscall. */
1911 if (!relayd) {
1912 /* This won't block, but will start writeout asynchronously */
1913 lttng_sync_file_range(outfd, stream->out_fd_offset, ret_splice,
1914 SYNC_FILE_RANGE_WRITE);
1915 stream->out_fd_offset += ret_splice;
1916 }
1917 stream->output_written += ret_splice;
1918 written += ret_splice;
1919 }
1920 lttng_consumer_sync_trace_file(stream, orig_offset);
1921 goto end;
1922
1923 write_error:
1924 /*
1925 * This is a special case that the relayd has closed its socket. Let's
1926 * cleanup the relayd object and all associated streams.
1927 */
1928 if (relayd && relayd_hang_up) {
1929 cleanup_relayd(relayd, ctx);
1930 /* Skip splice error so the consumer does not fail */
1931 goto end;
1932 }
1933
1934 splice_error:
1935 /* send the appropriate error description to sessiond */
1936 switch (ret) {
1937 case EINVAL:
1938 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_EINVAL);
1939 break;
1940 case ENOMEM:
1941 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_ENOMEM);
1942 break;
1943 case ESPIPE:
1944 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_ESPIPE);
1945 break;
1946 }
1947
1948 end:
1949 if (relayd && stream->metadata_flag) {
1950 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
1951 }
1952
1953 rcu_read_unlock();
1954 return written;
1955 }
1956
1957 /*
1958 * Take a snapshot for a specific fd
1959 *
1960 * Returns 0 on success, < 0 on error
1961 */
1962 int lttng_consumer_take_snapshot(struct lttng_consumer_stream *stream)
1963 {
1964 switch (consumer_data.type) {
1965 case LTTNG_CONSUMER_KERNEL:
1966 return lttng_kconsumer_take_snapshot(stream);
1967 case LTTNG_CONSUMER32_UST:
1968 case LTTNG_CONSUMER64_UST:
1969 return lttng_ustconsumer_take_snapshot(stream);
1970 default:
1971 ERR("Unknown consumer_data type");
1972 assert(0);
1973 return -ENOSYS;
1974 }
1975 }
1976
1977 /*
1978 * Get the produced position
1979 *
1980 * Returns 0 on success, < 0 on error
1981 */
1982 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream *stream,
1983 unsigned long *pos)
1984 {
1985 switch (consumer_data.type) {
1986 case LTTNG_CONSUMER_KERNEL:
1987 return lttng_kconsumer_get_produced_snapshot(stream, pos);
1988 case LTTNG_CONSUMER32_UST:
1989 case LTTNG_CONSUMER64_UST:
1990 return lttng_ustconsumer_get_produced_snapshot(stream, pos);
1991 default:
1992 ERR("Unknown consumer_data type");
1993 assert(0);
1994 return -ENOSYS;
1995 }
1996 }
1997
1998 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data *ctx,
1999 int sock, struct pollfd *consumer_sockpoll)
2000 {
2001 switch (consumer_data.type) {
2002 case LTTNG_CONSUMER_KERNEL:
2003 return lttng_kconsumer_recv_cmd(ctx, sock, consumer_sockpoll);
2004 case LTTNG_CONSUMER32_UST:
2005 case LTTNG_CONSUMER64_UST:
2006 return lttng_ustconsumer_recv_cmd(ctx, sock, consumer_sockpoll);
2007 default:
2008 ERR("Unknown consumer_data type");
2009 assert(0);
2010 return -ENOSYS;
2011 }
2012 }
2013
2014 void lttng_consumer_close_all_metadata(void)
2015 {
2016 switch (consumer_data.type) {
2017 case LTTNG_CONSUMER_KERNEL:
2018 /*
2019 * The Kernel consumer has a different metadata scheme so we don't
2020 * close anything because the stream will be closed by the session
2021 * daemon.
2022 */
2023 break;
2024 case LTTNG_CONSUMER32_UST:
2025 case LTTNG_CONSUMER64_UST:
2026 /*
2027 * Close all metadata streams. The metadata hash table is passed and
2028 * this call iterates over it by closing all wakeup fd. This is safe
2029 * because at this point we are sure that the metadata producer is
2030 * either dead or blocked.
2031 */
2032 lttng_ustconsumer_close_all_metadata(metadata_ht);
2033 break;
2034 default:
2035 ERR("Unknown consumer_data type");
2036 assert(0);
2037 }
2038 }
2039
2040 /*
2041 * Clean up a metadata stream and free its memory.
2042 */
2043 void consumer_del_metadata_stream(struct lttng_consumer_stream *stream,
2044 struct lttng_ht *ht)
2045 {
2046 struct lttng_consumer_channel *free_chan = NULL;
2047
2048 assert(stream);
2049 /*
2050 * This call should NEVER receive regular stream. It must always be
2051 * metadata stream and this is crucial for data structure synchronization.
2052 */
2053 assert(stream->metadata_flag);
2054
2055 DBG3("Consumer delete metadata stream %d", stream->wait_fd);
2056
2057 pthread_mutex_lock(&consumer_data.lock);
2058 pthread_mutex_lock(&stream->chan->lock);
2059 pthread_mutex_lock(&stream->lock);
2060
2061 /* Remove any reference to that stream. */
2062 consumer_stream_delete(stream, ht);
2063
2064 /* Close down everything including the relayd if one. */
2065 consumer_stream_close(stream);
2066 /* Destroy tracer buffers of the stream. */
2067 consumer_stream_destroy_buffers(stream);
2068
2069 /* Atomically decrement channel refcount since other threads can use it. */
2070 if (!uatomic_sub_return(&stream->chan->refcount, 1)
2071 && !uatomic_read(&stream->chan->nb_init_stream_left)) {
2072 /* Go for channel deletion! */
2073 free_chan = stream->chan;
2074 }
2075
2076 /*
2077 * Nullify the stream reference so it is not used after deletion. The
2078 * channel lock MUST be acquired before being able to check for a NULL
2079 * pointer value.
2080 */
2081 stream->chan->metadata_stream = NULL;
2082
2083 pthread_mutex_unlock(&stream->lock);
2084 pthread_mutex_unlock(&stream->chan->lock);
2085 pthread_mutex_unlock(&consumer_data.lock);
2086
2087 if (free_chan) {
2088 consumer_del_channel(free_chan);
2089 }
2090
2091 consumer_stream_free(stream);
2092 }
2093
2094 /*
2095 * Action done with the metadata stream when adding it to the consumer internal
2096 * data structures to handle it.
2097 */
2098 int consumer_add_metadata_stream(struct lttng_consumer_stream *stream)
2099 {
2100 struct lttng_ht *ht = metadata_ht;
2101 int ret = 0;
2102 struct lttng_ht_iter iter;
2103 struct lttng_ht_node_u64 *node;
2104
2105 assert(stream);
2106 assert(ht);
2107
2108 DBG3("Adding metadata stream %" PRIu64 " to hash table", stream->key);
2109
2110 pthread_mutex_lock(&consumer_data.lock);
2111 pthread_mutex_lock(&stream->chan->lock);
2112 pthread_mutex_lock(&stream->chan->timer_lock);
2113 pthread_mutex_lock(&stream->lock);
2114
2115 /*
2116 * From here, refcounts are updated so be _careful_ when returning an error
2117 * after this point.
2118 */
2119
2120 rcu_read_lock();
2121
2122 /*
2123 * Lookup the stream just to make sure it does not exist in our internal
2124 * state. This should NEVER happen.
2125 */
2126 lttng_ht_lookup(ht, &stream->key, &iter);
2127 node = lttng_ht_iter_get_node_u64(&iter);
2128 assert(!node);
2129
2130 /*
2131 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2132 * in terms of destroying the associated channel, because the action that
2133 * causes the count to become 0 also causes a stream to be added. The
2134 * channel deletion will thus be triggered by the following removal of this
2135 * stream.
2136 */
2137 if (uatomic_read(&stream->chan->nb_init_stream_left) > 0) {
2138 /* Increment refcount before decrementing nb_init_stream_left */
2139 cmm_smp_wmb();
2140 uatomic_dec(&stream->chan->nb_init_stream_left);
2141 }
2142
2143 lttng_ht_add_unique_u64(ht, &stream->node);
2144
2145 lttng_ht_add_unique_u64(consumer_data.stream_per_chan_id_ht,
2146 &stream->node_channel_id);
2147
2148 /*
2149 * Add stream to the stream_list_ht of the consumer data. No need to steal
2150 * the key since the HT does not use it and we allow to add redundant keys
2151 * into this table.
2152 */
2153 lttng_ht_add_u64(consumer_data.stream_list_ht, &stream->node_session_id);
2154
2155 rcu_read_unlock();
2156
2157 pthread_mutex_unlock(&stream->lock);
2158 pthread_mutex_unlock(&stream->chan->lock);
2159 pthread_mutex_unlock(&stream->chan->timer_lock);
2160 pthread_mutex_unlock(&consumer_data.lock);
2161 return ret;
2162 }
2163
2164 /*
2165 * Delete data stream that are flagged for deletion (endpoint_status).
2166 */
2167 static void validate_endpoint_status_data_stream(void)
2168 {
2169 struct lttng_ht_iter iter;
2170 struct lttng_consumer_stream *stream;
2171
2172 DBG("Consumer delete flagged data stream");
2173
2174 rcu_read_lock();
2175 cds_lfht_for_each_entry(data_ht->ht, &iter.iter, stream, node.node) {
2176 /* Validate delete flag of the stream */
2177 if (stream->endpoint_status == CONSUMER_ENDPOINT_ACTIVE) {
2178 continue;
2179 }
2180 /* Delete it right now */
2181 consumer_del_stream(stream, data_ht);
2182 }
2183 rcu_read_unlock();
2184 }
2185
2186 /*
2187 * Delete metadata stream that are flagged for deletion (endpoint_status).
2188 */
2189 static void validate_endpoint_status_metadata_stream(
2190 struct lttng_poll_event *pollset)
2191 {
2192 struct lttng_ht_iter iter;
2193 struct lttng_consumer_stream *stream;
2194
2195 DBG("Consumer delete flagged metadata stream");
2196
2197 assert(pollset);
2198
2199 rcu_read_lock();
2200 cds_lfht_for_each_entry(metadata_ht->ht, &iter.iter, stream, node.node) {
2201 /* Validate delete flag of the stream */
2202 if (stream->endpoint_status == CONSUMER_ENDPOINT_ACTIVE) {
2203 continue;
2204 }
2205 /*
2206 * Remove from pollset so the metadata thread can continue without
2207 * blocking on a deleted stream.
2208 */
2209 lttng_poll_del(pollset, stream->wait_fd);
2210
2211 /* Delete it right now */
2212 consumer_del_metadata_stream(stream, metadata_ht);
2213 }
2214 rcu_read_unlock();
2215 }
2216
2217 /*
2218 * Thread polls on metadata file descriptor and write them on disk or on the
2219 * network.
2220 */
2221 void *consumer_thread_metadata_poll(void *data)
2222 {
2223 int ret, i, pollfd, err = -1;
2224 uint32_t revents, nb_fd;
2225 struct lttng_consumer_stream *stream = NULL;
2226 struct lttng_ht_iter iter;
2227 struct lttng_ht_node_u64 *node;
2228 struct lttng_poll_event events;
2229 struct lttng_consumer_local_data *ctx = data;
2230 ssize_t len;
2231
2232 rcu_register_thread();
2233
2234 health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_METADATA);
2235
2236 if (testpoint(consumerd_thread_metadata)) {
2237 goto error_testpoint;
2238 }
2239
2240 health_code_update();
2241
2242 DBG("Thread metadata poll started");
2243
2244 /* Size is set to 1 for the consumer_metadata pipe */
2245 ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC);
2246 if (ret < 0) {
2247 ERR("Poll set creation failed");
2248 goto end_poll;
2249 }
2250
2251 ret = lttng_poll_add(&events,
2252 lttng_pipe_get_readfd(ctx->consumer_metadata_pipe), LPOLLIN);
2253 if (ret < 0) {
2254 goto end;
2255 }
2256
2257 /* Main loop */
2258 DBG("Metadata main loop started");
2259
2260 while (1) {
2261 restart:
2262 health_code_update();
2263 health_poll_entry();
2264 DBG("Metadata poll wait");
2265 ret = lttng_poll_wait(&events, -1);
2266 DBG("Metadata poll return from wait with %d fd(s)",
2267 LTTNG_POLL_GETNB(&events));
2268 health_poll_exit();
2269 DBG("Metadata event caught in thread");
2270 if (ret < 0) {
2271 if (errno == EINTR) {
2272 ERR("Poll EINTR caught");
2273 goto restart;
2274 }
2275 if (LTTNG_POLL_GETNB(&events) == 0) {
2276 err = 0; /* All is OK */
2277 }
2278 goto end;
2279 }
2280
2281 nb_fd = ret;
2282
2283 /* From here, the event is a metadata wait fd */
2284 for (i = 0; i < nb_fd; i++) {
2285 health_code_update();
2286
2287 revents = LTTNG_POLL_GETEV(&events, i);
2288 pollfd = LTTNG_POLL_GETFD(&events, i);
2289
2290 if (!revents) {
2291 /* No activity for this FD (poll implementation). */
2292 continue;
2293 }
2294
2295 if (pollfd == lttng_pipe_get_readfd(ctx->consumer_metadata_pipe)) {
2296 if (revents & LPOLLIN) {
2297 ssize_t pipe_len;
2298
2299 pipe_len = lttng_pipe_read(ctx->consumer_metadata_pipe,
2300 &stream, sizeof(stream));
2301 if (pipe_len < sizeof(stream)) {
2302 if (pipe_len < 0) {
2303 PERROR("read metadata stream");
2304 }
2305 /*
2306 * Remove the pipe from the poll set and continue the loop
2307 * since their might be data to consume.
2308 */
2309 lttng_poll_del(&events,
2310 lttng_pipe_get_readfd(ctx->consumer_metadata_pipe));
2311 lttng_pipe_read_close(ctx->consumer_metadata_pipe);
2312 continue;
2313 }
2314
2315 /* A NULL stream means that the state has changed. */
2316 if (stream == NULL) {
2317 /* Check for deleted streams. */
2318 validate_endpoint_status_metadata_stream(&events);
2319 goto restart;
2320 }
2321
2322 DBG("Adding metadata stream %d to poll set",
2323 stream->wait_fd);
2324
2325 /* Add metadata stream to the global poll events list */
2326 lttng_poll_add(&events, stream->wait_fd,
2327 LPOLLIN | LPOLLPRI | LPOLLHUP);
2328 } else if (revents & (LPOLLERR | LPOLLHUP)) {
2329 DBG("Metadata thread pipe hung up");
2330 /*
2331 * Remove the pipe from the poll set and continue the loop
2332 * since their might be data to consume.
2333 */
2334 lttng_poll_del(&events,
2335 lttng_pipe_get_readfd(ctx->consumer_metadata_pipe));
2336 lttng_pipe_read_close(ctx->consumer_metadata_pipe);
2337 continue;
2338 } else {
2339 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
2340 goto end;
2341 }
2342
2343 /* Handle other stream */
2344 continue;
2345 }
2346
2347 rcu_read_lock();
2348 {
2349 uint64_t tmp_id = (uint64_t) pollfd;
2350
2351 lttng_ht_lookup(metadata_ht, &tmp_id, &iter);
2352 }
2353 node = lttng_ht_iter_get_node_u64(&iter);
2354 assert(node);
2355
2356 stream = caa_container_of(node, struct lttng_consumer_stream,
2357 node);
2358
2359 if (revents & (LPOLLIN | LPOLLPRI)) {
2360 /* Get the data out of the metadata file descriptor */
2361 DBG("Metadata available on fd %d", pollfd);
2362 assert(stream->wait_fd == pollfd);
2363
2364 do {
2365 health_code_update();
2366
2367 len = ctx->on_buffer_ready(stream, ctx);
2368 /*
2369 * We don't check the return value here since if we get
2370 * a negative len, it means an error occured thus we
2371 * simply remove it from the poll set and free the
2372 * stream.
2373 */
2374 } while (len > 0);
2375
2376 /* It's ok to have an unavailable sub-buffer */
2377 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2378 /* Clean up stream from consumer and free it. */
2379 lttng_poll_del(&events, stream->wait_fd);
2380 consumer_del_metadata_stream(stream, metadata_ht);
2381 }
2382 } else if (revents & (LPOLLERR | LPOLLHUP)) {
2383 DBG("Metadata fd %d is hup|err.", pollfd);
2384 if (!stream->hangup_flush_done
2385 && (consumer_data.type == LTTNG_CONSUMER32_UST
2386 || consumer_data.type == LTTNG_CONSUMER64_UST)) {
2387 DBG("Attempting to flush and consume the UST buffers");
2388 lttng_ustconsumer_on_stream_hangup(stream);
2389
2390 /* We just flushed the stream now read it. */
2391 do {
2392 health_code_update();
2393
2394 len = ctx->on_buffer_ready(stream, ctx);
2395 /*
2396 * We don't check the return value here since if we get
2397 * a negative len, it means an error occured thus we
2398 * simply remove it from the poll set and free the
2399 * stream.
2400 */
2401 } while (len > 0);
2402 }
2403
2404 lttng_poll_del(&events, stream->wait_fd);
2405 /*
2406 * This call update the channel states, closes file descriptors
2407 * and securely free the stream.
2408 */
2409 consumer_del_metadata_stream(stream, metadata_ht);
2410 } else {
2411 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
2412 rcu_read_unlock();
2413 goto end;
2414 }
2415 /* Release RCU lock for the stream looked up */
2416 rcu_read_unlock();
2417 }
2418 }
2419
2420 /* All is OK */
2421 err = 0;
2422 end:
2423 DBG("Metadata poll thread exiting");
2424
2425 lttng_poll_clean(&events);
2426 end_poll:
2427 error_testpoint:
2428 if (err) {
2429 health_error();
2430 ERR("Health error occurred in %s", __func__);
2431 }
2432 health_unregister(health_consumerd);
2433 rcu_unregister_thread();
2434 return NULL;
2435 }
2436
2437 /*
2438 * This thread polls the fds in the set to consume the data and write
2439 * it to tracefile if necessary.
2440 */
2441 void *consumer_thread_data_poll(void *data)
2442 {
2443 int num_rdy, num_hup, high_prio, ret, i, err = -1;
2444 struct pollfd *pollfd = NULL;
2445 /* local view of the streams */
2446 struct lttng_consumer_stream **local_stream = NULL, *new_stream = NULL;
2447 /* local view of consumer_data.fds_count */
2448 int nb_fd = 0;
2449 struct lttng_consumer_local_data *ctx = data;
2450 ssize_t len;
2451
2452 rcu_register_thread();
2453
2454 health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_DATA);
2455
2456 if (testpoint(consumerd_thread_data)) {
2457 goto error_testpoint;
2458 }
2459
2460 health_code_update();
2461
2462 local_stream = zmalloc(sizeof(struct lttng_consumer_stream *));
2463 if (local_stream == NULL) {
2464 PERROR("local_stream malloc");
2465 goto end;
2466 }
2467
2468 while (1) {
2469 health_code_update();
2470
2471 high_prio = 0;
2472 num_hup = 0;
2473
2474 /*
2475 * the fds set has been updated, we need to update our
2476 * local array as well
2477 */
2478 pthread_mutex_lock(&consumer_data.lock);
2479 if (consumer_data.need_update) {
2480 free(pollfd);
2481 pollfd = NULL;
2482
2483 free(local_stream);
2484 local_stream = NULL;
2485
2486 /*
2487 * Allocate for all fds +1 for the consumer_data_pipe and +1 for
2488 * wake up pipe.
2489 */
2490 pollfd = zmalloc((consumer_data.stream_count + 2) * sizeof(struct pollfd));
2491 if (pollfd == NULL) {
2492 PERROR("pollfd malloc");
2493 pthread_mutex_unlock(&consumer_data.lock);
2494 goto end;
2495 }
2496
2497 local_stream = zmalloc((consumer_data.stream_count + 2) *
2498 sizeof(struct lttng_consumer_stream *));
2499 if (local_stream == NULL) {
2500 PERROR("local_stream malloc");
2501 pthread_mutex_unlock(&consumer_data.lock);
2502 goto end;
2503 }
2504 ret = update_poll_array(ctx, &pollfd, local_stream,
2505 data_ht);
2506 if (ret < 0) {
2507 ERR("Error in allocating pollfd or local_outfds");
2508 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
2509 pthread_mutex_unlock(&consumer_data.lock);
2510 goto end;
2511 }
2512 nb_fd = ret;
2513 consumer_data.need_update = 0;
2514 }
2515 pthread_mutex_unlock(&consumer_data.lock);
2516
2517 /* No FDs and consumer_quit, consumer_cleanup the thread */
2518 if (nb_fd == 0 && consumer_quit == 1) {
2519 err = 0; /* All is OK */
2520 goto end;
2521 }
2522 /* poll on the array of fds */
2523 restart:
2524 DBG("polling on %d fd", nb_fd + 2);
2525 health_poll_entry();
2526 num_rdy = poll(pollfd, nb_fd + 2, -1);
2527 health_poll_exit();
2528 DBG("poll num_rdy : %d", num_rdy);
2529 if (num_rdy == -1) {
2530 /*
2531 * Restart interrupted system call.
2532 */
2533 if (errno == EINTR) {
2534 goto restart;
2535 }
2536 PERROR("Poll error");
2537 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
2538 goto end;
2539 } else if (num_rdy == 0) {
2540 DBG("Polling thread timed out");
2541 goto end;
2542 }
2543
2544 /*
2545 * If the consumer_data_pipe triggered poll go directly to the
2546 * beginning of the loop to update the array. We want to prioritize
2547 * array update over low-priority reads.
2548 */
2549 if (pollfd[nb_fd].revents & (POLLIN | POLLPRI)) {
2550 ssize_t pipe_readlen;
2551
2552 DBG("consumer_data_pipe wake up");
2553 pipe_readlen = lttng_pipe_read(ctx->consumer_data_pipe,
2554 &new_stream, sizeof(new_stream));
2555 if (pipe_readlen < sizeof(new_stream)) {
2556 PERROR("Consumer data pipe");
2557 /* Continue so we can at least handle the current stream(s). */
2558 continue;
2559 }
2560
2561 /*
2562 * If the stream is NULL, just ignore it. It's also possible that
2563 * the sessiond poll thread changed the consumer_quit state and is
2564 * waking us up to test it.
2565 */
2566 if (new_stream == NULL) {
2567 validate_endpoint_status_data_stream();
2568 continue;
2569 }
2570
2571 /* Continue to update the local streams and handle prio ones */
2572 continue;
2573 }
2574
2575 /* Handle wakeup pipe. */
2576 if (pollfd[nb_fd + 1].revents & (POLLIN | POLLPRI)) {
2577 char dummy;
2578 ssize_t pipe_readlen;
2579
2580 pipe_readlen = lttng_pipe_read(ctx->consumer_wakeup_pipe, &dummy,
2581 sizeof(dummy));
2582 if (pipe_readlen < 0) {
2583 PERROR("Consumer data wakeup pipe");
2584 }
2585 /* We've been awakened to handle stream(s). */
2586 ctx->has_wakeup = 0;
2587 }
2588
2589 /* Take care of high priority channels first. */
2590 for (i = 0; i < nb_fd; i++) {
2591 health_code_update();
2592
2593 if (local_stream[i] == NULL) {
2594 continue;
2595 }
2596 if (pollfd[i].revents & POLLPRI) {
2597 DBG("Urgent read on fd %d", pollfd[i].fd);
2598 high_prio = 1;
2599 len = ctx->on_buffer_ready(local_stream[i], ctx);
2600 /* it's ok to have an unavailable sub-buffer */
2601 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2602 /* Clean the stream and free it. */
2603 consumer_del_stream(local_stream[i], data_ht);
2604 local_stream[i] = NULL;
2605 } else if (len > 0) {
2606 local_stream[i]->data_read = 1;
2607 }
2608 }
2609 }
2610
2611 /*
2612 * If we read high prio channel in this loop, try again
2613 * for more high prio data.
2614 */
2615 if (high_prio) {
2616 continue;
2617 }
2618
2619 /* Take care of low priority channels. */
2620 for (i = 0; i < nb_fd; i++) {
2621 health_code_update();
2622
2623 if (local_stream[i] == NULL) {
2624 continue;
2625 }
2626 if ((pollfd[i].revents & POLLIN) ||
2627 local_stream[i]->hangup_flush_done ||
2628 local_stream[i]->has_data) {
2629 DBG("Normal read on fd %d", pollfd[i].fd);
2630 len = ctx->on_buffer_ready(local_stream[i], ctx);
2631 /* it's ok to have an unavailable sub-buffer */
2632 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2633 /* Clean the stream and free it. */
2634 consumer_del_stream(local_stream[i], data_ht);
2635 local_stream[i] = NULL;
2636 } else if (len > 0) {
2637 local_stream[i]->data_read = 1;
2638 }
2639 }
2640 }
2641
2642 /* Handle hangup and errors */
2643 for (i = 0; i < nb_fd; i++) {
2644 health_code_update();
2645
2646 if (local_stream[i] == NULL) {
2647 continue;
2648 }
2649 if (!local_stream[i]->hangup_flush_done
2650 && (pollfd[i].revents & (POLLHUP | POLLERR | POLLNVAL))
2651 && (consumer_data.type == LTTNG_CONSUMER32_UST
2652 || consumer_data.type == LTTNG_CONSUMER64_UST)) {
2653 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2654 pollfd[i].fd);
2655 lttng_ustconsumer_on_stream_hangup(local_stream[i]);
2656 /* Attempt read again, for the data we just flushed. */
2657 local_stream[i]->data_read = 1;
2658 }
2659 /*
2660 * If the poll flag is HUP/ERR/NVAL and we have
2661 * read no data in this pass, we can remove the
2662 * stream from its hash table.
2663 */
2664 if ((pollfd[i].revents & POLLHUP)) {
2665 DBG("Polling fd %d tells it has hung up.", pollfd[i].fd);
2666 if (!local_stream[i]->data_read) {
2667 consumer_del_stream(local_stream[i], data_ht);
2668 local_stream[i] = NULL;
2669 num_hup++;
2670 }
2671 } else if (pollfd[i].revents & POLLERR) {
2672 ERR("Error returned in polling fd %d.", pollfd[i].fd);
2673 if (!local_stream[i]->data_read) {
2674 consumer_del_stream(local_stream[i], data_ht);
2675 local_stream[i] = NULL;
2676 num_hup++;
2677 }
2678 } else if (pollfd[i].revents & POLLNVAL) {
2679 ERR("Polling fd %d tells fd is not open.", pollfd[i].fd);
2680 if (!local_stream[i]->data_read) {
2681 consumer_del_stream(local_stream[i], data_ht);
2682 local_stream[i] = NULL;
2683 num_hup++;
2684 }
2685 }
2686 if (local_stream[i] != NULL) {
2687 local_stream[i]->data_read = 0;
2688 }
2689 }
2690 }
2691 /* All is OK */
2692 err = 0;
2693 end:
2694 DBG("polling thread exiting");
2695 free(pollfd);
2696 free(local_stream);
2697
2698 /*
2699 * Close the write side of the pipe so epoll_wait() in
2700 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2701 * read side of the pipe. If we close them both, epoll_wait strangely does
2702 * not return and could create a endless wait period if the pipe is the
2703 * only tracked fd in the poll set. The thread will take care of closing
2704 * the read side.
2705 */
2706 (void) lttng_pipe_write_close(ctx->consumer_metadata_pipe);
2707
2708 error_testpoint:
2709 if (err) {
2710 health_error();
2711 ERR("Health error occurred in %s", __func__);
2712 }
2713 health_unregister(health_consumerd);
2714
2715 rcu_unregister_thread();
2716 return NULL;
2717 }
2718
2719 /*
2720 * Close wake-up end of each stream belonging to the channel. This will
2721 * allow the poll() on the stream read-side to detect when the
2722 * write-side (application) finally closes them.
2723 */
2724 static
2725 void consumer_close_channel_streams(struct lttng_consumer_channel *channel)
2726 {
2727 struct lttng_ht *ht;
2728 struct lttng_consumer_stream *stream;
2729 struct lttng_ht_iter iter;
2730
2731 ht = consumer_data.stream_per_chan_id_ht;
2732
2733 rcu_read_lock();
2734 cds_lfht_for_each_entry_duplicate(ht->ht,
2735 ht->hash_fct(&channel->key, lttng_ht_seed),
2736 ht->match_fct, &channel->key,
2737 &iter.iter, stream, node_channel_id.node) {
2738 /*
2739 * Protect against teardown with mutex.
2740 */
2741 pthread_mutex_lock(&stream->lock);
2742 if (cds_lfht_is_node_deleted(&stream->node.node)) {
2743 goto next;
2744 }
2745 switch (consumer_data.type) {
2746 case LTTNG_CONSUMER_KERNEL:
2747 break;
2748 case LTTNG_CONSUMER32_UST:
2749 case LTTNG_CONSUMER64_UST:
2750 if (stream->metadata_flag) {
2751 /* Safe and protected by the stream lock. */
2752 lttng_ustconsumer_close_metadata(stream->chan);
2753 } else {
2754 /*
2755 * Note: a mutex is taken internally within
2756 * liblttng-ust-ctl to protect timer wakeup_fd
2757 * use from concurrent close.
2758 */
2759 lttng_ustconsumer_close_stream_wakeup(stream);
2760 }
2761 break;
2762 default:
2763 ERR("Unknown consumer_data type");
2764 assert(0);
2765 }
2766 next:
2767 pthread_mutex_unlock(&stream->lock);
2768 }
2769 rcu_read_unlock();
2770 }
2771
2772 static void destroy_channel_ht(struct lttng_ht *ht)
2773 {
2774 struct lttng_ht_iter iter;
2775 struct lttng_consumer_channel *channel;
2776 int ret;
2777
2778 if (ht == NULL) {
2779 return;
2780 }
2781
2782 rcu_read_lock();
2783 cds_lfht_for_each_entry(ht->ht, &iter.iter, channel, wait_fd_node.node) {
2784 ret = lttng_ht_del(ht, &iter);
2785 assert(ret != 0);
2786 }
2787 rcu_read_unlock();
2788
2789 lttng_ht_destroy(ht);
2790 }
2791
2792 /*
2793 * This thread polls the channel fds to detect when they are being
2794 * closed. It closes all related streams if the channel is detected as
2795 * closed. It is currently only used as a shim layer for UST because the
2796 * consumerd needs to keep the per-stream wakeup end of pipes open for
2797 * periodical flush.
2798 */
2799 void *consumer_thread_channel_poll(void *data)
2800 {
2801 int ret, i, pollfd, err = -1;
2802 uint32_t revents, nb_fd;
2803 struct lttng_consumer_channel *chan = NULL;
2804 struct lttng_ht_iter iter;
2805 struct lttng_ht_node_u64 *node;
2806 struct lttng_poll_event events;
2807 struct lttng_consumer_local_data *ctx = data;
2808 struct lttng_ht *channel_ht;
2809
2810 rcu_register_thread();
2811
2812 health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_CHANNEL);
2813
2814 if (testpoint(consumerd_thread_channel)) {
2815 goto error_testpoint;
2816 }
2817
2818 health_code_update();
2819
2820 channel_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
2821 if (!channel_ht) {
2822 /* ENOMEM at this point. Better to bail out. */
2823 goto end_ht;
2824 }
2825
2826 DBG("Thread channel poll started");
2827
2828 /* Size is set to 1 for the consumer_channel pipe */
2829 ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC);
2830 if (ret < 0) {
2831 ERR("Poll set creation failed");
2832 goto end_poll;
2833 }
2834
2835 ret = lttng_poll_add(&events, ctx->consumer_channel_pipe[0], LPOLLIN);
2836 if (ret < 0) {
2837 goto end;
2838 }
2839
2840 /* Main loop */
2841 DBG("Channel main loop started");
2842
2843 while (1) {
2844 restart:
2845 health_code_update();
2846 DBG("Channel poll wait");
2847 health_poll_entry();
2848 ret = lttng_poll_wait(&events, -1);
2849 DBG("Channel poll return from wait with %d fd(s)",
2850 LTTNG_POLL_GETNB(&events));
2851 health_poll_exit();
2852 DBG("Channel event caught in thread");
2853 if (ret < 0) {
2854 if (errno == EINTR) {
2855 ERR("Poll EINTR caught");
2856 goto restart;
2857 }
2858 if (LTTNG_POLL_GETNB(&events) == 0) {
2859 err = 0; /* All is OK */
2860 }
2861 goto end;
2862 }
2863
2864 nb_fd = ret;
2865
2866 /* From here, the event is a channel wait fd */
2867 for (i = 0; i < nb_fd; i++) {
2868 health_code_update();
2869
2870 revents = LTTNG_POLL_GETEV(&events, i);
2871 pollfd = LTTNG_POLL_GETFD(&events, i);
2872
2873 if (!revents) {
2874 /* No activity for this FD (poll implementation). */
2875 continue;
2876 }
2877
2878 if (pollfd == ctx->consumer_channel_pipe[0]) {
2879 if (revents & LPOLLIN) {
2880 enum consumer_channel_action action;
2881 uint64_t key;
2882
2883 ret = read_channel_pipe(ctx, &chan, &key, &action);
2884 if (ret <= 0) {
2885 if (ret < 0) {
2886 ERR("Error reading channel pipe");
2887 }
2888 lttng_poll_del(&events, ctx->consumer_channel_pipe[0]);
2889 continue;
2890 }
2891
2892 switch (action) {
2893 case CONSUMER_CHANNEL_ADD:
2894 DBG("Adding channel %d to poll set",
2895 chan->wait_fd);
2896
2897 lttng_ht_node_init_u64(&chan->wait_fd_node,
2898 chan->wait_fd);
2899 rcu_read_lock();
2900 lttng_ht_add_unique_u64(channel_ht,
2901 &chan->wait_fd_node);
2902 rcu_read_unlock();
2903 /* Add channel to the global poll events list */
2904 lttng_poll_add(&events, chan->wait_fd,
2905 LPOLLERR | LPOLLHUP);
2906 break;
2907 case CONSUMER_CHANNEL_DEL:
2908 {
2909 /*
2910 * This command should never be called if the channel
2911 * has streams monitored by either the data or metadata
2912 * thread. The consumer only notify this thread with a
2913 * channel del. command if it receives a destroy
2914 * channel command from the session daemon that send it
2915 * if a command prior to the GET_CHANNEL failed.
2916 */
2917
2918 rcu_read_lock();
2919 chan = consumer_find_channel(key);
2920 if (!chan) {
2921 rcu_read_unlock();
2922 ERR("UST consumer get channel key %" PRIu64 " not found for del channel", key);
2923 break;
2924 }
2925 lttng_poll_del(&events, chan->wait_fd);
2926 iter.iter.node = &chan->wait_fd_node.node;
2927 ret = lttng_ht_del(channel_ht, &iter);
2928 assert(ret == 0);
2929
2930 switch (consumer_data.type) {
2931 case LTTNG_CONSUMER_KERNEL:
2932 break;
2933 case LTTNG_CONSUMER32_UST:
2934 case LTTNG_CONSUMER64_UST:
2935 health_code_update();
2936 /* Destroy streams that might have been left in the stream list. */
2937 clean_channel_stream_list(chan);
2938 break;
2939 default:
2940 ERR("Unknown consumer_data type");
2941 assert(0);
2942 }
2943
2944 /*
2945 * Release our own refcount. Force channel deletion even if
2946 * streams were not initialized.
2947 */
2948 if (!uatomic_sub_return(&chan->refcount, 1)) {
2949 consumer_del_channel(chan);
2950 }
2951 rcu_read_unlock();
2952 goto restart;
2953 }
2954 case CONSUMER_CHANNEL_QUIT:
2955 /*
2956 * Remove the pipe from the poll set and continue the loop
2957 * since their might be data to consume.
2958 */
2959 lttng_poll_del(&events, ctx->consumer_channel_pipe[0]);
2960 continue;
2961 default:
2962 ERR("Unknown action");
2963 break;
2964 }
2965 } else if (revents & (LPOLLERR | LPOLLHUP)) {
2966 DBG("Channel thread pipe hung up");
2967 /*
2968 * Remove the pipe from the poll set and continue the loop
2969 * since their might be data to consume.
2970 */
2971 lttng_poll_del(&events, ctx->consumer_channel_pipe[0]);
2972 continue;
2973 } else {
2974 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
2975 goto end;
2976 }
2977
2978 /* Handle other stream */
2979 continue;
2980 }
2981
2982 rcu_read_lock();
2983 {
2984 uint64_t tmp_id = (uint64_t) pollfd;
2985
2986 lttng_ht_lookup(channel_ht, &tmp_id, &iter);
2987 }
2988 node = lttng_ht_iter_get_node_u64(&iter);
2989 assert(node);
2990
2991 chan = caa_container_of(node, struct lttng_consumer_channel,
2992 wait_fd_node);
2993
2994 /* Check for error event */
2995 if (revents & (LPOLLERR | LPOLLHUP)) {
2996 DBG("Channel fd %d is hup|err.", pollfd);
2997
2998 lttng_poll_del(&events, chan->wait_fd);
2999 ret = lttng_ht_del(channel_ht, &iter);
3000 assert(ret == 0);
3001
3002 /*
3003 * This will close the wait fd for each stream associated to
3004 * this channel AND monitored by the data/metadata thread thus
3005 * will be clean by the right thread.
3006 */
3007 consumer_close_channel_streams(chan);
3008
3009 /* Release our own refcount */
3010 if (!uatomic_sub_return(&chan->refcount, 1)
3011 && !uatomic_read(&chan->nb_init_stream_left)) {
3012 consumer_del_channel(chan);
3013 }
3014 } else {
3015 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
3016 rcu_read_unlock();
3017 goto end;
3018 }
3019
3020 /* Release RCU lock for the channel looked up */
3021 rcu_read_unlock();
3022 }
3023 }
3024
3025 /* All is OK */
3026 err = 0;
3027 end:
3028 lttng_poll_clean(&events);
3029 end_poll:
3030 destroy_channel_ht(channel_ht);
3031 end_ht:
3032 error_testpoint:
3033 DBG("Channel poll thread exiting");
3034 if (err) {
3035 health_error();
3036 ERR("Health error occurred in %s", __func__);
3037 }
3038 health_unregister(health_consumerd);
3039 rcu_unregister_thread();
3040 return NULL;
3041 }
3042
3043 static int set_metadata_socket(struct lttng_consumer_local_data *ctx,
3044 struct pollfd *sockpoll, int client_socket)
3045 {
3046 int ret;
3047
3048 assert(ctx);
3049 assert(sockpoll);
3050
3051 ret = lttng_consumer_poll_socket(sockpoll);
3052 if (ret) {
3053 goto error;
3054 }
3055 DBG("Metadata connection on client_socket");
3056
3057 /* Blocking call, waiting for transmission */
3058 ctx->consumer_metadata_socket = lttcomm_accept_unix_sock(client_socket);
3059 if (ctx->consumer_metadata_socket < 0) {
3060 WARN("On accept metadata");
3061 ret = -1;
3062 goto error;
3063 }
3064 ret = 0;
3065
3066 error:
3067 return ret;
3068 }
3069
3070 /*
3071 * This thread listens on the consumerd socket and receives the file
3072 * descriptors from the session daemon.
3073 */
3074 void *consumer_thread_sessiond_poll(void *data)
3075 {
3076 int sock = -1, client_socket, ret, err = -1;
3077 /*
3078 * structure to poll for incoming data on communication socket avoids
3079 * making blocking sockets.
3080 */
3081 struct pollfd consumer_sockpoll[2];
3082 struct lttng_consumer_local_data *ctx = data;
3083
3084 rcu_register_thread();
3085
3086 health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_SESSIOND);
3087
3088 if (testpoint(consumerd_thread_sessiond)) {
3089 goto error_testpoint;
3090 }
3091
3092 health_code_update();
3093
3094 DBG("Creating command socket %s", ctx->consumer_command_sock_path);
3095 unlink(ctx->consumer_command_sock_path);
3096 client_socket = lttcomm_create_unix_sock(ctx->consumer_command_sock_path);
3097 if (client_socket < 0) {
3098 ERR("Cannot create command socket");
3099 goto end;
3100 }
3101
3102 ret = lttcomm_listen_unix_sock(client_socket);
3103 if (ret < 0) {
3104 goto end;
3105 }
3106
3107 DBG("Sending ready command to lttng-sessiond");
3108 ret = lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY);
3109 /* return < 0 on error, but == 0 is not fatal */
3110 if (ret < 0) {
3111 ERR("Error sending ready command to lttng-sessiond");
3112 goto end;
3113 }
3114
3115 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3116 consumer_sockpoll[0].fd = ctx->consumer_should_quit[0];
3117 consumer_sockpoll[0].events = POLLIN | POLLPRI;
3118 consumer_sockpoll[1].fd = client_socket;
3119 consumer_sockpoll[1].events = POLLIN | POLLPRI;
3120
3121 ret = lttng_consumer_poll_socket(consumer_sockpoll);
3122 if (ret) {
3123 if (ret > 0) {
3124 /* should exit */
3125 err = 0;
3126 }
3127 goto end;
3128 }
3129 DBG("Connection on client_socket");
3130
3131 /* Blocking call, waiting for transmission */
3132 sock = lttcomm_accept_unix_sock(client_socket);
3133 if (sock < 0) {
3134 WARN("On accept");
3135 goto end;
3136 }
3137
3138 /*
3139 * Setup metadata socket which is the second socket connection on the
3140 * command unix socket.
3141 */
3142 ret = set_metadata_socket(ctx, consumer_sockpoll, client_socket);
3143 if (ret) {
3144 if (ret > 0) {
3145 /* should exit */
3146 err = 0;
3147 }
3148 goto end;
3149 }
3150
3151 /* This socket is not useful anymore. */
3152 ret = close(client_socket);
3153 if (ret < 0) {
3154 PERROR("close client_socket");
3155 }
3156 client_socket = -1;
3157
3158 /* update the polling structure to poll on the established socket */
3159 consumer_sockpoll[1].fd = sock;
3160 consumer_sockpoll[1].events = POLLIN | POLLPRI;
3161
3162 while (1) {
3163 health_code_update();
3164
3165 health_poll_entry();
3166 ret = lttng_consumer_poll_socket(consumer_sockpoll);
3167 health_poll_exit();
3168 if (ret) {
3169 if (ret > 0) {
3170 /* should exit */
3171 err = 0;
3172 }
3173 goto end;
3174 }
3175 DBG("Incoming command on sock");
3176 ret = lttng_consumer_recv_cmd(ctx, sock, consumer_sockpoll);
3177 if (ret <= 0) {
3178 /*
3179 * This could simply be a session daemon quitting. Don't output
3180 * ERR() here.
3181 */
3182 DBG("Communication interrupted on command socket");
3183 err = 0;
3184 goto end;
3185 }
3186 if (consumer_quit) {
3187 DBG("consumer_thread_receive_fds received quit from signal");
3188 err = 0; /* All is OK */
3189 goto end;
3190 }
3191 DBG("received command on sock");
3192 }
3193 /* All is OK */
3194 err = 0;
3195
3196 end:
3197 DBG("Consumer thread sessiond poll exiting");
3198
3199 /*
3200 * Close metadata streams since the producer is the session daemon which
3201 * just died.
3202 *
3203 * NOTE: for now, this only applies to the UST tracer.
3204 */
3205 lttng_consumer_close_all_metadata();
3206
3207 /*
3208 * when all fds have hung up, the polling thread
3209 * can exit cleanly
3210 */
3211 consumer_quit = 1;
3212
3213 /*
3214 * Notify the data poll thread to poll back again and test the
3215 * consumer_quit state that we just set so to quit gracefully.
3216 */
3217 notify_thread_lttng_pipe(ctx->consumer_data_pipe);
3218
3219 notify_channel_pipe(ctx, NULL, -1, CONSUMER_CHANNEL_QUIT);
3220
3221 notify_health_quit_pipe(health_quit_pipe);
3222
3223 /* Cleaning up possibly open sockets. */
3224 if (sock >= 0) {
3225 ret = close(sock);
3226 if (ret < 0) {