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