Fix: update next_net_seq_num after sending header
[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 /*
51 * Flag to inform the polling thread to quit when all fd hung up. Updated by
52 * the consumer_thread_receive_fds when it notices that all fds has hung up.
53 * Also updated by the signal handler (consumer_should_exit()). Read by the
54 * polling threads.
55 */
56 volatile int consumer_quit;
57
58 /*
59 * Global hash table containing respectively metadata and data streams. The
60 * stream element in this ht should only be updated by the metadata poll thread
61 * for the metadata and the data poll thread for the data.
62 */
63 static struct lttng_ht *metadata_ht;
64 static struct lttng_ht *data_ht;
65
66 /*
67 * Notify a thread pipe to poll back again. This usually means that some global
68 * state has changed so we just send back the thread in a poll wait call.
69 */
70 static void notify_thread_pipe(int wpipe)
71 {
72 int ret;
73
74 do {
75 struct lttng_consumer_stream *null_stream = NULL;
76
77 ret = write(wpipe, &null_stream, sizeof(null_stream));
78 } while (ret < 0 && errno == EINTR);
79 }
80
81 /*
82 * Find a stream. The consumer_data.lock must be locked during this
83 * call.
84 */
85 static struct lttng_consumer_stream *consumer_find_stream(int key,
86 struct lttng_ht *ht)
87 {
88 struct lttng_ht_iter iter;
89 struct lttng_ht_node_ulong *node;
90 struct lttng_consumer_stream *stream = NULL;
91
92 assert(ht);
93
94 /* Negative keys are lookup failures */
95 if (key < 0) {
96 return NULL;
97 }
98
99 rcu_read_lock();
100
101 lttng_ht_lookup(ht, (void *)((unsigned long) key), &iter);
102 node = lttng_ht_iter_get_node_ulong(&iter);
103 if (node != NULL) {
104 stream = caa_container_of(node, struct lttng_consumer_stream, node);
105 }
106
107 rcu_read_unlock();
108
109 return stream;
110 }
111
112 void consumer_steal_stream_key(int key, struct lttng_ht *ht)
113 {
114 struct lttng_consumer_stream *stream;
115
116 rcu_read_lock();
117 stream = consumer_find_stream(key, ht);
118 if (stream) {
119 stream->key = -1;
120 /*
121 * We don't want the lookup to match, but we still need
122 * to iterate on this stream when iterating over the hash table. Just
123 * change the node key.
124 */
125 stream->node.key = -1;
126 }
127 rcu_read_unlock();
128 }
129
130 /*
131 * Return a channel object for the given key.
132 *
133 * RCU read side lock MUST be acquired before calling this function and
134 * protects the channel ptr.
135 */
136 static struct lttng_consumer_channel *consumer_find_channel(int key)
137 {
138 struct lttng_ht_iter iter;
139 struct lttng_ht_node_ulong *node;
140 struct lttng_consumer_channel *channel = NULL;
141
142 /* Negative keys are lookup failures */
143 if (key < 0) {
144 return NULL;
145 }
146
147 lttng_ht_lookup(consumer_data.channel_ht, (void *)((unsigned long) key),
148 &iter);
149 node = lttng_ht_iter_get_node_ulong(&iter);
150 if (node != NULL) {
151 channel = caa_container_of(node, struct lttng_consumer_channel, node);
152 }
153
154 return channel;
155 }
156
157 static void consumer_steal_channel_key(int key)
158 {
159 struct lttng_consumer_channel *channel;
160
161 rcu_read_lock();
162 channel = consumer_find_channel(key);
163 if (channel) {
164 channel->key = -1;
165 /*
166 * We don't want the lookup to match, but we still need
167 * to iterate on this channel when iterating over the hash table. Just
168 * change the node key.
169 */
170 channel->node.key = -1;
171 }
172 rcu_read_unlock();
173 }
174
175 static
176 void consumer_free_stream(struct rcu_head *head)
177 {
178 struct lttng_ht_node_ulong *node =
179 caa_container_of(head, struct lttng_ht_node_ulong, head);
180 struct lttng_consumer_stream *stream =
181 caa_container_of(node, struct lttng_consumer_stream, node);
182
183 free(stream);
184 }
185
186 /*
187 * RCU protected relayd socket pair free.
188 */
189 static void consumer_rcu_free_relayd(struct rcu_head *head)
190 {
191 struct lttng_ht_node_ulong *node =
192 caa_container_of(head, struct lttng_ht_node_ulong, head);
193 struct consumer_relayd_sock_pair *relayd =
194 caa_container_of(node, struct consumer_relayd_sock_pair, node);
195
196 /*
197 * Close all sockets. This is done in the call RCU since we don't want the
198 * socket fds to be reassigned thus potentially creating bad state of the
199 * relayd object.
200 *
201 * We do not have to lock the control socket mutex here since at this stage
202 * there is no one referencing to this relayd object.
203 */
204 (void) relayd_close(&relayd->control_sock);
205 (void) relayd_close(&relayd->data_sock);
206
207 free(relayd);
208 }
209
210 /*
211 * Destroy and free relayd socket pair object.
212 *
213 * This function MUST be called with the consumer_data lock acquired.
214 */
215 static void destroy_relayd(struct consumer_relayd_sock_pair *relayd)
216 {
217 int ret;
218 struct lttng_ht_iter iter;
219
220 if (relayd == NULL) {
221 return;
222 }
223
224 DBG("Consumer destroy and close relayd socket pair");
225
226 iter.iter.node = &relayd->node.node;
227 ret = lttng_ht_del(consumer_data.relayd_ht, &iter);
228 if (ret != 0) {
229 /* We assume the relayd is being or is destroyed */
230 return;
231 }
232
233 /* RCU free() call */
234 call_rcu(&relayd->node.head, consumer_rcu_free_relayd);
235 }
236
237 /*
238 * Iterate over the relayd hash table and destroy each element. Finally,
239 * destroy the whole hash table.
240 */
241 static void cleanup_relayd_ht(void)
242 {
243 struct lttng_ht_iter iter;
244 struct consumer_relayd_sock_pair *relayd;
245
246 rcu_read_lock();
247
248 cds_lfht_for_each_entry(consumer_data.relayd_ht->ht, &iter.iter, relayd,
249 node.node) {
250 destroy_relayd(relayd);
251 }
252
253 lttng_ht_destroy(consumer_data.relayd_ht);
254
255 rcu_read_unlock();
256 }
257
258 /*
259 * Update the end point status of all streams having the given network sequence
260 * index (relayd index).
261 *
262 * It's atomically set without having the stream mutex locked which is fine
263 * because we handle the write/read race with a pipe wakeup for each thread.
264 */
265 static void update_endpoint_status_by_netidx(int net_seq_idx,
266 enum consumer_endpoint_status status)
267 {
268 struct lttng_ht_iter iter;
269 struct lttng_consumer_stream *stream;
270
271 DBG("Consumer set delete flag on stream by idx %d", net_seq_idx);
272
273 rcu_read_lock();
274
275 /* Let's begin with metadata */
276 cds_lfht_for_each_entry(metadata_ht->ht, &iter.iter, stream, node.node) {
277 if (stream->net_seq_idx == net_seq_idx) {
278 uatomic_set(&stream->endpoint_status, status);
279 DBG("Delete flag set to metadata stream %d", stream->wait_fd);
280 }
281 }
282
283 /* Follow up by the data streams */
284 cds_lfht_for_each_entry(data_ht->ht, &iter.iter, stream, node.node) {
285 if (stream->net_seq_idx == net_seq_idx) {
286 uatomic_set(&stream->endpoint_status, status);
287 DBG("Delete flag set to data stream %d", stream->wait_fd);
288 }
289 }
290 rcu_read_unlock();
291 }
292
293 /*
294 * Cleanup a relayd object by flagging every associated streams for deletion,
295 * destroying the object meaning removing it from the relayd hash table,
296 * closing the sockets and freeing the memory in a RCU call.
297 *
298 * If a local data context is available, notify the threads that the streams'
299 * state have changed.
300 */
301 static void cleanup_relayd(struct consumer_relayd_sock_pair *relayd,
302 struct lttng_consumer_local_data *ctx)
303 {
304 int netidx;
305
306 assert(relayd);
307
308 DBG("Cleaning up relayd sockets");
309
310 /* Save the net sequence index before destroying the object */
311 netidx = relayd->net_seq_idx;
312
313 /*
314 * Delete the relayd from the relayd hash table, close the sockets and free
315 * the object in a RCU call.
316 */
317 destroy_relayd(relayd);
318
319 /* Set inactive endpoint to all streams */
320 update_endpoint_status_by_netidx(netidx, CONSUMER_ENDPOINT_INACTIVE);
321
322 /*
323 * With a local data context, notify the threads that the streams' state
324 * have changed. The write() action on the pipe acts as an "implicit"
325 * memory barrier ordering the updates of the end point status from the
326 * read of this status which happens AFTER receiving this notify.
327 */
328 if (ctx) {
329 notify_thread_pipe(ctx->consumer_data_pipe[1]);
330 notify_thread_pipe(ctx->consumer_metadata_pipe[1]);
331 }
332 }
333
334 /*
335 * Flag a relayd socket pair for destruction. Destroy it if the refcount
336 * reaches zero.
337 *
338 * RCU read side lock MUST be aquired before calling this function.
339 */
340 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair *relayd)
341 {
342 assert(relayd);
343
344 /* Set destroy flag for this object */
345 uatomic_set(&relayd->destroy_flag, 1);
346
347 /* Destroy the relayd if refcount is 0 */
348 if (uatomic_read(&relayd->refcount) == 0) {
349 destroy_relayd(relayd);
350 }
351 }
352
353 /*
354 * Remove a stream from the global list protected by a mutex. This
355 * function is also responsible for freeing its data structures.
356 */
357 void consumer_del_stream(struct lttng_consumer_stream *stream,
358 struct lttng_ht *ht)
359 {
360 int ret;
361 struct lttng_ht_iter iter;
362 struct lttng_consumer_channel *free_chan = NULL;
363 struct consumer_relayd_sock_pair *relayd;
364
365 assert(stream);
366
367 DBG("Consumer del stream %d", stream->wait_fd);
368
369 if (ht == NULL) {
370 /* Means the stream was allocated but not successfully added */
371 goto free_stream;
372 }
373
374 pthread_mutex_lock(&consumer_data.lock);
375 pthread_mutex_lock(&stream->lock);
376
377 switch (consumer_data.type) {
378 case LTTNG_CONSUMER_KERNEL:
379 if (stream->mmap_base != NULL) {
380 ret = munmap(stream->mmap_base, stream->mmap_len);
381 if (ret != 0) {
382 PERROR("munmap");
383 }
384 }
385 break;
386 case LTTNG_CONSUMER32_UST:
387 case LTTNG_CONSUMER64_UST:
388 lttng_ustconsumer_del_stream(stream);
389 break;
390 default:
391 ERR("Unknown consumer_data type");
392 assert(0);
393 goto end;
394 }
395
396 rcu_read_lock();
397 iter.iter.node = &stream->node.node;
398 ret = lttng_ht_del(ht, &iter);
399 assert(!ret);
400
401 /* Remove node session id from the consumer_data stream ht */
402 iter.iter.node = &stream->node_session_id.node;
403 ret = lttng_ht_del(consumer_data.stream_list_ht, &iter);
404 assert(!ret);
405 rcu_read_unlock();
406
407 assert(consumer_data.stream_count > 0);
408 consumer_data.stream_count--;
409
410 if (stream->out_fd >= 0) {
411 ret = close(stream->out_fd);
412 if (ret) {
413 PERROR("close");
414 }
415 }
416 if (stream->wait_fd >= 0 && !stream->wait_fd_is_copy) {
417 ret = close(stream->wait_fd);
418 if (ret) {
419 PERROR("close");
420 }
421 }
422 if (stream->shm_fd >= 0 && stream->wait_fd != stream->shm_fd) {
423 ret = close(stream->shm_fd);
424 if (ret) {
425 PERROR("close");
426 }
427 }
428
429 /* Check and cleanup relayd */
430 rcu_read_lock();
431 relayd = consumer_find_relayd(stream->net_seq_idx);
432 if (relayd != NULL) {
433 uatomic_dec(&relayd->refcount);
434 assert(uatomic_read(&relayd->refcount) >= 0);
435
436 /* Closing streams requires to lock the control socket. */
437 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
438 ret = relayd_send_close_stream(&relayd->control_sock,
439 stream->relayd_stream_id,
440 stream->next_net_seq_num - 1);
441 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
442 if (ret < 0) {
443 DBG("Unable to close stream on the relayd. Continuing");
444 /*
445 * Continue here. There is nothing we can do for the relayd.
446 * Chances are that the relayd has closed the socket so we just
447 * continue cleaning up.
448 */
449 }
450
451 /* Both conditions are met, we destroy the relayd. */
452 if (uatomic_read(&relayd->refcount) == 0 &&
453 uatomic_read(&relayd->destroy_flag)) {
454 destroy_relayd(relayd);
455 }
456 }
457 rcu_read_unlock();
458
459 uatomic_dec(&stream->chan->refcount);
460 if (!uatomic_read(&stream->chan->refcount)
461 && !uatomic_read(&stream->chan->nb_init_streams)) {
462 free_chan = stream->chan;
463 }
464
465 end:
466 consumer_data.need_update = 1;
467 pthread_mutex_unlock(&stream->lock);
468 pthread_mutex_unlock(&consumer_data.lock);
469
470 if (free_chan) {
471 consumer_del_channel(free_chan);
472 }
473
474 free_stream:
475 call_rcu(&stream->node.head, consumer_free_stream);
476 }
477
478 struct lttng_consumer_stream *consumer_allocate_stream(
479 int channel_key, int stream_key,
480 int shm_fd, int wait_fd,
481 enum lttng_consumer_stream_state state,
482 uint64_t mmap_len,
483 enum lttng_event_output output,
484 const char *path_name,
485 uid_t uid,
486 gid_t gid,
487 int net_index,
488 int metadata_flag,
489 uint64_t session_id,
490 int *alloc_ret)
491 {
492 struct lttng_consumer_stream *stream;
493
494 stream = zmalloc(sizeof(*stream));
495 if (stream == NULL) {
496 PERROR("malloc struct lttng_consumer_stream");
497 *alloc_ret = -ENOMEM;
498 goto end;
499 }
500
501 rcu_read_lock();
502
503 /*
504 * Get stream's channel reference. Needed when adding the stream to the
505 * global hash table.
506 */
507 stream->chan = consumer_find_channel(channel_key);
508 if (!stream->chan) {
509 *alloc_ret = -ENOENT;
510 ERR("Unable to find channel for stream %d", stream_key);
511 goto error;
512 }
513
514 stream->key = stream_key;
515 stream->shm_fd = shm_fd;
516 stream->wait_fd = wait_fd;
517 stream->out_fd = -1;
518 stream->out_fd_offset = 0;
519 stream->state = state;
520 stream->mmap_len = mmap_len;
521 stream->mmap_base = NULL;
522 stream->output = output;
523 stream->uid = uid;
524 stream->gid = gid;
525 stream->net_seq_idx = net_index;
526 stream->metadata_flag = metadata_flag;
527 stream->session_id = session_id;
528 strncpy(stream->path_name, path_name, sizeof(stream->path_name));
529 stream->path_name[sizeof(stream->path_name) - 1] = '\0';
530 pthread_mutex_init(&stream->lock, NULL);
531
532 /*
533 * Index differently the metadata node because the thread is using an
534 * internal hash table to match streams in the metadata_ht to the epoll set
535 * file descriptor.
536 */
537 if (metadata_flag) {
538 lttng_ht_node_init_ulong(&stream->node, stream->wait_fd);
539 } else {
540 lttng_ht_node_init_ulong(&stream->node, stream->key);
541 }
542
543 /* Init session id node with the stream session id */
544 lttng_ht_node_init_ulong(&stream->node_session_id, stream->session_id);
545
546 /*
547 * The cpu number is needed before using any ustctl_* actions. Ignored for
548 * the kernel so the value does not matter.
549 */
550 pthread_mutex_lock(&consumer_data.lock);
551 stream->cpu = stream->chan->cpucount++;
552 pthread_mutex_unlock(&consumer_data.lock);
553
554 DBG3("Allocated stream %s (key %d, shm_fd %d, wait_fd %d, mmap_len %llu,"
555 " out_fd %d, net_seq_idx %d, session_id %" PRIu64,
556 stream->path_name, stream->key, stream->shm_fd, stream->wait_fd,
557 (unsigned long long) stream->mmap_len, stream->out_fd,
558 stream->net_seq_idx, stream->session_id);
559
560 rcu_read_unlock();
561 return stream;
562
563 error:
564 rcu_read_unlock();
565 free(stream);
566 end:
567 return NULL;
568 }
569
570 /*
571 * Add a stream to the global list protected by a mutex.
572 */
573 static int consumer_add_stream(struct lttng_consumer_stream *stream,
574 struct lttng_ht *ht)
575 {
576 int ret = 0;
577 struct consumer_relayd_sock_pair *relayd;
578
579 assert(stream);
580 assert(ht);
581
582 DBG3("Adding consumer stream %d", stream->key);
583
584 pthread_mutex_lock(&consumer_data.lock);
585 pthread_mutex_lock(&stream->lock);
586 rcu_read_lock();
587
588 /* Steal stream identifier to avoid having streams with the same key */
589 consumer_steal_stream_key(stream->key, ht);
590
591 lttng_ht_add_unique_ulong(ht, &stream->node);
592
593 /*
594 * Add stream to the stream_list_ht of the consumer data. No need to steal
595 * the key since the HT does not use it and we allow to add redundant keys
596 * into this table.
597 */
598 lttng_ht_add_ulong(consumer_data.stream_list_ht, &stream->node_session_id);
599
600 /* Check and cleanup relayd */
601 relayd = consumer_find_relayd(stream->net_seq_idx);
602 if (relayd != NULL) {
603 uatomic_inc(&relayd->refcount);
604 }
605
606 /* Update channel refcount once added without error(s). */
607 uatomic_inc(&stream->chan->refcount);
608
609 /*
610 * When nb_init_streams reaches 0, we don't need to trigger any action in
611 * terms of destroying the associated channel, because the action that
612 * causes the count to become 0 also causes a stream to be added. The
613 * channel deletion will thus be triggered by the following removal of this
614 * stream.
615 */
616 if (uatomic_read(&stream->chan->nb_init_streams) > 0) {
617 uatomic_dec(&stream->chan->nb_init_streams);
618 }
619
620 /* Update consumer data once the node is inserted. */
621 consumer_data.stream_count++;
622 consumer_data.need_update = 1;
623
624 rcu_read_unlock();
625 pthread_mutex_unlock(&stream->lock);
626 pthread_mutex_unlock(&consumer_data.lock);
627
628 return ret;
629 }
630
631 /*
632 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
633 * be acquired before calling this.
634 */
635 static int add_relayd(struct consumer_relayd_sock_pair *relayd)
636 {
637 int ret = 0;
638 struct lttng_ht_node_ulong *node;
639 struct lttng_ht_iter iter;
640
641 if (relayd == NULL) {
642 ret = -1;
643 goto end;
644 }
645
646 lttng_ht_lookup(consumer_data.relayd_ht,
647 (void *)((unsigned long) relayd->net_seq_idx), &iter);
648 node = lttng_ht_iter_get_node_ulong(&iter);
649 if (node != NULL) {
650 /* Relayd already exist. Ignore the insertion */
651 goto end;
652 }
653 lttng_ht_add_unique_ulong(consumer_data.relayd_ht, &relayd->node);
654
655 end:
656 return ret;
657 }
658
659 /*
660 * Allocate and return a consumer relayd socket.
661 */
662 struct consumer_relayd_sock_pair *consumer_allocate_relayd_sock_pair(
663 int net_seq_idx)
664 {
665 struct consumer_relayd_sock_pair *obj = NULL;
666
667 /* Negative net sequence index is a failure */
668 if (net_seq_idx < 0) {
669 goto error;
670 }
671
672 obj = zmalloc(sizeof(struct consumer_relayd_sock_pair));
673 if (obj == NULL) {
674 PERROR("zmalloc relayd sock");
675 goto error;
676 }
677
678 obj->net_seq_idx = net_seq_idx;
679 obj->refcount = 0;
680 obj->destroy_flag = 0;
681 lttng_ht_node_init_ulong(&obj->node, obj->net_seq_idx);
682 pthread_mutex_init(&obj->ctrl_sock_mutex, NULL);
683
684 error:
685 return obj;
686 }
687
688 /*
689 * Find a relayd socket pair in the global consumer data.
690 *
691 * Return the object if found else NULL.
692 * RCU read-side lock must be held across this call and while using the
693 * returned object.
694 */
695 struct consumer_relayd_sock_pair *consumer_find_relayd(int key)
696 {
697 struct lttng_ht_iter iter;
698 struct lttng_ht_node_ulong *node;
699 struct consumer_relayd_sock_pair *relayd = NULL;
700
701 /* Negative keys are lookup failures */
702 if (key < 0) {
703 goto error;
704 }
705
706 lttng_ht_lookup(consumer_data.relayd_ht, (void *)((unsigned long) key),
707 &iter);
708 node = lttng_ht_iter_get_node_ulong(&iter);
709 if (node != NULL) {
710 relayd = caa_container_of(node, struct consumer_relayd_sock_pair, node);
711 }
712
713 error:
714 return relayd;
715 }
716
717 /*
718 * Handle stream for relayd transmission if the stream applies for network
719 * streaming where the net sequence index is set.
720 *
721 * Return destination file descriptor or negative value on error.
722 */
723 static int write_relayd_stream_header(struct lttng_consumer_stream *stream,
724 size_t data_size, unsigned long padding,
725 struct consumer_relayd_sock_pair *relayd)
726 {
727 int outfd = -1, ret;
728 struct lttcomm_relayd_data_hdr data_hdr;
729
730 /* Safety net */
731 assert(stream);
732 assert(relayd);
733
734 /* Reset data header */
735 memset(&data_hdr, 0, sizeof(data_hdr));
736
737 if (stream->metadata_flag) {
738 /* Caller MUST acquire the relayd control socket lock */
739 ret = relayd_send_metadata(&relayd->control_sock, data_size);
740 if (ret < 0) {
741 goto error;
742 }
743
744 /* Metadata are always sent on the control socket. */
745 outfd = relayd->control_sock.fd;
746 } else {
747 /* Set header with stream information */
748 data_hdr.stream_id = htobe64(stream->relayd_stream_id);
749 data_hdr.data_size = htobe32(data_size);
750 data_hdr.padding_size = htobe32(padding);
751 /*
752 * Note that net_seq_num below is assigned with the *current* value of
753 * next_net_seq_num and only after that the next_net_seq_num will be
754 * increment. This is why when issuing a command on the relayd using
755 * this next value, 1 should always be substracted in order to compare
756 * the last seen sequence number on the relayd side to the last sent.
757 */
758 data_hdr.net_seq_num = htobe64(stream->next_net_seq_num);
759 /* Other fields are zeroed previously */
760
761 ret = relayd_send_data_hdr(&relayd->data_sock, &data_hdr,
762 sizeof(data_hdr));
763 if (ret < 0) {
764 goto error;
765 }
766
767 ++stream->next_net_seq_num;
768
769 /* Set to go on data socket */
770 outfd = relayd->data_sock.fd;
771 }
772
773 error:
774 return outfd;
775 }
776
777 static
778 void consumer_free_channel(struct rcu_head *head)
779 {
780 struct lttng_ht_node_ulong *node =
781 caa_container_of(head, struct lttng_ht_node_ulong, head);
782 struct lttng_consumer_channel *channel =
783 caa_container_of(node, struct lttng_consumer_channel, node);
784
785 free(channel);
786 }
787
788 /*
789 * Remove a channel from the global list protected by a mutex. This
790 * function is also responsible for freeing its data structures.
791 */
792 void consumer_del_channel(struct lttng_consumer_channel *channel)
793 {
794 int ret;
795 struct lttng_ht_iter iter;
796
797 DBG("Consumer delete channel key %d", channel->key);
798
799 pthread_mutex_lock(&consumer_data.lock);
800
801 switch (consumer_data.type) {
802 case LTTNG_CONSUMER_KERNEL:
803 break;
804 case LTTNG_CONSUMER32_UST:
805 case LTTNG_CONSUMER64_UST:
806 lttng_ustconsumer_del_channel(channel);
807 break;
808 default:
809 ERR("Unknown consumer_data type");
810 assert(0);
811 goto end;
812 }
813
814 rcu_read_lock();
815 iter.iter.node = &channel->node.node;
816 ret = lttng_ht_del(consumer_data.channel_ht, &iter);
817 assert(!ret);
818 rcu_read_unlock();
819
820 if (channel->mmap_base != NULL) {
821 ret = munmap(channel->mmap_base, channel->mmap_len);
822 if (ret != 0) {
823 PERROR("munmap");
824 }
825 }
826 if (channel->wait_fd >= 0 && !channel->wait_fd_is_copy) {
827 ret = close(channel->wait_fd);
828 if (ret) {
829 PERROR("close");
830 }
831 }
832 if (channel->shm_fd >= 0 && channel->wait_fd != channel->shm_fd) {
833 ret = close(channel->shm_fd);
834 if (ret) {
835 PERROR("close");
836 }
837 }
838
839 call_rcu(&channel->node.head, consumer_free_channel);
840 end:
841 pthread_mutex_unlock(&consumer_data.lock);
842 }
843
844 struct lttng_consumer_channel *consumer_allocate_channel(
845 int channel_key,
846 int shm_fd, int wait_fd,
847 uint64_t mmap_len,
848 uint64_t max_sb_size,
849 unsigned int nb_init_streams)
850 {
851 struct lttng_consumer_channel *channel;
852 int ret;
853
854 channel = zmalloc(sizeof(*channel));
855 if (channel == NULL) {
856 PERROR("malloc struct lttng_consumer_channel");
857 goto end;
858 }
859 channel->key = channel_key;
860 channel->shm_fd = shm_fd;
861 channel->wait_fd = wait_fd;
862 channel->mmap_len = mmap_len;
863 channel->max_sb_size = max_sb_size;
864 channel->refcount = 0;
865 channel->nb_init_streams = nb_init_streams;
866 lttng_ht_node_init_ulong(&channel->node, channel->key);
867
868 switch (consumer_data.type) {
869 case LTTNG_CONSUMER_KERNEL:
870 channel->mmap_base = NULL;
871 channel->mmap_len = 0;
872 break;
873 case LTTNG_CONSUMER32_UST:
874 case LTTNG_CONSUMER64_UST:
875 ret = lttng_ustconsumer_allocate_channel(channel);
876 if (ret) {
877 free(channel);
878 return NULL;
879 }
880 break;
881 default:
882 ERR("Unknown consumer_data type");
883 assert(0);
884 goto end;
885 }
886 DBG("Allocated channel (key %d, shm_fd %d, wait_fd %d, mmap_len %llu, max_sb_size %llu)",
887 channel->key, channel->shm_fd, channel->wait_fd,
888 (unsigned long long) channel->mmap_len,
889 (unsigned long long) channel->max_sb_size);
890 end:
891 return channel;
892 }
893
894 /*
895 * Add a channel to the global list protected by a mutex.
896 */
897 int consumer_add_channel(struct lttng_consumer_channel *channel)
898 {
899 struct lttng_ht_node_ulong *node;
900 struct lttng_ht_iter iter;
901
902 pthread_mutex_lock(&consumer_data.lock);
903 /* Steal channel identifier, for UST */
904 consumer_steal_channel_key(channel->key);
905 rcu_read_lock();
906
907 lttng_ht_lookup(consumer_data.channel_ht,
908 (void *)((unsigned long) channel->key), &iter);
909 node = lttng_ht_iter_get_node_ulong(&iter);
910 if (node != NULL) {
911 /* Channel already exist. Ignore the insertion */
912 goto end;
913 }
914
915 lttng_ht_add_unique_ulong(consumer_data.channel_ht, &channel->node);
916
917 end:
918 rcu_read_unlock();
919 pthread_mutex_unlock(&consumer_data.lock);
920
921 return 0;
922 }
923
924 /*
925 * Allocate the pollfd structure and the local view of the out fds to avoid
926 * doing a lookup in the linked list and concurrency issues when writing is
927 * needed. Called with consumer_data.lock held.
928 *
929 * Returns the number of fds in the structures.
930 */
931 static int consumer_update_poll_array(
932 struct lttng_consumer_local_data *ctx, struct pollfd **pollfd,
933 struct lttng_consumer_stream **local_stream, struct lttng_ht *ht)
934 {
935 int i = 0;
936 struct lttng_ht_iter iter;
937 struct lttng_consumer_stream *stream;
938
939 DBG("Updating poll fd array");
940 rcu_read_lock();
941 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
942 /*
943 * Only active streams with an active end point can be added to the
944 * poll set and local stream storage of the thread.
945 *
946 * There is a potential race here for endpoint_status to be updated
947 * just after the check. However, this is OK since the stream(s) will
948 * be deleted once the thread is notified that the end point state has
949 * changed where this function will be called back again.
950 */
951 if (stream->state != LTTNG_CONSUMER_ACTIVE_STREAM ||
952 stream->endpoint_status == CONSUMER_ENDPOINT_INACTIVE) {
953 continue;
954 }
955 DBG("Active FD %d", stream->wait_fd);
956 (*pollfd)[i].fd = stream->wait_fd;
957 (*pollfd)[i].events = POLLIN | POLLPRI;
958 local_stream[i] = stream;
959 i++;
960 }
961 rcu_read_unlock();
962
963 /*
964 * Insert the consumer_data_pipe at the end of the array and don't
965 * increment i so nb_fd is the number of real FD.
966 */
967 (*pollfd)[i].fd = ctx->consumer_data_pipe[0];
968 (*pollfd)[i].events = POLLIN | POLLPRI;
969 return i;
970 }
971
972 /*
973 * Poll on the should_quit pipe and the command socket return -1 on error and
974 * should exit, 0 if data is available on the command socket
975 */
976 int lttng_consumer_poll_socket(struct pollfd *consumer_sockpoll)
977 {
978 int num_rdy;
979
980 restart:
981 num_rdy = poll(consumer_sockpoll, 2, -1);
982 if (num_rdy == -1) {
983 /*
984 * Restart interrupted system call.
985 */
986 if (errno == EINTR) {
987 goto restart;
988 }
989 PERROR("Poll error");
990 goto exit;
991 }
992 if (consumer_sockpoll[0].revents & (POLLIN | POLLPRI)) {
993 DBG("consumer_should_quit wake up");
994 goto exit;
995 }
996 return 0;
997
998 exit:
999 return -1;
1000 }
1001
1002 /*
1003 * Set the error socket.
1004 */
1005 void lttng_consumer_set_error_sock(
1006 struct lttng_consumer_local_data *ctx, int sock)
1007 {
1008 ctx->consumer_error_socket = sock;
1009 }
1010
1011 /*
1012 * Set the command socket path.
1013 */
1014 void lttng_consumer_set_command_sock_path(
1015 struct lttng_consumer_local_data *ctx, char *sock)
1016 {
1017 ctx->consumer_command_sock_path = sock;
1018 }
1019
1020 /*
1021 * Send return code to the session daemon.
1022 * If the socket is not defined, we return 0, it is not a fatal error
1023 */
1024 int lttng_consumer_send_error(
1025 struct lttng_consumer_local_data *ctx, int cmd)
1026 {
1027 if (ctx->consumer_error_socket > 0) {
1028 return lttcomm_send_unix_sock(ctx->consumer_error_socket, &cmd,
1029 sizeof(enum lttcomm_sessiond_command));
1030 }
1031
1032 return 0;
1033 }
1034
1035 /*
1036 * Close all the tracefiles and stream fds and MUST be called when all
1037 * instances are destroyed i.e. when all threads were joined and are ended.
1038 */
1039 void lttng_consumer_cleanup(void)
1040 {
1041 struct lttng_ht_iter iter;
1042 struct lttng_ht_node_ulong *node;
1043
1044 rcu_read_lock();
1045
1046 cds_lfht_for_each_entry(consumer_data.channel_ht->ht, &iter.iter, node,
1047 node) {
1048 struct lttng_consumer_channel *channel =
1049 caa_container_of(node, struct lttng_consumer_channel, node);
1050 consumer_del_channel(channel);
1051 }
1052
1053 rcu_read_unlock();
1054
1055 lttng_ht_destroy(consumer_data.channel_ht);
1056
1057 cleanup_relayd_ht();
1058
1059 /*
1060 * This HT contains streams that are freed by either the metadata thread or
1061 * the data thread so we do *nothing* on the hash table and simply destroy
1062 * it.
1063 */
1064 lttng_ht_destroy(consumer_data.stream_list_ht);
1065 }
1066
1067 /*
1068 * Called from signal handler.
1069 */
1070 void lttng_consumer_should_exit(struct lttng_consumer_local_data *ctx)
1071 {
1072 int ret;
1073 consumer_quit = 1;
1074 do {
1075 ret = write(ctx->consumer_should_quit[1], "4", 1);
1076 } while (ret < 0 && errno == EINTR);
1077 if (ret < 0 || ret != 1) {
1078 PERROR("write consumer quit");
1079 }
1080
1081 DBG("Consumer flag that it should quit");
1082 }
1083
1084 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream *stream,
1085 off_t orig_offset)
1086 {
1087 int outfd = stream->out_fd;
1088
1089 /*
1090 * This does a blocking write-and-wait on any page that belongs to the
1091 * subbuffer prior to the one we just wrote.
1092 * Don't care about error values, as these are just hints and ways to
1093 * limit the amount of page cache used.
1094 */
1095 if (orig_offset < stream->chan->max_sb_size) {
1096 return;
1097 }
1098 lttng_sync_file_range(outfd, orig_offset - stream->chan->max_sb_size,
1099 stream->chan->max_sb_size,
1100 SYNC_FILE_RANGE_WAIT_BEFORE
1101 | SYNC_FILE_RANGE_WRITE
1102 | SYNC_FILE_RANGE_WAIT_AFTER);
1103 /*
1104 * Give hints to the kernel about how we access the file:
1105 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1106 * we write it.
1107 *
1108 * We need to call fadvise again after the file grows because the
1109 * kernel does not seem to apply fadvise to non-existing parts of the
1110 * file.
1111 *
1112 * Call fadvise _after_ having waited for the page writeback to
1113 * complete because the dirty page writeback semantic is not well
1114 * defined. So it can be expected to lead to lower throughput in
1115 * streaming.
1116 */
1117 posix_fadvise(outfd, orig_offset - stream->chan->max_sb_size,
1118 stream->chan->max_sb_size, POSIX_FADV_DONTNEED);
1119 }
1120
1121 /*
1122 * Initialise the necessary environnement :
1123 * - create a new context
1124 * - create the poll_pipe
1125 * - create the should_quit pipe (for signal handler)
1126 * - create the thread pipe (for splice)
1127 *
1128 * Takes a function pointer as argument, this function is called when data is
1129 * available on a buffer. This function is responsible to do the
1130 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1131 * buffer configuration and then kernctl_put_next_subbuf at the end.
1132 *
1133 * Returns a pointer to the new context or NULL on error.
1134 */
1135 struct lttng_consumer_local_data *lttng_consumer_create(
1136 enum lttng_consumer_type type,
1137 ssize_t (*buffer_ready)(struct lttng_consumer_stream *stream,
1138 struct lttng_consumer_local_data *ctx),
1139 int (*recv_channel)(struct lttng_consumer_channel *channel),
1140 int (*recv_stream)(struct lttng_consumer_stream *stream),
1141 int (*update_stream)(int stream_key, uint32_t state))
1142 {
1143 int ret, i;
1144 struct lttng_consumer_local_data *ctx;
1145
1146 assert(consumer_data.type == LTTNG_CONSUMER_UNKNOWN ||
1147 consumer_data.type == type);
1148 consumer_data.type = type;
1149
1150 ctx = zmalloc(sizeof(struct lttng_consumer_local_data));
1151 if (ctx == NULL) {
1152 PERROR("allocating context");
1153 goto error;
1154 }
1155
1156 ctx->consumer_error_socket = -1;
1157 /* assign the callbacks */
1158 ctx->on_buffer_ready = buffer_ready;
1159 ctx->on_recv_channel = recv_channel;
1160 ctx->on_recv_stream = recv_stream;
1161 ctx->on_update_stream = update_stream;
1162
1163 ret = pipe(ctx->consumer_data_pipe);
1164 if (ret < 0) {
1165 PERROR("Error creating poll pipe");
1166 goto error_poll_pipe;
1167 }
1168
1169 /* set read end of the pipe to non-blocking */
1170 ret = fcntl(ctx->consumer_data_pipe[0], F_SETFL, O_NONBLOCK);
1171 if (ret < 0) {
1172 PERROR("fcntl O_NONBLOCK");
1173 goto error_poll_fcntl;
1174 }
1175
1176 /* set write end of the pipe to non-blocking */
1177 ret = fcntl(ctx->consumer_data_pipe[1], F_SETFL, O_NONBLOCK);
1178 if (ret < 0) {
1179 PERROR("fcntl O_NONBLOCK");
1180 goto error_poll_fcntl;
1181 }
1182
1183 ret = pipe(ctx->consumer_should_quit);
1184 if (ret < 0) {
1185 PERROR("Error creating recv pipe");
1186 goto error_quit_pipe;
1187 }
1188
1189 ret = pipe(ctx->consumer_thread_pipe);
1190 if (ret < 0) {
1191 PERROR("Error creating thread pipe");
1192 goto error_thread_pipe;
1193 }
1194
1195 ret = utils_create_pipe(ctx->consumer_metadata_pipe);
1196 if (ret < 0) {
1197 goto error_metadata_pipe;
1198 }
1199
1200 ret = utils_create_pipe(ctx->consumer_splice_metadata_pipe);
1201 if (ret < 0) {
1202 goto error_splice_pipe;
1203 }
1204
1205 return ctx;
1206
1207 error_splice_pipe:
1208 utils_close_pipe(ctx->consumer_metadata_pipe);
1209 error_metadata_pipe:
1210 utils_close_pipe(ctx->consumer_thread_pipe);
1211 error_thread_pipe:
1212 for (i = 0; i < 2; i++) {
1213 int err;
1214
1215 err = close(ctx->consumer_should_quit[i]);
1216 if (err) {
1217 PERROR("close");
1218 }
1219 }
1220 error_poll_fcntl:
1221 error_quit_pipe:
1222 for (i = 0; i < 2; i++) {
1223 int err;
1224
1225 err = close(ctx->consumer_data_pipe[i]);
1226 if (err) {
1227 PERROR("close");
1228 }
1229 }
1230 error_poll_pipe:
1231 free(ctx);
1232 error:
1233 return NULL;
1234 }
1235
1236 /*
1237 * Close all fds associated with the instance and free the context.
1238 */
1239 void lttng_consumer_destroy(struct lttng_consumer_local_data *ctx)
1240 {
1241 int ret;
1242
1243 DBG("Consumer destroying it. Closing everything.");
1244
1245 ret = close(ctx->consumer_error_socket);
1246 if (ret) {
1247 PERROR("close");
1248 }
1249 ret = close(ctx->consumer_thread_pipe[0]);
1250 if (ret) {
1251 PERROR("close");
1252 }
1253 ret = close(ctx->consumer_thread_pipe[1]);
1254 if (ret) {
1255 PERROR("close");
1256 }
1257 ret = close(ctx->consumer_data_pipe[0]);
1258 if (ret) {
1259 PERROR("close");
1260 }
1261 ret = close(ctx->consumer_data_pipe[1]);
1262 if (ret) {
1263 PERROR("close");
1264 }
1265 ret = close(ctx->consumer_should_quit[0]);
1266 if (ret) {
1267 PERROR("close");
1268 }
1269 ret = close(ctx->consumer_should_quit[1]);
1270 if (ret) {
1271 PERROR("close");
1272 }
1273 utils_close_pipe(ctx->consumer_splice_metadata_pipe);
1274
1275 unlink(ctx->consumer_command_sock_path);
1276 free(ctx);
1277 }
1278
1279 /*
1280 * Write the metadata stream id on the specified file descriptor.
1281 */
1282 static int write_relayd_metadata_id(int fd,
1283 struct lttng_consumer_stream *stream,
1284 struct consumer_relayd_sock_pair *relayd,
1285 unsigned long padding)
1286 {
1287 int ret;
1288 struct lttcomm_relayd_metadata_payload hdr;
1289
1290 hdr.stream_id = htobe64(stream->relayd_stream_id);
1291 hdr.padding_size = htobe32(padding);
1292 do {
1293 ret = write(fd, (void *) &hdr, sizeof(hdr));
1294 } while (ret < 0 && errno == EINTR);
1295 if (ret < 0 || ret != sizeof(hdr)) {
1296 /*
1297 * This error means that the fd's end is closed so ignore the perror
1298 * not to clubber the error output since this can happen in a normal
1299 * code path.
1300 */
1301 if (errno != EPIPE) {
1302 PERROR("write metadata stream id");
1303 }
1304 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno);
1305 /*
1306 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1307 * handle writting the missing part so report that as an error and
1308 * don't lie to the caller.
1309 */
1310 ret = -1;
1311 goto end;
1312 }
1313 DBG("Metadata stream id %" PRIu64 " with padding %lu written before data",
1314 stream->relayd_stream_id, padding);
1315
1316 end:
1317 return ret;
1318 }
1319
1320 /*
1321 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1322 * core function for writing trace buffers to either the local filesystem or
1323 * the network.
1324 *
1325 * It must be called with the stream lock held.
1326 *
1327 * Careful review MUST be put if any changes occur!
1328 *
1329 * Returns the number of bytes written
1330 */
1331 ssize_t lttng_consumer_on_read_subbuffer_mmap(
1332 struct lttng_consumer_local_data *ctx,
1333 struct lttng_consumer_stream *stream, unsigned long len,
1334 unsigned long padding)
1335 {
1336 unsigned long mmap_offset;
1337 ssize_t ret = 0, written = 0;
1338 off_t orig_offset = stream->out_fd_offset;
1339 /* Default is on the disk */
1340 int outfd = stream->out_fd;
1341 struct consumer_relayd_sock_pair *relayd = NULL;
1342 unsigned int relayd_hang_up = 0;
1343
1344 /* RCU lock for the relayd pointer */
1345 rcu_read_lock();
1346
1347 /* Flag that the current stream if set for network streaming. */
1348 if (stream->net_seq_idx != -1) {
1349 relayd = consumer_find_relayd(stream->net_seq_idx);
1350 if (relayd == NULL) {
1351 goto end;
1352 }
1353 }
1354
1355 /* get the offset inside the fd to mmap */
1356 switch (consumer_data.type) {
1357 case LTTNG_CONSUMER_KERNEL:
1358 ret = kernctl_get_mmap_read_offset(stream->wait_fd, &mmap_offset);
1359 break;
1360 case LTTNG_CONSUMER32_UST:
1361 case LTTNG_CONSUMER64_UST:
1362 ret = lttng_ustctl_get_mmap_read_offset(stream->chan->handle,
1363 stream->buf, &mmap_offset);
1364 break;
1365 default:
1366 ERR("Unknown consumer_data type");
1367 assert(0);
1368 }
1369 if (ret != 0) {
1370 errno = -ret;
1371 PERROR("tracer ctl get_mmap_read_offset");
1372 written = ret;
1373 goto end;
1374 }
1375
1376 /* Handle stream on the relayd if the output is on the network */
1377 if (relayd) {
1378 unsigned long netlen = len;
1379
1380 /*
1381 * Lock the control socket for the complete duration of the function
1382 * since from this point on we will use the socket.
1383 */
1384 if (stream->metadata_flag) {
1385 /* Metadata requires the control socket. */
1386 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
1387 netlen += sizeof(struct lttcomm_relayd_metadata_payload);
1388 }
1389
1390 ret = write_relayd_stream_header(stream, netlen, padding, relayd);
1391 if (ret >= 0) {
1392 /* Use the returned socket. */
1393 outfd = ret;
1394
1395 /* Write metadata stream id before payload */
1396 if (stream->metadata_flag) {
1397 ret = write_relayd_metadata_id(outfd, stream, relayd, padding);
1398 if (ret < 0) {
1399 written = ret;
1400 /* Socket operation failed. We consider the relayd dead */
1401 if (ret == -EPIPE || ret == -EINVAL) {
1402 relayd_hang_up = 1;
1403 goto write_error;
1404 }
1405 goto end;
1406 }
1407 }
1408 } else {
1409 /* Socket operation failed. We consider the relayd dead */
1410 if (ret == -EPIPE || ret == -EINVAL) {
1411 relayd_hang_up = 1;
1412 goto write_error;
1413 }
1414 /* Else, use the default set before which is the filesystem. */
1415 }
1416 } else {
1417 /* No streaming, we have to set the len with the full padding */
1418 len += padding;
1419 }
1420
1421 while (len > 0) {
1422 do {
1423 ret = write(outfd, stream->mmap_base + mmap_offset, len);
1424 } while (ret < 0 && errno == EINTR);
1425 DBG("Consumer mmap write() ret %zd (len %lu)", ret, len);
1426 if (ret < 0) {
1427 /*
1428 * This is possible if the fd is closed on the other side (outfd)
1429 * or any write problem. It can be verbose a bit for a normal
1430 * execution if for instance the relayd is stopped abruptly. This
1431 * can happen so set this to a DBG statement.
1432 */
1433 DBG("Error in file write mmap");
1434 if (written == 0) {
1435 written = ret;
1436 }
1437 /* Socket operation failed. We consider the relayd dead */
1438 if (errno == EPIPE || errno == EINVAL) {
1439 relayd_hang_up = 1;
1440 goto write_error;
1441 }
1442 goto end;
1443 } else if (ret > len) {
1444 PERROR("Error in file write (ret %zd > len %lu)", ret, len);
1445 written += ret;
1446 goto end;
1447 } else {
1448 len -= ret;
1449 mmap_offset += ret;
1450 }
1451
1452 /* This call is useless on a socket so better save a syscall. */
1453 if (!relayd) {
1454 /* This won't block, but will start writeout asynchronously */
1455 lttng_sync_file_range(outfd, stream->out_fd_offset, ret,
1456 SYNC_FILE_RANGE_WRITE);
1457 stream->out_fd_offset += ret;
1458 }
1459 written += ret;
1460 }
1461 lttng_consumer_sync_trace_file(stream, orig_offset);
1462
1463 write_error:
1464 /*
1465 * This is a special case that the relayd has closed its socket. Let's
1466 * cleanup the relayd object and all associated streams.
1467 */
1468 if (relayd && relayd_hang_up) {
1469 cleanup_relayd(relayd, ctx);
1470 }
1471
1472 end:
1473 /* Unlock only if ctrl socket used */
1474 if (relayd && stream->metadata_flag) {
1475 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
1476 }
1477
1478 rcu_read_unlock();
1479 return written;
1480 }
1481
1482 /*
1483 * Splice the data from the ring buffer to the tracefile.
1484 *
1485 * It must be called with the stream lock held.
1486 *
1487 * Returns the number of bytes spliced.
1488 */
1489 ssize_t lttng_consumer_on_read_subbuffer_splice(
1490 struct lttng_consumer_local_data *ctx,
1491 struct lttng_consumer_stream *stream, unsigned long len,
1492 unsigned long padding)
1493 {
1494 ssize_t ret = 0, written = 0, ret_splice = 0;
1495 loff_t offset = 0;
1496 off_t orig_offset = stream->out_fd_offset;
1497 int fd = stream->wait_fd;
1498 /* Default is on the disk */
1499 int outfd = stream->out_fd;
1500 struct consumer_relayd_sock_pair *relayd = NULL;
1501 int *splice_pipe;
1502 unsigned int relayd_hang_up = 0;
1503
1504 switch (consumer_data.type) {
1505 case LTTNG_CONSUMER_KERNEL:
1506 break;
1507 case LTTNG_CONSUMER32_UST:
1508 case LTTNG_CONSUMER64_UST:
1509 /* Not supported for user space tracing */
1510 return -ENOSYS;
1511 default:
1512 ERR("Unknown consumer_data type");
1513 assert(0);
1514 }
1515
1516 /* RCU lock for the relayd pointer */
1517 rcu_read_lock();
1518
1519 /* Flag that the current stream if set for network streaming. */
1520 if (stream->net_seq_idx != -1) {
1521 relayd = consumer_find_relayd(stream->net_seq_idx);
1522 if (relayd == NULL) {
1523 goto end;
1524 }
1525 }
1526
1527 /*
1528 * Choose right pipe for splice. Metadata and trace data are handled by
1529 * different threads hence the use of two pipes in order not to race or
1530 * corrupt the written data.
1531 */
1532 if (stream->metadata_flag) {
1533 splice_pipe = ctx->consumer_splice_metadata_pipe;
1534 } else {
1535 splice_pipe = ctx->consumer_thread_pipe;
1536 }
1537
1538 /* Write metadata stream id before payload */
1539 if (relayd) {
1540 int total_len = len;
1541
1542 if (stream->metadata_flag) {
1543 /*
1544 * Lock the control socket for the complete duration of the function
1545 * since from this point on we will use the socket.
1546 */
1547 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
1548
1549 ret = write_relayd_metadata_id(splice_pipe[1], stream, relayd,
1550 padding);
1551 if (ret < 0) {
1552 written = ret;
1553 /* Socket operation failed. We consider the relayd dead */
1554 if (ret == -EBADF) {
1555 WARN("Remote relayd disconnected. Stopping");
1556 relayd_hang_up = 1;
1557 goto write_error;
1558 }
1559 goto end;
1560 }
1561
1562 total_len += sizeof(struct lttcomm_relayd_metadata_payload);
1563 }
1564
1565 ret = write_relayd_stream_header(stream, total_len, padding, relayd);
1566 if (ret >= 0) {
1567 /* Use the returned socket. */
1568 outfd = ret;
1569 } else {
1570 /* Socket operation failed. We consider the relayd dead */
1571 if (ret == -EBADF) {
1572 WARN("Remote relayd disconnected. Stopping");
1573 relayd_hang_up = 1;
1574 goto write_error;
1575 }
1576 goto end;
1577 }
1578 } else {
1579 /* No streaming, we have to set the len with the full padding */
1580 len += padding;
1581 }
1582
1583 while (len > 0) {
1584 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1585 (unsigned long)offset, len, fd, splice_pipe[1]);
1586 ret_splice = splice(fd, &offset, splice_pipe[1], NULL, len,
1587 SPLICE_F_MOVE | SPLICE_F_MORE);
1588 DBG("splice chan to pipe, ret %zd", ret_splice);
1589 if (ret_splice < 0) {
1590 PERROR("Error in relay splice");
1591 if (written == 0) {
1592 written = ret_splice;
1593 }
1594 ret = errno;
1595 goto splice_error;
1596 }
1597
1598 /* Handle stream on the relayd if the output is on the network */
1599 if (relayd) {
1600 if (stream->metadata_flag) {
1601 size_t metadata_payload_size =
1602 sizeof(struct lttcomm_relayd_metadata_payload);
1603
1604 /* Update counter to fit the spliced data */
1605 ret_splice += metadata_payload_size;
1606 len += metadata_payload_size;
1607 /*
1608 * We do this so the return value can match the len passed as
1609 * argument to this function.
1610 */
1611 written -= metadata_payload_size;
1612 }
1613 }
1614
1615 /* Splice data out */
1616 ret_splice = splice(splice_pipe[0], NULL, outfd, NULL,
1617 ret_splice, SPLICE_F_MOVE | SPLICE_F_MORE);
1618 DBG("Consumer splice pipe to file, ret %zd", ret_splice);
1619 if (ret_splice < 0) {
1620 PERROR("Error in file splice");
1621 if (written == 0) {
1622 written = ret_splice;
1623 }
1624 /* Socket operation failed. We consider the relayd dead */
1625 if (errno == EBADF || errno == EPIPE) {
1626 WARN("Remote relayd disconnected. Stopping");
1627 relayd_hang_up = 1;
1628 goto write_error;
1629 }
1630 ret = errno;
1631 goto splice_error;
1632 } else if (ret_splice > len) {
1633 errno = EINVAL;
1634 PERROR("Wrote more data than requested %zd (len: %lu)",
1635 ret_splice, len);
1636 written += ret_splice;
1637 ret = errno;
1638 goto splice_error;
1639 }
1640 len -= ret_splice;
1641
1642 /* This call is useless on a socket so better save a syscall. */
1643 if (!relayd) {
1644 /* This won't block, but will start writeout asynchronously */
1645 lttng_sync_file_range(outfd, stream->out_fd_offset, ret_splice,
1646 SYNC_FILE_RANGE_WRITE);
1647 stream->out_fd_offset += ret_splice;
1648 }
1649 written += ret_splice;
1650 }
1651 lttng_consumer_sync_trace_file(stream, orig_offset);
1652
1653 ret = ret_splice;
1654
1655 goto end;
1656
1657 write_error:
1658 /*
1659 * This is a special case that the relayd has closed its socket. Let's
1660 * cleanup the relayd object and all associated streams.
1661 */
1662 if (relayd && relayd_hang_up) {
1663 cleanup_relayd(relayd, ctx);
1664 /* Skip splice error so the consumer does not fail */
1665 goto end;
1666 }
1667
1668 splice_error:
1669 /* send the appropriate error description to sessiond */
1670 switch (ret) {
1671 case EINVAL:
1672 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_EINVAL);
1673 break;
1674 case ENOMEM:
1675 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_ENOMEM);
1676 break;
1677 case ESPIPE:
1678 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_ESPIPE);
1679 break;
1680 }
1681
1682 end:
1683 if (relayd && stream->metadata_flag) {
1684 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
1685 }
1686
1687 rcu_read_unlock();
1688 return written;
1689 }
1690
1691 /*
1692 * Take a snapshot for a specific fd
1693 *
1694 * Returns 0 on success, < 0 on error
1695 */
1696 int lttng_consumer_take_snapshot(struct lttng_consumer_local_data *ctx,
1697 struct lttng_consumer_stream *stream)
1698 {
1699 switch (consumer_data.type) {
1700 case LTTNG_CONSUMER_KERNEL:
1701 return lttng_kconsumer_take_snapshot(ctx, stream);
1702 case LTTNG_CONSUMER32_UST:
1703 case LTTNG_CONSUMER64_UST:
1704 return lttng_ustconsumer_take_snapshot(ctx, stream);
1705 default:
1706 ERR("Unknown consumer_data type");
1707 assert(0);
1708 return -ENOSYS;
1709 }
1710
1711 }
1712
1713 /*
1714 * Get the produced position
1715 *
1716 * Returns 0 on success, < 0 on error
1717 */
1718 int lttng_consumer_get_produced_snapshot(
1719 struct lttng_consumer_local_data *ctx,
1720 struct lttng_consumer_stream *stream,
1721 unsigned long *pos)
1722 {
1723 switch (consumer_data.type) {
1724 case LTTNG_CONSUMER_KERNEL:
1725 return lttng_kconsumer_get_produced_snapshot(ctx, stream, pos);
1726 case LTTNG_CONSUMER32_UST:
1727 case LTTNG_CONSUMER64_UST:
1728 return lttng_ustconsumer_get_produced_snapshot(ctx, stream, pos);
1729 default:
1730 ERR("Unknown consumer_data type");
1731 assert(0);
1732 return -ENOSYS;
1733 }
1734 }
1735
1736 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data *ctx,
1737 int sock, struct pollfd *consumer_sockpoll)
1738 {
1739 switch (consumer_data.type) {
1740 case LTTNG_CONSUMER_KERNEL:
1741 return lttng_kconsumer_recv_cmd(ctx, sock, consumer_sockpoll);
1742 case LTTNG_CONSUMER32_UST:
1743 case LTTNG_CONSUMER64_UST:
1744 return lttng_ustconsumer_recv_cmd(ctx, sock, consumer_sockpoll);
1745 default:
1746 ERR("Unknown consumer_data type");
1747 assert(0);
1748 return -ENOSYS;
1749 }
1750 }
1751
1752 /*
1753 * Iterate over all streams of the hashtable and free them properly.
1754 *
1755 * WARNING: *MUST* be used with data stream only.
1756 */
1757 static void destroy_data_stream_ht(struct lttng_ht *ht)
1758 {
1759 struct lttng_ht_iter iter;
1760 struct lttng_consumer_stream *stream;
1761
1762 if (ht == NULL) {
1763 return;
1764 }
1765
1766 rcu_read_lock();
1767 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
1768 /*
1769 * Ignore return value since we are currently cleaning up so any error
1770 * can't be handled.
1771 */
1772 (void) consumer_del_stream(stream, ht);
1773 }
1774 rcu_read_unlock();
1775
1776 lttng_ht_destroy(ht);
1777 }
1778
1779 /*
1780 * Iterate over all streams of the hashtable and free them properly.
1781 *
1782 * XXX: Should not be only for metadata stream or else use an other name.
1783 */
1784 static void destroy_stream_ht(struct lttng_ht *ht)
1785 {
1786 struct lttng_ht_iter iter;
1787 struct lttng_consumer_stream *stream;
1788
1789 if (ht == NULL) {
1790 return;
1791 }
1792
1793 rcu_read_lock();
1794 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
1795 /*
1796 * Ignore return value since we are currently cleaning up so any error
1797 * can't be handled.
1798 */
1799 (void) consumer_del_metadata_stream(stream, ht);
1800 }
1801 rcu_read_unlock();
1802
1803 lttng_ht_destroy(ht);
1804 }
1805
1806 /*
1807 * Clean up a metadata stream and free its memory.
1808 */
1809 void consumer_del_metadata_stream(struct lttng_consumer_stream *stream,
1810 struct lttng_ht *ht)
1811 {
1812 int ret;
1813 struct lttng_ht_iter iter;
1814 struct lttng_consumer_channel *free_chan = NULL;
1815 struct consumer_relayd_sock_pair *relayd;
1816
1817 assert(stream);
1818 /*
1819 * This call should NEVER receive regular stream. It must always be
1820 * metadata stream and this is crucial for data structure synchronization.
1821 */
1822 assert(stream->metadata_flag);
1823
1824 DBG3("Consumer delete metadata stream %d", stream->wait_fd);
1825
1826 if (ht == NULL) {
1827 /* Means the stream was allocated but not successfully added */
1828 goto free_stream;
1829 }
1830
1831 pthread_mutex_lock(&consumer_data.lock);
1832 pthread_mutex_lock(&stream->lock);
1833
1834 switch (consumer_data.type) {
1835 case LTTNG_CONSUMER_KERNEL:
1836 if (stream->mmap_base != NULL) {
1837 ret = munmap(stream->mmap_base, stream->mmap_len);
1838 if (ret != 0) {
1839 PERROR("munmap metadata stream");
1840 }
1841 }
1842 break;
1843 case LTTNG_CONSUMER32_UST:
1844 case LTTNG_CONSUMER64_UST:
1845 lttng_ustconsumer_del_stream(stream);
1846 break;
1847 default:
1848 ERR("Unknown consumer_data type");
1849 assert(0);
1850 goto end;
1851 }
1852
1853 rcu_read_lock();
1854 iter.iter.node = &stream->node.node;
1855 ret = lttng_ht_del(ht, &iter);
1856 assert(!ret);
1857
1858 /* Remove node session id from the consumer_data stream ht */
1859 iter.iter.node = &stream->node_session_id.node;
1860 ret = lttng_ht_del(consumer_data.stream_list_ht, &iter);
1861 assert(!ret);
1862 rcu_read_unlock();
1863
1864 if (stream->out_fd >= 0) {
1865 ret = close(stream->out_fd);
1866 if (ret) {
1867 PERROR("close");
1868 }
1869 }
1870
1871 if (stream->wait_fd >= 0 && !stream->wait_fd_is_copy) {
1872 ret = close(stream->wait_fd);
1873 if (ret) {
1874 PERROR("close");
1875 }
1876 }
1877
1878 if (stream->shm_fd >= 0 && stream->wait_fd != stream->shm_fd) {
1879 ret = close(stream->shm_fd);
1880 if (ret) {
1881 PERROR("close");
1882 }
1883 }
1884
1885 /* Check and cleanup relayd */
1886 rcu_read_lock();
1887 relayd = consumer_find_relayd(stream->net_seq_idx);
1888 if (relayd != NULL) {
1889 uatomic_dec(&relayd->refcount);
1890 assert(uatomic_read(&relayd->refcount) >= 0);
1891
1892 /* Closing streams requires to lock the control socket. */
1893 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
1894 ret = relayd_send_close_stream(&relayd->control_sock,
1895 stream->relayd_stream_id, stream->next_net_seq_num - 1);
1896 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
1897 if (ret < 0) {
1898 DBG("Unable to close stream on the relayd. Continuing");
1899 /*
1900 * Continue here. There is nothing we can do for the relayd.
1901 * Chances are that the relayd has closed the socket so we just
1902 * continue cleaning up.
1903 */
1904 }
1905
1906 /* Both conditions are met, we destroy the relayd. */
1907 if (uatomic_read(&relayd->refcount) == 0 &&
1908 uatomic_read(&relayd->destroy_flag)) {
1909 destroy_relayd(relayd);
1910 }
1911 }
1912 rcu_read_unlock();
1913
1914 /* Atomically decrement channel refcount since other threads can use it. */
1915 uatomic_dec(&stream->chan->refcount);
1916 if (!uatomic_read(&stream->chan->refcount)
1917 && !uatomic_read(&stream->chan->nb_init_streams)) {
1918 /* Go for channel deletion! */
1919 free_chan = stream->chan;
1920 }
1921
1922 end:
1923 pthread_mutex_unlock(&stream->lock);
1924 pthread_mutex_unlock(&consumer_data.lock);
1925
1926 if (free_chan) {
1927 consumer_del_channel(free_chan);
1928 }
1929
1930 free_stream:
1931 call_rcu(&stream->node.head, consumer_free_stream);
1932 }
1933
1934 /*
1935 * Action done with the metadata stream when adding it to the consumer internal
1936 * data structures to handle it.
1937 */
1938 static int consumer_add_metadata_stream(struct lttng_consumer_stream *stream,
1939 struct lttng_ht *ht)
1940 {
1941 int ret = 0;
1942 struct consumer_relayd_sock_pair *relayd;
1943 struct lttng_ht_iter iter;
1944 struct lttng_ht_node_ulong *node;
1945
1946 assert(stream);
1947 assert(ht);
1948
1949 DBG3("Adding metadata stream %d to hash table", stream->wait_fd);
1950
1951 pthread_mutex_lock(&consumer_data.lock);
1952 pthread_mutex_lock(&stream->lock);
1953
1954 /*
1955 * From here, refcounts are updated so be _careful_ when returning an error
1956 * after this point.
1957 */
1958
1959 rcu_read_lock();
1960
1961 /*
1962 * Lookup the stream just to make sure it does not exist in our internal
1963 * state. This should NEVER happen.
1964 */
1965 lttng_ht_lookup(ht, (void *)((unsigned long) stream->wait_fd), &iter);
1966 node = lttng_ht_iter_get_node_ulong(&iter);
1967 assert(!node);
1968
1969 /* Find relayd and, if one is found, increment refcount. */
1970 relayd = consumer_find_relayd(stream->net_seq_idx);
1971 if (relayd != NULL) {
1972 uatomic_inc(&relayd->refcount);
1973 }
1974
1975 /* Update channel refcount once added without error(s). */
1976 uatomic_inc(&stream->chan->refcount);
1977
1978 /*
1979 * When nb_init_streams reaches 0, we don't need to trigger any action in
1980 * terms of destroying the associated channel, because the action that
1981 * causes the count to become 0 also causes a stream to be added. The
1982 * channel deletion will thus be triggered by the following removal of this
1983 * stream.
1984 */
1985 if (uatomic_read(&stream->chan->nb_init_streams) > 0) {
1986 uatomic_dec(&stream->chan->nb_init_streams);
1987 }
1988
1989 lttng_ht_add_unique_ulong(ht, &stream->node);
1990
1991 /*
1992 * Add stream to the stream_list_ht of the consumer data. No need to steal
1993 * the key since the HT does not use it and we allow to add redundant keys
1994 * into this table.
1995 */
1996 lttng_ht_add_ulong(consumer_data.stream_list_ht, &stream->node_session_id);
1997
1998 rcu_read_unlock();
1999
2000 pthread_mutex_unlock(&stream->lock);
2001 pthread_mutex_unlock(&consumer_data.lock);
2002 return ret;
2003 }
2004
2005 /*
2006 * Delete data stream that are flagged for deletion (endpoint_status).
2007 */
2008 static void validate_endpoint_status_data_stream(void)
2009 {
2010 struct lttng_ht_iter iter;
2011 struct lttng_consumer_stream *stream;
2012
2013 DBG("Consumer delete flagged data stream");
2014
2015 rcu_read_lock();
2016 cds_lfht_for_each_entry(data_ht->ht, &iter.iter, stream, node.node) {
2017 /* Validate delete flag of the stream */
2018 if (stream->endpoint_status == CONSUMER_ENDPOINT_ACTIVE) {
2019 continue;
2020 }
2021 /* Delete it right now */
2022 consumer_del_stream(stream, data_ht);
2023 }
2024 rcu_read_unlock();
2025 }
2026
2027 /*
2028 * Delete metadata stream that are flagged for deletion (endpoint_status).
2029 */
2030 static void validate_endpoint_status_metadata_stream(
2031 struct lttng_poll_event *pollset)
2032 {
2033 struct lttng_ht_iter iter;
2034 struct lttng_consumer_stream *stream;
2035
2036 DBG("Consumer delete flagged metadata stream");
2037
2038 assert(pollset);
2039
2040 rcu_read_lock();
2041 cds_lfht_for_each_entry(metadata_ht->ht, &iter.iter, stream, node.node) {
2042 /* Validate delete flag of the stream */
2043 if (stream->endpoint_status == CONSUMER_ENDPOINT_ACTIVE) {
2044 continue;
2045 }
2046 /*
2047 * Remove from pollset so the metadata thread can continue without
2048 * blocking on a deleted stream.
2049 */
2050 lttng_poll_del(pollset, stream->wait_fd);
2051
2052 /* Delete it right now */
2053 consumer_del_metadata_stream(stream, metadata_ht);
2054 }
2055 rcu_read_unlock();
2056 }
2057
2058 /*
2059 * Thread polls on metadata file descriptor and write them on disk or on the
2060 * network.
2061 */
2062 void *consumer_thread_metadata_poll(void *data)
2063 {
2064 int ret, i, pollfd;
2065 uint32_t revents, nb_fd;
2066 struct lttng_consumer_stream *stream = NULL;
2067 struct lttng_ht_iter iter;
2068 struct lttng_ht_node_ulong *node;
2069 struct lttng_poll_event events;
2070 struct lttng_consumer_local_data *ctx = data;
2071 ssize_t len;
2072
2073 rcu_register_thread();
2074
2075 metadata_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG);
2076 if (!metadata_ht) {
2077 /* ENOMEM at this point. Better to bail out. */
2078 goto error;
2079 }
2080
2081 DBG("Thread metadata poll started");
2082
2083 /* Size is set to 1 for the consumer_metadata pipe */
2084 ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC);
2085 if (ret < 0) {
2086 ERR("Poll set creation failed");
2087 goto end;
2088 }
2089
2090 ret = lttng_poll_add(&events, ctx->consumer_metadata_pipe[0], LPOLLIN);
2091 if (ret < 0) {
2092 goto end;
2093 }
2094
2095 /* Main loop */
2096 DBG("Metadata main loop started");
2097
2098 while (1) {
2099 /* Only the metadata pipe is set */
2100 if (LTTNG_POLL_GETNB(&events) == 0 && consumer_quit == 1) {
2101 goto end;
2102 }
2103
2104 restart:
2105 DBG("Metadata poll wait with %d fd(s)", LTTNG_POLL_GETNB(&events));
2106 ret = lttng_poll_wait(&events, -1);
2107 DBG("Metadata event catched in thread");
2108 if (ret < 0) {
2109 if (errno == EINTR) {
2110 ERR("Poll EINTR catched");
2111 goto restart;
2112 }
2113 goto error;
2114 }
2115
2116 nb_fd = ret;
2117
2118 /* From here, the event is a metadata wait fd */
2119 for (i = 0; i < nb_fd; i++) {
2120 revents = LTTNG_POLL_GETEV(&events, i);
2121 pollfd = LTTNG_POLL_GETFD(&events, i);
2122
2123 /* Just don't waste time if no returned events for the fd */
2124 if (!revents) {
2125 continue;
2126 }
2127
2128 if (pollfd == ctx->consumer_metadata_pipe[0]) {
2129 if (revents & (LPOLLERR | LPOLLHUP )) {
2130 DBG("Metadata thread pipe hung up");
2131 /*
2132 * Remove the pipe from the poll set and continue the loop
2133 * since their might be data to consume.
2134 */
2135 lttng_poll_del(&events, ctx->consumer_metadata_pipe[0]);
2136 ret = close(ctx->consumer_metadata_pipe[0]);
2137 if (ret < 0) {
2138 PERROR("close metadata pipe");
2139 }
2140 continue;
2141 } else if (revents & LPOLLIN) {
2142 do {
2143 /* Get the stream pointer received */
2144 ret = read(pollfd, &stream, sizeof(stream));
2145 } while (ret < 0 && errno == EINTR);
2146 if (ret < 0 ||
2147 ret < sizeof(struct lttng_consumer_stream *)) {
2148 PERROR("read metadata stream");
2149 /*
2150 * Let's continue here and hope we can still work
2151 * without stopping the consumer. XXX: Should we?
2152 */
2153 continue;
2154 }
2155
2156 /* A NULL stream means that the state has changed. */
2157 if (stream == NULL) {
2158 /* Check for deleted streams. */
2159 validate_endpoint_status_metadata_stream(&events);
2160 goto restart;
2161 }
2162
2163 DBG("Adding metadata stream %d to poll set",
2164 stream->wait_fd);
2165
2166 ret = consumer_add_metadata_stream(stream, metadata_ht);
2167 if (ret) {
2168 ERR("Unable to add metadata stream");
2169 /* Stream was not setup properly. Continuing. */
2170 consumer_del_metadata_stream(stream, NULL);
2171 continue;
2172 }
2173
2174 /* Add metadata stream to the global poll events list */
2175 lttng_poll_add(&events, stream->wait_fd,
2176 LPOLLIN | LPOLLPRI);
2177 }
2178
2179 /* Handle other stream */
2180 continue;
2181 }
2182
2183 rcu_read_lock();
2184 lttng_ht_lookup(metadata_ht, (void *)((unsigned long) pollfd),
2185 &iter);
2186 node = lttng_ht_iter_get_node_ulong(&iter);
2187 assert(node);
2188
2189 stream = caa_container_of(node, struct lttng_consumer_stream,
2190 node);
2191
2192 /* Check for error event */
2193 if (revents & (LPOLLERR | LPOLLHUP)) {
2194 DBG("Metadata fd %d is hup|err.", pollfd);
2195 if (!stream->hangup_flush_done
2196 && (consumer_data.type == LTTNG_CONSUMER32_UST
2197 || consumer_data.type == LTTNG_CONSUMER64_UST)) {
2198 DBG("Attempting to flush and consume the UST buffers");
2199 lttng_ustconsumer_on_stream_hangup(stream);
2200
2201 /* We just flushed the stream now read it. */
2202 do {
2203 len = ctx->on_buffer_ready(stream, ctx);
2204 /*
2205 * We don't check the return value here since if we get
2206 * a negative len, it means an error occured thus we
2207 * simply remove it from the poll set and free the
2208 * stream.
2209 */
2210 } while (len > 0);
2211 }
2212
2213 lttng_poll_del(&events, stream->wait_fd);
2214 /*
2215 * This call update the channel states, closes file descriptors
2216 * and securely free the stream.
2217 */
2218 consumer_del_metadata_stream(stream, metadata_ht);
2219 } else if (revents & (LPOLLIN | LPOLLPRI)) {
2220 /* Get the data out of the metadata file descriptor */
2221 DBG("Metadata available on fd %d", pollfd);
2222 assert(stream->wait_fd == pollfd);
2223
2224 len = ctx->on_buffer_ready(stream, ctx);
2225 /* It's ok to have an unavailable sub-buffer */
2226 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2227 /* Clean up stream from consumer and free it. */
2228 lttng_poll_del(&events, stream->wait_fd);
2229 consumer_del_metadata_stream(stream, metadata_ht);
2230 } else if (len > 0) {
2231 stream->data_read = 1;
2232 }
2233 }
2234
2235 /* Release RCU lock for the stream looked up */
2236 rcu_read_unlock();
2237 }
2238 }
2239
2240 error:
2241 end:
2242 DBG("Metadata poll thread exiting");
2243 lttng_poll_clean(&events);
2244
2245 destroy_stream_ht(metadata_ht);
2246
2247 rcu_unregister_thread();
2248 return NULL;
2249 }
2250
2251 /*
2252 * This thread polls the fds in the set to consume the data and write
2253 * it to tracefile if necessary.
2254 */
2255 void *consumer_thread_data_poll(void *data)
2256 {
2257 int num_rdy, num_hup, high_prio, ret, i;
2258 struct pollfd *pollfd = NULL;
2259 /* local view of the streams */
2260 struct lttng_consumer_stream **local_stream = NULL, *new_stream = NULL;
2261 /* local view of consumer_data.fds_count */
2262 int nb_fd = 0;
2263 struct lttng_consumer_local_data *ctx = data;
2264 ssize_t len;
2265
2266 rcu_register_thread();
2267
2268 data_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG);
2269 if (data_ht == NULL) {
2270 /* ENOMEM at this point. Better to bail out. */
2271 goto end;
2272 }
2273
2274 local_stream = zmalloc(sizeof(struct lttng_consumer_stream));
2275
2276 while (1) {
2277 high_prio = 0;
2278 num_hup = 0;
2279
2280 /*
2281 * the fds set has been updated, we need to update our
2282 * local array as well
2283 */
2284 pthread_mutex_lock(&consumer_data.lock);
2285 if (consumer_data.need_update) {
2286 if (pollfd != NULL) {
2287 free(pollfd);
2288 pollfd = NULL;
2289 }
2290 if (local_stream != NULL) {
2291 free(local_stream);
2292 local_stream = NULL;
2293 }
2294
2295 /* allocate for all fds + 1 for the consumer_data_pipe */
2296 pollfd = zmalloc((consumer_data.stream_count + 1) * sizeof(struct pollfd));
2297 if (pollfd == NULL) {
2298 PERROR("pollfd malloc");
2299 pthread_mutex_unlock(&consumer_data.lock);
2300 goto end;
2301 }
2302
2303 /* allocate for all fds + 1 for the consumer_data_pipe */
2304 local_stream = zmalloc((consumer_data.stream_count + 1) *
2305 sizeof(struct lttng_consumer_stream));
2306 if (local_stream == NULL) {
2307 PERROR("local_stream malloc");
2308 pthread_mutex_unlock(&consumer_data.lock);
2309 goto end;
2310 }
2311 ret = consumer_update_poll_array(ctx, &pollfd, local_stream,
2312 data_ht);
2313 if (ret < 0) {
2314 ERR("Error in allocating pollfd or local_outfds");
2315 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
2316 pthread_mutex_unlock(&consumer_data.lock);
2317 goto end;
2318 }
2319 nb_fd = ret;
2320 consumer_data.need_update = 0;
2321 }
2322 pthread_mutex_unlock(&consumer_data.lock);
2323
2324 /* No FDs and consumer_quit, consumer_cleanup the thread */
2325 if (nb_fd == 0 && consumer_quit == 1) {
2326 goto end;
2327 }
2328 /* poll on the array of fds */
2329 restart:
2330 DBG("polling on %d fd", nb_fd + 1);
2331 num_rdy = poll(pollfd, nb_fd + 1, -1);
2332 DBG("poll num_rdy : %d", num_rdy);
2333 if (num_rdy == -1) {
2334 /*
2335 * Restart interrupted system call.
2336 */
2337 if (errno == EINTR) {
2338 goto restart;
2339 }
2340 PERROR("Poll error");
2341 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
2342 goto end;
2343 } else if (num_rdy == 0) {
2344 DBG("Polling thread timed out");
2345 goto end;
2346 }
2347
2348 /*
2349 * If the consumer_data_pipe triggered poll go directly to the
2350 * beginning of the loop to update the array. We want to prioritize
2351 * array update over low-priority reads.
2352 */
2353 if (pollfd[nb_fd].revents & (POLLIN | POLLPRI)) {
2354 size_t pipe_readlen;
2355
2356 DBG("consumer_data_pipe wake up");
2357 /* Consume 1 byte of pipe data */
2358 do {
2359 pipe_readlen = read(ctx->consumer_data_pipe[0], &new_stream,
2360 sizeof(new_stream));
2361 } while (pipe_readlen == -1 && errno == EINTR);
2362 if (pipe_readlen < 0) {
2363 PERROR("read consumer data pipe");
2364 /* Continue so we can at least handle the current stream(s). */
2365 continue;
2366 }
2367
2368 /*
2369 * If the stream is NULL, just ignore it. It's also possible that
2370 * the sessiond poll thread changed the consumer_quit state and is
2371 * waking us up to test it.
2372 */
2373 if (new_stream == NULL) {
2374 validate_endpoint_status_data_stream();
2375 continue;
2376 }
2377
2378 ret = consumer_add_stream(new_stream, data_ht);
2379 if (ret) {
2380 ERR("Consumer add stream %d failed. Continuing",
2381 new_stream->key);
2382 /*
2383 * At this point, if the add_stream fails, it is not in the
2384 * hash table thus passing the NULL value here.
2385 */
2386 consumer_del_stream(new_stream, NULL);
2387 }
2388
2389 /* Continue to update the local streams and handle prio ones */
2390 continue;
2391 }
2392
2393 /* Take care of high priority channels first. */
2394 for (i = 0; i < nb_fd; i++) {
2395 if (local_stream[i] == NULL) {
2396 continue;
2397 }
2398 if (pollfd[i].revents & POLLPRI) {
2399 DBG("Urgent read on fd %d", pollfd[i].fd);
2400 high_prio = 1;
2401 len = ctx->on_buffer_ready(local_stream[i], ctx);
2402 /* it's ok to have an unavailable sub-buffer */
2403 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2404 /* Clean the stream and free it. */
2405 consumer_del_stream(local_stream[i], data_ht);
2406 local_stream[i] = NULL;
2407 } else if (len > 0) {
2408 local_stream[i]->data_read = 1;
2409 }
2410 }
2411 }
2412
2413 /*
2414 * If we read high prio channel in this loop, try again
2415 * for more high prio data.
2416 */
2417 if (high_prio) {
2418 continue;
2419 }
2420
2421 /* Take care of low priority channels. */
2422 for (i = 0; i < nb_fd; i++) {
2423 if (local_stream[i] == NULL) {
2424 continue;
2425 }
2426 if ((pollfd[i].revents & POLLIN) ||
2427 local_stream[i]->hangup_flush_done) {
2428 DBG("Normal read on fd %d", pollfd[i].fd);
2429 len = ctx->on_buffer_ready(local_stream[i], ctx);
2430 /* it's ok to have an unavailable sub-buffer */
2431 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2432 /* Clean the stream and free it. */
2433 consumer_del_stream(local_stream[i], data_ht);
2434 local_stream[i] = NULL;
2435 } else if (len > 0) {
2436 local_stream[i]->data_read = 1;
2437 }
2438 }
2439 }
2440
2441 /* Handle hangup and errors */
2442 for (i = 0; i < nb_fd; i++) {
2443 if (local_stream[i] == NULL) {
2444 continue;
2445 }
2446 if (!local_stream[i]->hangup_flush_done
2447 && (pollfd[i].revents & (POLLHUP | POLLERR | POLLNVAL))
2448 && (consumer_data.type == LTTNG_CONSUMER32_UST
2449 || consumer_data.type == LTTNG_CONSUMER64_UST)) {
2450 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2451 pollfd[i].fd);
2452 lttng_ustconsumer_on_stream_hangup(local_stream[i]);
2453 /* Attempt read again, for the data we just flushed. */
2454 local_stream[i]->data_read = 1;
2455 }
2456 /*
2457 * If the poll flag is HUP/ERR/NVAL and we have
2458 * read no data in this pass, we can remove the
2459 * stream from its hash table.
2460 */
2461 if ((pollfd[i].revents & POLLHUP)) {
2462 DBG("Polling fd %d tells it has hung up.", pollfd[i].fd);
2463 if (!local_stream[i]->data_read) {
2464 consumer_del_stream(local_stream[i], data_ht);
2465 local_stream[i] = NULL;
2466 num_hup++;
2467 }
2468 } else if (pollfd[i].revents & POLLERR) {
2469 ERR("Error returned in polling fd %d.", pollfd[i].fd);
2470 if (!local_stream[i]->data_read) {
2471 consumer_del_stream(local_stream[i], data_ht);
2472 local_stream[i] = NULL;
2473 num_hup++;
2474 }
2475 } else if (pollfd[i].revents & POLLNVAL) {
2476 ERR("Polling fd %d tells fd is not open.", pollfd[i].fd);
2477 if (!local_stream[i]->data_read) {
2478 consumer_del_stream(local_stream[i], data_ht);
2479 local_stream[i] = NULL;
2480 num_hup++;
2481 }
2482 }
2483 if (local_stream[i] != NULL) {
2484 local_stream[i]->data_read = 0;
2485 }
2486 }
2487 }
2488 end:
2489 DBG("polling thread exiting");
2490 if (pollfd != NULL) {
2491 free(pollfd);
2492 pollfd = NULL;
2493 }
2494 if (local_stream != NULL) {
2495 free(local_stream);
2496 local_stream = NULL;
2497 }
2498
2499 /*
2500 * Close the write side of the pipe so epoll_wait() in
2501 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2502 * read side of the pipe. If we close them both, epoll_wait strangely does
2503 * not return and could create a endless wait period if the pipe is the
2504 * only tracked fd in the poll set. The thread will take care of closing
2505 * the read side.
2506 */
2507 ret = close(ctx->consumer_metadata_pipe[1]);
2508 if (ret < 0) {
2509 PERROR("close data pipe");
2510 }
2511
2512 destroy_data_stream_ht(data_ht);
2513
2514 rcu_unregister_thread();
2515 return NULL;
2516 }
2517
2518 /*
2519 * This thread listens on the consumerd socket and receives the file
2520 * descriptors from the session daemon.
2521 */
2522 void *consumer_thread_sessiond_poll(void *data)
2523 {
2524 int sock = -1, client_socket, ret;
2525 /*
2526 * structure to poll for incoming data on communication socket avoids
2527 * making blocking sockets.
2528 */
2529 struct pollfd consumer_sockpoll[2];
2530 struct lttng_consumer_local_data *ctx = data;
2531
2532 rcu_register_thread();
2533
2534 DBG("Creating command socket %s", ctx->consumer_command_sock_path);
2535 unlink(ctx->consumer_command_sock_path);
2536 client_socket = lttcomm_create_unix_sock(ctx->consumer_command_sock_path);
2537 if (client_socket < 0) {
2538 ERR("Cannot create command socket");
2539 goto end;
2540 }
2541
2542 ret = lttcomm_listen_unix_sock(client_socket);
2543 if (ret < 0) {
2544 goto end;
2545 }
2546
2547 DBG("Sending ready command to lttng-sessiond");
2548 ret = lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY);
2549 /* return < 0 on error, but == 0 is not fatal */
2550 if (ret < 0) {
2551 ERR("Error sending ready command to lttng-sessiond");
2552 goto end;
2553 }
2554
2555 ret = fcntl(client_socket, F_SETFL, O_NONBLOCK);
2556 if (ret < 0) {
2557 PERROR("fcntl O_NONBLOCK");
2558 goto end;
2559 }
2560
2561 /* prepare the FDs to poll : to client socket and the should_quit pipe */
2562 consumer_sockpoll[0].fd = ctx->consumer_should_quit[0];
2563 consumer_sockpoll[0].events = POLLIN | POLLPRI;
2564 consumer_sockpoll[1].fd = client_socket;
2565 consumer_sockpoll[1].events = POLLIN | POLLPRI;
2566
2567 if (lttng_consumer_poll_socket(consumer_sockpoll) < 0) {
2568 goto end;
2569 }
2570 DBG("Connection on client_socket");
2571
2572 /* Blocking call, waiting for transmission */
2573 sock = lttcomm_accept_unix_sock(client_socket);
2574 if (sock < 0) {
2575 WARN("On accept");
2576 goto end;
2577 }
2578 ret = fcntl(sock, F_SETFL, O_NONBLOCK);
2579 if (ret < 0) {
2580 PERROR("fcntl O_NONBLOCK");
2581 goto end;
2582 }
2583
2584 /* This socket is not useful anymore. */
2585 ret = close(client_socket);
2586 if (ret < 0) {
2587 PERROR("close client_socket");
2588 }
2589 client_socket = -1;
2590
2591 /* update the polling structure to poll on the established socket */
2592 consumer_sockpoll[1].fd = sock;
2593 consumer_sockpoll[1].events = POLLIN | POLLPRI;
2594
2595 while (1) {
2596 if (lttng_consumer_poll_socket(consumer_sockpoll) < 0) {
2597 goto end;
2598 }
2599 DBG("Incoming command on sock");
2600 ret = lttng_consumer_recv_cmd(ctx, sock, consumer_sockpoll);
2601 if (ret == -ENOENT) {
2602 DBG("Received STOP command");
2603 goto end;
2604 }
2605 if (ret <= 0) {
2606 /*
2607 * This could simply be a session daemon quitting. Don't output
2608 * ERR() here.
2609 */
2610 DBG("Communication interrupted on command socket");
2611 goto end;
2612 }
2613 if (consumer_quit) {
2614 DBG("consumer_thread_receive_fds received quit from signal");
2615 goto end;
2616 }
2617 DBG("received fds on sock");
2618 }
2619 end:
2620 DBG("consumer_thread_receive_fds exiting");
2621
2622 /*
2623 * when all fds have hung up, the polling thread
2624 * can exit cleanly
2625 */
2626 consumer_quit = 1;
2627
2628 /*
2629 * Notify the data poll thread to poll back again and test the
2630 * consumer_quit state that we just set so to quit gracefully.
2631 */
2632 notify_thread_pipe(ctx->consumer_data_pipe[1]);
2633
2634 /* Cleaning up possibly open sockets. */
2635 if (sock >= 0) {
2636 ret = close(sock);
2637 if (ret < 0) {
2638 PERROR("close sock sessiond poll");
2639 }
2640 }
2641 if (client_socket >= 0) {
2642 ret = close(sock);
2643 if (ret < 0) {
2644 PERROR("close client_socket sessiond poll");
2645 }
2646 }
2647
2648 rcu_unregister_thread();
2649 return NULL;
2650 }
2651
2652 ssize_t lttng_consumer_read_subbuffer(struct lttng_consumer_stream *stream,
2653 struct lttng_consumer_local_data *ctx)
2654 {
2655 ssize_t ret;
2656
2657 pthread_mutex_lock(&stream->lock);
2658
2659 switch (consumer_data.type) {
2660 case LTTNG_CONSUMER_KERNEL:
2661 ret = lttng_kconsumer_read_subbuffer(stream, ctx);
2662 break;
2663 case LTTNG_CONSUMER32_UST:
2664 case LTTNG_CONSUMER64_UST:
2665 ret = lttng_ustconsumer_read_subbuffer(stream, ctx);
2666 break;
2667 default:
2668 ERR("Unknown consumer_data type");
2669 assert(0);
2670 ret = -ENOSYS;
2671 break;
2672 }
2673
2674 pthread_mutex_unlock(&stream->lock);
2675 return ret;
2676 }
2677
2678 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream *stream)
2679 {
2680 switch (consumer_data.type) {
2681 case LTTNG_CONSUMER_KERNEL:
2682 return lttng_kconsumer_on_recv_stream(stream);
2683 case LTTNG_CONSUMER32_UST:
2684 case LTTNG_CONSUMER64_UST:
2685 return lttng_ustconsumer_on_recv_stream(stream);
2686 default:
2687 ERR("Unknown consumer_data type");
2688 assert(0);
2689 return -ENOSYS;
2690 }
2691 }
2692
2693 /*
2694 * Allocate and set consumer data hash tables.
2695 */
2696 void lttng_consumer_init(void)
2697 {
2698 consumer_data.channel_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG);
2699 consumer_data.relayd_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG);
2700 consumer_data.stream_list_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG);
2701 }
2702
2703 /*
2704 * Process the ADD_RELAYD command receive by a consumer.
2705 *
2706 * This will create a relayd socket pair and add it to the relayd hash table.
2707 * The caller MUST acquire a RCU read side lock before calling it.
2708 */
2709 int consumer_add_relayd_socket(int net_seq_idx, int sock_type,
2710 struct lttng_consumer_local_data *ctx, int sock,
2711 struct pollfd *consumer_sockpoll, struct lttcomm_sock *relayd_sock,
2712 unsigned int sessiond_id)
2713 {
2714 int fd = -1, ret = -1, relayd_created = 0;
2715 enum lttng_error_code ret_code = LTTNG_OK;
2716 struct consumer_relayd_sock_pair *relayd;
2717
2718 DBG("Consumer adding relayd socket (idx: %d)", net_seq_idx);
2719
2720 /* First send a status message before receiving the fds. */
2721 ret = consumer_send_status_msg(sock, ret_code);
2722 if (ret < 0) {
2723 /* Somehow, the session daemon is not responding anymore. */
2724 goto error;
2725 }
2726
2727 /* Get relayd reference if exists. */
2728 relayd = consumer_find_relayd(net_seq_idx);
2729 if (relayd == NULL) {
2730 /* Not found. Allocate one. */
2731 relayd = consumer_allocate_relayd_sock_pair(net_seq_idx);
2732 if (relayd == NULL) {
2733 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_OUTFD_ERROR);
2734 ret = -1;
2735 goto error;
2736 }
2737 relayd->sessiond_session_id = (uint64_t) sessiond_id;
2738 relayd_created = 1;
2739 }
2740
2741 /* Poll on consumer socket. */
2742 if (lttng_consumer_poll_socket(consumer_sockpoll) < 0) {
2743 ret = -EINTR;
2744 goto error;
2745 }
2746
2747 /* Get relayd socket from session daemon */
2748 ret = lttcomm_recv_fds_unix_sock(sock, &fd, 1);
2749 if (ret != sizeof(fd)) {
2750 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_ERROR_RECV_FD);
2751 ret = -1;
2752 fd = -1; /* Just in case it gets set with an invalid value. */
2753 goto error;
2754 }
2755
2756 /* We have the fds without error. Send status back. */
2757 ret = consumer_send_status_msg(sock, ret_code);
2758 if (ret < 0) {
2759 /* Somehow, the session daemon is not responding anymore. */
2760 goto error;
2761 }
2762
2763 /* Copy socket information and received FD */
2764 switch (sock_type) {
2765 case LTTNG_STREAM_CONTROL:
2766 /* Copy received lttcomm socket */
2767 lttcomm_copy_sock(&relayd->control_sock, relayd_sock);
2768 ret = lttcomm_create_sock(&relayd->control_sock);
2769 /* Immediately try to close the created socket if valid. */
2770 if (relayd->control_sock.fd >= 0) {
2771 if (close(relayd->control_sock.fd)) {
2772 PERROR("close relayd control socket");
2773 }
2774 }
2775 /* Handle create_sock error. */
2776 if (ret < 0) {
2777 goto error;
2778 }
2779
2780 /* Assign new file descriptor */
2781 relayd->control_sock.fd = fd;
2782
2783 /*
2784 * Create a session on the relayd and store the returned id. Lock the
2785 * control socket mutex if the relayd was NOT created before.
2786 */
2787 if (!relayd_created) {
2788 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
2789 }
2790 ret = relayd_create_session(&relayd->control_sock,
2791 &relayd->relayd_session_id);
2792 if (!relayd_created) {
2793 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
2794 }
2795 if (ret < 0) {
2796 goto error;
2797 }
2798
2799 break;
2800 case LTTNG_STREAM_DATA:
2801 /* Copy received lttcomm socket */
2802 lttcomm_copy_sock(&relayd->data_sock, relayd_sock);
2803 ret = lttcomm_create_sock(&relayd->data_sock);
2804 /* Immediately try to close the created socket if valid. */
2805 if (relayd->data_sock.fd >= 0) {
2806 if (close(relayd->data_sock.fd)) {
2807 PERROR("close relayd data socket");
2808 }
2809 }
2810 /* Handle create_sock error. */
2811 if (ret < 0) {
2812 goto error;
2813 }
2814
2815 /* Assign new file descriptor */
2816 relayd->data_sock.fd = fd;
2817 break;
2818 default:
2819 ERR("Unknown relayd socket type (%d)", sock_type);
2820 ret = -1;
2821 goto error;
2822 }
2823
2824 DBG("Consumer %s socket created successfully with net idx %d (fd: %d)",
2825 sock_type == LTTNG_STREAM_CONTROL ? "control" : "data",
2826 relayd->net_seq_idx, fd);
2827
2828 /*
2829 * Add relayd socket pair to consumer data hashtable. If object already
2830 * exists or on error, the function gracefully returns.
2831 */
2832 add_relayd(relayd);
2833
2834 /* All good! */
2835 return 0;
2836
2837 error:
2838 /* Close received socket if valid. */
2839 if (fd >= 0) {
2840 if (close(fd)) {
2841 PERROR("close received socket");
2842 }
2843 }
2844
2845 if (relayd_created) {
2846 /* We just want to cleanup. Ignore ret value. */
2847 (void) relayd_close(&relayd->control_sock);
2848 (void) relayd_close(&relayd->data_sock);
2849 free(relayd);
2850 }
2851
2852 return ret;
2853 }
2854
2855 /*
2856 * Try to lock the stream mutex.
2857 *
2858 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
2859 */
2860 static int stream_try_lock(struct lttng_consumer_stream *stream)
2861 {
2862 int ret;
2863
2864 assert(stream);
2865
2866 /*
2867 * Try to lock the stream mutex. On failure, we know that the stream is
2868 * being used else where hence there is data still being extracted.
2869 */
2870 ret = pthread_mutex_trylock(&stream->lock);
2871 if (ret) {
2872 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
2873 ret = 0;
2874 goto end;
2875 }
2876
2877 ret = 1;
2878
2879 end:
2880 return ret;
2881 }
2882
2883 /*
2884 * Search for a relayd associated to the session id and return the reference.
2885 *
2886 * A rcu read side lock MUST be acquire before calling this function and locked
2887 * until the relayd object is no longer necessary.
2888 */
2889 static struct consumer_relayd_sock_pair *find_relayd_by_session_id(uint64_t id)
2890 {
2891 struct lttng_ht_iter iter;
2892 struct consumer_relayd_sock_pair *relayd = NULL;
2893
2894 /* Iterate over all relayd since they are indexed by net_seq_idx. */
2895 cds_lfht_for_each_entry(consumer_data.relayd_ht->ht, &iter.iter, relayd,
2896 node.node) {
2897 /*
2898 * Check by sessiond id which is unique here where the relayd session
2899 * id might not be when having multiple relayd.
2900 */
2901 if (relayd->sessiond_session_id == id) {
2902 /* Found the relayd. There can be only one per id. */
2903 goto found;
2904 }
2905 }
2906
2907 return NULL;
2908
2909 found:
2910 return relayd;
2911 }
2912
2913 /*
2914 * Check if for a given session id there is still data needed to be extract
2915 * from the buffers.
2916 *
2917 * Return 1 if data is pending or else 0 meaning ready to be read.
2918 */
2919 int consumer_data_pending(uint64_t id)
2920 {
2921 int ret;
2922 struct lttng_ht_iter iter;
2923 struct lttng_ht *ht;
2924 struct lttng_consumer_stream *stream;
2925 struct consumer_relayd_sock_pair *relayd = NULL;
2926 int (*data_pending)(struct lttng_consumer_stream *);
2927
2928 DBG("Consumer data pending command on session id %" PRIu64, id);
2929
2930 rcu_read_lock();
2931 pthread_mutex_lock(&consumer_data.lock);
2932
2933 switch (consumer_data.type) {
2934 case LTTNG_CONSUMER_KERNEL:
2935 data_pending = lttng_kconsumer_data_pending;
2936 break;
2937 case LTTNG_CONSUMER32_UST:
2938 case LTTNG_CONSUMER64_UST:
2939 data_pending = lttng_ustconsumer_data_pending;
2940 break;
2941 default:
2942 ERR("Unknown consumer data type");
2943 assert(0);
2944 }
2945
2946 /* Ease our life a bit */
2947 ht = consumer_data.stream_list_ht;
2948
2949 relayd = find_relayd_by_session_id(id);
2950 if (relayd) {
2951 /* Send init command for data pending. */
2952 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
2953 ret = relayd_begin_data_pending(&relayd->control_sock,
2954 relayd->relayd_session_id);
2955 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
2956 if (ret < 0) {
2957 /* Communication error thus the relayd so no data pending. */
2958 goto data_not_pending;
2959 }
2960 }
2961
2962 cds_lfht_for_each_entry_duplicate(ht->ht,
2963 ht->hash_fct((void *)((unsigned long) id), lttng_ht_seed),
2964 ht->match_fct, (void *)((unsigned long) id),
2965 &iter.iter, stream, node_session_id.node) {
2966 /* If this call fails, the stream is being used hence data pending. */
2967 ret = stream_try_lock(stream);
2968 if (!ret) {
2969 goto data_pending;
2970 }
2971
2972 /*
2973 * A removed node from the hash table indicates that the stream has
2974 * been deleted thus having a guarantee that the buffers are closed
2975 * on the consumer side. However, data can still be transmitted
2976 * over the network so don't skip the relayd check.
2977 */
2978 ret = cds_lfht_is_node_deleted(&stream->node.node);
2979 if (!ret) {
2980 /* Check the stream if there is data in the buffers. */
2981 ret = data_pending(stream);
2982 if (ret == 1) {
2983 pthread_mutex_unlock(&stream->lock);
2984 goto data_pending;
2985 }
2986 }
2987
2988 /* Relayd check */
2989 if (relayd) {
2990 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
2991 if (stream->metadata_flag) {
2992 ret = relayd_quiescent_control(&relayd->control_sock,
2993 stream->relayd_stream_id);
2994 } else {
2995 ret = relayd_data_pending(&relayd->control_sock,
2996 stream->relayd_stream_id,
2997 stream->next_net_seq_num - 1);
2998 }
2999 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
3000 if (ret == 1) {
3001 pthread_mutex_unlock(&stream->lock);
3002 goto data_pending;
3003 }
3004 }
3005 pthread_mutex_unlock(&stream->lock);
3006 }
3007
3008 if (relayd) {
3009 unsigned int is_data_inflight = 0;
3010
3011 /* Send init command for data pending. */
3012 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
3013 ret = relayd_end_data_pending(&relayd->control_sock,
3014 relayd->relayd_session_id, &is_data_inflight);
3015 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
3016 if (ret < 0) {
3017 goto data_not_pending;
3018 }
3019 if (is_data_inflight) {
3020 goto data_pending;
3021 }
3022 }
3023
3024 /*
3025 * Finding _no_ node in the hash table and no inflight data means that the
3026 * stream(s) have been removed thus data is guaranteed to be available for
3027 * analysis from the trace files.
3028 */
3029
3030 data_not_pending:
3031 /* Data is available to be read by a viewer. */
3032 pthread_mutex_unlock(&consumer_data.lock);
3033 rcu_read_unlock();
3034 return 0;
3035
3036 data_pending:
3037 /* Data is still being extracted from buffers. */
3038 pthread_mutex_unlock(&consumer_data.lock);
3039 rcu_read_unlock();
3040 return 1;
3041 }
3042
3043 /*
3044 * Send a ret code status message to the sessiond daemon.
3045 *
3046 * Return the sendmsg() return value.
3047 */
3048 int consumer_send_status_msg(int sock, int ret_code)
3049 {
3050 struct lttcomm_consumer_status_msg msg;
3051
3052 msg.ret_code = ret_code;
3053
3054 return lttcomm_send_unix_sock(sock, &msg, sizeof(msg));
3055 }
This page took 0.165785 seconds and 4 git commands to generate.