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