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