Fix: close indexes when rotating the trace files in mmap 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 = close(stream->index_fd);
1600 if (ret < 0) {
1601 PERROR("Closing index");
1602 goto end;
1603 }
1604 stream->index_fd = -1;
1605 ret = index_create_file(stream->chan->pathname,
1606 stream->name, stream->uid, stream->gid,
1607 stream->chan->tracefile_size,
1608 stream->tracefile_count_current);
1609 if (ret < 0) {
1610 goto end;
1611 }
1612 stream->index_fd = ret;
1613 }
1614
1615 /* Reset current size because we just perform a rotation. */
1616 stream->tracefile_size_current = 0;
1617 stream->out_fd_offset = 0;
1618 orig_offset = 0;
1619 }
1620 stream->tracefile_size_current += len;
1621 if (index) {
1622 index->offset = htobe64(stream->out_fd_offset);
1623 }
1624 }
1625
1626 /*
1627 * This call guarantee that len or less is returned. It's impossible to
1628 * receive a ret value that is bigger than len.
1629 */
1630 ret = lttng_write(outfd, mmap_base + mmap_offset, len);
1631 DBG("Consumer mmap write() ret %zd (len %lu)", ret, len);
1632 if (ret < 0 || ((size_t) ret != len)) {
1633 /*
1634 * Report error to caller if nothing was written else at least send the
1635 * amount written.
1636 */
1637 if (ret < 0) {
1638 ret = -errno;
1639 }
1640 relayd_hang_up = 1;
1641
1642 /* Socket operation failed. We consider the relayd dead */
1643 if (errno == EPIPE || errno == EINVAL || errno == EBADF) {
1644 /*
1645 * This is possible if the fd is closed on the other side
1646 * (outfd) or any write problem. It can be verbose a bit for a
1647 * normal execution if for instance the relayd is stopped
1648 * abruptly. This can happen so set this to a DBG statement.
1649 */
1650 DBG("Consumer mmap write detected relayd hang up");
1651 } else {
1652 /* Unhandled error, print it and stop function right now. */
1653 PERROR("Error in write mmap (ret %zd != len %lu)", ret, len);
1654 }
1655 goto write_error;
1656 }
1657 stream->output_written += ret;
1658
1659 /* This call is useless on a socket so better save a syscall. */
1660 if (!relayd) {
1661 /* This won't block, but will start writeout asynchronously */
1662 lttng_sync_file_range(outfd, stream->out_fd_offset, len,
1663 SYNC_FILE_RANGE_WRITE);
1664 stream->out_fd_offset += len;
1665 }
1666 lttng_consumer_sync_trace_file(stream, orig_offset);
1667
1668 write_error:
1669 /*
1670 * This is a special case that the relayd has closed its socket. Let's
1671 * cleanup the relayd object and all associated streams.
1672 */
1673 if (relayd && relayd_hang_up) {
1674 cleanup_relayd(relayd, ctx);
1675 }
1676
1677 end:
1678 /* Unlock only if ctrl socket used */
1679 if (relayd && stream->metadata_flag) {
1680 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
1681 }
1682
1683 rcu_read_unlock();
1684 return ret;
1685 }
1686
1687 /*
1688 * Splice the data from the ring buffer to the tracefile.
1689 *
1690 * It must be called with the stream lock held.
1691 *
1692 * Returns the number of bytes spliced.
1693 */
1694 ssize_t lttng_consumer_on_read_subbuffer_splice(
1695 struct lttng_consumer_local_data *ctx,
1696 struct lttng_consumer_stream *stream, unsigned long len,
1697 unsigned long padding,
1698 struct ctf_packet_index *index)
1699 {
1700 ssize_t ret = 0, written = 0, ret_splice = 0;
1701 loff_t offset = 0;
1702 off_t orig_offset = stream->out_fd_offset;
1703 int fd = stream->wait_fd;
1704 /* Default is on the disk */
1705 int outfd = stream->out_fd;
1706 struct consumer_relayd_sock_pair *relayd = NULL;
1707 int *splice_pipe;
1708 unsigned int relayd_hang_up = 0;
1709
1710 switch (consumer_data.type) {
1711 case LTTNG_CONSUMER_KERNEL:
1712 break;
1713 case LTTNG_CONSUMER32_UST:
1714 case LTTNG_CONSUMER64_UST:
1715 /* Not supported for user space tracing */
1716 return -ENOSYS;
1717 default:
1718 ERR("Unknown consumer_data type");
1719 assert(0);
1720 }
1721
1722 /* RCU lock for the relayd pointer */
1723 rcu_read_lock();
1724
1725 /* Flag that the current stream if set for network streaming. */
1726 if (stream->net_seq_idx != (uint64_t) -1ULL) {
1727 relayd = consumer_find_relayd(stream->net_seq_idx);
1728 if (relayd == NULL) {
1729 written = -ret;
1730 goto end;
1731 }
1732 }
1733 splice_pipe = stream->splice_pipe;
1734
1735 /* Write metadata stream id before payload */
1736 if (relayd) {
1737 unsigned long total_len = len;
1738
1739 if (stream->metadata_flag) {
1740 /*
1741 * Lock the control socket for the complete duration of the function
1742 * since from this point on we will use the socket.
1743 */
1744 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
1745
1746 ret = write_relayd_metadata_id(splice_pipe[1], stream, relayd,
1747 padding);
1748 if (ret < 0) {
1749 written = ret;
1750 relayd_hang_up = 1;
1751 goto write_error;
1752 }
1753
1754 total_len += sizeof(struct lttcomm_relayd_metadata_payload);
1755 }
1756
1757 ret = write_relayd_stream_header(stream, total_len, padding, relayd);
1758 if (ret < 0) {
1759 written = ret;
1760 relayd_hang_up = 1;
1761 goto write_error;
1762 }
1763 /* Use the returned socket. */
1764 outfd = ret;
1765 } else {
1766 /* No streaming, we have to set the len with the full padding */
1767 len += padding;
1768
1769 /*
1770 * Check if we need to change the tracefile before writing the packet.
1771 */
1772 if (stream->chan->tracefile_size > 0 &&
1773 (stream->tracefile_size_current + len) >
1774 stream->chan->tracefile_size) {
1775 ret = utils_rotate_stream_file(stream->chan->pathname,
1776 stream->name, stream->chan->tracefile_size,
1777 stream->chan->tracefile_count, stream->uid, stream->gid,
1778 stream->out_fd, &(stream->tracefile_count_current),
1779 &stream->out_fd);
1780 if (ret < 0) {
1781 written = ret;
1782 ERR("Rotating output file");
1783 goto end;
1784 }
1785 outfd = stream->out_fd;
1786
1787 if (stream->index_fd >= 0) {
1788 ret = close(stream->index_fd);
1789 if (ret < 0) {
1790 PERROR("Closing index");
1791 goto end;
1792 }
1793 stream->index_fd = -1;
1794 ret = index_create_file(stream->chan->pathname,
1795 stream->name, stream->uid, stream->gid,
1796 stream->chan->tracefile_size,
1797 stream->tracefile_count_current);
1798 if (ret < 0) {
1799 written = ret;
1800 goto end;
1801 }
1802 stream->index_fd = ret;
1803 }
1804
1805 /* Reset current size because we just perform a rotation. */
1806 stream->tracefile_size_current = 0;
1807 stream->out_fd_offset = 0;
1808 orig_offset = 0;
1809 }
1810 stream->tracefile_size_current += len;
1811 index->offset = htobe64(stream->out_fd_offset);
1812 }
1813
1814 while (len > 0) {
1815 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1816 (unsigned long)offset, len, fd, splice_pipe[1]);
1817 ret_splice = splice(fd, &offset, splice_pipe[1], NULL, len,
1818 SPLICE_F_MOVE | SPLICE_F_MORE);
1819 DBG("splice chan to pipe, ret %zd", ret_splice);
1820 if (ret_splice < 0) {
1821 ret = errno;
1822 written = -ret;
1823 PERROR("Error in relay splice");
1824 goto splice_error;
1825 }
1826
1827 /* Handle stream on the relayd if the output is on the network */
1828 if (relayd && stream->metadata_flag) {
1829 size_t metadata_payload_size =
1830 sizeof(struct lttcomm_relayd_metadata_payload);
1831
1832 /* Update counter to fit the spliced data */
1833 ret_splice += metadata_payload_size;
1834 len += metadata_payload_size;
1835 /*
1836 * We do this so the return value can match the len passed as
1837 * argument to this function.
1838 */
1839 written -= metadata_payload_size;
1840 }
1841
1842 /* Splice data out */
1843 ret_splice = splice(splice_pipe[0], NULL, outfd, NULL,
1844 ret_splice, SPLICE_F_MOVE | SPLICE_F_MORE);
1845 DBG("Consumer splice pipe to file (out_fd: %d), ret %zd",
1846 outfd, ret_splice);
1847 if (ret_splice < 0) {
1848 ret = errno;
1849 written = -ret;
1850 relayd_hang_up = 1;
1851 goto write_error;
1852 } else if (ret_splice > len) {
1853 /*
1854 * We don't expect this code path to be executed but you never know
1855 * so this is an extra protection agains a buggy splice().
1856 */
1857 ret = errno;
1858 written += ret_splice;
1859 PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice,
1860 len);
1861 goto splice_error;
1862 } else {
1863 /* All good, update current len and continue. */
1864 len -= ret_splice;
1865 }
1866
1867 /* This call is useless on a socket so better save a syscall. */
1868 if (!relayd) {
1869 /* This won't block, but will start writeout asynchronously */
1870 lttng_sync_file_range(outfd, stream->out_fd_offset, ret_splice,
1871 SYNC_FILE_RANGE_WRITE);
1872 stream->out_fd_offset += ret_splice;
1873 }
1874 stream->output_written += ret_splice;
1875 written += ret_splice;
1876 }
1877 lttng_consumer_sync_trace_file(stream, orig_offset);
1878 goto end;
1879
1880 write_error:
1881 /*
1882 * This is a special case that the relayd has closed its socket. Let's
1883 * cleanup the relayd object and all associated streams.
1884 */
1885 if (relayd && relayd_hang_up) {
1886 cleanup_relayd(relayd, ctx);
1887 /* Skip splice error so the consumer does not fail */
1888 goto end;
1889 }
1890
1891 splice_error:
1892 /* send the appropriate error description to sessiond */
1893 switch (ret) {
1894 case EINVAL:
1895 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_EINVAL);
1896 break;
1897 case ENOMEM:
1898 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_ENOMEM);
1899 break;
1900 case ESPIPE:
1901 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_ESPIPE);
1902 break;
1903 }
1904
1905 end:
1906 if (relayd && stream->metadata_flag) {
1907 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
1908 }
1909
1910 rcu_read_unlock();
1911 return written;
1912 }
1913
1914 /*
1915 * Take a snapshot for a specific fd
1916 *
1917 * Returns 0 on success, < 0 on error
1918 */
1919 int lttng_consumer_take_snapshot(struct lttng_consumer_stream *stream)
1920 {
1921 switch (consumer_data.type) {
1922 case LTTNG_CONSUMER_KERNEL:
1923 return lttng_kconsumer_take_snapshot(stream);
1924 case LTTNG_CONSUMER32_UST:
1925 case LTTNG_CONSUMER64_UST:
1926 return lttng_ustconsumer_take_snapshot(stream);
1927 default:
1928 ERR("Unknown consumer_data type");
1929 assert(0);
1930 return -ENOSYS;
1931 }
1932 }
1933
1934 /*
1935 * Get the produced position
1936 *
1937 * Returns 0 on success, < 0 on error
1938 */
1939 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream *stream,
1940 unsigned long *pos)
1941 {
1942 switch (consumer_data.type) {
1943 case LTTNG_CONSUMER_KERNEL:
1944 return lttng_kconsumer_get_produced_snapshot(stream, pos);
1945 case LTTNG_CONSUMER32_UST:
1946 case LTTNG_CONSUMER64_UST:
1947 return lttng_ustconsumer_get_produced_snapshot(stream, pos);
1948 default:
1949 ERR("Unknown consumer_data type");
1950 assert(0);
1951 return -ENOSYS;
1952 }
1953 }
1954
1955 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data *ctx,
1956 int sock, struct pollfd *consumer_sockpoll)
1957 {
1958 switch (consumer_data.type) {
1959 case LTTNG_CONSUMER_KERNEL:
1960 return lttng_kconsumer_recv_cmd(ctx, sock, consumer_sockpoll);
1961 case LTTNG_CONSUMER32_UST:
1962 case LTTNG_CONSUMER64_UST:
1963 return lttng_ustconsumer_recv_cmd(ctx, sock, consumer_sockpoll);
1964 default:
1965 ERR("Unknown consumer_data type");
1966 assert(0);
1967 return -ENOSYS;
1968 }
1969 }
1970
1971 void lttng_consumer_close_all_metadata(void)
1972 {
1973 switch (consumer_data.type) {
1974 case LTTNG_CONSUMER_KERNEL:
1975 /*
1976 * The Kernel consumer has a different metadata scheme so we don't
1977 * close anything because the stream will be closed by the session
1978 * daemon.
1979 */
1980 break;
1981 case LTTNG_CONSUMER32_UST:
1982 case LTTNG_CONSUMER64_UST:
1983 /*
1984 * Close all metadata streams. The metadata hash table is passed and
1985 * this call iterates over it by closing all wakeup fd. This is safe
1986 * because at this point we are sure that the metadata producer is
1987 * either dead or blocked.
1988 */
1989 lttng_ustconsumer_close_all_metadata(metadata_ht);
1990 break;
1991 default:
1992 ERR("Unknown consumer_data type");
1993 assert(0);
1994 }
1995 }
1996
1997 /*
1998 * Clean up a metadata stream and free its memory.
1999 */
2000 void consumer_del_metadata_stream(struct lttng_consumer_stream *stream,
2001 struct lttng_ht *ht)
2002 {
2003 struct lttng_consumer_channel *free_chan = NULL;
2004
2005 assert(stream);
2006 /*
2007 * This call should NEVER receive regular stream. It must always be
2008 * metadata stream and this is crucial for data structure synchronization.
2009 */
2010 assert(stream->metadata_flag);
2011
2012 DBG3("Consumer delete metadata stream %d", stream->wait_fd);
2013
2014 pthread_mutex_lock(&consumer_data.lock);
2015 pthread_mutex_lock(&stream->chan->lock);
2016 pthread_mutex_lock(&stream->lock);
2017
2018 /* Remove any reference to that stream. */
2019 consumer_stream_delete(stream, ht);
2020
2021 /* Close down everything including the relayd if one. */
2022 consumer_stream_close(stream);
2023 /* Destroy tracer buffers of the stream. */
2024 consumer_stream_destroy_buffers(stream);
2025
2026 /* Atomically decrement channel refcount since other threads can use it. */
2027 if (!uatomic_sub_return(&stream->chan->refcount, 1)
2028 && !uatomic_read(&stream->chan->nb_init_stream_left)) {
2029 /* Go for channel deletion! */
2030 free_chan = stream->chan;
2031 }
2032
2033 /*
2034 * Nullify the stream reference so it is not used after deletion. The
2035 * channel lock MUST be acquired before being able to check for a NULL
2036 * pointer value.
2037 */
2038 stream->chan->metadata_stream = NULL;
2039
2040 pthread_mutex_unlock(&stream->lock);
2041 pthread_mutex_unlock(&stream->chan->lock);
2042 pthread_mutex_unlock(&consumer_data.lock);
2043
2044 if (free_chan) {
2045 consumer_del_channel(free_chan);
2046 }
2047
2048 consumer_stream_free(stream);
2049 }
2050
2051 /*
2052 * Action done with the metadata stream when adding it to the consumer internal
2053 * data structures to handle it.
2054 */
2055 int consumer_add_metadata_stream(struct lttng_consumer_stream *stream)
2056 {
2057 struct lttng_ht *ht = metadata_ht;
2058 int ret = 0;
2059 struct lttng_ht_iter iter;
2060 struct lttng_ht_node_u64 *node;
2061
2062 assert(stream);
2063 assert(ht);
2064
2065 DBG3("Adding metadata stream %" PRIu64 " to hash table", stream->key);
2066
2067 pthread_mutex_lock(&consumer_data.lock);
2068 pthread_mutex_lock(&stream->chan->lock);
2069 pthread_mutex_lock(&stream->chan->timer_lock);
2070 pthread_mutex_lock(&stream->lock);
2071
2072 /*
2073 * From here, refcounts are updated so be _careful_ when returning an error
2074 * after this point.
2075 */
2076
2077 rcu_read_lock();
2078
2079 /*
2080 * Lookup the stream just to make sure it does not exist in our internal
2081 * state. This should NEVER happen.
2082 */
2083 lttng_ht_lookup(ht, &stream->key, &iter);
2084 node = lttng_ht_iter_get_node_u64(&iter);
2085 assert(!node);
2086
2087 /*
2088 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2089 * in terms of destroying the associated channel, because the action that
2090 * causes the count to become 0 also causes a stream to be added. The
2091 * channel deletion will thus be triggered by the following removal of this
2092 * stream.
2093 */
2094 if (uatomic_read(&stream->chan->nb_init_stream_left) > 0) {
2095 /* Increment refcount before decrementing nb_init_stream_left */
2096 cmm_smp_wmb();
2097 uatomic_dec(&stream->chan->nb_init_stream_left);
2098 }
2099
2100 lttng_ht_add_unique_u64(ht, &stream->node);
2101
2102 lttng_ht_add_unique_u64(consumer_data.stream_per_chan_id_ht,
2103 &stream->node_channel_id);
2104
2105 /*
2106 * Add stream to the stream_list_ht of the consumer data. No need to steal
2107 * the key since the HT does not use it and we allow to add redundant keys
2108 * into this table.
2109 */
2110 lttng_ht_add_u64(consumer_data.stream_list_ht, &stream->node_session_id);
2111
2112 rcu_read_unlock();
2113
2114 pthread_mutex_unlock(&stream->lock);
2115 pthread_mutex_unlock(&stream->chan->lock);
2116 pthread_mutex_unlock(&stream->chan->timer_lock);
2117 pthread_mutex_unlock(&consumer_data.lock);
2118 return ret;
2119 }
2120
2121 /*
2122 * Delete data stream that are flagged for deletion (endpoint_status).
2123 */
2124 static void validate_endpoint_status_data_stream(void)
2125 {
2126 struct lttng_ht_iter iter;
2127 struct lttng_consumer_stream *stream;
2128
2129 DBG("Consumer delete flagged data stream");
2130
2131 rcu_read_lock();
2132 cds_lfht_for_each_entry(data_ht->ht, &iter.iter, stream, node.node) {
2133 /* Validate delete flag of the stream */
2134 if (stream->endpoint_status == CONSUMER_ENDPOINT_ACTIVE) {
2135 continue;
2136 }
2137 /* Delete it right now */
2138 consumer_del_stream(stream, data_ht);
2139 }
2140 rcu_read_unlock();
2141 }
2142
2143 /*
2144 * Delete metadata stream that are flagged for deletion (endpoint_status).
2145 */
2146 static void validate_endpoint_status_metadata_stream(
2147 struct lttng_poll_event *pollset)
2148 {
2149 struct lttng_ht_iter iter;
2150 struct lttng_consumer_stream *stream;
2151
2152 DBG("Consumer delete flagged metadata stream");
2153
2154 assert(pollset);
2155
2156 rcu_read_lock();
2157 cds_lfht_for_each_entry(metadata_ht->ht, &iter.iter, stream, node.node) {
2158 /* Validate delete flag of the stream */
2159 if (stream->endpoint_status == CONSUMER_ENDPOINT_ACTIVE) {
2160 continue;
2161 }
2162 /*
2163 * Remove from pollset so the metadata thread can continue without
2164 * blocking on a deleted stream.
2165 */
2166 lttng_poll_del(pollset, stream->wait_fd);
2167
2168 /* Delete it right now */
2169 consumer_del_metadata_stream(stream, metadata_ht);
2170 }
2171 rcu_read_unlock();
2172 }
2173
2174 /*
2175 * Thread polls on metadata file descriptor and write them on disk or on the
2176 * network.
2177 */
2178 void *consumer_thread_metadata_poll(void *data)
2179 {
2180 int ret, i, pollfd, err = -1;
2181 uint32_t revents, nb_fd;
2182 struct lttng_consumer_stream *stream = NULL;
2183 struct lttng_ht_iter iter;
2184 struct lttng_ht_node_u64 *node;
2185 struct lttng_poll_event events;
2186 struct lttng_consumer_local_data *ctx = data;
2187 ssize_t len;
2188
2189 rcu_register_thread();
2190
2191 health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_METADATA);
2192
2193 if (testpoint(consumerd_thread_metadata)) {
2194 goto error_testpoint;
2195 }
2196
2197 health_code_update();
2198
2199 DBG("Thread metadata poll started");
2200
2201 /* Size is set to 1 for the consumer_metadata pipe */
2202 ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC);
2203 if (ret < 0) {
2204 ERR("Poll set creation failed");
2205 goto end_poll;
2206 }
2207
2208 ret = lttng_poll_add(&events,
2209 lttng_pipe_get_readfd(ctx->consumer_metadata_pipe), LPOLLIN);
2210 if (ret < 0) {
2211 goto end;
2212 }
2213
2214 /* Main loop */
2215 DBG("Metadata main loop started");
2216
2217 while (1) {
2218 restart:
2219 health_code_update();
2220 health_poll_entry();
2221 DBG("Metadata poll wait");
2222 ret = lttng_poll_wait(&events, -1);
2223 DBG("Metadata poll return from wait with %d fd(s)",
2224 LTTNG_POLL_GETNB(&events));
2225 health_poll_exit();
2226 DBG("Metadata event catched in thread");
2227 if (ret < 0) {
2228 if (errno == EINTR) {
2229 ERR("Poll EINTR catched");
2230 goto restart;
2231 }
2232 if (LTTNG_POLL_GETNB(&events) == 0) {
2233 err = 0; /* All is OK */
2234 }
2235 goto end;
2236 }
2237
2238 nb_fd = ret;
2239
2240 /* From here, the event is a metadata wait fd */
2241 for (i = 0; i < nb_fd; i++) {
2242 health_code_update();
2243
2244 revents = LTTNG_POLL_GETEV(&events, i);
2245 pollfd = LTTNG_POLL_GETFD(&events, i);
2246
2247 if (!revents) {
2248 /* No activity for this FD (poll implementation). */
2249 continue;
2250 }
2251
2252 if (pollfd == lttng_pipe_get_readfd(ctx->consumer_metadata_pipe)) {
2253 if (revents & LPOLLIN) {
2254 ssize_t pipe_len;
2255
2256 pipe_len = lttng_pipe_read(ctx->consumer_metadata_pipe,
2257 &stream, sizeof(stream));
2258 if (pipe_len < sizeof(stream)) {
2259 if (pipe_len < 0) {
2260 PERROR("read metadata stream");
2261 }
2262 /*
2263 * Remove the pipe from the poll set and continue the loop
2264 * since their might be data to consume.
2265 */
2266 lttng_poll_del(&events,
2267 lttng_pipe_get_readfd(ctx->consumer_metadata_pipe));
2268 lttng_pipe_read_close(ctx->consumer_metadata_pipe);
2269 continue;
2270 }
2271
2272 /* A NULL stream means that the state has changed. */
2273 if (stream == NULL) {
2274 /* Check for deleted streams. */
2275 validate_endpoint_status_metadata_stream(&events);
2276 goto restart;
2277 }
2278
2279 DBG("Adding metadata stream %d to poll set",
2280 stream->wait_fd);
2281
2282 /* Add metadata stream to the global poll events list */
2283 lttng_poll_add(&events, stream->wait_fd,
2284 LPOLLIN | LPOLLPRI | LPOLLHUP);
2285 } else if (revents & (LPOLLERR | LPOLLHUP)) {
2286 DBG("Metadata thread pipe hung up");
2287 /*
2288 * Remove the pipe from the poll set and continue the loop
2289 * since their might be data to consume.
2290 */
2291 lttng_poll_del(&events,
2292 lttng_pipe_get_readfd(ctx->consumer_metadata_pipe));
2293 lttng_pipe_read_close(ctx->consumer_metadata_pipe);
2294 continue;
2295 } else {
2296 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
2297 goto end;
2298 }
2299
2300 /* Handle other stream */
2301 continue;
2302 }
2303
2304 rcu_read_lock();
2305 {
2306 uint64_t tmp_id = (uint64_t) pollfd;
2307
2308 lttng_ht_lookup(metadata_ht, &tmp_id, &iter);
2309 }
2310 node = lttng_ht_iter_get_node_u64(&iter);
2311 assert(node);
2312
2313 stream = caa_container_of(node, struct lttng_consumer_stream,
2314 node);
2315
2316 if (revents & (LPOLLIN | LPOLLPRI)) {
2317 /* Get the data out of the metadata file descriptor */
2318 DBG("Metadata available on fd %d", pollfd);
2319 assert(stream->wait_fd == pollfd);
2320
2321 do {
2322 health_code_update();
2323
2324 len = ctx->on_buffer_ready(stream, ctx);
2325 /*
2326 * We don't check the return value here since if we get
2327 * a negative len, it means an error occured thus we
2328 * simply remove it from the poll set and free the
2329 * stream.
2330 */
2331 } while (len > 0);
2332
2333 /* It's ok to have an unavailable sub-buffer */
2334 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2335 /* Clean up stream from consumer and free it. */
2336 lttng_poll_del(&events, stream->wait_fd);
2337 consumer_del_metadata_stream(stream, metadata_ht);
2338 }
2339 } else if (revents & (LPOLLERR | LPOLLHUP)) {
2340 DBG("Metadata fd %d is hup|err.", pollfd);
2341 if (!stream->hangup_flush_done
2342 && (consumer_data.type == LTTNG_CONSUMER32_UST
2343 || consumer_data.type == LTTNG_CONSUMER64_UST)) {
2344 DBG("Attempting to flush and consume the UST buffers");
2345 lttng_ustconsumer_on_stream_hangup(stream);
2346
2347 /* We just flushed the stream now read it. */
2348 do {
2349 health_code_update();
2350
2351 len = ctx->on_buffer_ready(stream, ctx);
2352 /*
2353 * We don't check the return value here since if we get
2354 * a negative len, it means an error occured thus we
2355 * simply remove it from the poll set and free the
2356 * stream.
2357 */
2358 } while (len > 0);
2359 }
2360
2361 lttng_poll_del(&events, stream->wait_fd);
2362 /*
2363 * This call update the channel states, closes file descriptors
2364 * and securely free the stream.
2365 */
2366 consumer_del_metadata_stream(stream, metadata_ht);
2367 } else {
2368 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
2369 rcu_read_unlock();
2370 goto end;
2371 }
2372 /* Release RCU lock for the stream looked up */
2373 rcu_read_unlock();
2374 }
2375 }
2376
2377 /* All is OK */
2378 err = 0;
2379 end:
2380 DBG("Metadata poll thread exiting");
2381
2382 lttng_poll_clean(&events);
2383 end_poll:
2384 error_testpoint:
2385 if (err) {
2386 health_error();
2387 ERR("Health error occurred in %s", __func__);
2388 }
2389 health_unregister(health_consumerd);
2390 rcu_unregister_thread();
2391 return NULL;
2392 }
2393
2394 /*
2395 * This thread polls the fds in the set to consume the data and write
2396 * it to tracefile if necessary.
2397 */
2398 void *consumer_thread_data_poll(void *data)
2399 {
2400 int num_rdy, num_hup, high_prio, ret, i, err = -1;
2401 struct pollfd *pollfd = NULL;
2402 /* local view of the streams */
2403 struct lttng_consumer_stream **local_stream = NULL, *new_stream = NULL;
2404 /* local view of consumer_data.fds_count */
2405 int nb_fd = 0;
2406 struct lttng_consumer_local_data *ctx = data;
2407 ssize_t len;
2408
2409 rcu_register_thread();
2410
2411 health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_DATA);
2412
2413 if (testpoint(consumerd_thread_data)) {
2414 goto error_testpoint;
2415 }
2416
2417 health_code_update();
2418
2419 local_stream = zmalloc(sizeof(struct lttng_consumer_stream *));
2420 if (local_stream == NULL) {
2421 PERROR("local_stream malloc");
2422 goto end;
2423 }
2424
2425 while (1) {
2426 health_code_update();
2427
2428 high_prio = 0;
2429 num_hup = 0;
2430
2431 /*
2432 * the fds set has been updated, we need to update our
2433 * local array as well
2434 */
2435 pthread_mutex_lock(&consumer_data.lock);
2436 if (consumer_data.need_update) {
2437 free(pollfd);
2438 pollfd = NULL;
2439
2440 free(local_stream);
2441 local_stream = NULL;
2442
2443 /*
2444 * Allocate for all fds +1 for the consumer_data_pipe and +1 for
2445 * wake up pipe.
2446 */
2447 pollfd = zmalloc((consumer_data.stream_count + 2) * sizeof(struct pollfd));
2448 if (pollfd == NULL) {
2449 PERROR("pollfd malloc");
2450 pthread_mutex_unlock(&consumer_data.lock);
2451 goto end;
2452 }
2453
2454 local_stream = zmalloc((consumer_data.stream_count + 2) *
2455 sizeof(struct lttng_consumer_stream *));
2456 if (local_stream == NULL) {
2457 PERROR("local_stream malloc");
2458 pthread_mutex_unlock(&consumer_data.lock);
2459 goto end;
2460 }
2461 ret = update_poll_array(ctx, &pollfd, local_stream,
2462 data_ht);
2463 if (ret < 0) {
2464 ERR("Error in allocating pollfd or local_outfds");
2465 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
2466 pthread_mutex_unlock(&consumer_data.lock);
2467 goto end;
2468 }
2469 nb_fd = ret;
2470 consumer_data.need_update = 0;
2471 }
2472 pthread_mutex_unlock(&consumer_data.lock);
2473
2474 /* No FDs and consumer_quit, consumer_cleanup the thread */
2475 if (nb_fd == 0 && consumer_quit == 1) {
2476 err = 0; /* All is OK */
2477 goto end;
2478 }
2479 /* poll on the array of fds */
2480 restart:
2481 DBG("polling on %d fd", nb_fd + 2);
2482 health_poll_entry();
2483 num_rdy = poll(pollfd, nb_fd + 2, -1);
2484 health_poll_exit();
2485 DBG("poll num_rdy : %d", num_rdy);
2486 if (num_rdy == -1) {
2487 /*
2488 * Restart interrupted system call.
2489 */
2490 if (errno == EINTR) {
2491 goto restart;
2492 }
2493 PERROR("Poll error");
2494 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
2495 goto end;
2496 } else if (num_rdy == 0) {
2497 DBG("Polling thread timed out");
2498 goto end;
2499 }
2500
2501 /*
2502 * If the consumer_data_pipe triggered poll go directly to the
2503 * beginning of the loop to update the array. We want to prioritize
2504 * array update over low-priority reads.
2505 */
2506 if (pollfd[nb_fd].revents & (POLLIN | POLLPRI)) {
2507 ssize_t pipe_readlen;
2508
2509 DBG("consumer_data_pipe wake up");
2510 pipe_readlen = lttng_pipe_read(ctx->consumer_data_pipe,
2511 &new_stream, sizeof(new_stream));
2512 if (pipe_readlen < sizeof(new_stream)) {
2513 PERROR("Consumer data pipe");
2514 /* Continue so we can at least handle the current stream(s). */
2515 continue;
2516 }
2517
2518 /*
2519 * If the stream is NULL, just ignore it. It's also possible that
2520 * the sessiond poll thread changed the consumer_quit state and is
2521 * waking us up to test it.
2522 */
2523 if (new_stream == NULL) {
2524 validate_endpoint_status_data_stream();
2525 continue;
2526 }
2527
2528 /* Continue to update the local streams and handle prio ones */
2529 continue;
2530 }
2531
2532 /* Handle wakeup pipe. */
2533 if (pollfd[nb_fd + 1].revents & (POLLIN | POLLPRI)) {
2534 char dummy;
2535 ssize_t pipe_readlen;
2536
2537 pipe_readlen = lttng_pipe_read(ctx->consumer_wakeup_pipe, &dummy,
2538 sizeof(dummy));
2539 if (pipe_readlen < 0) {
2540 PERROR("Consumer data wakeup pipe");
2541 }
2542 /* We've been awakened to handle stream(s). */
2543 ctx->has_wakeup = 0;
2544 }
2545
2546 /* Take care of high priority channels first. */
2547 for (i = 0; i < nb_fd; i++) {
2548 health_code_update();
2549
2550 if (local_stream[i] == NULL) {
2551 continue;
2552 }
2553 if (pollfd[i].revents & POLLPRI) {
2554 DBG("Urgent read on fd %d", pollfd[i].fd);
2555 high_prio = 1;
2556 len = ctx->on_buffer_ready(local_stream[i], ctx);
2557 /* it's ok to have an unavailable sub-buffer */
2558 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2559 /* Clean the stream and free it. */
2560 consumer_del_stream(local_stream[i], data_ht);
2561 local_stream[i] = NULL;
2562 } else if (len > 0) {
2563 local_stream[i]->data_read = 1;
2564 }
2565 }
2566 }
2567
2568 /*
2569 * If we read high prio channel in this loop, try again
2570 * for more high prio data.
2571 */
2572 if (high_prio) {
2573 continue;
2574 }
2575
2576 /* Take care of low priority channels. */
2577 for (i = 0; i < nb_fd; i++) {
2578 health_code_update();
2579
2580 if (local_stream[i] == NULL) {
2581 continue;
2582 }
2583 if ((pollfd[i].revents & POLLIN) ||
2584 local_stream[i]->hangup_flush_done ||
2585 local_stream[i]->has_data) {
2586 DBG("Normal read on fd %d", pollfd[i].fd);
2587 len = ctx->on_buffer_ready(local_stream[i], ctx);
2588 /* it's ok to have an unavailable sub-buffer */
2589 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2590 /* Clean the stream and free it. */
2591 consumer_del_stream(local_stream[i], data_ht);
2592 local_stream[i] = NULL;
2593 } else if (len > 0) {
2594 local_stream[i]->data_read = 1;
2595 }
2596 }
2597 }
2598
2599 /* Handle hangup and errors */
2600 for (i = 0; i < nb_fd; i++) {
2601 health_code_update();
2602
2603 if (local_stream[i] == NULL) {
2604 continue;
2605 }
2606 if (!local_stream[i]->hangup_flush_done
2607 && (pollfd[i].revents & (POLLHUP | POLLERR | POLLNVAL))
2608 && (consumer_data.type == LTTNG_CONSUMER32_UST
2609 || consumer_data.type == LTTNG_CONSUMER64_UST)) {
2610 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2611 pollfd[i].fd);
2612 lttng_ustconsumer_on_stream_hangup(local_stream[i]);
2613 /* Attempt read again, for the data we just flushed. */
2614 local_stream[i]->data_read = 1;
2615 }
2616 /*
2617 * If the poll flag is HUP/ERR/NVAL and we have
2618 * read no data in this pass, we can remove the
2619 * stream from its hash table.
2620 */
2621 if ((pollfd[i].revents & POLLHUP)) {
2622 DBG("Polling fd %d tells it has hung up.", pollfd[i].fd);
2623 if (!local_stream[i]->data_read) {
2624 consumer_del_stream(local_stream[i], data_ht);
2625 local_stream[i] = NULL;
2626 num_hup++;
2627 }
2628 } else if (pollfd[i].revents & POLLERR) {
2629 ERR("Error returned in polling fd %d.", pollfd[i].fd);
2630 if (!local_stream[i]->data_read) {
2631 consumer_del_stream(local_stream[i], data_ht);
2632 local_stream[i] = NULL;
2633 num_hup++;
2634 }
2635 } else if (pollfd[i].revents & POLLNVAL) {
2636 ERR("Polling fd %d tells fd is not open.", pollfd[i].fd);
2637 if (!local_stream[i]->data_read) {
2638 consumer_del_stream(local_stream[i], data_ht);
2639 local_stream[i] = NULL;
2640 num_hup++;
2641 }
2642 }
2643 if (local_stream[i] != NULL) {
2644 local_stream[i]->data_read = 0;
2645 }
2646 }
2647 }
2648 /* All is OK */
2649 err = 0;
2650 end:
2651 DBG("polling thread exiting");
2652 free(pollfd);
2653 free(local_stream);
2654
2655 /*
2656 * Close the write side of the pipe so epoll_wait() in
2657 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2658 * read side of the pipe. If we close them both, epoll_wait strangely does
2659 * not return and could create a endless wait period if the pipe is the
2660 * only tracked fd in the poll set. The thread will take care of closing
2661 * the read side.
2662 */
2663 (void) lttng_pipe_write_close(ctx->consumer_metadata_pipe);
2664
2665 error_testpoint:
2666 if (err) {
2667 health_error();
2668 ERR("Health error occurred in %s", __func__);
2669 }
2670 health_unregister(health_consumerd);
2671
2672 rcu_unregister_thread();
2673 return NULL;
2674 }
2675
2676 /*
2677 * Close wake-up end of each stream belonging to the channel. This will
2678 * allow the poll() on the stream read-side to detect when the
2679 * write-side (application) finally closes them.
2680 */
2681 static
2682 void consumer_close_channel_streams(struct lttng_consumer_channel *channel)
2683 {
2684 struct lttng_ht *ht;
2685 struct lttng_consumer_stream *stream;
2686 struct lttng_ht_iter iter;
2687
2688 ht = consumer_data.stream_per_chan_id_ht;
2689
2690 rcu_read_lock();
2691 cds_lfht_for_each_entry_duplicate(ht->ht,
2692 ht->hash_fct(&channel->key, lttng_ht_seed),
2693 ht->match_fct, &channel->key,
2694 &iter.iter, stream, node_channel_id.node) {
2695 /*
2696 * Protect against teardown with mutex.
2697 */
2698 pthread_mutex_lock(&stream->lock);
2699 if (cds_lfht_is_node_deleted(&stream->node.node)) {
2700 goto next;
2701 }
2702 switch (consumer_data.type) {
2703 case LTTNG_CONSUMER_KERNEL:
2704 break;
2705 case LTTNG_CONSUMER32_UST:
2706 case LTTNG_CONSUMER64_UST:
2707 if (stream->metadata_flag) {
2708 /* Safe and protected by the stream lock. */
2709 lttng_ustconsumer_close_metadata(stream->chan);
2710 } else {
2711 /*
2712 * Note: a mutex is taken internally within
2713 * liblttng-ust-ctl to protect timer wakeup_fd
2714 * use from concurrent close.
2715 */
2716 lttng_ustconsumer_close_stream_wakeup(stream);
2717 }
2718 break;
2719 default:
2720 ERR("Unknown consumer_data type");
2721 assert(0);
2722 }
2723 next:
2724 pthread_mutex_unlock(&stream->lock);
2725 }
2726 rcu_read_unlock();
2727 }
2728
2729 static void destroy_channel_ht(struct lttng_ht *ht)
2730 {
2731 struct lttng_ht_iter iter;
2732 struct lttng_consumer_channel *channel;
2733 int ret;
2734
2735 if (ht == NULL) {
2736 return;
2737 }
2738
2739 rcu_read_lock();
2740 cds_lfht_for_each_entry(ht->ht, &iter.iter, channel, wait_fd_node.node) {
2741 ret = lttng_ht_del(ht, &iter);
2742 assert(ret != 0);
2743 }
2744 rcu_read_unlock();
2745
2746 lttng_ht_destroy(ht);
2747 }
2748
2749 /*
2750 * This thread polls the channel fds to detect when they are being
2751 * closed. It closes all related streams if the channel is detected as
2752 * closed. It is currently only used as a shim layer for UST because the
2753 * consumerd needs to keep the per-stream wakeup end of pipes open for
2754 * periodical flush.
2755 */
2756 void *consumer_thread_channel_poll(void *data)
2757 {
2758 int ret, i, pollfd, err = -1;
2759 uint32_t revents, nb_fd;
2760 struct lttng_consumer_channel *chan = NULL;
2761 struct lttng_ht_iter iter;
2762 struct lttng_ht_node_u64 *node;
2763 struct lttng_poll_event events;
2764 struct lttng_consumer_local_data *ctx = data;
2765 struct lttng_ht *channel_ht;
2766
2767 rcu_register_thread();
2768
2769 health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_CHANNEL);
2770
2771 if (testpoint(consumerd_thread_channel)) {
2772 goto error_testpoint;
2773 }
2774
2775 health_code_update();
2776
2777 channel_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
2778 if (!channel_ht) {
2779 /* ENOMEM at this point. Better to bail out. */
2780 goto end_ht;
2781 }
2782
2783 DBG("Thread channel poll started");
2784
2785 /* Size is set to 1 for the consumer_channel pipe */
2786 ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC);
2787 if (ret < 0) {
2788 ERR("Poll set creation failed");
2789 goto end_poll;
2790 }
2791
2792 ret = lttng_poll_add(&events, ctx->consumer_channel_pipe[0], LPOLLIN);
2793 if (ret < 0) {
2794 goto end;
2795 }
2796
2797 /* Main loop */
2798 DBG("Channel main loop started");
2799
2800 while (1) {
2801 restart:
2802 health_code_update();
2803 DBG("Channel poll wait");
2804 health_poll_entry();
2805 ret = lttng_poll_wait(&events, -1);
2806 DBG("Channel poll return from wait with %d fd(s)",
2807 LTTNG_POLL_GETNB(&events));
2808 health_poll_exit();
2809 DBG("Channel event catched in thread");
2810 if (ret < 0) {
2811 if (errno == EINTR) {
2812 ERR("Poll EINTR catched");
2813 goto restart;
2814 }
2815 if (LTTNG_POLL_GETNB(&events) == 0) {
2816 err = 0; /* All is OK */
2817 }
2818 goto end;
2819 }
2820
2821 nb_fd = ret;
2822
2823 /* From here, the event is a channel wait fd */
2824 for (i = 0; i < nb_fd; i++) {
2825 health_code_update();
2826
2827 revents = LTTNG_POLL_GETEV(&events, i);
2828 pollfd = LTTNG_POLL_GETFD(&events, i);
2829
2830 if (!revents) {
2831 /* No activity for this FD (poll implementation). */
2832 continue;
2833 }
2834
2835 if (pollfd == ctx->consumer_channel_pipe[0]) {
2836 if (revents & LPOLLIN) {
2837 enum consumer_channel_action action;
2838 uint64_t key;
2839
2840 ret = read_channel_pipe(ctx, &chan, &key, &action);
2841 if (ret <= 0) {
2842 if (ret < 0) {
2843 ERR("Error reading channel pipe");
2844 }
2845 lttng_poll_del(&events, ctx->consumer_channel_pipe[0]);
2846 continue;
2847 }
2848
2849 switch (action) {
2850 case CONSUMER_CHANNEL_ADD:
2851 DBG("Adding channel %d to poll set",
2852 chan->wait_fd);
2853
2854 lttng_ht_node_init_u64(&chan->wait_fd_node,
2855 chan->wait_fd);
2856 rcu_read_lock();
2857 lttng_ht_add_unique_u64(channel_ht,
2858 &chan->wait_fd_node);
2859 rcu_read_unlock();
2860 /* Add channel to the global poll events list */
2861 lttng_poll_add(&events, chan->wait_fd,
2862 LPOLLERR | LPOLLHUP);
2863 break;
2864 case CONSUMER_CHANNEL_DEL:
2865 {
2866 /*
2867 * This command should never be called if the channel
2868 * has streams monitored by either the data or metadata
2869 * thread. The consumer only notify this thread with a
2870 * channel del. command if it receives a destroy
2871 * channel command from the session daemon that send it
2872 * if a command prior to the GET_CHANNEL failed.
2873 */
2874
2875 rcu_read_lock();
2876 chan = consumer_find_channel(key);
2877 if (!chan) {
2878 rcu_read_unlock();
2879 ERR("UST consumer get channel key %" PRIu64 " not found for del channel", key);
2880 break;
2881 }
2882 lttng_poll_del(&events, chan->wait_fd);
2883 iter.iter.node = &chan->wait_fd_node.node;
2884 ret = lttng_ht_del(channel_ht, &iter);
2885 assert(ret == 0);
2886
2887 switch (consumer_data.type) {
2888 case LTTNG_CONSUMER_KERNEL:
2889 break;
2890 case LTTNG_CONSUMER32_UST:
2891 case LTTNG_CONSUMER64_UST:
2892 health_code_update();
2893 /* Destroy streams that might have been left in the stream list. */
2894 clean_channel_stream_list(chan);
2895 break;
2896 default:
2897 ERR("Unknown consumer_data type");
2898 assert(0);
2899 }
2900
2901 /*
2902 * Release our own refcount. Force channel deletion even if
2903 * streams were not initialized.
2904 */
2905 if (!uatomic_sub_return(&chan->refcount, 1)) {
2906 consumer_del_channel(chan);
2907 }
2908 rcu_read_unlock();
2909 goto restart;
2910 }
2911 case CONSUMER_CHANNEL_QUIT:
2912 /*
2913 * Remove the pipe from the poll set and continue the loop
2914 * since their might be data to consume.
2915 */
2916 lttng_poll_del(&events, ctx->consumer_channel_pipe[0]);
2917 continue;
2918 default:
2919 ERR("Unknown action");
2920 break;
2921 }
2922 } else if (revents & (LPOLLERR | LPOLLHUP)) {
2923 DBG("Channel thread pipe hung up");
2924 /*
2925 * Remove the pipe from the poll set and continue the loop
2926 * since their might be data to consume.
2927 */
2928 lttng_poll_del(&events, ctx->consumer_channel_pipe[0]);
2929 continue;
2930 } else {
2931 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
2932 goto end;
2933 }
2934
2935 /* Handle other stream */
2936 continue;
2937 }
2938
2939 rcu_read_lock();
2940 {
2941 uint64_t tmp_id = (uint64_t) pollfd;
2942
2943 lttng_ht_lookup(channel_ht, &tmp_id, &iter);
2944 }
2945 node = lttng_ht_iter_get_node_u64(&iter);
2946 assert(node);
2947
2948 chan = caa_container_of(node, struct lttng_consumer_channel,
2949 wait_fd_node);
2950
2951 /* Check for error event */
2952 if (revents & (LPOLLERR | LPOLLHUP)) {
2953 DBG("Channel fd %d is hup|err.", pollfd);
2954
2955 lttng_poll_del(&events, chan->wait_fd);
2956 ret = lttng_ht_del(channel_ht, &iter);
2957 assert(ret == 0);
2958
2959 /*
2960 * This will close the wait fd for each stream associated to
2961 * this channel AND monitored by the data/metadata thread thus
2962 * will be clean by the right thread.
2963 */
2964 consumer_close_channel_streams(chan);
2965
2966 /* Release our own refcount */
2967 if (!uatomic_sub_return(&chan->refcount, 1)
2968 && !uatomic_read(&chan->nb_init_stream_left)) {
2969 consumer_del_channel(chan);
2970 }
2971 } else {
2972 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
2973 rcu_read_unlock();
2974 goto end;
2975 }
2976
2977 /* Release RCU lock for the channel looked up */
2978 rcu_read_unlock();
2979 }
2980 }
2981
2982 /* All is OK */
2983 err = 0;
2984 end:
2985 lttng_poll_clean(&events);
2986 end_poll:
2987 destroy_channel_ht(channel_ht);
2988 end_ht:
2989 error_testpoint:
2990 DBG("Channel poll thread exiting");
2991 if (err) {
2992 health_error();
2993 ERR("Health error occurred in %s", __func__);
2994 }
2995 health_unregister(health_consumerd);
2996 rcu_unregister_thread();
2997 return NULL;
2998 }
2999
3000 static int set_metadata_socket(struct lttng_consumer_local_data *ctx,
3001 struct pollfd *sockpoll, int client_socket)
3002 {
3003 int ret;
3004
3005 assert(ctx);
3006 assert(sockpoll);
3007
3008 ret = lttng_consumer_poll_socket(sockpoll);
3009 if (ret) {
3010 goto error;
3011 }
3012 DBG("Metadata connection on client_socket");
3013
3014 /* Blocking call, waiting for transmission */
3015 ctx->consumer_metadata_socket = lttcomm_accept_unix_sock(client_socket);
3016 if (ctx->consumer_metadata_socket < 0) {
3017 WARN("On accept metadata");
3018 ret = -1;
3019 goto error;
3020 }
3021 ret = 0;
3022
3023 error:
3024 return ret;
3025 }
3026
3027 /*
3028 * This thread listens on the consumerd socket and receives the file
3029 * descriptors from the session daemon.
3030 */
3031 void *consumer_thread_sessiond_poll(void *data)
3032 {
3033 int sock = -1, client_socket, ret, err = -1;
3034 /*
3035 * structure to poll for incoming data on communication socket avoids
3036 * making blocking sockets.
3037 */
3038 struct pollfd consumer_sockpoll[2];
3039 struct lttng_consumer_local_data *ctx = data;
3040
3041 rcu_register_thread();
3042
3043 health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_SESSIOND);
3044
3045 if (testpoint(consumerd_thread_sessiond)) {
3046 goto error_testpoint;
3047 }
3048
3049 health_code_update();
3050
3051 DBG("Creating command socket %s", ctx->consumer_command_sock_path);
3052 unlink(ctx->consumer_command_sock_path);
3053 client_socket = lttcomm_create_unix_sock(ctx->consumer_command_sock_path);
3054 if (client_socket < 0) {
3055 ERR("Cannot create command socket");
3056 goto end;
3057 }
3058
3059 ret = lttcomm_listen_unix_sock(client_socket);
3060 if (ret < 0) {
3061 goto end;
3062 }
3063
3064 DBG("Sending ready command to lttng-sessiond");
3065 ret = lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY);
3066 /* return < 0 on error, but == 0 is not fatal */
3067 if (ret < 0) {
3068 ERR("Error sending ready command to lttng-sessiond");
3069 goto end;
3070 }
3071
3072 /* prepare the FDs to poll : to client socket and the should_quit pipe */
3073 consumer_sockpoll[0].fd = ctx->consumer_should_quit[0];
3074 consumer_sockpoll[0].events = POLLIN | POLLPRI;
3075 consumer_sockpoll[1].fd = client_socket;
3076 consumer_sockpoll[1].events = POLLIN | POLLPRI;
3077
3078 ret = lttng_consumer_poll_socket(consumer_sockpoll);
3079 if (ret) {
3080 if (ret > 0) {
3081 /* should exit */
3082 err = 0;
3083 }
3084 goto end;
3085 }
3086 DBG("Connection on client_socket");
3087
3088 /* Blocking call, waiting for transmission */
3089 sock = lttcomm_accept_unix_sock(client_socket);
3090 if (sock < 0) {
3091 WARN("On accept");
3092 goto end;
3093 }
3094
3095 /*
3096 * Setup metadata socket which is the second socket connection on the
3097 * command unix socket.
3098 */
3099 ret = set_metadata_socket(ctx, consumer_sockpoll, client_socket);
3100 if (ret) {
3101 if (ret > 0) {
3102 /* should exit */
3103 err = 0;
3104 }
3105 goto end;
3106 }
3107
3108 /* This socket is not useful anymore. */
3109 ret = close(client_socket);
3110 if (ret < 0) {
3111 PERROR("close client_socket");
3112 }
3113 client_socket = -1;
3114
3115 /* update the polling structure to poll on the established socket */
3116 consumer_sockpoll[1].fd = sock;
3117 consumer_sockpoll[1].events = POLLIN | POLLPRI;
3118
3119 while (1) {
3120 health_code_update();
3121
3122 health_poll_entry();
3123 ret = lttng_consumer_poll_socket(consumer_sockpoll);
3124 health_poll_exit();
3125 if (ret) {
3126 if (ret > 0) {
3127 /* should exit */
3128 err = 0;
3129 }
3130 goto end;
3131 }
3132 DBG("Incoming command on sock");
3133 ret = lttng_consumer_recv_cmd(ctx, sock, consumer_sockpoll);
3134 if (ret <= 0) {
3135 /*
3136 * This could simply be a session daemon quitting. Don't output
3137 * ERR() here.
3138 */
3139 DBG("Communication interrupted on command socket");
3140 err = 0;
3141 goto end;
3142 }
3143 if (consumer_quit) {
3144 DBG("consumer_thread_receive_fds received quit from signal");
3145 err = 0; /* All is OK */
3146 goto end;
3147 }
3148 DBG("received command on sock");
3149 }
3150 /* All is OK */
3151 err = 0;
3152
3153 end:
3154 DBG("Consumer thread sessiond poll exiting");
3155
3156 /*
3157 * Close metadata streams since the producer is the session daemon which
3158 * just died.
3159 *
3160 * NOTE: for now, this only applies to the UST tracer.
3161 */
3162 lttng_consumer_close_all_metadata();
3163
3164 /*
3165 * when all fds have hung up, the polling thread
3166 * can exit cleanly
3167 */
3168 consumer_quit = 1;
3169
3170 /*
3171 * Notify the data poll thread to poll back again and test the
3172 * consumer_quit state that we just set so to quit gracefully.
3173 */
3174 notify_thread_lttng_pipe(ctx->consumer_data_pipe);
3175
3176 notify_channel_pipe(ctx, NULL, -1, CONSUMER_CHANNEL_QUIT);
3177
3178 notify_health_quit_pipe(health_quit_pipe);
3179
3180 /* Cleaning up possibly open sockets. */
3181 if (sock >= 0) {
3182 ret = close(sock);
3183 if (ret < 0) {
3184 PERROR("close sock sessiond poll");
3185 }
3186 }
3187 if (client_socket >= 0) {
3188 ret = close(client_socket);
3189 if (ret < 0) {
3190 PERROR("close client_socket sessiond poll");
3191 }
3192 }
3193
3194 error_testpoint:
3195 if (err) {
3196 health_error();
3197 ERR("Health error occurred in %s", __func__);
3198 }
3199 health_unregister(health_consumerd);
3200
3201 rcu_unregister_thread();
3202 return NULL;
3203 }
3204
3205 ssize_t lttng_consumer_read_subbuffer(struct lttng_consumer_stream *stream,
3206 struct lttng_consumer_local_data *ctx)
3207 {
3208 ssize_t ret;
3209
3210 pthread_mutex_lock(&stream->lock);
3211 if (stream->metadata_flag) {
3212 pthread_mutex_lock(&stream->metadata_rdv_lock);
3213 }
3214
3215 switch (consumer_data.type) {
3216 case LTTNG_CONSUMER_KERNEL:
3217 ret = lttng_kconsumer_read_subbuffer(stream, ctx);
3218 break;
3219 case LTTNG_CONSUMER32_UST:
3220 case LTTNG_CONSUMER64_UST:
3221 ret = lttng_ustconsumer_read_subbuffer(stream, ctx);
3222 break;
3223 default:
3224 ERR("Unknown consumer_data type");
3225 assert(0);
3226 ret = -ENOSYS;
3227 break;
3228 }
3229
3230 if (stream->metadata_flag) {
3231 pthread_cond_broadcast(&stream->metadata_rdv);
3232 pthread_mutex_unlock