8d1a34025687da37b7c3c1555d80441c93d8761d
[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
32 #include <common/common.h>
33 #include <common/utils.h>
34 #include <common/compat/poll.h>
35 #include <common/kernel-ctl/kernel-ctl.h>
36 #include <common/sessiond-comm/relayd.h>
37 #include <common/sessiond-comm/sessiond-comm.h>
38 #include <common/kernel-consumer/kernel-consumer.h>
39 #include <common/relayd/relayd.h>
40 #include <common/ust-consumer/ust-consumer.h>
41
42 #include "consumer.h"
43
44 struct lttng_consumer_global_data consumer_data = {
45 .stream_count = 0,
46 .need_update = 1,
47 .type = LTTNG_CONSUMER_UNKNOWN,
48 };
49
50 /* timeout parameter, to control the polling thread grace period. */
51 int consumer_poll_timeout = -1;
52
53 /*
54 * Flag to inform the polling thread to quit when all fd hung up. Updated by
55 * the consumer_thread_receive_fds when it notices that all fds has hung up.
56 * Also updated by the signal handler (consumer_should_exit()). Read by the
57 * polling threads.
58 */
59 volatile int consumer_quit = 0;
60
61 /*
62 * The following two hash tables are visible by all threads which are separated
63 * in different source files.
64 *
65 * Global hash table containing respectively metadata and data streams. The
66 * stream element in this ht should only be updated by the metadata poll thread
67 * for the metadata and the data poll thread for the data.
68 */
69 struct lttng_ht *metadata_ht = NULL;
70 struct lttng_ht *data_ht = NULL;
71
72 /*
73 * Find a stream. The consumer_data.lock must be locked during this
74 * call.
75 */
76 static struct lttng_consumer_stream *consumer_find_stream(int key,
77 struct lttng_ht *ht)
78 {
79 struct lttng_ht_iter iter;
80 struct lttng_ht_node_ulong *node;
81 struct lttng_consumer_stream *stream = NULL;
82
83 assert(ht);
84
85 /* Negative keys are lookup failures */
86 if (key < 0) {
87 return NULL;
88 }
89
90 rcu_read_lock();
91
92 lttng_ht_lookup(ht, (void *)((unsigned long) key), &iter);
93 node = lttng_ht_iter_get_node_ulong(&iter);
94 if (node != NULL) {
95 stream = caa_container_of(node, struct lttng_consumer_stream, node);
96 }
97
98 rcu_read_unlock();
99
100 return stream;
101 }
102
103 void consumer_steal_stream_key(int key, struct lttng_ht *ht)
104 {
105 struct lttng_consumer_stream *stream;
106
107 rcu_read_lock();
108 stream = consumer_find_stream(key, ht);
109 if (stream) {
110 stream->key = -1;
111 /*
112 * We don't want the lookup to match, but we still need
113 * to iterate on this stream when iterating over the hash table. Just
114 * change the node key.
115 */
116 stream->node.key = -1;
117 }
118 rcu_read_unlock();
119 }
120
121 static struct lttng_consumer_channel *consumer_find_channel(int key)
122 {
123 struct lttng_ht_iter iter;
124 struct lttng_ht_node_ulong *node;
125 struct lttng_consumer_channel *channel = NULL;
126
127 /* Negative keys are lookup failures */
128 if (key < 0) {
129 return NULL;
130 }
131
132 rcu_read_lock();
133
134 lttng_ht_lookup(consumer_data.channel_ht, (void *)((unsigned long) key),
135 &iter);
136 node = lttng_ht_iter_get_node_ulong(&iter);
137 if (node != NULL) {
138 channel = caa_container_of(node, struct lttng_consumer_channel, node);
139 }
140
141 rcu_read_unlock();
142
143 return channel;
144 }
145
146 static void consumer_steal_channel_key(int key)
147 {
148 struct lttng_consumer_channel *channel;
149
150 rcu_read_lock();
151 channel = consumer_find_channel(key);
152 if (channel) {
153 channel->key = -1;
154 /*
155 * We don't want the lookup to match, but we still need
156 * to iterate on this channel when iterating over the hash table. Just
157 * change the node key.
158 */
159 channel->node.key = -1;
160 }
161 rcu_read_unlock();
162 }
163
164 static
165 void consumer_free_stream(struct rcu_head *head)
166 {
167 struct lttng_ht_node_ulong *node =
168 caa_container_of(head, struct lttng_ht_node_ulong, head);
169 struct lttng_consumer_stream *stream =
170 caa_container_of(node, struct lttng_consumer_stream, node);
171
172 free(stream);
173 }
174
175 /*
176 * RCU protected relayd socket pair free.
177 */
178 static void consumer_rcu_free_relayd(struct rcu_head *head)
179 {
180 struct lttng_ht_node_ulong *node =
181 caa_container_of(head, struct lttng_ht_node_ulong, head);
182 struct consumer_relayd_sock_pair *relayd =
183 caa_container_of(node, struct consumer_relayd_sock_pair, node);
184
185 free(relayd);
186 }
187
188 /*
189 * Destroy and free relayd socket pair object.
190 *
191 * This function MUST be called with the consumer_data lock acquired.
192 */
193 static void destroy_relayd(struct consumer_relayd_sock_pair *relayd)
194 {
195 int ret;
196 struct lttng_ht_iter iter;
197
198 if (relayd == NULL) {
199 return;
200 }
201
202 DBG("Consumer destroy and close relayd socket pair");
203
204 iter.iter.node = &relayd->node.node;
205 ret = lttng_ht_del(consumer_data.relayd_ht, &iter);
206 if (ret != 0) {
207 /* We assume the relayd was already destroyed */
208 return;
209 }
210
211 /* Close all sockets */
212 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
213 (void) relayd_close(&relayd->control_sock);
214 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
215 (void) relayd_close(&relayd->data_sock);
216
217 /* RCU free() call */
218 call_rcu(&relayd->node.head, consumer_rcu_free_relayd);
219 }
220
221 /*
222 * Flag a relayd socket pair for destruction. Destroy it if the refcount
223 * reaches zero.
224 *
225 * RCU read side lock MUST be aquired before calling this function.
226 */
227 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair *relayd)
228 {
229 assert(relayd);
230
231 /* Set destroy flag for this object */
232 uatomic_set(&relayd->destroy_flag, 1);
233
234 /* Destroy the relayd if refcount is 0 */
235 if (uatomic_read(&relayd->refcount) == 0) {
236 destroy_relayd(relayd);
237 }
238 }
239
240 /*
241 * Remove a stream from the global list protected by a mutex. This
242 * function is also responsible for freeing its data structures.
243 */
244 void consumer_del_stream(struct lttng_consumer_stream *stream,
245 struct lttng_ht *ht)
246 {
247 int ret;
248 struct lttng_ht_iter iter;
249 struct lttng_consumer_channel *free_chan = NULL;
250 struct consumer_relayd_sock_pair *relayd;
251
252 assert(stream);
253
254 if (ht == NULL) {
255 /* Means the stream was allocated but not successfully added */
256 goto free_stream;
257 }
258
259 pthread_mutex_lock(&consumer_data.lock);
260
261 switch (consumer_data.type) {
262 case LTTNG_CONSUMER_KERNEL:
263 if (stream->mmap_base != NULL) {
264 ret = munmap(stream->mmap_base, stream->mmap_len);
265 if (ret != 0) {
266 PERROR("munmap");
267 }
268 }
269 break;
270 case LTTNG_CONSUMER32_UST:
271 case LTTNG_CONSUMER64_UST:
272 lttng_ustconsumer_del_stream(stream);
273 break;
274 default:
275 ERR("Unknown consumer_data type");
276 assert(0);
277 goto end;
278 }
279
280 rcu_read_lock();
281 iter.iter.node = &stream->node.node;
282 ret = lttng_ht_del(ht, &iter);
283 assert(!ret);
284
285 /* Remove node session id from the consumer_data stream ht */
286 iter.iter.node = &stream->node_session_id.node;
287 ret = lttng_ht_del(consumer_data.stream_list_ht, &iter);
288 assert(!ret);
289 rcu_read_unlock();
290
291 assert(consumer_data.stream_count > 0);
292 consumer_data.stream_count--;
293
294 if (stream->out_fd >= 0) {
295 ret = close(stream->out_fd);
296 if (ret) {
297 PERROR("close");
298 }
299 }
300 if (stream->wait_fd >= 0 && !stream->wait_fd_is_copy) {
301 ret = close(stream->wait_fd);
302 if (ret) {
303 PERROR("close");
304 }
305 }
306 if (stream->shm_fd >= 0 && stream->wait_fd != stream->shm_fd) {
307 ret = close(stream->shm_fd);
308 if (ret) {
309 PERROR("close");
310 }
311 }
312
313 /* Check and cleanup relayd */
314 rcu_read_lock();
315 relayd = consumer_find_relayd(stream->net_seq_idx);
316 if (relayd != NULL) {
317 uatomic_dec(&relayd->refcount);
318 assert(uatomic_read(&relayd->refcount) >= 0);
319
320 /* Closing streams requires to lock the control socket. */
321 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
322 ret = relayd_send_close_stream(&relayd->control_sock,
323 stream->relayd_stream_id,
324 stream->next_net_seq_num - 1);
325 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
326 if (ret < 0) {
327 DBG("Unable to close stream on the relayd. Continuing");
328 /*
329 * Continue here. There is nothing we can do for the relayd.
330 * Chances are that the relayd has closed the socket so we just
331 * continue cleaning up.
332 */
333 }
334
335 /* Both conditions are met, we destroy the relayd. */
336 if (uatomic_read(&relayd->refcount) == 0 &&
337 uatomic_read(&relayd->destroy_flag)) {
338 destroy_relayd(relayd);
339 }
340 }
341 rcu_read_unlock();
342
343 uatomic_dec(&stream->chan->refcount);
344 if (!uatomic_read(&stream->chan->refcount)
345 && !uatomic_read(&stream->chan->nb_init_streams)) {
346 free_chan = stream->chan;
347 }
348
349 end:
350 consumer_data.need_update = 1;
351 pthread_mutex_unlock(&consumer_data.lock);
352
353 if (free_chan) {
354 consumer_del_channel(free_chan);
355 }
356
357 free_stream:
358 call_rcu(&stream->node.head, consumer_free_stream);
359 }
360
361 struct lttng_consumer_stream *consumer_allocate_stream(
362 int channel_key, int stream_key,
363 int shm_fd, int wait_fd,
364 enum lttng_consumer_stream_state state,
365 uint64_t mmap_len,
366 enum lttng_event_output output,
367 const char *path_name,
368 uid_t uid,
369 gid_t gid,
370 int net_index,
371 int metadata_flag,
372 uint64_t session_id,
373 int *alloc_ret)
374 {
375 struct lttng_consumer_stream *stream;
376
377 stream = zmalloc(sizeof(*stream));
378 if (stream == NULL) {
379 PERROR("malloc struct lttng_consumer_stream");
380 *alloc_ret = -ENOMEM;
381 goto end;
382 }
383
384 /*
385 * Get stream's channel reference. Needed when adding the stream to the
386 * global hash table.
387 */
388 stream->chan = consumer_find_channel(channel_key);
389 if (!stream->chan) {
390 *alloc_ret = -ENOENT;
391 ERR("Unable to find channel for stream %d", stream_key);
392 goto error;
393 }
394
395 stream->key = stream_key;
396 stream->shm_fd = shm_fd;
397 stream->wait_fd = wait_fd;
398 stream->out_fd = -1;
399 stream->out_fd_offset = 0;
400 stream->state = state;
401 stream->mmap_len = mmap_len;
402 stream->mmap_base = NULL;
403 stream->output = output;
404 stream->uid = uid;
405 stream->gid = gid;
406 stream->net_seq_idx = net_index;
407 stream->metadata_flag = metadata_flag;
408 stream->session_id = session_id;
409 strncpy(stream->path_name, path_name, sizeof(stream->path_name));
410 stream->path_name[sizeof(stream->path_name) - 1] = '\0';
411 pthread_mutex_init(&stream->lock, NULL);
412
413 /*
414 * Index differently the metadata node because the thread is using an
415 * internal hash table to match streams in the metadata_ht to the epoll set
416 * file descriptor.
417 */
418 if (metadata_flag) {
419 lttng_ht_node_init_ulong(&stream->node, stream->wait_fd);
420 } else {
421 lttng_ht_node_init_ulong(&stream->node, stream->key);
422 }
423
424 /* Init session id node with the stream session id */
425 lttng_ht_node_init_ulong(&stream->node_session_id, stream->session_id);
426
427 /*
428 * The cpu number is needed before using any ustctl_* actions. Ignored for
429 * the kernel so the value does not matter.
430 */
431 pthread_mutex_lock(&consumer_data.lock);
432 stream->cpu = stream->chan->cpucount++;
433 pthread_mutex_unlock(&consumer_data.lock);
434
435 DBG3("Allocated stream %s (key %d, shm_fd %d, wait_fd %d, mmap_len %llu,"
436 " out_fd %d, net_seq_idx %d, session_id %" PRIu64,
437 stream->path_name, stream->key, stream->shm_fd, stream->wait_fd,
438 (unsigned long long) stream->mmap_len, stream->out_fd,
439 stream->net_seq_idx, stream->session_id);
440 return stream;
441
442 error:
443 free(stream);
444 end:
445 return NULL;
446 }
447
448 /*
449 * Add a stream to the global list protected by a mutex.
450 */
451 static int consumer_add_stream(struct lttng_consumer_stream *stream,
452 struct lttng_ht *ht)
453 {
454 int ret = 0;
455 struct consumer_relayd_sock_pair *relayd;
456
457 assert(stream);
458 assert(ht);
459
460 DBG3("Adding consumer stream %d", stream->key);
461
462 pthread_mutex_lock(&consumer_data.lock);
463 rcu_read_lock();
464
465 /* Steal stream identifier to avoid having streams with the same key */
466 consumer_steal_stream_key(stream->key, ht);
467
468 lttng_ht_add_unique_ulong(ht, &stream->node);
469
470 /*
471 * Add stream to the stream_list_ht of the consumer data. No need to steal
472 * the key since the HT does not use it and we allow to add redundant keys
473 * into this table.
474 */
475 lttng_ht_add_ulong(consumer_data.stream_list_ht, &stream->node_session_id);
476
477 /* Check and cleanup relayd */
478 relayd = consumer_find_relayd(stream->net_seq_idx);
479 if (relayd != NULL) {
480 uatomic_inc(&relayd->refcount);
481 }
482
483 /* Update channel refcount once added without error(s). */
484 uatomic_inc(&stream->chan->refcount);
485
486 /*
487 * When nb_init_streams reaches 0, we don't need to trigger any action in
488 * terms of destroying the associated channel, because the action that
489 * causes the count to become 0 also causes a stream to be added. The
490 * channel deletion will thus be triggered by the following removal of this
491 * stream.
492 */
493 if (uatomic_read(&stream->chan->nb_init_streams) > 0) {
494 uatomic_dec(&stream->chan->nb_init_streams);
495 }
496
497 /* Update consumer data once the node is inserted. */
498 consumer_data.stream_count++;
499 consumer_data.need_update = 1;
500
501 rcu_read_unlock();
502 pthread_mutex_unlock(&consumer_data.lock);
503
504 return ret;
505 }
506
507 /*
508 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
509 * be acquired before calling this.
510 */
511 static int add_relayd(struct consumer_relayd_sock_pair *relayd)
512 {
513 int ret = 0;
514 struct lttng_ht_node_ulong *node;
515 struct lttng_ht_iter iter;
516
517 if (relayd == NULL) {
518 ret = -1;
519 goto end;
520 }
521
522 lttng_ht_lookup(consumer_data.relayd_ht,
523 (void *)((unsigned long) relayd->net_seq_idx), &iter);
524 node = lttng_ht_iter_get_node_ulong(&iter);
525 if (node != NULL) {
526 /* Relayd already exist. Ignore the insertion */
527 goto end;
528 }
529 lttng_ht_add_unique_ulong(consumer_data.relayd_ht, &relayd->node);
530
531 end:
532 return ret;
533 }
534
535 /*
536 * Allocate and return a consumer relayd socket.
537 */
538 struct consumer_relayd_sock_pair *consumer_allocate_relayd_sock_pair(
539 int net_seq_idx)
540 {
541 struct consumer_relayd_sock_pair *obj = NULL;
542
543 /* Negative net sequence index is a failure */
544 if (net_seq_idx < 0) {
545 goto error;
546 }
547
548 obj = zmalloc(sizeof(struct consumer_relayd_sock_pair));
549 if (obj == NULL) {
550 PERROR("zmalloc relayd sock");
551 goto error;
552 }
553
554 obj->net_seq_idx = net_seq_idx;
555 obj->refcount = 0;
556 obj->destroy_flag = 0;
557 lttng_ht_node_init_ulong(&obj->node, obj->net_seq_idx);
558 pthread_mutex_init(&obj->ctrl_sock_mutex, NULL);
559
560 error:
561 return obj;
562 }
563
564 /*
565 * Find a relayd socket pair in the global consumer data.
566 *
567 * Return the object if found else NULL.
568 * RCU read-side lock must be held across this call and while using the
569 * returned object.
570 */
571 struct consumer_relayd_sock_pair *consumer_find_relayd(int key)
572 {
573 struct lttng_ht_iter iter;
574 struct lttng_ht_node_ulong *node;
575 struct consumer_relayd_sock_pair *relayd = NULL;
576
577 /* Negative keys are lookup failures */
578 if (key < 0) {
579 goto error;
580 }
581
582 lttng_ht_lookup(consumer_data.relayd_ht, (void *)((unsigned long) key),
583 &iter);
584 node = lttng_ht_iter_get_node_ulong(&iter);
585 if (node != NULL) {
586 relayd = caa_container_of(node, struct consumer_relayd_sock_pair, node);
587 }
588
589 error:
590 return relayd;
591 }
592
593 /*
594 * Handle stream for relayd transmission if the stream applies for network
595 * streaming where the net sequence index is set.
596 *
597 * Return destination file descriptor or negative value on error.
598 */
599 static int write_relayd_stream_header(struct lttng_consumer_stream *stream,
600 size_t data_size, unsigned long padding,
601 struct consumer_relayd_sock_pair *relayd)
602 {
603 int outfd = -1, ret;
604 struct lttcomm_relayd_data_hdr data_hdr;
605
606 /* Safety net */
607 assert(stream);
608 assert(relayd);
609
610 /* Reset data header */
611 memset(&data_hdr, 0, sizeof(data_hdr));
612
613 if (stream->metadata_flag) {
614 /* Caller MUST acquire the relayd control socket lock */
615 ret = relayd_send_metadata(&relayd->control_sock, data_size);
616 if (ret < 0) {
617 goto error;
618 }
619
620 /* Metadata are always sent on the control socket. */
621 outfd = relayd->control_sock.fd;
622 } else {
623 /* Set header with stream information */
624 data_hdr.stream_id = htobe64(stream->relayd_stream_id);
625 data_hdr.data_size = htobe32(data_size);
626 data_hdr.padding_size = htobe32(padding);
627 data_hdr.net_seq_num = htobe64(stream->next_net_seq_num++);
628 /* Other fields are zeroed previously */
629
630 ret = relayd_send_data_hdr(&relayd->data_sock, &data_hdr,
631 sizeof(data_hdr));
632 if (ret < 0) {
633 goto error;
634 }
635
636 /* Set to go on data socket */
637 outfd = relayd->data_sock.fd;
638 }
639
640 error:
641 return outfd;
642 }
643
644 static
645 void consumer_free_channel(struct rcu_head *head)
646 {
647 struct lttng_ht_node_ulong *node =
648 caa_container_of(head, struct lttng_ht_node_ulong, head);
649 struct lttng_consumer_channel *channel =
650 caa_container_of(node, struct lttng_consumer_channel, node);
651
652 free(channel);
653 }
654
655 /*
656 * Remove a channel from the global list protected by a mutex. This
657 * function is also responsible for freeing its data structures.
658 */
659 void consumer_del_channel(struct lttng_consumer_channel *channel)
660 {
661 int ret;
662 struct lttng_ht_iter iter;
663
664 pthread_mutex_lock(&consumer_data.lock);
665
666 switch (consumer_data.type) {
667 case LTTNG_CONSUMER_KERNEL:
668 break;
669 case LTTNG_CONSUMER32_UST:
670 case LTTNG_CONSUMER64_UST:
671 lttng_ustconsumer_del_channel(channel);
672 break;
673 default:
674 ERR("Unknown consumer_data type");
675 assert(0);
676 goto end;
677 }
678
679 rcu_read_lock();
680 iter.iter.node = &channel->node.node;
681 ret = lttng_ht_del(consumer_data.channel_ht, &iter);
682 assert(!ret);
683 rcu_read_unlock();
684
685 if (channel->mmap_base != NULL) {
686 ret = munmap(channel->mmap_base, channel->mmap_len);
687 if (ret != 0) {
688 PERROR("munmap");
689 }
690 }
691 if (channel->wait_fd >= 0 && !channel->wait_fd_is_copy) {
692 ret = close(channel->wait_fd);
693 if (ret) {
694 PERROR("close");
695 }
696 }
697 if (channel->shm_fd >= 0 && channel->wait_fd != channel->shm_fd) {
698 ret = close(channel->shm_fd);
699 if (ret) {
700 PERROR("close");
701 }
702 }
703
704 call_rcu(&channel->node.head, consumer_free_channel);
705 end:
706 pthread_mutex_unlock(&consumer_data.lock);
707 }
708
709 struct lttng_consumer_channel *consumer_allocate_channel(
710 int channel_key,
711 int shm_fd, int wait_fd,
712 uint64_t mmap_len,
713 uint64_t max_sb_size,
714 unsigned int nb_init_streams)
715 {
716 struct lttng_consumer_channel *channel;
717 int ret;
718
719 channel = zmalloc(sizeof(*channel));
720 if (channel == NULL) {
721 PERROR("malloc struct lttng_consumer_channel");
722 goto end;
723 }
724 channel->key = channel_key;
725 channel->shm_fd = shm_fd;
726 channel->wait_fd = wait_fd;
727 channel->mmap_len = mmap_len;
728 channel->max_sb_size = max_sb_size;
729 channel->refcount = 0;
730 channel->nb_init_streams = nb_init_streams;
731 lttng_ht_node_init_ulong(&channel->node, channel->key);
732
733 switch (consumer_data.type) {
734 case LTTNG_CONSUMER_KERNEL:
735 channel->mmap_base = NULL;
736 channel->mmap_len = 0;
737 break;
738 case LTTNG_CONSUMER32_UST:
739 case LTTNG_CONSUMER64_UST:
740 ret = lttng_ustconsumer_allocate_channel(channel);
741 if (ret) {
742 free(channel);
743 return NULL;
744 }
745 break;
746 default:
747 ERR("Unknown consumer_data type");
748 assert(0);
749 goto end;
750 }
751 DBG("Allocated channel (key %d, shm_fd %d, wait_fd %d, mmap_len %llu, max_sb_size %llu)",
752 channel->key, channel->shm_fd, channel->wait_fd,
753 (unsigned long long) channel->mmap_len,
754 (unsigned long long) channel->max_sb_size);
755 end:
756 return channel;
757 }
758
759 /*
760 * Add a channel to the global list protected by a mutex.
761 */
762 int consumer_add_channel(struct lttng_consumer_channel *channel)
763 {
764 struct lttng_ht_node_ulong *node;
765 struct lttng_ht_iter iter;
766
767 pthread_mutex_lock(&consumer_data.lock);
768 /* Steal channel identifier, for UST */
769 consumer_steal_channel_key(channel->key);
770 rcu_read_lock();
771
772 lttng_ht_lookup(consumer_data.channel_ht,
773 (void *)((unsigned long) channel->key), &iter);
774 node = lttng_ht_iter_get_node_ulong(&iter);
775 if (node != NULL) {
776 /* Channel already exist. Ignore the insertion */
777 goto end;
778 }
779
780 lttng_ht_add_unique_ulong(consumer_data.channel_ht, &channel->node);
781
782 end:
783 rcu_read_unlock();
784 pthread_mutex_unlock(&consumer_data.lock);
785
786 return 0;
787 }
788
789 /*
790 * Allocate the pollfd structure and the local view of the out fds to avoid
791 * doing a lookup in the linked list and concurrency issues when writing is
792 * needed. Called with consumer_data.lock held.
793 *
794 * Returns the number of fds in the structures.
795 */
796 static int consumer_update_poll_array(
797 struct lttng_consumer_local_data *ctx, struct pollfd **pollfd,
798 struct lttng_consumer_stream **local_stream, struct lttng_ht *ht)
799 {
800 int i = 0;
801 struct lttng_ht_iter iter;
802 struct lttng_consumer_stream *stream;
803
804 DBG("Updating poll fd array");
805 rcu_read_lock();
806 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
807 if (stream->state != LTTNG_CONSUMER_ACTIVE_STREAM) {
808 continue;
809 }
810 DBG("Active FD %d", stream->wait_fd);
811 (*pollfd)[i].fd = stream->wait_fd;
812 (*pollfd)[i].events = POLLIN | POLLPRI;
813 local_stream[i] = stream;
814 i++;
815 }
816 rcu_read_unlock();
817
818 /*
819 * Insert the consumer_data_pipe at the end of the array and don't
820 * increment i so nb_fd is the number of real FD.
821 */
822 (*pollfd)[i].fd = ctx->consumer_data_pipe[0];
823 (*pollfd)[i].events = POLLIN | POLLPRI;
824 return i;
825 }
826
827 /*
828 * Poll on the should_quit pipe and the command socket return -1 on error and
829 * should exit, 0 if data is available on the command socket
830 */
831 int lttng_consumer_poll_socket(struct pollfd *consumer_sockpoll)
832 {
833 int num_rdy;
834
835 restart:
836 num_rdy = poll(consumer_sockpoll, 2, -1);
837 if (num_rdy == -1) {
838 /*
839 * Restart interrupted system call.
840 */
841 if (errno == EINTR) {
842 goto restart;
843 }
844 PERROR("Poll error");
845 goto exit;
846 }
847 if (consumer_sockpoll[0].revents & (POLLIN | POLLPRI)) {
848 DBG("consumer_should_quit wake up");
849 goto exit;
850 }
851 return 0;
852
853 exit:
854 return -1;
855 }
856
857 /*
858 * Set the error socket.
859 */
860 void lttng_consumer_set_error_sock(
861 struct lttng_consumer_local_data *ctx, int sock)
862 {
863 ctx->consumer_error_socket = sock;
864 }
865
866 /*
867 * Set the command socket path.
868 */
869 void lttng_consumer_set_command_sock_path(
870 struct lttng_consumer_local_data *ctx, char *sock)
871 {
872 ctx->consumer_command_sock_path = sock;
873 }
874
875 /*
876 * Send return code to the session daemon.
877 * If the socket is not defined, we return 0, it is not a fatal error
878 */
879 int lttng_consumer_send_error(
880 struct lttng_consumer_local_data *ctx, int cmd)
881 {
882 if (ctx->consumer_error_socket > 0) {
883 return lttcomm_send_unix_sock(ctx->consumer_error_socket, &cmd,
884 sizeof(enum lttcomm_sessiond_command));
885 }
886
887 return 0;
888 }
889
890 /*
891 * Close all the tracefiles and stream fds, should be called when all instances
892 * are destroyed.
893 */
894 void lttng_consumer_cleanup(void)
895 {
896 struct lttng_ht_iter iter;
897 struct lttng_ht_node_ulong *node;
898
899 rcu_read_lock();
900
901 cds_lfht_for_each_entry(consumer_data.channel_ht->ht, &iter.iter, node,
902 node) {
903 struct lttng_consumer_channel *channel =
904 caa_container_of(node, struct lttng_consumer_channel, node);
905 consumer_del_channel(channel);
906 }
907
908 rcu_read_unlock();
909
910 lttng_ht_destroy(consumer_data.channel_ht);
911 }
912
913 /*
914 * Called from signal handler.
915 */
916 void lttng_consumer_should_exit(struct lttng_consumer_local_data *ctx)
917 {
918 int ret;
919 consumer_quit = 1;
920 do {
921 ret = write(ctx->consumer_should_quit[1], "4", 1);
922 } while (ret < 0 && errno == EINTR);
923 if (ret < 0) {
924 PERROR("write consumer quit");
925 }
926 }
927
928 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream *stream,
929 off_t orig_offset)
930 {
931 int outfd = stream->out_fd;
932
933 /*
934 * This does a blocking write-and-wait on any page that belongs to the
935 * subbuffer prior to the one we just wrote.
936 * Don't care about error values, as these are just hints and ways to
937 * limit the amount of page cache used.
938 */
939 if (orig_offset < stream->chan->max_sb_size) {
940 return;
941 }
942 lttng_sync_file_range(outfd, orig_offset - stream->chan->max_sb_size,
943 stream->chan->max_sb_size,
944 SYNC_FILE_RANGE_WAIT_BEFORE
945 | SYNC_FILE_RANGE_WRITE
946 | SYNC_FILE_RANGE_WAIT_AFTER);
947 /*
948 * Give hints to the kernel about how we access the file:
949 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
950 * we write it.
951 *
952 * We need to call fadvise again after the file grows because the
953 * kernel does not seem to apply fadvise to non-existing parts of the
954 * file.
955 *
956 * Call fadvise _after_ having waited for the page writeback to
957 * complete because the dirty page writeback semantic is not well
958 * defined. So it can be expected to lead to lower throughput in
959 * streaming.
960 */
961 posix_fadvise(outfd, orig_offset - stream->chan->max_sb_size,
962 stream->chan->max_sb_size, POSIX_FADV_DONTNEED);
963 }
964
965 /*
966 * Initialise the necessary environnement :
967 * - create a new context
968 * - create the poll_pipe
969 * - create the should_quit pipe (for signal handler)
970 * - create the thread pipe (for splice)
971 *
972 * Takes a function pointer as argument, this function is called when data is
973 * available on a buffer. This function is responsible to do the
974 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
975 * buffer configuration and then kernctl_put_next_subbuf at the end.
976 *
977 * Returns a pointer to the new context or NULL on error.
978 */
979 struct lttng_consumer_local_data *lttng_consumer_create(
980 enum lttng_consumer_type type,
981 ssize_t (*buffer_ready)(struct lttng_consumer_stream *stream,
982 struct lttng_consumer_local_data *ctx),
983 int (*recv_channel)(struct lttng_consumer_channel *channel),
984 int (*recv_stream)(struct lttng_consumer_stream *stream),
985 int (*update_stream)(int stream_key, uint32_t state))
986 {
987 int ret, i;
988 struct lttng_consumer_local_data *ctx;
989
990 assert(consumer_data.type == LTTNG_CONSUMER_UNKNOWN ||
991 consumer_data.type == type);
992 consumer_data.type = type;
993
994 ctx = zmalloc(sizeof(struct lttng_consumer_local_data));
995 if (ctx == NULL) {
996 PERROR("allocating context");
997 goto error;
998 }
999
1000 ctx->consumer_error_socket = -1;
1001 /* assign the callbacks */
1002 ctx->on_buffer_ready = buffer_ready;
1003 ctx->on_recv_channel = recv_channel;
1004 ctx->on_recv_stream = recv_stream;
1005 ctx->on_update_stream = update_stream;
1006
1007 ret = pipe(ctx->consumer_data_pipe);
1008 if (ret < 0) {
1009 PERROR("Error creating poll pipe");
1010 goto error_poll_pipe;
1011 }
1012
1013 /* set read end of the pipe to non-blocking */
1014 ret = fcntl(ctx->consumer_data_pipe[0], F_SETFL, O_NONBLOCK);
1015 if (ret < 0) {
1016 PERROR("fcntl O_NONBLOCK");
1017 goto error_poll_fcntl;
1018 }
1019
1020 /* set write end of the pipe to non-blocking */
1021 ret = fcntl(ctx->consumer_data_pipe[1], F_SETFL, O_NONBLOCK);
1022 if (ret < 0) {
1023 PERROR("fcntl O_NONBLOCK");
1024 goto error_poll_fcntl;
1025 }
1026
1027 ret = pipe(ctx->consumer_should_quit);
1028 if (ret < 0) {
1029 PERROR("Error creating recv pipe");
1030 goto error_quit_pipe;
1031 }
1032
1033 ret = pipe(ctx->consumer_thread_pipe);
1034 if (ret < 0) {
1035 PERROR("Error creating thread pipe");
1036 goto error_thread_pipe;
1037 }
1038
1039 ret = utils_create_pipe(ctx->consumer_metadata_pipe);
1040 if (ret < 0) {
1041 goto error_metadata_pipe;
1042 }
1043
1044 ret = utils_create_pipe(ctx->consumer_splice_metadata_pipe);
1045 if (ret < 0) {
1046 goto error_splice_pipe;
1047 }
1048
1049 return ctx;
1050
1051 error_splice_pipe:
1052 utils_close_pipe(ctx->consumer_metadata_pipe);
1053 error_metadata_pipe:
1054 utils_close_pipe(ctx->consumer_thread_pipe);
1055 error_thread_pipe:
1056 for (i = 0; i < 2; i++) {
1057 int err;
1058
1059 err = close(ctx->consumer_should_quit[i]);
1060 if (err) {
1061 PERROR("close");
1062 }
1063 }
1064 error_poll_fcntl:
1065 error_quit_pipe:
1066 for (i = 0; i < 2; i++) {
1067 int err;
1068
1069 err = close(ctx->consumer_data_pipe[i]);
1070 if (err) {
1071 PERROR("close");
1072 }
1073 }
1074 error_poll_pipe:
1075 free(ctx);
1076 error:
1077 return NULL;
1078 }
1079
1080 /*
1081 * Close all fds associated with the instance and free the context.
1082 */
1083 void lttng_consumer_destroy(struct lttng_consumer_local_data *ctx)
1084 {
1085 int ret;
1086
1087 ret = close(ctx->consumer_error_socket);
1088 if (ret) {
1089 PERROR("close");
1090 }
1091 ret = close(ctx->consumer_thread_pipe[0]);
1092 if (ret) {
1093 PERROR("close");
1094 }
1095 ret = close(ctx->consumer_thread_pipe[1]);
1096 if (ret) {
1097 PERROR("close");
1098 }
1099 ret = close(ctx->consumer_data_pipe[0]);
1100 if (ret) {
1101 PERROR("close");
1102 }
1103 ret = close(ctx->consumer_data_pipe[1]);
1104 if (ret) {
1105 PERROR("close");
1106 }
1107 ret = close(ctx->consumer_should_quit[0]);
1108 if (ret) {
1109 PERROR("close");
1110 }
1111 ret = close(ctx->consumer_should_quit[1]);
1112 if (ret) {
1113 PERROR("close");
1114 }
1115 utils_close_pipe(ctx->consumer_splice_metadata_pipe);
1116
1117 unlink(ctx->consumer_command_sock_path);
1118 free(ctx);
1119 }
1120
1121 /*
1122 * Write the metadata stream id on the specified file descriptor.
1123 */
1124 static int write_relayd_metadata_id(int fd,
1125 struct lttng_consumer_stream *stream,
1126 struct consumer_relayd_sock_pair *relayd,
1127 unsigned long padding)
1128 {
1129 int ret;
1130 struct lttcomm_relayd_metadata_payload hdr;
1131
1132 hdr.stream_id = htobe64(stream->relayd_stream_id);
1133 hdr.padding_size = htobe32(padding);
1134 do {
1135 ret = write(fd, (void *) &hdr, sizeof(hdr));
1136 } while (ret < 0 && errno == EINTR);
1137 if (ret < 0) {
1138 PERROR("write metadata stream id");
1139 goto end;
1140 }
1141 DBG("Metadata stream id %" PRIu64 " with padding %lu written before data",
1142 stream->relayd_stream_id, padding);
1143
1144 end:
1145 return ret;
1146 }
1147
1148 /*
1149 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1150 * core function for writing trace buffers to either the local filesystem or
1151 * the network.
1152 *
1153 * Careful review MUST be put if any changes occur!
1154 *
1155 * Returns the number of bytes written
1156 */
1157 ssize_t lttng_consumer_on_read_subbuffer_mmap(
1158 struct lttng_consumer_local_data *ctx,
1159 struct lttng_consumer_stream *stream, unsigned long len,
1160 unsigned long padding)
1161 {
1162 unsigned long mmap_offset;
1163 ssize_t ret = 0, written = 0;
1164 off_t orig_offset = stream->out_fd_offset;
1165 /* Default is on the disk */
1166 int outfd = stream->out_fd;
1167 struct consumer_relayd_sock_pair *relayd = NULL;
1168
1169 /* RCU lock for the relayd pointer */
1170 rcu_read_lock();
1171
1172 /* Flag that the current stream if set for network streaming. */
1173 if (stream->net_seq_idx != -1) {
1174 relayd = consumer_find_relayd(stream->net_seq_idx);
1175 if (relayd == NULL) {
1176 goto end;
1177 }
1178 }
1179
1180 /* get the offset inside the fd to mmap */
1181 switch (consumer_data.type) {
1182 case LTTNG_CONSUMER_KERNEL:
1183 ret = kernctl_get_mmap_read_offset(stream->wait_fd, &mmap_offset);
1184 break;
1185 case LTTNG_CONSUMER32_UST:
1186 case LTTNG_CONSUMER64_UST:
1187 ret = lttng_ustctl_get_mmap_read_offset(stream->chan->handle,
1188 stream->buf, &mmap_offset);
1189 break;
1190 default:
1191 ERR("Unknown consumer_data type");
1192 assert(0);
1193 }
1194 if (ret != 0) {
1195 errno = -ret;
1196 PERROR("tracer ctl get_mmap_read_offset");
1197 written = ret;
1198 goto end;
1199 }
1200
1201 /* Handle stream on the relayd if the output is on the network */
1202 if (relayd) {
1203 unsigned long netlen = len;
1204
1205 /*
1206 * Lock the control socket for the complete duration of the function
1207 * since from this point on we will use the socket.
1208 */
1209 if (stream->metadata_flag) {
1210 /* Metadata requires the control socket. */
1211 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
1212 netlen += sizeof(struct lttcomm_relayd_metadata_payload);
1213 }
1214
1215 ret = write_relayd_stream_header(stream, netlen, padding, relayd);
1216 if (ret >= 0) {
1217 /* Use the returned socket. */
1218 outfd = ret;
1219
1220 /* Write metadata stream id before payload */
1221 if (stream->metadata_flag) {
1222 ret = write_relayd_metadata_id(outfd, stream, relayd, padding);
1223 if (ret < 0) {
1224 written = ret;
1225 goto end;
1226 }
1227 }
1228 }
1229 /* Else, use the default set before which is the filesystem. */
1230 } else {
1231 /* No streaming, we have to set the len with the full padding */
1232 len += padding;
1233 }
1234
1235 while (len > 0) {
1236 do {
1237 ret = write(outfd, stream->mmap_base + mmap_offset, len);
1238 } while (ret < 0 && errno == EINTR);
1239 DBG("Consumer mmap write() ret %zd (len %lu)", ret, len);
1240 if (ret < 0) {
1241 PERROR("Error in file write");
1242 if (written == 0) {
1243 written = ret;
1244 }
1245 goto end;
1246 } else if (ret > len) {
1247 PERROR("Error in file write (ret %zd > len %lu)", ret, len);
1248 written += ret;
1249 goto end;
1250 } else {
1251 len -= ret;
1252 mmap_offset += ret;
1253 }
1254
1255 /* This call is useless on a socket so better save a syscall. */
1256 if (!relayd) {
1257 /* This won't block, but will start writeout asynchronously */
1258 lttng_sync_file_range(outfd, stream->out_fd_offset, ret,
1259 SYNC_FILE_RANGE_WRITE);
1260 stream->out_fd_offset += ret;
1261 }
1262 written += ret;
1263 }
1264 lttng_consumer_sync_trace_file(stream, orig_offset);
1265
1266 end:
1267 /* Unlock only if ctrl socket used */
1268 if (relayd && stream->metadata_flag) {
1269 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
1270 }
1271
1272 rcu_read_unlock();
1273 return written;
1274 }
1275
1276 /*
1277 * Splice the data from the ring buffer to the tracefile.
1278 *
1279 * Returns the number of bytes spliced.
1280 */
1281 ssize_t lttng_consumer_on_read_subbuffer_splice(
1282 struct lttng_consumer_local_data *ctx,
1283 struct lttng_consumer_stream *stream, unsigned long len,
1284 unsigned long padding)
1285 {
1286 ssize_t ret = 0, written = 0, ret_splice = 0;
1287 loff_t offset = 0;
1288 off_t orig_offset = stream->out_fd_offset;
1289 int fd = stream->wait_fd;
1290 /* Default is on the disk */
1291 int outfd = stream->out_fd;
1292 struct consumer_relayd_sock_pair *relayd = NULL;
1293 int *splice_pipe;
1294
1295 switch (consumer_data.type) {
1296 case LTTNG_CONSUMER_KERNEL:
1297 break;
1298 case LTTNG_CONSUMER32_UST:
1299 case LTTNG_CONSUMER64_UST:
1300 /* Not supported for user space tracing */
1301 return -ENOSYS;
1302 default:
1303 ERR("Unknown consumer_data type");
1304 assert(0);
1305 }
1306
1307 /* RCU lock for the relayd pointer */
1308 rcu_read_lock();
1309
1310 /* Flag that the current stream if set for network streaming. */
1311 if (stream->net_seq_idx != -1) {
1312 relayd = consumer_find_relayd(stream->net_seq_idx);
1313 if (relayd == NULL) {
1314 goto end;
1315 }
1316 }
1317
1318 /*
1319 * Choose right pipe for splice. Metadata and trace data are handled by
1320 * different threads hence the use of two pipes in order not to race or
1321 * corrupt the written data.
1322 */
1323 if (stream->metadata_flag) {
1324 splice_pipe = ctx->consumer_splice_metadata_pipe;
1325 } else {
1326 splice_pipe = ctx->consumer_thread_pipe;
1327 }
1328
1329 /* Write metadata stream id before payload */
1330 if (relayd) {
1331 int total_len = len;
1332
1333 if (stream->metadata_flag) {
1334 /*
1335 * Lock the control socket for the complete duration of the function
1336 * since from this point on we will use the socket.
1337 */
1338 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
1339
1340 ret = write_relayd_metadata_id(splice_pipe[1], stream, relayd,
1341 padding);
1342 if (ret < 0) {
1343 written = ret;
1344 goto end;
1345 }
1346
1347 total_len += sizeof(struct lttcomm_relayd_metadata_payload);
1348 }
1349
1350 ret = write_relayd_stream_header(stream, total_len, padding, relayd);
1351 if (ret >= 0) {
1352 /* Use the returned socket. */
1353 outfd = ret;
1354 } else {
1355 ERR("Remote relayd disconnected. Stopping");
1356 goto end;
1357 }
1358 } else {
1359 /* No streaming, we have to set the len with the full padding */
1360 len += padding;
1361 }
1362
1363 while (len > 0) {
1364 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1365 (unsigned long)offset, len, fd, splice_pipe[1]);
1366 ret_splice = splice(fd, &offset, splice_pipe[1], NULL, len,
1367 SPLICE_F_MOVE | SPLICE_F_MORE);
1368 DBG("splice chan to pipe, ret %zd", ret_splice);
1369 if (ret_splice < 0) {
1370 PERROR("Error in relay splice");
1371 if (written == 0) {
1372 written = ret_splice;
1373 }
1374 ret = errno;
1375 goto splice_error;
1376 }
1377
1378 /* Handle stream on the relayd if the output is on the network */
1379 if (relayd) {
1380 if (stream->metadata_flag) {
1381 size_t metadata_payload_size =
1382 sizeof(struct lttcomm_relayd_metadata_payload);
1383
1384 /* Update counter to fit the spliced data */
1385 ret_splice += metadata_payload_size;
1386 len += metadata_payload_size;
1387 /*
1388 * We do this so the return value can match the len passed as
1389 * argument to this function.
1390 */
1391 written -= metadata_payload_size;
1392 }
1393 }
1394
1395 /* Splice data out */
1396 ret_splice = splice(splice_pipe[0], NULL, outfd, NULL,
1397 ret_splice, SPLICE_F_MOVE | SPLICE_F_MORE);
1398 DBG("Consumer splice pipe to file, ret %zd", ret_splice);
1399 if (ret_splice < 0) {
1400 PERROR("Error in file splice");
1401 if (written == 0) {
1402 written = ret_splice;
1403 }
1404 ret = errno;
1405 goto splice_error;
1406 } else if (ret_splice > len) {
1407 errno = EINVAL;
1408 PERROR("Wrote more data than requested %zd (len: %lu)",
1409 ret_splice, len);
1410 written += ret_splice;
1411 ret = errno;
1412 goto splice_error;
1413 }
1414 len -= ret_splice;
1415
1416 /* This call is useless on a socket so better save a syscall. */
1417 if (!relayd) {
1418 /* This won't block, but will start writeout asynchronously */
1419 lttng_sync_file_range(outfd, stream->out_fd_offset, ret_splice,
1420 SYNC_FILE_RANGE_WRITE);
1421 stream->out_fd_offset += ret_splice;
1422 }
1423 written += ret_splice;
1424 }
1425 lttng_consumer_sync_trace_file(stream, orig_offset);
1426
1427 ret = ret_splice;
1428
1429 goto end;
1430
1431 splice_error:
1432 /* send the appropriate error description to sessiond */
1433 switch (ret) {
1434 case EBADF:
1435 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_EBADF);
1436 break;
1437 case EINVAL:
1438 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_EINVAL);
1439 break;
1440 case ENOMEM:
1441 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_ENOMEM);
1442 break;
1443 case ESPIPE:
1444 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_ESPIPE);
1445 break;
1446 }
1447
1448 end:
1449 if (relayd && stream->metadata_flag) {
1450 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
1451 }
1452
1453 rcu_read_unlock();
1454 return written;
1455 }
1456
1457 /*
1458 * Take a snapshot for a specific fd
1459 *
1460 * Returns 0 on success, < 0 on error
1461 */
1462 int lttng_consumer_take_snapshot(struct lttng_consumer_local_data *ctx,
1463 struct lttng_consumer_stream *stream)
1464 {
1465 switch (consumer_data.type) {
1466 case LTTNG_CONSUMER_KERNEL:
1467 return lttng_kconsumer_take_snapshot(ctx, stream);
1468 case LTTNG_CONSUMER32_UST:
1469 case LTTNG_CONSUMER64_UST:
1470 return lttng_ustconsumer_take_snapshot(ctx, stream);
1471 default:
1472 ERR("Unknown consumer_data type");
1473 assert(0);
1474 return -ENOSYS;
1475 }
1476
1477 }
1478
1479 /*
1480 * Get the produced position
1481 *
1482 * Returns 0 on success, < 0 on error
1483 */
1484 int lttng_consumer_get_produced_snapshot(
1485 struct lttng_consumer_local_data *ctx,
1486 struct lttng_consumer_stream *stream,
1487 unsigned long *pos)
1488 {
1489 switch (consumer_data.type) {
1490 case LTTNG_CONSUMER_KERNEL:
1491 return lttng_kconsumer_get_produced_snapshot(ctx, stream, pos);
1492 case LTTNG_CONSUMER32_UST:
1493 case LTTNG_CONSUMER64_UST:
1494 return lttng_ustconsumer_get_produced_snapshot(ctx, stream, pos);
1495 default:
1496 ERR("Unknown consumer_data type");
1497 assert(0);
1498 return -ENOSYS;
1499 }
1500 }
1501
1502 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data *ctx,
1503 int sock, struct pollfd *consumer_sockpoll)
1504 {
1505 switch (consumer_data.type) {
1506 case LTTNG_CONSUMER_KERNEL:
1507 return lttng_kconsumer_recv_cmd(ctx, sock, consumer_sockpoll);
1508 case LTTNG_CONSUMER32_UST:
1509 case LTTNG_CONSUMER64_UST:
1510 return lttng_ustconsumer_recv_cmd(ctx, sock, consumer_sockpoll);
1511 default:
1512 ERR("Unknown consumer_data type");
1513 assert(0);
1514 return -ENOSYS;
1515 }
1516 }
1517
1518 /*
1519 * Iterate over all streams of the hashtable and free them properly.
1520 *
1521 * WARNING: *MUST* be used with data stream only.
1522 */
1523 static void destroy_data_stream_ht(struct lttng_ht *ht)
1524 {
1525 int ret;
1526 struct lttng_ht_iter iter;
1527 struct lttng_consumer_stream *stream;
1528
1529 if (ht == NULL) {
1530 return;
1531 }
1532
1533 rcu_read_lock();
1534 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
1535 ret = lttng_ht_del(ht, &iter);
1536 assert(!ret);
1537
1538 call_rcu(&stream->node.head, consumer_free_stream);
1539 }
1540 rcu_read_unlock();
1541
1542 lttng_ht_destroy(ht);
1543 }
1544
1545 /*
1546 * Iterate over all streams of the hashtable and free them properly.
1547 *
1548 * XXX: Should not be only for metadata stream or else use an other name.
1549 */
1550 static void destroy_stream_ht(struct lttng_ht *ht)
1551 {
1552 int ret;
1553 struct lttng_ht_iter iter;
1554 struct lttng_consumer_stream *stream;
1555
1556 if (ht == NULL) {
1557 return;
1558 }
1559
1560 rcu_read_lock();
1561 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
1562 ret = lttng_ht_del(ht, &iter);
1563 assert(!ret);
1564
1565 call_rcu(&stream->node.head, consumer_free_stream);
1566 }
1567 rcu_read_unlock();
1568
1569 lttng_ht_destroy(ht);
1570 }
1571
1572 /*
1573 * Clean up a metadata stream and free its memory.
1574 */
1575 void consumer_del_metadata_stream(struct lttng_consumer_stream *stream,
1576 struct lttng_ht *ht)
1577 {
1578 int ret;
1579 struct lttng_ht_iter iter;
1580 struct lttng_consumer_channel *free_chan = NULL;
1581 struct consumer_relayd_sock_pair *relayd;
1582
1583 assert(stream);
1584 /*
1585 * This call should NEVER receive regular stream. It must always be
1586 * metadata stream and this is crucial for data structure synchronization.
1587 */
1588 assert(stream->metadata_flag);
1589
1590 DBG3("Consumer delete metadata stream %d", stream->wait_fd);
1591
1592 if (ht == NULL) {
1593 /* Means the stream was allocated but not successfully added */
1594 goto free_stream;
1595 }
1596
1597 pthread_mutex_lock(&consumer_data.lock);
1598 switch (consumer_data.type) {
1599 case LTTNG_CONSUMER_KERNEL:
1600 if (stream->mmap_base != NULL) {
1601 ret = munmap(stream->mmap_base, stream->mmap_len);
1602 if (ret != 0) {
1603 PERROR("munmap metadata stream");
1604 }
1605 }
1606 break;
1607 case LTTNG_CONSUMER32_UST:
1608 case LTTNG_CONSUMER64_UST:
1609 lttng_ustconsumer_del_stream(stream);
1610 break;
1611 default:
1612 ERR("Unknown consumer_data type");
1613 assert(0);
1614 goto end;
1615 }
1616
1617 rcu_read_lock();
1618 iter.iter.node = &stream->node.node;
1619 ret = lttng_ht_del(ht, &iter);
1620 assert(!ret);
1621
1622 /* Remove node session id from the consumer_data stream ht */
1623 iter.iter.node = &stream->node_session_id.node;
1624 ret = lttng_ht_del(consumer_data.stream_list_ht, &iter);
1625 assert(!ret);
1626 rcu_read_unlock();
1627
1628 if (stream->out_fd >= 0) {
1629 ret = close(stream->out_fd);
1630 if (ret) {
1631 PERROR("close");
1632 }
1633 }
1634
1635 if (stream->wait_fd >= 0 && !stream->wait_fd_is_copy) {
1636 ret = close(stream->wait_fd);
1637 if (ret) {
1638 PERROR("close");
1639 }
1640 }
1641
1642 if (stream->shm_fd >= 0 && stream->wait_fd != stream->shm_fd) {
1643 ret = close(stream->shm_fd);
1644 if (ret) {
1645 PERROR("close");
1646 }
1647 }
1648
1649 /* Check and cleanup relayd */
1650 rcu_read_lock();
1651 relayd = consumer_find_relayd(stream->net_seq_idx);
1652 if (relayd != NULL) {
1653 uatomic_dec(&relayd->refcount);
1654 assert(uatomic_read(&relayd->refcount) >= 0);
1655
1656 /* Closing streams requires to lock the control socket. */
1657 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
1658 ret = relayd_send_close_stream(&relayd->control_sock,
1659 stream->relayd_stream_id, stream->next_net_seq_num - 1);
1660 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
1661 if (ret < 0) {
1662 DBG("Unable to close stream on the relayd. Continuing");
1663 /*
1664 * Continue here. There is nothing we can do for the relayd.
1665 * Chances are that the relayd has closed the socket so we just
1666 * continue cleaning up.
1667 */
1668 }
1669
1670 /* Both conditions are met, we destroy the relayd. */
1671 if (uatomic_read(&relayd->refcount) == 0 &&
1672 uatomic_read(&relayd->destroy_flag)) {
1673 destroy_relayd(relayd);
1674 }
1675 }
1676 rcu_read_unlock();
1677
1678 /* Atomically decrement channel refcount since other threads can use it. */
1679 uatomic_dec(&stream->chan->refcount);
1680 if (!uatomic_read(&stream->chan->refcount)
1681 && !uatomic_read(&stream->chan->nb_init_streams)) {
1682 /* Go for channel deletion! */
1683 free_chan = stream->chan;
1684 }
1685
1686 end:
1687 pthread_mutex_unlock(&consumer_data.lock);
1688
1689 if (free_chan) {
1690 consumer_del_channel(free_chan);
1691 }
1692
1693 free_stream:
1694 call_rcu(&stream->node.head, consumer_free_stream);
1695 }
1696
1697 /*
1698 * Action done with the metadata stream when adding it to the consumer internal
1699 * data structures to handle it.
1700 */
1701 static int consumer_add_metadata_stream(struct lttng_consumer_stream *stream,
1702 struct lttng_ht *ht)
1703 {
1704 int ret = 0;
1705 struct consumer_relayd_sock_pair *relayd;
1706
1707 assert(stream);
1708 assert(ht);
1709
1710 DBG3("Adding metadata stream %d to hash table", stream->wait_fd);
1711
1712 pthread_mutex_lock(&consumer_data.lock);
1713
1714 /*
1715 * From here, refcounts are updated so be _careful_ when returning an error
1716 * after this point.
1717 */
1718
1719 rcu_read_lock();
1720 /* Find relayd and, if one is found, increment refcount. */
1721 relayd = consumer_find_relayd(stream->net_seq_idx);
1722 if (relayd != NULL) {
1723 uatomic_inc(&relayd->refcount);
1724 }
1725
1726 /* Update channel refcount once added without error(s). */
1727 uatomic_inc(&stream->chan->refcount);
1728
1729 /*
1730 * When nb_init_streams reaches 0, we don't need to trigger any action in
1731 * terms of destroying the associated channel, because the action that
1732 * causes the count to become 0 also causes a stream to be added. The
1733 * channel deletion will thus be triggered by the following removal of this
1734 * stream.
1735 */
1736 if (uatomic_read(&stream->chan->nb_init_streams) > 0) {
1737 uatomic_dec(&stream->chan->nb_init_streams);
1738 }
1739
1740 /* Steal stream identifier to avoid having streams with the same key */
1741 consumer_steal_stream_key(stream->key, ht);
1742
1743 lttng_ht_add_unique_ulong(ht, &stream->node);
1744
1745 /*
1746 * Add stream to the stream_list_ht of the consumer data. No need to steal
1747 * the key since the HT does not use it and we allow to add redundant keys
1748 * into this table.
1749 */
1750 lttng_ht_add_ulong(consumer_data.stream_list_ht, &stream->node_session_id);
1751
1752 rcu_read_unlock();
1753
1754 pthread_mutex_unlock(&consumer_data.lock);
1755 return ret;
1756 }
1757
1758 /*
1759 * Thread polls on metadata file descriptor and write them on disk or on the
1760 * network.
1761 */
1762 void *consumer_thread_metadata_poll(void *data)
1763 {
1764 int ret, i, pollfd;
1765 uint32_t revents, nb_fd;
1766 struct lttng_consumer_stream *stream = NULL;
1767 struct lttng_ht_iter iter;
1768 struct lttng_ht_node_ulong *node;
1769 struct lttng_poll_event events;
1770 struct lttng_consumer_local_data *ctx = data;
1771 ssize_t len;
1772
1773 rcu_register_thread();
1774
1775 DBG("Thread metadata poll started");
1776
1777 /* Size is set to 1 for the consumer_metadata pipe */
1778 ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC);
1779 if (ret < 0) {
1780 ERR("Poll set creation failed");
1781 goto end;
1782 }
1783
1784 ret = lttng_poll_add(&events, ctx->consumer_metadata_pipe[0], LPOLLIN);
1785 if (ret < 0) {
1786 goto end;
1787 }
1788
1789 /* Main loop */
1790 DBG("Metadata main loop started");
1791
1792 while (1) {
1793 lttng_poll_reset(&events);
1794
1795 nb_fd = LTTNG_POLL_GETNB(&events);
1796
1797 /* Only the metadata pipe is set */
1798 if (nb_fd == 0 && consumer_quit == 1) {
1799 goto end;
1800 }
1801
1802 restart:
1803 DBG("Metadata poll wait with %d fd(s)", nb_fd);
1804 ret = lttng_poll_wait(&events, -1);
1805 DBG("Metadata event catched in thread");
1806 if (ret < 0) {
1807 if (errno == EINTR) {
1808 ERR("Poll EINTR catched");
1809 goto restart;
1810 }
1811 goto error;
1812 }
1813
1814 /* From here, the event is a metadata wait fd */
1815 for (i = 0; i < nb_fd; i++) {
1816 revents = LTTNG_POLL_GETEV(&events, i);
1817 pollfd = LTTNG_POLL_GETFD(&events, i);
1818
1819 /* Just don't waste time if no returned events for the fd */
1820 if (!revents) {
1821 continue;
1822 }
1823
1824 if (pollfd == ctx->consumer_metadata_pipe[0]) {
1825 if (revents & (LPOLLERR | LPOLLHUP )) {
1826 DBG("Metadata thread pipe hung up");
1827 /*
1828 * Remove the pipe from the poll set and continue the loop
1829 * since their might be data to consume.
1830 */
1831 lttng_poll_del(&events, ctx->consumer_metadata_pipe[0]);
1832 close(ctx->consumer_metadata_pipe[0]);
1833 continue;
1834 } else if (revents & LPOLLIN) {
1835 do {
1836 /* Get the stream pointer received */
1837 ret = read(pollfd, &stream, sizeof(stream));
1838 } while (ret < 0 && errno == EINTR);
1839 if (ret < 0 ||
1840 ret < sizeof(struct lttng_consumer_stream *)) {
1841 PERROR("read metadata stream");
1842 /*
1843 * Let's continue here and hope we can still work
1844 * without stopping the consumer. XXX: Should we?
1845 */
1846 continue;
1847 }
1848
1849 DBG("Adding metadata stream %d to poll set",
1850 stream->wait_fd);
1851
1852 ret = consumer_add_metadata_stream(stream, metadata_ht);
1853 if (ret) {
1854 ERR("Unable to add metadata stream");
1855 /* Stream was not setup properly. Continuing. */
1856 consumer_del_metadata_stream(stream, NULL);
1857 continue;
1858 }
1859
1860 /* Add metadata stream to the global poll events list */
1861 lttng_poll_add(&events, stream->wait_fd,
1862 LPOLLIN | LPOLLPRI);
1863 }
1864
1865 /* Handle other stream */
1866 continue;
1867 }
1868
1869 rcu_read_lock();
1870 lttng_ht_lookup(metadata_ht, (void *)((unsigned long) pollfd),
1871 &iter);
1872 node = lttng_ht_iter_get_node_ulong(&iter);
1873 assert(node);
1874
1875 stream = caa_container_of(node, struct lttng_consumer_stream,
1876 node);
1877
1878 /* Check for error event */
1879 if (revents & (LPOLLERR | LPOLLHUP)) {
1880 DBG("Metadata fd %d is hup|err.", pollfd);
1881 if (!stream->hangup_flush_done
1882 && (consumer_data.type == LTTNG_CONSUMER32_UST
1883 || consumer_data.type == LTTNG_CONSUMER64_UST)) {
1884 DBG("Attempting to flush and consume the UST buffers");
1885 lttng_ustconsumer_on_stream_hangup(stream);
1886
1887 /* We just flushed the stream now read it. */
1888 do {
1889 len = ctx->on_buffer_ready(stream, ctx);
1890 /*
1891 * We don't check the return value here since if we get
1892 * a negative len, it means an error occured thus we
1893 * simply remove it from the poll set and free the
1894 * stream.
1895 */
1896 } while (len > 0);
1897 }
1898
1899 lttng_poll_del(&events, stream->wait_fd);
1900 /*
1901 * This call update the channel states, closes file descriptors
1902 * and securely free the stream.
1903 */
1904 consumer_del_metadata_stream(stream, metadata_ht);
1905 } else if (revents & (LPOLLIN | LPOLLPRI)) {
1906 /* Get the data out of the metadata file descriptor */
1907 DBG("Metadata available on fd %d", pollfd);
1908 assert(stream->wait_fd == pollfd);
1909
1910 len = ctx->on_buffer_ready(stream, ctx);
1911 /* It's ok to have an unavailable sub-buffer */
1912 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
1913 rcu_read_unlock();
1914 goto end;
1915 } else if (len > 0) {
1916 stream->data_read = 1;
1917 }
1918 }
1919
1920 /* Release RCU lock for the stream looked up */
1921 rcu_read_unlock();
1922 }
1923 }
1924
1925 error:
1926 end:
1927 DBG("Metadata poll thread exiting");
1928 lttng_poll_clean(&events);
1929
1930 if (metadata_ht) {
1931 destroy_stream_ht(metadata_ht);
1932 }
1933
1934 rcu_unregister_thread();
1935 return NULL;
1936 }
1937
1938 /*
1939 * This thread polls the fds in the set to consume the data and write
1940 * it to tracefile if necessary.
1941 */
1942 void *consumer_thread_data_poll(void *data)
1943 {
1944 int num_rdy, num_hup, high_prio, ret, i;
1945 struct pollfd *pollfd = NULL;
1946 /* local view of the streams */
1947 struct lttng_consumer_stream **local_stream = NULL, *new_stream = NULL;
1948 /* local view of consumer_data.fds_count */
1949 int nb_fd = 0;
1950 struct lttng_consumer_local_data *ctx = data;
1951 ssize_t len;
1952
1953 rcu_register_thread();
1954
1955 data_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG);
1956 if (data_ht == NULL) {
1957 goto end;
1958 }
1959
1960 local_stream = zmalloc(sizeof(struct lttng_consumer_stream));
1961
1962 while (1) {
1963 high_prio = 0;
1964 num_hup = 0;
1965
1966 /*
1967 * the fds set has been updated, we need to update our
1968 * local array as well
1969 */
1970 pthread_mutex_lock(&consumer_data.lock);
1971 if (consumer_data.need_update) {
1972 if (pollfd != NULL) {
1973 free(pollfd);
1974 pollfd = NULL;
1975 }
1976 if (local_stream != NULL) {
1977 free(local_stream);
1978 local_stream = NULL;
1979 }
1980
1981 /* allocate for all fds + 1 for the consumer_data_pipe */
1982 pollfd = zmalloc((consumer_data.stream_count + 1) * sizeof(struct pollfd));
1983 if (pollfd == NULL) {
1984 PERROR("pollfd malloc");
1985 pthread_mutex_unlock(&consumer_data.lock);
1986 goto end;
1987 }
1988
1989 /* allocate for all fds + 1 for the consumer_data_pipe */
1990 local_stream = zmalloc((consumer_data.stream_count + 1) *
1991 sizeof(struct lttng_consumer_stream));
1992 if (local_stream == NULL) {
1993 PERROR("local_stream malloc");
1994 pthread_mutex_unlock(&consumer_data.lock);
1995 goto end;
1996 }
1997 ret = consumer_update_poll_array(ctx, &pollfd, local_stream,
1998 data_ht);
1999 if (ret < 0) {
2000 ERR("Error in allocating pollfd or local_outfds");
2001 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
2002 pthread_mutex_unlock(&consumer_data.lock);
2003 goto end;
2004 }
2005 nb_fd = ret;
2006 consumer_data.need_update = 0;
2007 }
2008 pthread_mutex_unlock(&consumer_data.lock);
2009
2010 /* No FDs and consumer_quit, consumer_cleanup the thread */
2011 if (nb_fd == 0 && consumer_quit == 1) {
2012 goto end;
2013 }
2014 /* poll on the array of fds */
2015 restart:
2016 DBG("polling on %d fd", nb_fd + 1);
2017 num_rdy = poll(pollfd, nb_fd + 1, consumer_poll_timeout);
2018 DBG("poll num_rdy : %d", num_rdy);
2019 if (num_rdy == -1) {
2020 /*
2021 * Restart interrupted system call.
2022 */
2023 if (errno == EINTR) {
2024 goto restart;
2025 }
2026 PERROR("Poll error");
2027 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
2028 goto end;
2029 } else if (num_rdy == 0) {
2030 DBG("Polling thread timed out");
2031 goto end;
2032 }
2033
2034 /*
2035 * If the consumer_data_pipe triggered poll go directly to the
2036 * beginning of the loop to update the array. We want to prioritize
2037 * array update over low-priority reads.
2038 */
2039 if (pollfd[nb_fd].revents & (POLLIN | POLLPRI)) {
2040 size_t pipe_readlen;
2041
2042 DBG("consumer_data_pipe wake up");
2043 /* Consume 1 byte of pipe data */
2044 do {
2045 pipe_readlen = read(ctx->consumer_data_pipe[0], &new_stream,
2046 sizeof(new_stream));
2047 } while (pipe_readlen == -1 && errno == EINTR);
2048
2049 /*
2050 * If the stream is NULL, just ignore it. It's also possible that
2051 * the sessiond poll thread changed the consumer_quit state and is
2052 * waking us up to test it.
2053 */
2054 if (new_stream == NULL) {
2055 continue;
2056 }
2057
2058 ret = consumer_add_stream(new_stream, data_ht);
2059 if (ret) {
2060 ERR("Consumer add stream %d failed. Continuing",
2061 new_stream->key);
2062 /*
2063 * At this point, if the add_stream fails, it is not in the
2064 * hash table thus passing the NULL value here.
2065 */
2066 consumer_del_stream(new_stream, NULL);
2067 }
2068
2069 /* Continue to update the local streams and handle prio ones */
2070 continue;
2071 }
2072
2073 /* Take care of high priority channels first. */
2074 for (i = 0; i < nb_fd; i++) {
2075 if (pollfd[i].revents & POLLPRI) {
2076 DBG("Urgent read on fd %d", pollfd[i].fd);
2077 high_prio = 1;
2078 len = ctx->on_buffer_ready(local_stream[i], ctx);
2079 /* it's ok to have an unavailable sub-buffer */
2080 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2081 goto end;
2082 } else if (len > 0) {
2083 local_stream[i]->data_read = 1;
2084 }
2085 }
2086 }
2087
2088 /*
2089 * If we read high prio channel in this loop, try again
2090 * for more high prio data.
2091 */
2092 if (high_prio) {
2093 continue;
2094 }
2095
2096 /* Take care of low priority channels. */
2097 for (i = 0; i < nb_fd; i++) {
2098 if ((pollfd[i].revents & POLLIN) ||
2099 local_stream[i]->hangup_flush_done) {
2100 DBG("Normal read on fd %d", pollfd[i].fd);
2101 len = ctx->on_buffer_ready(local_stream[i], ctx);
2102 /* it's ok to have an unavailable sub-buffer */
2103 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2104 goto end;
2105 } else if (len > 0) {
2106 local_stream[i]->data_read = 1;
2107 }
2108 }
2109 }
2110
2111 /* Handle hangup and errors */
2112 for (i = 0; i < nb_fd; i++) {
2113 if (!local_stream[i]->hangup_flush_done
2114 && (pollfd[i].revents & (POLLHUP | POLLERR | POLLNVAL))
2115 && (consumer_data.type == LTTNG_CONSUMER32_UST
2116 || consumer_data.type == LTTNG_CONSUMER64_UST)) {
2117 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2118 pollfd[i].fd);
2119 lttng_ustconsumer_on_stream_hangup(local_stream[i]);
2120 /* Attempt read again, for the data we just flushed. */
2121 local_stream[i]->data_read = 1;
2122 }
2123 /*
2124 * If the poll flag is HUP/ERR/NVAL and we have
2125 * read no data in this pass, we can remove the
2126 * stream from its hash table.
2127 */
2128 if ((pollfd[i].revents & POLLHUP)) {
2129 DBG("Polling fd %d tells it has hung up.", pollfd[i].fd);
2130 if (!local_stream[i]->data_read) {
2131 consumer_del_stream(local_stream[i], data_ht);
2132 num_hup++;
2133 }
2134 } else if (pollfd[i].revents & POLLERR) {
2135 ERR("Error returned in polling fd %d.", pollfd[i].fd);
2136 if (!local_stream[i]->data_read) {
2137 consumer_del_stream(local_stream[i], data_ht);
2138 num_hup++;
2139 }
2140 } else if (pollfd[i].revents & POLLNVAL) {
2141 ERR("Polling fd %d tells fd is not open.", pollfd[i].fd);
2142 if (!local_stream[i]->data_read) {
2143 consumer_del_stream(local_stream[i], data_ht);
2144 num_hup++;
2145 }
2146 }
2147 local_stream[i]->data_read = 0;
2148 }
2149 }
2150 end:
2151 DBG("polling thread exiting");
2152 if (pollfd != NULL) {
2153 free(pollfd);
2154 pollfd = NULL;
2155 }
2156 if (local_stream != NULL) {
2157 free(local_stream);
2158 local_stream = NULL;
2159 }
2160
2161 /*
2162 * Close the write side of the pipe so epoll_wait() in
2163 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2164 * read side of the pipe. If we close them both, epoll_wait strangely does
2165 * not return and could create a endless wait period if the pipe is the
2166 * only tracked fd in the poll set. The thread will take care of closing
2167 * the read side.
2168 */
2169 close(ctx->consumer_metadata_pipe[1]);
2170
2171 if (data_ht) {
2172 destroy_data_stream_ht(data_ht);
2173 }
2174
2175 rcu_unregister_thread();
2176 return NULL;
2177 }
2178
2179 /*
2180 * This thread listens on the consumerd socket and receives the file
2181 * descriptors from the session daemon.
2182 */
2183 void *consumer_thread_sessiond_poll(void *data)
2184 {
2185 int sock, client_socket, ret;
2186 /*
2187 * structure to poll for incoming data on communication socket avoids
2188 * making blocking sockets.
2189 */
2190 struct pollfd consumer_sockpoll[2];
2191 struct lttng_consumer_local_data *ctx = data;
2192
2193 rcu_register_thread();
2194
2195 DBG("Creating command socket %s", ctx->consumer_command_sock_path);
2196 unlink(ctx->consumer_command_sock_path);
2197 client_socket = lttcomm_create_unix_sock(ctx->consumer_command_sock_path);
2198 if (client_socket < 0) {
2199 ERR("Cannot create command socket");
2200 goto end;
2201 }
2202
2203 ret = lttcomm_listen_unix_sock(client_socket);
2204 if (ret < 0) {
2205 goto end;
2206 }
2207
2208 DBG("Sending ready command to lttng-sessiond");
2209 ret = lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY);
2210 /* return < 0 on error, but == 0 is not fatal */
2211 if (ret < 0) {
2212 ERR("Error sending ready command to lttng-sessiond");
2213 goto end;
2214 }
2215
2216 ret = fcntl(client_socket, F_SETFL, O_NONBLOCK);
2217 if (ret < 0) {
2218 PERROR("fcntl O_NONBLOCK");
2219 goto end;
2220 }
2221
2222 /* prepare the FDs to poll : to client socket and the should_quit pipe */
2223 consumer_sockpoll[0].fd = ctx->consumer_should_quit[0];
2224 consumer_sockpoll[0].events = POLLIN | POLLPRI;
2225 consumer_sockpoll[1].fd = client_socket;
2226 consumer_sockpoll[1].events = POLLIN | POLLPRI;
2227
2228 if (lttng_consumer_poll_socket(consumer_sockpoll) < 0) {
2229 goto end;
2230 }
2231 DBG("Connection on client_socket");
2232
2233 /* Blocking call, waiting for transmission */
2234 sock = lttcomm_accept_unix_sock(client_socket);
2235 if (sock <= 0) {
2236 WARN("On accept");
2237 goto end;
2238 }
2239 ret = fcntl(sock, F_SETFL, O_NONBLOCK);
2240 if (ret < 0) {
2241 PERROR("fcntl O_NONBLOCK");
2242 goto end;
2243 }
2244
2245 /* update the polling structure to poll on the established socket */
2246 consumer_sockpoll[1].fd = sock;
2247 consumer_sockpoll[1].events = POLLIN | POLLPRI;
2248
2249 while (1) {
2250 if (lttng_consumer_poll_socket(consumer_sockpoll) < 0) {
2251 goto end;
2252 }
2253 DBG("Incoming command on sock");
2254 ret = lttng_consumer_recv_cmd(ctx, sock, consumer_sockpoll);
2255 if (ret == -ENOENT) {
2256 DBG("Received STOP command");
2257 goto end;
2258 }
2259 if (ret <= 0) {
2260 /*
2261 * This could simply be a session daemon quitting. Don't output
2262 * ERR() here.
2263 */
2264 DBG("Communication interrupted on command socket");
2265 goto end;
2266 }
2267 if (consumer_quit) {
2268 DBG("consumer_thread_receive_fds received quit from signal");
2269 goto end;
2270 }
2271 DBG("received fds on sock");
2272 }
2273 end:
2274 DBG("consumer_thread_receive_fds exiting");
2275
2276 /*
2277 * when all fds have hung up, the polling thread
2278 * can exit cleanly
2279 */
2280 consumer_quit = 1;
2281
2282 /*
2283 * 2s of grace period, if no polling events occur during
2284 * this period, the polling thread will exit even if there
2285 * are still open FDs (should not happen, but safety mechanism).
2286 */
2287 consumer_poll_timeout = LTTNG_CONSUMER_POLL_TIMEOUT;
2288
2289 /*
2290 * Notify the data poll thread to poll back again and test the
2291 * consumer_quit state to quit gracefully.
2292 */
2293 do {
2294 struct lttng_consumer_stream *null_stream = NULL;
2295
2296 ret = write(ctx->consumer_data_pipe[1], &null_stream,
2297 sizeof(null_stream));
2298 } while (ret < 0 && errno == EINTR);
2299
2300 rcu_unregister_thread();
2301 return NULL;
2302 }
2303
2304 ssize_t lttng_consumer_read_subbuffer(struct lttng_consumer_stream *stream,
2305 struct lttng_consumer_local_data *ctx)
2306 {
2307 switch (consumer_data.type) {
2308 case LTTNG_CONSUMER_KERNEL:
2309 return lttng_kconsumer_read_subbuffer(stream, ctx);
2310 case LTTNG_CONSUMER32_UST:
2311 case LTTNG_CONSUMER64_UST:
2312 return lttng_ustconsumer_read_subbuffer(stream, ctx);
2313 default:
2314 ERR("Unknown consumer_data type");
2315 assert(0);
2316 return -ENOSYS;
2317 }
2318 }
2319
2320 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream *stream)
2321 {
2322 switch (consumer_data.type) {
2323 case LTTNG_CONSUMER_KERNEL:
2324 return lttng_kconsumer_on_recv_stream(stream);
2325 case LTTNG_CONSUMER32_UST:
2326 case LTTNG_CONSUMER64_UST:
2327 return lttng_ustconsumer_on_recv_stream(stream);
2328 default:
2329 ERR("Unknown consumer_data type");
2330 assert(0);
2331 return -ENOSYS;
2332 }
2333 }
2334
2335 /*
2336 * Allocate and set consumer data hash tables.
2337 */
2338 void lttng_consumer_init(void)
2339 {
2340 consumer_data.channel_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG);
2341 consumer_data.relayd_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG);
2342 consumer_data.stream_list_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG);
2343
2344 metadata_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG);
2345 assert(metadata_ht);
2346 data_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG);
2347 assert(data_ht);
2348 }
2349
2350 /*
2351 * Process the ADD_RELAYD command receive by a consumer.
2352 *
2353 * This will create a relayd socket pair and add it to the relayd hash table.
2354 * The caller MUST acquire a RCU read side lock before calling it.
2355 */
2356 int consumer_add_relayd_socket(int net_seq_idx, int sock_type,
2357 struct lttng_consumer_local_data *ctx, int sock,
2358 struct pollfd *consumer_sockpoll, struct lttcomm_sock *relayd_sock)
2359 {
2360 int fd, ret = -1;
2361 struct consumer_relayd_sock_pair *relayd;
2362
2363 DBG("Consumer adding relayd socket (idx: %d)", net_seq_idx);
2364
2365 /* Get relayd reference if exists. */
2366 relayd = consumer_find_relayd(net_seq_idx);
2367 if (relayd == NULL) {
2368 /* Not found. Allocate one. */
2369 relayd = consumer_allocate_relayd_sock_pair(net_seq_idx);
2370 if (relayd == NULL) {
2371 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_OUTFD_ERROR);
2372 goto error;
2373 }
2374 }
2375
2376 /* Poll on consumer socket. */
2377 if (lttng_consumer_poll_socket(consumer_sockpoll) < 0) {
2378 ret = -EINTR;
2379 goto error;
2380 }
2381
2382 /* Get relayd socket from session daemon */
2383 ret = lttcomm_recv_fds_unix_sock(sock, &fd, 1);
2384 if (ret != sizeof(fd)) {
2385 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_ERROR_RECV_FD);
2386 ret = -1;
2387 goto error;
2388 }
2389
2390 /* Copy socket information and received FD */
2391 switch (sock_type) {
2392 case LTTNG_STREAM_CONTROL:
2393 /* Copy received lttcomm socket */
2394 lttcomm_copy_sock(&relayd->control_sock, relayd_sock);
2395 ret = lttcomm_create_sock(&relayd->control_sock);
2396 if (ret < 0) {
2397 goto error;
2398 }
2399
2400 /* Close the created socket fd which is useless */
2401 close(relayd->control_sock.fd);
2402
2403 /* Assign new file descriptor */
2404 relayd->control_sock.fd = fd;
2405 break;
2406 case LTTNG_STREAM_DATA:
2407 /* Copy received lttcomm socket */
2408 lttcomm_copy_sock(&relayd->data_sock, relayd_sock);
2409 ret = lttcomm_create_sock(&relayd->data_sock);
2410 if (ret < 0) {
2411 goto error;
2412 }
2413
2414 /* Close the created socket fd which is useless */
2415 close(relayd->data_sock.fd);
2416
2417 /* Assign new file descriptor */
2418 relayd->data_sock.fd = fd;
2419 break;
2420 default:
2421 ERR("Unknown relayd socket type (%d)", sock_type);
2422 goto error;
2423 }
2424
2425 DBG("Consumer %s socket created successfully with net idx %d (fd: %d)",
2426 sock_type == LTTNG_STREAM_CONTROL ? "control" : "data",
2427 relayd->net_seq_idx, fd);
2428
2429 /*
2430 * Add relayd socket pair to consumer data hashtable. If object already
2431 * exists or on error, the function gracefully returns.
2432 */
2433 add_relayd(relayd);
2434
2435 /* All good! */
2436 ret = 0;
2437
2438 error:
2439 return ret;
2440 }
2441
2442 /*
2443 * Check if for a given session id there is still data needed to be extract
2444 * from the buffers.
2445 *
2446 * Return 1 if data is in fact available to be read or else 0.
2447 */
2448 int consumer_data_available(uint64_t id)
2449 {
2450 int ret;
2451 struct lttng_ht_iter iter;
2452 struct lttng_ht *ht;
2453 struct lttng_consumer_stream *stream;
2454 int (*data_available)(struct lttng_consumer_stream *);
2455
2456 DBG("Consumer data available command on session id %" PRIu64, id);
2457
2458 pthread_mutex_lock(&consumer_data.lock);
2459
2460 switch (consumer_data.type) {
2461 case LTTNG_CONSUMER_KERNEL:
2462 data_available = lttng_kconsumer_data_available;
2463 break;
2464 case LTTNG_CONSUMER32_UST:
2465 case LTTNG_CONSUMER64_UST:
2466 data_available = lttng_ustconsumer_data_available;
2467 break;
2468 default:
2469 ERR("Unknown consumer data type");
2470 assert(0);
2471 }
2472
2473 /* Ease our life a bit */
2474 ht = consumer_data.stream_list_ht;
2475
2476 cds_lfht_for_each_entry_duplicate(ht->ht, (long unsigned int) ht->hash_fct,
2477 ht->match_fct, (void *)((unsigned long) id),
2478 &iter.iter, stream, node_session_id.node) {
2479 /* Check the stream for data. */
2480 ret = data_available(stream);
2481 if (ret == 0) {
2482 goto data_not_available;
2483 }
2484 }
2485
2486 /* TODO: Support to ask the relayd if the streams are remote */
2487
2488 /*
2489 * Finding _no_ node in the hash table means that the stream(s) have been
2490 * removed thus data is guaranteed to be available for analysis from the
2491 * trace files. This is *only* true for local consumer and not network
2492 * streaming.
2493 */
2494
2495 /* Data is available to be read by a viewer. */
2496 pthread_mutex_unlock(&consumer_data.lock);
2497 return 1;
2498
2499 data_not_available:
2500 /* Data is still being extracted from buffers. */
2501 pthread_mutex_unlock(&consumer_data.lock);
2502 return 0;
2503 }
This page took 0.135451 seconds and 3 git commands to generate.