910f386d932f5a71924bd637b4a5a467abcd6e2f
[lttng-tools.git] / src / common / consumer.c
1 /*
2 * Copyright (C) 2011 - Julien Desfossez <julien.desfossez@polymtl.ca>
3 * Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
4 * 2012 - David Goulet <dgoulet@efficios.com>
5 *
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License, version 2 only,
8 * as published by the Free Software Foundation.
9 *
10 * This program is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
13 * more details.
14 *
15 * You should have received a copy of the GNU General Public License along
16 * with this program; if not, write to the Free Software Foundation, Inc.,
17 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
18 */
19
20 #define _GNU_SOURCE
21 #include <assert.h>
22 #include <poll.h>
23 #include <pthread.h>
24 #include <stdlib.h>
25 #include <string.h>
26 #include <sys/mman.h>
27 #include <sys/socket.h>
28 #include <sys/types.h>
29 #include <unistd.h>
30 #include <inttypes.h>
31 #include <signal.h>
32
33 #include <common/common.h>
34 #include <common/utils.h>
35 #include <common/compat/poll.h>
36 #include <common/kernel-ctl/kernel-ctl.h>
37 #include <common/sessiond-comm/relayd.h>
38 #include <common/sessiond-comm/sessiond-comm.h>
39 #include <common/kernel-consumer/kernel-consumer.h>
40 #include <common/relayd/relayd.h>
41 #include <common/ust-consumer/ust-consumer.h>
42
43 #include "consumer.h"
44 #include "consumer-stream.h"
45
46 struct lttng_consumer_global_data consumer_data = {
47 .stream_count = 0,
48 .need_update = 1,
49 .type = LTTNG_CONSUMER_UNKNOWN,
50 };
51
52 enum consumer_channel_action {
53 CONSUMER_CHANNEL_ADD,
54 CONSUMER_CHANNEL_DEL,
55 CONSUMER_CHANNEL_QUIT,
56 };
57
58 struct consumer_channel_msg {
59 enum consumer_channel_action action;
60 struct lttng_consumer_channel *chan; /* add */
61 uint64_t key; /* del */
62 };
63
64 /*
65 * Flag to inform the polling thread to quit when all fd hung up. Updated by
66 * the consumer_thread_receive_fds when it notices that all fds has hung up.
67 * Also updated by the signal handler (consumer_should_exit()). Read by the
68 * polling threads.
69 */
70 volatile int consumer_quit;
71
72 /*
73 * Global hash table containing respectively metadata and data streams. The
74 * stream element in this ht should only be updated by the metadata poll thread
75 * for the metadata and the data poll thread for the data.
76 */
77 static struct lttng_ht *metadata_ht;
78 static struct lttng_ht *data_ht;
79
80 /*
81 * Notify a thread lttng pipe to poll back again. This usually means that some
82 * global state has changed so we just send back the thread in a poll wait
83 * call.
84 */
85 static void notify_thread_lttng_pipe(struct lttng_pipe *pipe)
86 {
87 struct lttng_consumer_stream *null_stream = NULL;
88
89 assert(pipe);
90
91 (void) lttng_pipe_write(pipe, &null_stream, sizeof(null_stream));
92 }
93
94 static void notify_channel_pipe(struct lttng_consumer_local_data *ctx,
95 struct lttng_consumer_channel *chan,
96 uint64_t key,
97 enum consumer_channel_action action)
98 {
99 struct consumer_channel_msg msg;
100 int ret;
101
102 memset(&msg, 0, sizeof(msg));
103
104 msg.action = action;
105 msg.chan = chan;
106 msg.key = key;
107 do {
108 ret = write(ctx->consumer_channel_pipe[1], &msg, sizeof(msg));
109 } while (ret < 0 && errno == EINTR);
110 }
111
112 void notify_thread_del_channel(struct lttng_consumer_local_data *ctx,
113 uint64_t key)
114 {
115 notify_channel_pipe(ctx, NULL, key, CONSUMER_CHANNEL_DEL);
116 }
117
118 static int read_channel_pipe(struct lttng_consumer_local_data *ctx,
119 struct lttng_consumer_channel **chan,
120 uint64_t *key,
121 enum consumer_channel_action *action)
122 {
123 struct consumer_channel_msg msg;
124 int ret;
125
126 do {
127 ret = read(ctx->consumer_channel_pipe[0], &msg, sizeof(msg));
128 } while (ret < 0 && errno == EINTR);
129 if (ret > 0) {
130 *action = msg.action;
131 *chan = msg.chan;
132 *key = msg.key;
133 }
134 return ret;
135 }
136
137 /*
138 * Find a stream. The consumer_data.lock must be locked during this
139 * call.
140 */
141 static struct lttng_consumer_stream *find_stream(uint64_t key,
142 struct lttng_ht *ht)
143 {
144 struct lttng_ht_iter iter;
145 struct lttng_ht_node_u64 *node;
146 struct lttng_consumer_stream *stream = NULL;
147
148 assert(ht);
149
150 /* -1ULL keys are lookup failures */
151 if (key == (uint64_t) -1ULL) {
152 return NULL;
153 }
154
155 rcu_read_lock();
156
157 lttng_ht_lookup(ht, &key, &iter);
158 node = lttng_ht_iter_get_node_u64(&iter);
159 if (node != NULL) {
160 stream = caa_container_of(node, struct lttng_consumer_stream, node);
161 }
162
163 rcu_read_unlock();
164
165 return stream;
166 }
167
168 static void steal_stream_key(uint64_t key, struct lttng_ht *ht)
169 {
170 struct lttng_consumer_stream *stream;
171
172 rcu_read_lock();
173 stream = find_stream(key, ht);
174 if (stream) {
175 stream->key = (uint64_t) -1ULL;
176 /*
177 * We don't want the lookup to match, but we still need
178 * to iterate on this stream when iterating over the hash table. Just
179 * change the node key.
180 */
181 stream->node.key = (uint64_t) -1ULL;
182 }
183 rcu_read_unlock();
184 }
185
186 /*
187 * Return a channel object for the given key.
188 *
189 * RCU read side lock MUST be acquired before calling this function and
190 * protects the channel ptr.
191 */
192 struct lttng_consumer_channel *consumer_find_channel(uint64_t key)
193 {
194 struct lttng_ht_iter iter;
195 struct lttng_ht_node_u64 *node;
196 struct lttng_consumer_channel *channel = NULL;
197
198 /* -1ULL keys are lookup failures */
199 if (key == (uint64_t) -1ULL) {
200 return NULL;
201 }
202
203 lttng_ht_lookup(consumer_data.channel_ht, &key, &iter);
204 node = lttng_ht_iter_get_node_u64(&iter);
205 if (node != NULL) {
206 channel = caa_container_of(node, struct lttng_consumer_channel, node);
207 }
208
209 return channel;
210 }
211
212 static void free_stream_rcu(struct rcu_head *head)
213 {
214 struct lttng_ht_node_u64 *node =
215 caa_container_of(head, struct lttng_ht_node_u64, head);
216 struct lttng_consumer_stream *stream =
217 caa_container_of(node, struct lttng_consumer_stream, node);
218
219 free(stream);
220 }
221
222 static void free_channel_rcu(struct rcu_head *head)
223 {
224 struct lttng_ht_node_u64 *node =
225 caa_container_of(head, struct lttng_ht_node_u64, head);
226 struct lttng_consumer_channel *channel =
227 caa_container_of(node, struct lttng_consumer_channel, node);
228
229 free(channel);
230 }
231
232 /*
233 * RCU protected relayd socket pair free.
234 */
235 static void free_relayd_rcu(struct rcu_head *head)
236 {
237 struct lttng_ht_node_u64 *node =
238 caa_container_of(head, struct lttng_ht_node_u64, head);
239 struct consumer_relayd_sock_pair *relayd =
240 caa_container_of(node, struct consumer_relayd_sock_pair, node);
241
242 /*
243 * Close all sockets. This is done in the call RCU since we don't want the
244 * socket fds to be reassigned thus potentially creating bad state of the
245 * relayd object.
246 *
247 * We do not have to lock the control socket mutex here since at this stage
248 * there is no one referencing to this relayd object.
249 */
250 (void) relayd_close(&relayd->control_sock);
251 (void) relayd_close(&relayd->data_sock);
252
253 free(relayd);
254 }
255
256 /*
257 * Destroy and free relayd socket pair object.
258 */
259 void consumer_destroy_relayd(struct consumer_relayd_sock_pair *relayd)
260 {
261 int ret;
262 struct lttng_ht_iter iter;
263
264 if (relayd == NULL) {
265 return;
266 }
267
268 DBG("Consumer destroy and close relayd socket pair");
269
270 iter.iter.node = &relayd->node.node;
271 ret = lttng_ht_del(consumer_data.relayd_ht, &iter);
272 if (ret != 0) {
273 /* We assume the relayd is being or is destroyed */
274 return;
275 }
276
277 /* RCU free() call */
278 call_rcu(&relayd->node.head, free_relayd_rcu);
279 }
280
281 /*
282 * Remove a channel from the global list protected by a mutex. This function is
283 * also responsible for freeing its data structures.
284 */
285 void consumer_del_channel(struct lttng_consumer_channel *channel)
286 {
287 int ret;
288 struct lttng_ht_iter iter;
289 struct lttng_consumer_stream *stream, *stmp;
290
291 DBG("Consumer delete channel key %" PRIu64, channel->key);
292
293 pthread_mutex_lock(&consumer_data.lock);
294
295 /* Delete streams that might have been left in the stream list. */
296 cds_list_for_each_entry_safe(stream, stmp, &channel->streams.head,
297 send_node) {
298 cds_list_del(&stream->send_node);
299 /*
300 * Once a stream is added to this list, the buffers were created so
301 * we have a guarantee that this call will succeed.
302 */
303 consumer_stream_destroy(stream, NULL);
304 }
305
306 switch (consumer_data.type) {
307 case LTTNG_CONSUMER_KERNEL:
308 break;
309 case LTTNG_CONSUMER32_UST:
310 case LTTNG_CONSUMER64_UST:
311 lttng_ustconsumer_del_channel(channel);
312 break;
313 default:
314 ERR("Unknown consumer_data type");
315 assert(0);
316 goto end;
317 }
318
319 rcu_read_lock();
320 iter.iter.node = &channel->node.node;
321 ret = lttng_ht_del(consumer_data.channel_ht, &iter);
322 assert(!ret);
323 rcu_read_unlock();
324
325 call_rcu(&channel->node.head, free_channel_rcu);
326 end:
327 pthread_mutex_unlock(&consumer_data.lock);
328 }
329
330 /*
331 * Iterate over the relayd hash table and destroy each element. Finally,
332 * destroy the whole hash table.
333 */
334 static void cleanup_relayd_ht(void)
335 {
336 struct lttng_ht_iter iter;
337 struct consumer_relayd_sock_pair *relayd;
338
339 rcu_read_lock();
340
341 cds_lfht_for_each_entry(consumer_data.relayd_ht->ht, &iter.iter, relayd,
342 node.node) {
343 consumer_destroy_relayd(relayd);
344 }
345
346 rcu_read_unlock();
347
348 lttng_ht_destroy(consumer_data.relayd_ht);
349 }
350
351 /*
352 * Update the end point status of all streams having the given network sequence
353 * index (relayd index).
354 *
355 * It's atomically set without having the stream mutex locked which is fine
356 * because we handle the write/read race with a pipe wakeup for each thread.
357 */
358 static void update_endpoint_status_by_netidx(uint64_t net_seq_idx,
359 enum consumer_endpoint_status status)
360 {
361 struct lttng_ht_iter iter;
362 struct lttng_consumer_stream *stream;
363
364 DBG("Consumer set delete flag on stream by idx %" PRIu64, net_seq_idx);
365
366 rcu_read_lock();
367
368 /* Let's begin with metadata */
369 cds_lfht_for_each_entry(metadata_ht->ht, &iter.iter, stream, node.node) {
370 if (stream->net_seq_idx == net_seq_idx) {
371 uatomic_set(&stream->endpoint_status, status);
372 DBG("Delete flag set to metadata stream %d", stream->wait_fd);
373 }
374 }
375
376 /* Follow up by the data streams */
377 cds_lfht_for_each_entry(data_ht->ht, &iter.iter, stream, node.node) {
378 if (stream->net_seq_idx == net_seq_idx) {
379 uatomic_set(&stream->endpoint_status, status);
380 DBG("Delete flag set to data stream %d", stream->wait_fd);
381 }
382 }
383 rcu_read_unlock();
384 }
385
386 /*
387 * Cleanup a relayd object by flagging every associated streams for deletion,
388 * destroying the object meaning removing it from the relayd hash table,
389 * closing the sockets and freeing the memory in a RCU call.
390 *
391 * If a local data context is available, notify the threads that the streams'
392 * state have changed.
393 */
394 static void cleanup_relayd(struct consumer_relayd_sock_pair *relayd,
395 struct lttng_consumer_local_data *ctx)
396 {
397 uint64_t netidx;
398
399 assert(relayd);
400
401 DBG("Cleaning up relayd sockets");
402
403 /* Save the net sequence index before destroying the object */
404 netidx = relayd->net_seq_idx;
405
406 /*
407 * Delete the relayd from the relayd hash table, close the sockets and free
408 * the object in a RCU call.
409 */
410 consumer_destroy_relayd(relayd);
411
412 /* Set inactive endpoint to all streams */
413 update_endpoint_status_by_netidx(netidx, CONSUMER_ENDPOINT_INACTIVE);
414
415 /*
416 * With a local data context, notify the threads that the streams' state
417 * have changed. The write() action on the pipe acts as an "implicit"
418 * memory barrier ordering the updates of the end point status from the
419 * read of this status which happens AFTER receiving this notify.
420 */
421 if (ctx) {
422 notify_thread_lttng_pipe(ctx->consumer_data_pipe);
423 notify_thread_lttng_pipe(ctx->consumer_metadata_pipe);
424 }
425 }
426
427 /*
428 * Flag a relayd socket pair for destruction. Destroy it if the refcount
429 * reaches zero.
430 *
431 * RCU read side lock MUST be aquired before calling this function.
432 */
433 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair *relayd)
434 {
435 assert(relayd);
436
437 /* Set destroy flag for this object */
438 uatomic_set(&relayd->destroy_flag, 1);
439
440 /* Destroy the relayd if refcount is 0 */
441 if (uatomic_read(&relayd->refcount) == 0) {
442 consumer_destroy_relayd(relayd);
443 }
444 }
445
446 /*
447 * Completly destroy stream from every visiable data structure and the given
448 * hash table if one.
449 *
450 * One this call returns, the stream object is not longer usable nor visible.
451 */
452 void consumer_del_stream(struct lttng_consumer_stream *stream,
453 struct lttng_ht *ht)
454 {
455 consumer_stream_destroy(stream, ht);
456 }
457
458 struct lttng_consumer_stream *consumer_allocate_stream(uint64_t channel_key,
459 uint64_t stream_key,
460 enum lttng_consumer_stream_state state,
461 const char *channel_name,
462 uid_t uid,
463 gid_t gid,
464 uint64_t relayd_id,
465 uint64_t session_id,
466 int cpu,
467 int *alloc_ret,
468 enum consumer_channel_type type,
469 unsigned int monitor)
470 {
471 int ret;
472 struct lttng_consumer_stream *stream;
473
474 stream = zmalloc(sizeof(*stream));
475 if (stream == NULL) {
476 PERROR("malloc struct lttng_consumer_stream");
477 ret = -ENOMEM;
478 goto end;
479 }
480
481 rcu_read_lock();
482
483 stream->key = stream_key;
484 stream->out_fd = -1;
485 stream->out_fd_offset = 0;
486 stream->state = state;
487 stream->uid = uid;
488 stream->gid = gid;
489 stream->net_seq_idx = relayd_id;
490 stream->session_id = session_id;
491 stream->monitor = monitor;
492 pthread_mutex_init(&stream->lock, NULL);
493
494 /* If channel is the metadata, flag this stream as metadata. */
495 if (type == CONSUMER_CHANNEL_TYPE_METADATA) {
496 stream->metadata_flag = 1;
497 /* Metadata is flat out. */
498 strncpy(stream->name, DEFAULT_METADATA_NAME, sizeof(stream->name));
499 } else {
500 /* Format stream name to <channel_name>_<cpu_number> */
501 ret = snprintf(stream->name, sizeof(stream->name), "%s_%d",
502 channel_name, cpu);
503 if (ret < 0) {
504 PERROR("snprintf stream name");
505 goto error;
506 }
507 }
508
509 /* Key is always the wait_fd for streams. */
510 lttng_ht_node_init_u64(&stream->node, stream->key);
511
512 /* Init node per channel id key */
513 lttng_ht_node_init_u64(&stream->node_channel_id, channel_key);
514
515 /* Init session id node with the stream session id */
516 lttng_ht_node_init_u64(&stream->node_session_id, stream->session_id);
517
518 DBG3("Allocated stream %s (key %" PRIu64 ", chan_key %" PRIu64
519 " relayd_id %" PRIu64 ", session_id %" PRIu64,
520 stream->name, stream->key, channel_key,
521 stream->net_seq_idx, stream->session_id);
522
523 rcu_read_unlock();
524 return stream;
525
526 error:
527 rcu_read_unlock();
528 free(stream);
529 end:
530 if (alloc_ret) {
531 *alloc_ret = ret;
532 }
533 return NULL;
534 }
535
536 /*
537 * Add a stream to the global list protected by a mutex.
538 */
539 static int add_stream(struct lttng_consumer_stream *stream,
540 struct lttng_ht *ht)
541 {
542 int ret = 0;
543
544 assert(stream);
545 assert(ht);
546
547 DBG3("Adding consumer stream %" PRIu64, stream->key);
548
549 pthread_mutex_lock(&consumer_data.lock);
550 pthread_mutex_lock(&stream->lock);
551 rcu_read_lock();
552
553 /* Steal stream identifier to avoid having streams with the same key */
554 steal_stream_key(stream->key, ht);
555
556 lttng_ht_add_unique_u64(ht, &stream->node);
557
558 lttng_ht_add_u64(consumer_data.stream_per_chan_id_ht,
559 &stream->node_channel_id);
560
561 /*
562 * Add stream to the stream_list_ht of the consumer data. No need to steal
563 * the key since the HT does not use it and we allow to add redundant keys
564 * into this table.
565 */
566 lttng_ht_add_u64(consumer_data.stream_list_ht, &stream->node_session_id);
567
568 /*
569 * When nb_init_stream_left reaches 0, we don't need to trigger any action
570 * in terms of destroying the associated channel, because the action that
571 * causes the count to become 0 also causes a stream to be added. The
572 * channel deletion will thus be triggered by the following removal of this
573 * stream.
574 */
575 if (uatomic_read(&stream->chan->nb_init_stream_left) > 0) {
576 /* Increment refcount before decrementing nb_init_stream_left */
577 cmm_smp_wmb();
578 uatomic_dec(&stream->chan->nb_init_stream_left);
579 }
580
581 /* Update consumer data once the node is inserted. */
582 consumer_data.stream_count++;
583 consumer_data.need_update = 1;
584
585 rcu_read_unlock();
586 pthread_mutex_unlock(&stream->lock);
587 pthread_mutex_unlock(&consumer_data.lock);
588
589 return ret;
590 }
591
592 /*
593 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
594 * be acquired before calling this.
595 */
596 static int add_relayd(struct consumer_relayd_sock_pair *relayd)
597 {
598 int ret = 0;
599 struct lttng_ht_node_u64 *node;
600 struct lttng_ht_iter iter;
601
602 assert(relayd);
603
604 lttng_ht_lookup(consumer_data.relayd_ht,
605 &relayd->net_seq_idx, &iter);
606 node = lttng_ht_iter_get_node_u64(&iter);
607 if (node != NULL) {
608 goto end;
609 }
610 lttng_ht_add_unique_u64(consumer_data.relayd_ht, &relayd->node);
611
612 end:
613 return ret;
614 }
615
616 /*
617 * Allocate and return a consumer relayd socket.
618 */
619 struct consumer_relayd_sock_pair *consumer_allocate_relayd_sock_pair(
620 uint64_t net_seq_idx)
621 {
622 struct consumer_relayd_sock_pair *obj = NULL;
623
624 /* net sequence index of -1 is a failure */
625 if (net_seq_idx == (uint64_t) -1ULL) {
626 goto error;
627 }
628
629 obj = zmalloc(sizeof(struct consumer_relayd_sock_pair));
630 if (obj == NULL) {
631 PERROR("zmalloc relayd sock");
632 goto error;
633 }
634
635 obj->net_seq_idx = net_seq_idx;
636 obj->refcount = 0;
637 obj->destroy_flag = 0;
638 obj->control_sock.sock.fd = -1;
639 obj->data_sock.sock.fd = -1;
640 lttng_ht_node_init_u64(&obj->node, obj->net_seq_idx);
641 pthread_mutex_init(&obj->ctrl_sock_mutex, NULL);
642
643 error:
644 return obj;
645 }
646
647 /*
648 * Find a relayd socket pair in the global consumer data.
649 *
650 * Return the object if found else NULL.
651 * RCU read-side lock must be held across this call and while using the
652 * returned object.
653 */
654 struct consumer_relayd_sock_pair *consumer_find_relayd(uint64_t key)
655 {
656 struct lttng_ht_iter iter;
657 struct lttng_ht_node_u64 *node;
658 struct consumer_relayd_sock_pair *relayd = NULL;
659
660 /* Negative keys are lookup failures */
661 if (key == (uint64_t) -1ULL) {
662 goto error;
663 }
664
665 lttng_ht_lookup(consumer_data.relayd_ht, &key,
666 &iter);
667 node = lttng_ht_iter_get_node_u64(&iter);
668 if (node != NULL) {
669 relayd = caa_container_of(node, struct consumer_relayd_sock_pair, node);
670 }
671
672 error:
673 return relayd;
674 }
675
676 /*
677 * Find a relayd and send the stream
678 *
679 * Returns 0 on success, < 0 on error
680 */
681 int consumer_send_relayd_stream(struct lttng_consumer_stream *stream,
682 char *path)
683 {
684 int ret = 0;
685 struct consumer_relayd_sock_pair *relayd;
686
687 assert(stream);
688 assert(stream->net_seq_idx != -1ULL);
689 assert(path);
690
691 /* The stream is not metadata. Get relayd reference if exists. */
692 rcu_read_lock();
693 relayd = consumer_find_relayd(stream->net_seq_idx);
694 if (relayd != NULL) {
695 /* Add stream on the relayd */
696 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
697 ret = relayd_add_stream(&relayd->control_sock, stream->name,
698 path, &stream->relayd_stream_id,
699 stream->chan->tracefile_size, stream->chan->tracefile_count);
700 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
701 if (ret < 0) {
702 goto end;
703 }
704 uatomic_inc(&relayd->refcount);
705 stream->sent_to_relayd = 1;
706 } else {
707 ERR("Stream %" PRIu64 " relayd ID %" PRIu64 " unknown. Can't send it.",
708 stream->key, stream->net_seq_idx);
709 ret = -1;
710 goto end;
711 }
712
713 DBG("Stream %s with key %" PRIu64 " sent to relayd id %" PRIu64,
714 stream->name, stream->key, stream->net_seq_idx);
715
716 end:
717 rcu_read_unlock();
718 return ret;
719 }
720
721 /*
722 * Find a relayd and close the stream
723 */
724 void close_relayd_stream(struct lttng_consumer_stream *stream)
725 {
726 struct consumer_relayd_sock_pair *relayd;
727
728 /* The stream is not metadata. Get relayd reference if exists. */
729 rcu_read_lock();
730 relayd = consumer_find_relayd(stream->net_seq_idx);
731 if (relayd) {
732 consumer_stream_relayd_close(stream, relayd);
733 }
734 rcu_read_unlock();
735 }
736
737 /*
738 * Handle stream for relayd transmission if the stream applies for network
739 * streaming where the net sequence index is set.
740 *
741 * Return destination file descriptor or negative value on error.
742 */
743 static int write_relayd_stream_header(struct lttng_consumer_stream *stream,
744 size_t data_size, unsigned long padding,
745 struct consumer_relayd_sock_pair *relayd)
746 {
747 int outfd = -1, ret;
748 struct lttcomm_relayd_data_hdr data_hdr;
749
750 /* Safety net */
751 assert(stream);
752 assert(relayd);
753
754 /* Reset data header */
755 memset(&data_hdr, 0, sizeof(data_hdr));
756
757 if (stream->metadata_flag) {
758 /* Caller MUST acquire the relayd control socket lock */
759 ret = relayd_send_metadata(&relayd->control_sock, data_size);
760 if (ret < 0) {
761 goto error;
762 }
763
764 /* Metadata are always sent on the control socket. */
765 outfd = relayd->control_sock.sock.fd;
766 } else {
767 /* Set header with stream information */
768 data_hdr.stream_id = htobe64(stream->relayd_stream_id);
769 data_hdr.data_size = htobe32(data_size);
770 data_hdr.padding_size = htobe32(padding);
771 /*
772 * Note that net_seq_num below is assigned with the *current* value of
773 * next_net_seq_num and only after that the next_net_seq_num will be
774 * increment. This is why when issuing a command on the relayd using
775 * this next value, 1 should always be substracted in order to compare
776 * the last seen sequence number on the relayd side to the last sent.
777 */
778 data_hdr.net_seq_num = htobe64(stream->next_net_seq_num);
779 /* Other fields are zeroed previously */
780
781 ret = relayd_send_data_hdr(&relayd->data_sock, &data_hdr,
782 sizeof(data_hdr));
783 if (ret < 0) {
784 goto error;
785 }
786
787 ++stream->next_net_seq_num;
788
789 /* Set to go on data socket */
790 outfd = relayd->data_sock.sock.fd;
791 }
792
793 error:
794 return outfd;
795 }
796
797 /*
798 * Allocate and return a new lttng_consumer_channel object using the given key
799 * to initialize the hash table node.
800 *
801 * On error, return NULL.
802 */
803 struct lttng_consumer_channel *consumer_allocate_channel(uint64_t key,
804 uint64_t session_id,
805 const char *pathname,
806 const char *name,
807 uid_t uid,
808 gid_t gid,
809 uint64_t relayd_id,
810 enum lttng_event_output output,
811 uint64_t tracefile_size,
812 uint64_t tracefile_count,
813 uint64_t session_id_per_pid,
814 unsigned int monitor)
815 {
816 struct lttng_consumer_channel *channel;
817
818 channel = zmalloc(sizeof(*channel));
819 if (channel == NULL) {
820 PERROR("malloc struct lttng_consumer_channel");
821 goto end;
822 }
823
824 channel->key = key;
825 channel->refcount = 0;
826 channel->session_id = session_id;
827 channel->session_id_per_pid = session_id_per_pid;
828 channel->uid = uid;
829 channel->gid = gid;
830 channel->relayd_id = relayd_id;
831 channel->output = output;
832 channel->tracefile_size = tracefile_size;
833 channel->tracefile_count = tracefile_count;
834 channel->monitor = monitor;
835
836 /*
837 * In monitor mode, the streams associated with the channel will be put in
838 * a special list ONLY owned by this channel. So, the refcount is set to 1
839 * here meaning that the channel itself has streams that are referenced.
840 *
841 * On a channel deletion, once the channel is no longer visible, the
842 * refcount is decremented and checked for a zero value to delete it. With
843 * streams in no monitor mode, it will now be safe to destroy the channel.
844 */
845 if (!channel->monitor) {
846 channel->refcount = 1;
847 }
848
849 strncpy(channel->pathname, pathname, sizeof(channel->pathname));
850 channel->pathname[sizeof(channel->pathname) - 1] = '\0';
851
852 strncpy(channel->name, name, sizeof(channel->name));
853 channel->name[sizeof(channel->name) - 1] = '\0';
854
855 lttng_ht_node_init_u64(&channel->node, channel->key);
856
857 channel->wait_fd = -1;
858
859 CDS_INIT_LIST_HEAD(&channel->streams.head);
860
861 DBG("Allocated channel (key %" PRIu64 ")", channel->key)
862
863 end:
864 return channel;
865 }
866
867 /*
868 * Add a channel to the global list protected by a mutex.
869 *
870 * On success 0 is returned else a negative value.
871 */
872 int consumer_add_channel(struct lttng_consumer_channel *channel,
873 struct lttng_consumer_local_data *ctx)
874 {
875 int ret = 0;
876 struct lttng_ht_node_u64 *node;
877 struct lttng_ht_iter iter;
878
879 pthread_mutex_lock(&consumer_data.lock);
880 rcu_read_lock();
881
882 lttng_ht_lookup(consumer_data.channel_ht, &channel->key, &iter);
883 node = lttng_ht_iter_get_node_u64(&iter);
884 if (node != NULL) {
885 /* Channel already exist. Ignore the insertion */
886 ERR("Consumer add channel key %" PRIu64 " already exists!",
887 channel->key);
888 ret = -EEXIST;
889 goto end;
890 }
891
892 lttng_ht_add_unique_u64(consumer_data.channel_ht, &channel->node);
893
894 end:
895 rcu_read_unlock();
896 pthread_mutex_unlock(&consumer_data.lock);
897
898 if (!ret && channel->wait_fd != -1 &&
899 channel->type == CONSUMER_CHANNEL_TYPE_DATA) {
900 notify_channel_pipe(ctx, channel, -1, CONSUMER_CHANNEL_ADD);
901 }
902 return ret;
903 }
904
905 /*
906 * Allocate the pollfd structure and the local view of the out fds to avoid
907 * doing a lookup in the linked list and concurrency issues when writing is
908 * needed. Called with consumer_data.lock held.
909 *
910 * Returns the number of fds in the structures.
911 */
912 static int update_poll_array(struct lttng_consumer_local_data *ctx,
913 struct pollfd **pollfd, struct lttng_consumer_stream **local_stream,
914 struct lttng_ht *ht)
915 {
916 int i = 0;
917 struct lttng_ht_iter iter;
918 struct lttng_consumer_stream *stream;
919
920 assert(ctx);
921 assert(ht);
922 assert(pollfd);
923 assert(local_stream);
924
925 DBG("Updating poll fd array");
926 rcu_read_lock();
927 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
928 /*
929 * Only active streams with an active end point can be added to the
930 * poll set and local stream storage of the thread.
931 *
932 * There is a potential race here for endpoint_status to be updated
933 * just after the check. However, this is OK since the stream(s) will
934 * be deleted once the thread is notified that the end point state has
935 * changed where this function will be called back again.
936 */
937 if (stream->state != LTTNG_CONSUMER_ACTIVE_STREAM ||
938 stream->endpoint_status == CONSUMER_ENDPOINT_INACTIVE) {
939 continue;
940 }
941 /*
942 * This clobbers way too much the debug output. Uncomment that if you
943 * need it for debugging purposes.
944 *
945 * DBG("Active FD %d", stream->wait_fd);
946 */
947 (*pollfd)[i].fd = stream->wait_fd;
948 (*pollfd)[i].events = POLLIN | POLLPRI;
949 local_stream[i] = stream;
950 i++;
951 }
952 rcu_read_unlock();
953
954 /*
955 * Insert the consumer_data_pipe at the end of the array and don't
956 * increment i so nb_fd is the number of real FD.
957 */
958 (*pollfd)[i].fd = lttng_pipe_get_readfd(ctx->consumer_data_pipe);
959 (*pollfd)[i].events = POLLIN | POLLPRI;
960 return i;
961 }
962
963 /*
964 * Poll on the should_quit pipe and the command socket return -1 on error and
965 * should exit, 0 if data is available on the command socket
966 */
967 int lttng_consumer_poll_socket(struct pollfd *consumer_sockpoll)
968 {
969 int num_rdy;
970
971 restart:
972 num_rdy = poll(consumer_sockpoll, 2, -1);
973 if (num_rdy == -1) {
974 /*
975 * Restart interrupted system call.
976 */
977 if (errno == EINTR) {
978 goto restart;
979 }
980 PERROR("Poll error");
981 goto exit;
982 }
983 if (consumer_sockpoll[0].revents & (POLLIN | POLLPRI)) {
984 DBG("consumer_should_quit wake up");
985 goto exit;
986 }
987 return 0;
988
989 exit:
990 return -1;
991 }
992
993 /*
994 * Set the error socket.
995 */
996 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data *ctx,
997 int sock)
998 {
999 ctx->consumer_error_socket = sock;
1000 }
1001
1002 /*
1003 * Set the command socket path.
1004 */
1005 void lttng_consumer_set_command_sock_path(
1006 struct lttng_consumer_local_data *ctx, char *sock)
1007 {
1008 ctx->consumer_command_sock_path = sock;
1009 }
1010
1011 /*
1012 * Send return code to the session daemon.
1013 * If the socket is not defined, we return 0, it is not a fatal error
1014 */
1015 int lttng_consumer_send_error(struct lttng_consumer_local_data *ctx, int cmd)
1016 {
1017 if (ctx->consumer_error_socket > 0) {
1018 return lttcomm_send_unix_sock(ctx->consumer_error_socket, &cmd,
1019 sizeof(enum lttcomm_sessiond_command));
1020 }
1021
1022 return 0;
1023 }
1024
1025 /*
1026 * Close all the tracefiles and stream fds and MUST be called when all
1027 * instances are destroyed i.e. when all threads were joined and are ended.
1028 */
1029 void lttng_consumer_cleanup(void)
1030 {
1031 struct lttng_ht_iter iter;
1032 struct lttng_consumer_channel *channel;
1033
1034 rcu_read_lock();
1035
1036 cds_lfht_for_each_entry(consumer_data.channel_ht->ht, &iter.iter, channel,
1037 node.node) {
1038 consumer_del_channel(channel);
1039 }
1040
1041 rcu_read_unlock();
1042
1043 lttng_ht_destroy(consumer_data.channel_ht);
1044
1045 cleanup_relayd_ht();
1046
1047 lttng_ht_destroy(consumer_data.stream_per_chan_id_ht);
1048
1049 /*
1050 * This HT contains streams that are freed by either the metadata thread or
1051 * the data thread so we do *nothing* on the hash table and simply destroy
1052 * it.
1053 */
1054 lttng_ht_destroy(consumer_data.stream_list_ht);
1055 }
1056
1057 /*
1058 * Called from signal handler.
1059 */
1060 void lttng_consumer_should_exit(struct lttng_consumer_local_data *ctx)
1061 {
1062 int ret;
1063 consumer_quit = 1;
1064 do {
1065 ret = write(ctx->consumer_should_quit[1], "4", 1);
1066 } while (ret < 0 && errno == EINTR);
1067 if (ret < 0 || ret != 1) {
1068 PERROR("write consumer quit");
1069 }
1070
1071 DBG("Consumer flag that it should quit");
1072 }
1073
1074 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream *stream,
1075 off_t orig_offset)
1076 {
1077 int outfd = stream->out_fd;
1078
1079 /*
1080 * This does a blocking write-and-wait on any page that belongs to the
1081 * subbuffer prior to the one we just wrote.
1082 * Don't care about error values, as these are just hints and ways to
1083 * limit the amount of page cache used.
1084 */
1085 if (orig_offset < stream->max_sb_size) {
1086 return;
1087 }
1088 lttng_sync_file_range(outfd, orig_offset - stream->max_sb_size,
1089 stream->max_sb_size,
1090 SYNC_FILE_RANGE_WAIT_BEFORE
1091 | SYNC_FILE_RANGE_WRITE
1092 | SYNC_FILE_RANGE_WAIT_AFTER);
1093 /*
1094 * Give hints to the kernel about how we access the file:
1095 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1096 * we write it.
1097 *
1098 * We need to call fadvise again after the file grows because the
1099 * kernel does not seem to apply fadvise to non-existing parts of the
1100 * file.
1101 *
1102 * Call fadvise _after_ having waited for the page writeback to
1103 * complete because the dirty page writeback semantic is not well
1104 * defined. So it can be expected to lead to lower throughput in
1105 * streaming.
1106 */
1107 posix_fadvise(outfd, orig_offset - stream->max_sb_size,
1108 stream->max_sb_size, POSIX_FADV_DONTNEED);
1109 }
1110
1111 /*
1112 * Initialise the necessary environnement :
1113 * - create a new context
1114 * - create the poll_pipe
1115 * - create the should_quit pipe (for signal handler)
1116 * - create the thread pipe (for splice)
1117 *
1118 * Takes a function pointer as argument, this function is called when data is
1119 * available on a buffer. This function is responsible to do the
1120 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1121 * buffer configuration and then kernctl_put_next_subbuf at the end.
1122 *
1123 * Returns a pointer to the new context or NULL on error.
1124 */
1125 struct lttng_consumer_local_data *lttng_consumer_create(
1126 enum lttng_consumer_type type,
1127 ssize_t (*buffer_ready)(struct lttng_consumer_stream *stream,
1128 struct lttng_consumer_local_data *ctx),
1129 int (*recv_channel)(struct lttng_consumer_channel *channel),
1130 int (*recv_stream)(struct lttng_consumer_stream *stream),
1131 int (*update_stream)(int stream_key, uint32_t state))
1132 {
1133 int ret;
1134 struct lttng_consumer_local_data *ctx;
1135
1136 assert(consumer_data.type == LTTNG_CONSUMER_UNKNOWN ||
1137 consumer_data.type == type);
1138 consumer_data.type = type;
1139
1140 ctx = zmalloc(sizeof(struct lttng_consumer_local_data));
1141 if (ctx == NULL) {
1142 PERROR("allocating context");
1143 goto error;
1144 }
1145
1146 ctx->consumer_error_socket = -1;
1147 ctx->consumer_metadata_socket = -1;
1148 /* assign the callbacks */
1149 ctx->on_buffer_ready = buffer_ready;
1150 ctx->on_recv_channel = recv_channel;
1151 ctx->on_recv_stream = recv_stream;
1152 ctx->on_update_stream = update_stream;
1153
1154 ctx->consumer_data_pipe = lttng_pipe_open(0);
1155 if (!ctx->consumer_data_pipe) {
1156 goto error_poll_pipe;
1157 }
1158
1159 ret = pipe(ctx->consumer_should_quit);
1160 if (ret < 0) {
1161 PERROR("Error creating recv pipe");
1162 goto error_quit_pipe;
1163 }
1164
1165 ret = pipe(ctx->consumer_thread_pipe);
1166 if (ret < 0) {
1167 PERROR("Error creating thread pipe");
1168 goto error_thread_pipe;
1169 }
1170
1171 ret = pipe(ctx->consumer_channel_pipe);
1172 if (ret < 0) {
1173 PERROR("Error creating channel pipe");
1174 goto error_channel_pipe;
1175 }
1176
1177 ctx->consumer_metadata_pipe = lttng_pipe_open(0);
1178 if (!ctx->consumer_metadata_pipe) {
1179 goto error_metadata_pipe;
1180 }
1181
1182 ret = utils_create_pipe(ctx->consumer_splice_metadata_pipe);
1183 if (ret < 0) {
1184 goto error_splice_pipe;
1185 }
1186
1187 return ctx;
1188
1189 error_splice_pipe:
1190 lttng_pipe_destroy(ctx->consumer_metadata_pipe);
1191 error_metadata_pipe:
1192 utils_close_pipe(ctx->consumer_channel_pipe);
1193 error_channel_pipe:
1194 utils_close_pipe(ctx->consumer_thread_pipe);
1195 error_thread_pipe:
1196 utils_close_pipe(ctx->consumer_should_quit);
1197 error_quit_pipe:
1198 lttng_pipe_destroy(ctx->consumer_data_pipe);
1199 error_poll_pipe:
1200 free(ctx);
1201 error:
1202 return NULL;
1203 }
1204
1205 /*
1206 * Close all fds associated with the instance and free the context.
1207 */
1208 void lttng_consumer_destroy(struct lttng_consumer_local_data *ctx)
1209 {
1210 int ret;
1211
1212 DBG("Consumer destroying it. Closing everything.");
1213
1214 ret = close(ctx->consumer_error_socket);
1215 if (ret) {
1216 PERROR("close");
1217 }
1218 ret = close(ctx->consumer_metadata_socket);
1219 if (ret) {
1220 PERROR("close");
1221 }
1222 utils_close_pipe(ctx->consumer_thread_pipe);
1223 utils_close_pipe(ctx->consumer_channel_pipe);
1224 lttng_pipe_destroy(ctx->consumer_data_pipe);
1225 lttng_pipe_destroy(ctx->consumer_metadata_pipe);
1226 utils_close_pipe(ctx->consumer_should_quit);
1227 utils_close_pipe(ctx->consumer_splice_metadata_pipe);
1228
1229 unlink(ctx->consumer_command_sock_path);
1230 free(ctx);
1231 }
1232
1233 /*
1234 * Write the metadata stream id on the specified file descriptor.
1235 */
1236 static int write_relayd_metadata_id(int fd,
1237 struct lttng_consumer_stream *stream,
1238 struct consumer_relayd_sock_pair *relayd, unsigned long padding)
1239 {
1240 int ret;
1241 struct lttcomm_relayd_metadata_payload hdr;
1242
1243 hdr.stream_id = htobe64(stream->relayd_stream_id);
1244 hdr.padding_size = htobe32(padding);
1245 do {
1246 ret = write(fd, (void *) &hdr, sizeof(hdr));
1247 } while (ret < 0 && errno == EINTR);
1248 if (ret < 0 || ret != sizeof(hdr)) {
1249 /*
1250 * This error means that the fd's end is closed so ignore the perror
1251 * not to clubber the error output since this can happen in a normal
1252 * code path.
1253 */
1254 if (errno != EPIPE) {
1255 PERROR("write metadata stream id");
1256 }
1257 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno);
1258 /*
1259 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1260 * handle writting the missing part so report that as an error and
1261 * don't lie to the caller.
1262 */
1263 ret = -1;
1264 goto end;
1265 }
1266 DBG("Metadata stream id %" PRIu64 " with padding %lu written before data",
1267 stream->relayd_stream_id, padding);
1268
1269 end:
1270 return ret;
1271 }
1272
1273 /*
1274 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1275 * core function for writing trace buffers to either the local filesystem or
1276 * the network.
1277 *
1278 * It must be called with the stream lock held.
1279 *
1280 * Careful review MUST be put if any changes occur!
1281 *
1282 * Returns the number of bytes written
1283 */
1284 ssize_t lttng_consumer_on_read_subbuffer_mmap(
1285 struct lttng_consumer_local_data *ctx,
1286 struct lttng_consumer_stream *stream, unsigned long len,
1287 unsigned long padding)
1288 {
1289 unsigned long mmap_offset;
1290 void *mmap_base;
1291 ssize_t ret = 0, written = 0;
1292 off_t orig_offset = stream->out_fd_offset;
1293 /* Default is on the disk */
1294 int outfd = stream->out_fd;
1295 struct consumer_relayd_sock_pair *relayd = NULL;
1296 unsigned int relayd_hang_up = 0;
1297
1298 /* RCU lock for the relayd pointer */
1299 rcu_read_lock();
1300
1301 /* Flag that the current stream if set for network streaming. */
1302 if (stream->net_seq_idx != (uint64_t) -1ULL) {
1303 relayd = consumer_find_relayd(stream->net_seq_idx);
1304 if (relayd == NULL) {
1305 goto end;
1306 }
1307 }
1308
1309 /* get the offset inside the fd to mmap */
1310 switch (consumer_data.type) {
1311 case LTTNG_CONSUMER_KERNEL:
1312 mmap_base = stream->mmap_base;
1313 ret = kernctl_get_mmap_read_offset(stream->wait_fd, &mmap_offset);
1314 break;
1315 case LTTNG_CONSUMER32_UST:
1316 case LTTNG_CONSUMER64_UST:
1317 mmap_base = lttng_ustctl_get_mmap_base(stream);
1318 if (!mmap_base) {
1319 ERR("read mmap get mmap base for stream %s", stream->name);
1320 written = -1;
1321 goto end;
1322 }
1323 ret = lttng_ustctl_get_mmap_read_offset(stream, &mmap_offset);
1324
1325 break;
1326 default:
1327 ERR("Unknown consumer_data type");
1328 assert(0);
1329 }
1330 if (ret != 0) {
1331 errno = -ret;
1332 PERROR("tracer ctl get_mmap_read_offset");
1333 written = ret;
1334 goto end;
1335 }
1336
1337 /* Handle stream on the relayd if the output is on the network */
1338 if (relayd) {
1339 unsigned long netlen = len;
1340
1341 /*
1342 * Lock the control socket for the complete duration of the function
1343 * since from this point on we will use the socket.
1344 */
1345 if (stream->metadata_flag) {
1346 /* Metadata requires the control socket. */
1347 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
1348 netlen += sizeof(struct lttcomm_relayd_metadata_payload);
1349 }
1350
1351 ret = write_relayd_stream_header(stream, netlen, padding, relayd);
1352 if (ret >= 0) {
1353 /* Use the returned socket. */
1354 outfd = ret;
1355
1356 /* Write metadata stream id before payload */
1357 if (stream->metadata_flag) {
1358 ret = write_relayd_metadata_id(outfd, stream, relayd, padding);
1359 if (ret < 0) {
1360 written = ret;
1361 /* Socket operation failed. We consider the relayd dead */
1362 if (ret == -EPIPE || ret == -EINVAL) {
1363 relayd_hang_up = 1;
1364 goto write_error;
1365 }
1366 goto end;
1367 }
1368 }
1369 } else {
1370 /* Socket operation failed. We consider the relayd dead */
1371 if (ret == -EPIPE || ret == -EINVAL) {
1372 relayd_hang_up = 1;
1373 goto write_error;
1374 }
1375 /* Else, use the default set before which is the filesystem. */
1376 }
1377 } else {
1378 /* No streaming, we have to set the len with the full padding */
1379 len += padding;
1380
1381 /*
1382 * Check if we need to change the tracefile before writing the packet.
1383 */
1384 if (stream->chan->tracefile_size > 0 &&
1385 (stream->tracefile_size_current + len) >
1386 stream->chan->tracefile_size) {
1387 ret = utils_rotate_stream_file(stream->chan->pathname,
1388 stream->name, stream->chan->tracefile_size,
1389 stream->chan->tracefile_count, stream->uid, stream->gid,
1390 stream->out_fd, &(stream->tracefile_count_current));
1391 if (ret < 0) {
1392 ERR("Rotating output file");
1393 goto end;
1394 }
1395 outfd = stream->out_fd = ret;
1396 /* Reset current size because we just perform a rotation. */
1397 stream->tracefile_size_current = 0;
1398 }
1399 stream->tracefile_size_current += len;
1400 }
1401
1402 while (len > 0) {
1403 do {
1404 ret = write(outfd, mmap_base + mmap_offset, len);
1405 } while (ret < 0 && errno == EINTR);
1406 DBG("Consumer mmap write() ret %zd (len %lu)", ret, len);
1407 if (ret < 0) {
1408 /*
1409 * This is possible if the fd is closed on the other side (outfd)
1410 * or any write problem. It can be verbose a bit for a normal
1411 * execution if for instance the relayd is stopped abruptly. This
1412 * can happen so set this to a DBG statement.
1413 */
1414 DBG("Error in file write mmap");
1415 if (written == 0) {
1416 written = ret;
1417 }
1418 /* Socket operation failed. We consider the relayd dead */
1419 if (errno == EPIPE || errno == EINVAL) {
1420 relayd_hang_up = 1;
1421 goto write_error;
1422 }
1423 goto end;
1424 } else if (ret > len) {
1425 PERROR("Error in file write (ret %zd > len %lu)", ret, len);
1426 written += ret;
1427 goto end;
1428 } else {
1429 len -= ret;
1430 mmap_offset += ret;
1431 }
1432
1433 /* This call is useless on a socket so better save a syscall. */
1434 if (!relayd) {
1435 /* This won't block, but will start writeout asynchronously */
1436 lttng_sync_file_range(outfd, stream->out_fd_offset, ret,
1437 SYNC_FILE_RANGE_WRITE);
1438 stream->out_fd_offset += ret;
1439 }
1440 written += ret;
1441 }
1442 lttng_consumer_sync_trace_file(stream, orig_offset);
1443
1444 write_error:
1445 /*
1446 * This is a special case that the relayd has closed its socket. Let's
1447 * cleanup the relayd object and all associated streams.
1448 */
1449 if (relayd && relayd_hang_up) {
1450 cleanup_relayd(relayd, ctx);
1451 }
1452
1453 end:
1454 /* Unlock only if ctrl socket used */
1455 if (relayd && stream->metadata_flag) {
1456 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
1457 }
1458
1459 rcu_read_unlock();
1460 return written;
1461 }
1462
1463 /*
1464 * Splice the data from the ring buffer to the tracefile.
1465 *
1466 * It must be called with the stream lock held.
1467 *
1468 * Returns the number of bytes spliced.
1469 */
1470 ssize_t lttng_consumer_on_read_subbuffer_splice(
1471 struct lttng_consumer_local_data *ctx,
1472 struct lttng_consumer_stream *stream, unsigned long len,
1473 unsigned long padding)
1474 {
1475 ssize_t ret = 0, written = 0, ret_splice = 0;
1476 loff_t offset = 0;
1477 off_t orig_offset = stream->out_fd_offset;
1478 int fd = stream->wait_fd;
1479 /* Default is on the disk */
1480 int outfd = stream->out_fd;
1481 struct consumer_relayd_sock_pair *relayd = NULL;
1482 int *splice_pipe;
1483 unsigned int relayd_hang_up = 0;
1484
1485 switch (consumer_data.type) {
1486 case LTTNG_CONSUMER_KERNEL:
1487 break;
1488 case LTTNG_CONSUMER32_UST:
1489 case LTTNG_CONSUMER64_UST:
1490 /* Not supported for user space tracing */
1491 return -ENOSYS;
1492 default:
1493 ERR("Unknown consumer_data type");
1494 assert(0);
1495 }
1496
1497 /* RCU lock for the relayd pointer */
1498 rcu_read_lock();
1499
1500 /* Flag that the current stream if set for network streaming. */
1501 if (stream->net_seq_idx != (uint64_t) -1ULL) {
1502 relayd = consumer_find_relayd(stream->net_seq_idx);
1503 if (relayd == NULL) {
1504 goto end;
1505 }
1506 }
1507
1508 /*
1509 * Choose right pipe for splice. Metadata and trace data are handled by
1510 * different threads hence the use of two pipes in order not to race or
1511 * corrupt the written data.
1512 */
1513 if (stream->metadata_flag) {
1514 splice_pipe = ctx->consumer_splice_metadata_pipe;
1515 } else {
1516 splice_pipe = ctx->consumer_thread_pipe;
1517 }
1518
1519 /* Write metadata stream id before payload */
1520 if (relayd) {
1521 int total_len = len;
1522
1523 if (stream->metadata_flag) {
1524 /*
1525 * Lock the control socket for the complete duration of the function
1526 * since from this point on we will use the socket.
1527 */
1528 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
1529
1530 ret = write_relayd_metadata_id(splice_pipe[1], stream, relayd,
1531 padding);
1532 if (ret < 0) {
1533 written = ret;
1534 /* Socket operation failed. We consider the relayd dead */
1535 if (ret == -EBADF) {
1536 WARN("Remote relayd disconnected. Stopping");
1537 relayd_hang_up = 1;
1538 goto write_error;
1539 }
1540 goto end;
1541 }
1542
1543 total_len += sizeof(struct lttcomm_relayd_metadata_payload);
1544 }
1545
1546 ret = write_relayd_stream_header(stream, total_len, padding, relayd);
1547 if (ret >= 0) {
1548 /* Use the returned socket. */
1549 outfd = ret;
1550 } else {
1551 /* Socket operation failed. We consider the relayd dead */
1552 if (ret == -EBADF) {
1553 WARN("Remote relayd disconnected. Stopping");
1554 relayd_hang_up = 1;
1555 goto write_error;
1556 }
1557 goto end;
1558 }
1559 } else {
1560 /* No streaming, we have to set the len with the full padding */
1561 len += padding;
1562
1563 /*
1564 * Check if we need to change the tracefile before writing the packet.
1565 */
1566 if (stream->chan->tracefile_size > 0 &&
1567 (stream->tracefile_size_current + len) >
1568 stream->chan->tracefile_size) {
1569 ret = utils_rotate_stream_file(stream->chan->pathname,
1570 stream->name, stream->chan->tracefile_size,
1571 stream->chan->tracefile_count, stream->uid, stream->gid,
1572 stream->out_fd, &(stream->tracefile_count_current));
1573 if (ret < 0) {
1574 ERR("Rotating output file");
1575 goto end;
1576 }
1577 outfd = stream->out_fd = ret;
1578 /* Reset current size because we just perform a rotation. */
1579 stream->tracefile_size_current = 0;
1580 }
1581 stream->tracefile_size_current += len;
1582 }
1583
1584 while (len > 0) {
1585 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1586 (unsigned long)offset, len, fd, splice_pipe[1]);
1587 ret_splice = splice(fd, &offset, splice_pipe[1], NULL, len,
1588 SPLICE_F_MOVE | SPLICE_F_MORE);
1589 DBG("splice chan to pipe, ret %zd", ret_splice);
1590 if (ret_splice < 0) {
1591 PERROR("Error in relay splice");
1592 if (written == 0) {
1593 written = ret_splice;
1594 }
1595 ret = errno;
1596 goto splice_error;
1597 }
1598
1599 /* Handle stream on the relayd if the output is on the network */
1600 if (relayd) {
1601 if (stream->metadata_flag) {
1602 size_t metadata_payload_size =
1603 sizeof(struct lttcomm_relayd_metadata_payload);
1604
1605 /* Update counter to fit the spliced data */
1606 ret_splice += metadata_payload_size;
1607 len += metadata_payload_size;
1608 /*
1609 * We do this so the return value can match the len passed as
1610 * argument to this function.
1611 */
1612 written -= metadata_payload_size;
1613 }
1614 }
1615
1616 /* Splice data out */
1617 ret_splice = splice(splice_pipe[0], NULL, outfd, NULL,
1618 ret_splice, SPLICE_F_MOVE | SPLICE_F_MORE);
1619 DBG("Consumer splice pipe to file, ret %zd", ret_splice);
1620 if (ret_splice < 0) {
1621 PERROR("Error in file splice");
1622 if (written == 0) {
1623 written = ret_splice;
1624 }
1625 /* Socket operation failed. We consider the relayd dead */
1626 if (errno == EBADF || errno == EPIPE) {
1627 WARN("Remote relayd disconnected. Stopping");
1628 relayd_hang_up = 1;
1629 goto write_error;
1630 }
1631 ret = errno;
1632 goto splice_error;
1633 } else if (ret_splice > len) {
1634 errno = EINVAL;
1635 PERROR("Wrote more data than requested %zd (len: %lu)",
1636 ret_splice, len);
1637 written += ret_splice;
1638 ret = errno;
1639 goto splice_error;
1640 }
1641 len -= ret_splice;
1642
1643 /* This call is useless on a socket so better save a syscall. */
1644 if (!relayd) {
1645 /* This won't block, but will start writeout asynchronously */
1646 lttng_sync_file_range(outfd, stream->out_fd_offset, ret_splice,
1647 SYNC_FILE_RANGE_WRITE);
1648 stream->out_fd_offset += ret_splice;
1649 }
1650 written += ret_splice;
1651 }
1652 lttng_consumer_sync_trace_file(stream, orig_offset);
1653
1654 ret = ret_splice;
1655
1656 goto end;
1657
1658 write_error:
1659 /*
1660 * This is a special case that the relayd has closed its socket. Let's
1661 * cleanup the relayd object and all associated streams.
1662 */
1663 if (relayd && relayd_hang_up) {
1664 cleanup_relayd(relayd, ctx);
1665 /* Skip splice error so the consumer does not fail */
1666 goto end;
1667 }
1668
1669 splice_error:
1670 /* send the appropriate error description to sessiond */
1671 switch (ret) {
1672 case EINVAL:
1673 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_EINVAL);
1674 break;
1675 case ENOMEM:
1676 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_ENOMEM);
1677 break;
1678 case ESPIPE:
1679 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_ESPIPE);
1680 break;
1681 }
1682
1683 end:
1684 if (relayd && stream->metadata_flag) {
1685 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
1686 }
1687
1688 rcu_read_unlock();
1689 return written;
1690 }
1691
1692 /*
1693 * Take a snapshot for a specific fd
1694 *
1695 * Returns 0 on success, < 0 on error
1696 */
1697 int lttng_consumer_take_snapshot(struct lttng_consumer_stream *stream)
1698 {
1699 switch (consumer_data.type) {
1700 case LTTNG_CONSUMER_KERNEL:
1701 return lttng_kconsumer_take_snapshot(stream);
1702 case LTTNG_CONSUMER32_UST:
1703 case LTTNG_CONSUMER64_UST:
1704 return lttng_ustconsumer_take_snapshot(stream);
1705 default:
1706 ERR("Unknown consumer_data type");
1707 assert(0);
1708 return -ENOSYS;
1709 }
1710 }
1711
1712 /*
1713 * Get the produced position
1714 *
1715 * Returns 0 on success, < 0 on error
1716 */
1717 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream *stream,
1718 unsigned long *pos)
1719 {
1720 switch (consumer_data.type) {
1721 case LTTNG_CONSUMER_KERNEL:
1722 return lttng_kconsumer_get_produced_snapshot(stream, pos);
1723 case LTTNG_CONSUMER32_UST:
1724 case LTTNG_CONSUMER64_UST:
1725 return lttng_ustconsumer_get_produced_snapshot(stream, pos);
1726 default:
1727 ERR("Unknown consumer_data type");
1728 assert(0);
1729 return -ENOSYS;
1730 }
1731 }
1732
1733 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data *ctx,
1734 int sock, struct pollfd *consumer_sockpoll)
1735 {
1736 switch (consumer_data.type) {
1737 case LTTNG_CONSUMER_KERNEL:
1738 return lttng_kconsumer_recv_cmd(ctx, sock, consumer_sockpoll);
1739 case LTTNG_CONSUMER32_UST:
1740 case LTTNG_CONSUMER64_UST:
1741 return lttng_ustconsumer_recv_cmd(ctx, sock, consumer_sockpoll);
1742 default:
1743 ERR("Unknown consumer_data type");
1744 assert(0);
1745 return -ENOSYS;
1746 }
1747 }
1748
1749 /*
1750 * Iterate over all streams of the hashtable and free them properly.
1751 *
1752 * WARNING: *MUST* be used with data stream only.
1753 */
1754 static void destroy_data_stream_ht(struct lttng_ht *ht)
1755 {
1756 struct lttng_ht_iter iter;
1757 struct lttng_consumer_stream *stream;
1758
1759 if (ht == NULL) {
1760 return;
1761 }
1762
1763 rcu_read_lock();
1764 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
1765 /*
1766 * Ignore return value since we are currently cleaning up so any error
1767 * can't be handled.
1768 */
1769 (void) consumer_del_stream(stream, ht);
1770 }
1771 rcu_read_unlock();
1772
1773 lttng_ht_destroy(ht);
1774 }
1775
1776 /*
1777 * Iterate over all streams of the hashtable and free them properly.
1778 *
1779 * XXX: Should not be only for metadata stream or else use an other name.
1780 */
1781 static void destroy_stream_ht(struct lttng_ht *ht)
1782 {
1783 struct lttng_ht_iter iter;
1784 struct lttng_consumer_stream *stream;
1785
1786 if (ht == NULL) {
1787 return;
1788 }
1789
1790 rcu_read_lock();
1791 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
1792 /*
1793 * Ignore return value since we are currently cleaning up so any error
1794 * can't be handled.
1795 */
1796 (void) consumer_del_metadata_stream(stream, ht);
1797 }
1798 rcu_read_unlock();
1799
1800 lttng_ht_destroy(ht);
1801 }
1802
1803 void lttng_consumer_close_metadata(void)
1804 {
1805 switch (consumer_data.type) {
1806 case LTTNG_CONSUMER_KERNEL:
1807 /*
1808 * The Kernel consumer has a different metadata scheme so we don't
1809 * close anything because the stream will be closed by the session
1810 * daemon.
1811 */
1812 break;
1813 case LTTNG_CONSUMER32_UST:
1814 case LTTNG_CONSUMER64_UST:
1815 /*
1816 * Close all metadata streams. The metadata hash table is passed and
1817 * this call iterates over it by closing all wakeup fd. This is safe
1818 * because at this point we are sure that the metadata producer is
1819 * either dead or blocked.
1820 */
1821 lttng_ustconsumer_close_metadata(metadata_ht);
1822 break;
1823 default:
1824 ERR("Unknown consumer_data type");
1825 assert(0);
1826 }
1827 }
1828
1829 /*
1830 * Clean up a metadata stream and free its memory.
1831 */
1832 void consumer_del_metadata_stream(struct lttng_consumer_stream *stream,
1833 struct lttng_ht *ht)
1834 {
1835 int ret;
1836 struct lttng_ht_iter iter;
1837 struct lttng_consumer_channel *free_chan = NULL;
1838 struct consumer_relayd_sock_pair *relayd;
1839
1840 assert(stream);
1841 /*
1842 * This call should NEVER receive regular stream. It must always be
1843 * metadata stream and this is crucial for data structure synchronization.
1844 */
1845 assert(stream->metadata_flag);
1846
1847 DBG3("Consumer delete metadata stream %d", stream->wait_fd);
1848
1849 if (ht == NULL) {
1850 /* Means the stream was allocated but not successfully added */
1851 goto free_stream_rcu;
1852 }
1853
1854 pthread_mutex_lock(&consumer_data.lock);
1855 pthread_mutex_lock(&stream->lock);
1856
1857 switch (consumer_data.type) {
1858 case LTTNG_CONSUMER_KERNEL:
1859 if (stream->mmap_base != NULL) {
1860 ret = munmap(stream->mmap_base, stream->mmap_len);
1861 if (ret != 0) {
1862 PERROR("munmap metadata stream");
1863 }
1864 }
1865 if (stream->wait_fd >= 0) {
1866 ret = close(stream->wait_fd);
1867 if (ret < 0) {
1868 PERROR("close kernel metadata wait_fd");
1869 }
1870 }
1871 break;
1872 case LTTNG_CONSUMER32_UST:
1873 case LTTNG_CONSUMER64_UST:
1874 lttng_ustconsumer_del_stream(stream);
1875 break;
1876 default:
1877 ERR("Unknown consumer_data type");
1878 assert(0);
1879 goto end;
1880 }
1881
1882 rcu_read_lock();
1883 iter.iter.node = &stream->node.node;
1884 ret = lttng_ht_del(ht, &iter);
1885 assert(!ret);
1886
1887 iter.iter.node = &stream->node_channel_id.node;
1888 ret = lttng_ht_del(consumer_data.stream_per_chan_id_ht, &iter);
1889 assert(!ret);
1890
1891 iter.iter.node = &stream->node_session_id.node;
1892 ret = lttng_ht_del(consumer_data.stream_list_ht, &iter);
1893 assert(!ret);
1894 rcu_read_unlock();
1895
1896 if (stream->out_fd >= 0) {
1897 ret = close(stream->out_fd);
1898 if (ret) {
1899 PERROR("close");
1900 }
1901 }
1902
1903 /* Check and cleanup relayd */
1904 rcu_read_lock();
1905 relayd = consumer_find_relayd(stream->net_seq_idx);
1906 if (relayd != NULL) {
1907 uatomic_dec(&relayd->refcount);
1908 assert(uatomic_read(&relayd->refcount) >= 0);
1909
1910 /* Closing streams requires to lock the control socket. */
1911 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
1912 ret = relayd_send_close_stream(&relayd->control_sock,
1913 stream->relayd_stream_id, stream->next_net_seq_num - 1);
1914 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
1915 if (ret < 0) {
1916 DBG("Unable to close stream on the relayd. Continuing");
1917 /*
1918 * Continue here. There is nothing we can do for the relayd.
1919 * Chances are that the relayd has closed the socket so we just
1920 * continue cleaning up.
1921 */
1922 }
1923
1924 /* Both conditions are met, we destroy the relayd. */
1925 if (uatomic_read(&relayd->refcount) == 0 &&
1926 uatomic_read(&relayd->destroy_flag)) {
1927 consumer_destroy_relayd(relayd);
1928 }
1929 }
1930 rcu_read_unlock();
1931
1932 /* Atomically decrement channel refcount since other threads can use it. */
1933 if (!uatomic_sub_return(&stream->chan->refcount, 1)
1934 && !uatomic_read(&stream->chan->nb_init_stream_left)) {
1935 /* Go for channel deletion! */
1936 free_chan = stream->chan;
1937 }
1938
1939 end:
1940 /*
1941 * Nullify the stream reference so it is not used after deletion. The
1942 * consumer data lock MUST be acquired before being able to check for a
1943 * NULL pointer value.
1944 */
1945 stream->chan->metadata_stream = NULL;
1946
1947 pthread_mutex_unlock(&stream->lock);
1948 pthread_mutex_unlock(&consumer_data.lock);
1949
1950 if (free_chan) {
1951 consumer_del_channel(free_chan);
1952 }
1953
1954 free_stream_rcu:
1955 call_rcu(&stream->node.head, free_stream_rcu);
1956 }
1957
1958 /*
1959 * Action done with the metadata stream when adding it to the consumer internal
1960 * data structures to handle it.
1961 */
1962 static int add_metadata_stream(struct lttng_consumer_stream *stream,
1963 struct lttng_ht *ht)
1964 {
1965 int ret = 0;
1966 struct lttng_ht_iter iter;
1967 struct lttng_ht_node_u64 *node;
1968
1969 assert(stream);
1970 assert(ht);
1971
1972 DBG3("Adding metadata stream %" PRIu64 " to hash table", stream->key);
1973
1974 pthread_mutex_lock(&consumer_data.lock);
1975 pthread_mutex_lock(&stream->lock);
1976
1977 /*
1978 * From here, refcounts are updated so be _careful_ when returning an error
1979 * after this point.
1980 */
1981
1982 rcu_read_lock();
1983
1984 /*
1985 * Lookup the stream just to make sure it does not exist in our internal
1986 * state. This should NEVER happen.
1987 */
1988 lttng_ht_lookup(ht, &stream->key, &iter);
1989 node = lttng_ht_iter_get_node_u64(&iter);
1990 assert(!node);
1991
1992 /*
1993 * When nb_init_stream_left reaches 0, we don't need to trigger any action
1994 * in terms of destroying the associated channel, because the action that
1995 * causes the count to become 0 also causes a stream to be added. The
1996 * channel deletion will thus be triggered by the following removal of this
1997 * stream.
1998 */
1999 if (uatomic_read(&stream->chan->nb_init_stream_left) > 0) {
2000 /* Increment refcount before decrementing nb_init_stream_left */
2001 cmm_smp_wmb();
2002 uatomic_dec(&stream->chan->nb_init_stream_left);
2003 }
2004
2005 lttng_ht_add_unique_u64(ht, &stream->node);
2006
2007 lttng_ht_add_unique_u64(consumer_data.stream_per_chan_id_ht,
2008 &stream->node_channel_id);
2009
2010 /*
2011 * Add stream to the stream_list_ht of the consumer data. No need to steal
2012 * the key since the HT does not use it and we allow to add redundant keys
2013 * into this table.
2014 */
2015 lttng_ht_add_u64(consumer_data.stream_list_ht, &stream->node_session_id);
2016
2017 rcu_read_unlock();
2018
2019 pthread_mutex_unlock(&stream->lock);
2020 pthread_mutex_unlock(&consumer_data.lock);
2021 return ret;
2022 }
2023
2024 /*
2025 * Delete data stream that are flagged for deletion (endpoint_status).
2026 */
2027 static void validate_endpoint_status_data_stream(void)
2028 {
2029 struct lttng_ht_iter iter;
2030 struct lttng_consumer_stream *stream;
2031
2032 DBG("Consumer delete flagged data stream");
2033
2034 rcu_read_lock();
2035 cds_lfht_for_each_entry(data_ht->ht, &iter.iter, stream, node.node) {
2036 /* Validate delete flag of the stream */
2037 if (stream->endpoint_status == CONSUMER_ENDPOINT_ACTIVE) {
2038 continue;
2039 }
2040 /* Delete it right now */
2041 consumer_del_stream(stream, data_ht);
2042 }
2043 rcu_read_unlock();
2044 }
2045
2046 /*
2047 * Delete metadata stream that are flagged for deletion (endpoint_status).
2048 */
2049 static void validate_endpoint_status_metadata_stream(
2050 struct lttng_poll_event *pollset)
2051 {
2052 struct lttng_ht_iter iter;
2053 struct lttng_consumer_stream *stream;
2054
2055 DBG("Consumer delete flagged metadata stream");
2056
2057 assert(pollset);
2058
2059 rcu_read_lock();
2060 cds_lfht_for_each_entry(metadata_ht->ht, &iter.iter, stream, node.node) {
2061 /* Validate delete flag of the stream */
2062 if (stream->endpoint_status == CONSUMER_ENDPOINT_ACTIVE) {
2063 continue;
2064 }
2065 /*
2066 * Remove from pollset so the metadata thread can continue without
2067 * blocking on a deleted stream.
2068 */
2069 lttng_poll_del(pollset, stream->wait_fd);
2070
2071 /* Delete it right now */
2072 consumer_del_metadata_stream(stream, metadata_ht);
2073 }
2074 rcu_read_unlock();
2075 }
2076
2077 /*
2078 * Thread polls on metadata file descriptor and write them on disk or on the
2079 * network.
2080 */
2081 void *consumer_thread_metadata_poll(void *data)
2082 {
2083 int ret, i, pollfd;
2084 uint32_t revents, nb_fd;
2085 struct lttng_consumer_stream *stream = NULL;
2086 struct lttng_ht_iter iter;
2087 struct lttng_ht_node_u64 *node;
2088 struct lttng_poll_event events;
2089 struct lttng_consumer_local_data *ctx = data;
2090 ssize_t len;
2091
2092 rcu_register_thread();
2093
2094 metadata_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
2095 if (!metadata_ht) {
2096 /* ENOMEM at this point. Better to bail out. */
2097 goto end_ht;
2098 }
2099
2100 DBG("Thread metadata poll started");
2101
2102 /* Size is set to 1 for the consumer_metadata pipe */
2103 ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC);
2104 if (ret < 0) {
2105 ERR("Poll set creation failed");
2106 goto end_poll;
2107 }
2108
2109 ret = lttng_poll_add(&events,
2110 lttng_pipe_get_readfd(ctx->consumer_metadata_pipe), LPOLLIN);
2111 if (ret < 0) {
2112 goto end;
2113 }
2114
2115 /* Main loop */
2116 DBG("Metadata main loop started");
2117
2118 while (1) {
2119 /* Only the metadata pipe is set */
2120 if (LTTNG_POLL_GETNB(&events) == 0 && consumer_quit == 1) {
2121 goto end;
2122 }
2123
2124 restart:
2125 DBG("Metadata poll wait with %d fd(s)", LTTNG_POLL_GETNB(&events));
2126 ret = lttng_poll_wait(&events, -1);
2127 DBG("Metadata event catched in thread");
2128 if (ret < 0) {
2129 if (errno == EINTR) {
2130 ERR("Poll EINTR catched");
2131 goto restart;
2132 }
2133 goto error;
2134 }
2135
2136 nb_fd = ret;
2137
2138 /* From here, the event is a metadata wait fd */
2139 for (i = 0; i < nb_fd; i++) {
2140 revents = LTTNG_POLL_GETEV(&events, i);
2141 pollfd = LTTNG_POLL_GETFD(&events, i);
2142
2143 /* Just don't waste time if no returned events for the fd */
2144 if (!revents) {
2145 continue;
2146 }
2147
2148 if (pollfd == lttng_pipe_get_readfd(ctx->consumer_metadata_pipe)) {
2149 if (revents & (LPOLLERR | LPOLLHUP )) {
2150 DBG("Metadata thread pipe hung up");
2151 /*
2152 * Remove the pipe from the poll set and continue the loop
2153 * since their might be data to consume.
2154 */
2155 lttng_poll_del(&events,
2156 lttng_pipe_get_readfd(ctx->consumer_metadata_pipe));
2157 lttng_pipe_read_close(ctx->consumer_metadata_pipe);
2158 continue;
2159 } else if (revents & LPOLLIN) {
2160 ssize_t pipe_len;
2161
2162 pipe_len = lttng_pipe_read(ctx->consumer_metadata_pipe,
2163 &stream, sizeof(stream));
2164 if (pipe_len < 0) {
2165 ERR("read metadata stream, ret: %ld", pipe_len);
2166 /*
2167 * Continue here to handle the rest of the streams.
2168 */
2169 continue;
2170 }
2171
2172 /* A NULL stream means that the state has changed. */
2173 if (stream == NULL) {
2174 /* Check for deleted streams. */
2175 validate_endpoint_status_metadata_stream(&events);
2176 goto restart;
2177 }
2178
2179 DBG("Adding metadata stream %d to poll set",
2180 stream->wait_fd);
2181
2182 ret = add_metadata_stream(stream, metadata_ht);
2183 if (ret) {
2184 ERR("Unable to add metadata stream");
2185 /* Stream was not setup properly. Continuing. */
2186 consumer_del_metadata_stream(stream, NULL);
2187 continue;
2188 }
2189
2190 /* Add metadata stream to the global poll events list */
2191 lttng_poll_add(&events, stream->wait_fd,
2192 LPOLLIN | LPOLLPRI);
2193 }
2194
2195 /* Handle other stream */
2196 continue;
2197 }
2198
2199 rcu_read_lock();
2200 {
2201 uint64_t tmp_id = (uint64_t) pollfd;
2202
2203 lttng_ht_lookup(metadata_ht, &tmp_id, &iter);
2204 }
2205 node = lttng_ht_iter_get_node_u64(&iter);
2206 assert(node);
2207
2208 stream = caa_container_of(node, struct lttng_consumer_stream,
2209 node);
2210
2211 /* Check for error event */
2212 if (revents & (LPOLLERR | LPOLLHUP)) {
2213 DBG("Metadata fd %d is hup|err.", pollfd);
2214 if (!stream->hangup_flush_done
2215 && (consumer_data.type == LTTNG_CONSUMER32_UST
2216 || consumer_data.type == LTTNG_CONSUMER64_UST)) {
2217 DBG("Attempting to flush and consume the UST buffers");
2218 lttng_ustconsumer_on_stream_hangup(stream);
2219
2220 /* We just flushed the stream now read it. */
2221 do {
2222 len = ctx->on_buffer_ready(stream, ctx);
2223 /*
2224 * We don't check the return value here since if we get
2225 * a negative len, it means an error occured thus we
2226 * simply remove it from the poll set and free the
2227 * stream.
2228 */
2229 } while (len > 0);
2230 }
2231
2232 lttng_poll_del(&events, stream->wait_fd);
2233 /*
2234 * This call update the channel states, closes file descriptors
2235 * and securely free the stream.
2236 */
2237 consumer_del_metadata_stream(stream, metadata_ht);
2238 } else if (revents & (LPOLLIN | LPOLLPRI)) {
2239 /* Get the data out of the metadata file descriptor */
2240 DBG("Metadata available on fd %d", pollfd);
2241 assert(stream->wait_fd == pollfd);
2242
2243 len = ctx->on_buffer_ready(stream, ctx);
2244 /* It's ok to have an unavailable sub-buffer */
2245 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2246 /* Clean up stream from consumer and free it. */
2247 lttng_poll_del(&events, stream->wait_fd);
2248 consumer_del_metadata_stream(stream, metadata_ht);
2249 } else if (len > 0) {
2250 stream->data_read = 1;
2251 }
2252 }
2253
2254 /* Release RCU lock for the stream looked up */
2255 rcu_read_unlock();
2256 }
2257 }
2258
2259 error:
2260 end:
2261 DBG("Metadata poll thread exiting");
2262
2263 lttng_poll_clean(&events);
2264 end_poll:
2265 destroy_stream_ht(metadata_ht);
2266 end_ht:
2267 rcu_unregister_thread();
2268 return NULL;
2269 }
2270
2271 /*
2272 * This thread polls the fds in the set to consume the data and write
2273 * it to tracefile if necessary.
2274 */
2275 void *consumer_thread_data_poll(void *data)
2276 {
2277 int num_rdy, num_hup, high_prio, ret, i;
2278 struct pollfd *pollfd = NULL;
2279 /* local view of the streams */
2280 struct lttng_consumer_stream **local_stream = NULL, *new_stream = NULL;
2281 /* local view of consumer_data.fds_count */
2282 int nb_fd = 0;
2283 struct lttng_consumer_local_data *ctx = data;
2284 ssize_t len;
2285
2286 rcu_register_thread();
2287
2288 data_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
2289 if (data_ht == NULL) {
2290 /* ENOMEM at this point. Better to bail out. */
2291 goto end;
2292 }
2293
2294 local_stream = zmalloc(sizeof(struct lttng_consumer_stream));
2295
2296 while (1) {
2297 high_prio = 0;
2298 num_hup = 0;
2299
2300 /*
2301 * the fds set has been updated, we need to update our
2302 * local array as well
2303 */
2304 pthread_mutex_lock(&consumer_data.lock);
2305 if (consumer_data.need_update) {
2306 free(pollfd);
2307 pollfd = NULL;
2308
2309 free(local_stream);
2310 local_stream = NULL;
2311
2312 /* allocate for all fds + 1 for the consumer_data_pipe */
2313 pollfd = zmalloc((consumer_data.stream_count + 1) * sizeof(struct pollfd));
2314 if (pollfd == NULL) {
2315 PERROR("pollfd malloc");
2316 pthread_mutex_unlock(&consumer_data.lock);
2317 goto end;
2318 }
2319
2320 /* allocate for all fds + 1 for the consumer_data_pipe */
2321 local_stream = zmalloc((consumer_data.stream_count + 1) *
2322 sizeof(struct lttng_consumer_stream *));
2323 if (local_stream == NULL) {
2324 PERROR("local_stream malloc");
2325 pthread_mutex_unlock(&consumer_data.lock);
2326 goto end;
2327 }
2328 ret = update_poll_array(ctx, &pollfd, local_stream,
2329 data_ht);
2330 if (ret < 0) {
2331 ERR("Error in allocating pollfd or local_outfds");
2332 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
2333 pthread_mutex_unlock(&consumer_data.lock);
2334 goto end;
2335 }
2336 nb_fd = ret;
2337 consumer_data.need_update = 0;
2338 }
2339 pthread_mutex_unlock(&consumer_data.lock);
2340
2341 /* No FDs and consumer_quit, consumer_cleanup the thread */
2342 if (nb_fd == 0 && consumer_quit == 1) {
2343 goto end;
2344 }
2345 /* poll on the array of fds */
2346 restart:
2347 DBG("polling on %d fd", nb_fd + 1);
2348 num_rdy = poll(pollfd, nb_fd + 1, -1);
2349 DBG("poll num_rdy : %d", num_rdy);
2350 if (num_rdy == -1) {
2351 /*
2352 * Restart interrupted system call.
2353 */
2354 if (errno == EINTR) {
2355 goto restart;
2356 }
2357 PERROR("Poll error");
2358 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
2359 goto end;
2360 } else if (num_rdy == 0) {
2361 DBG("Polling thread timed out");
2362 goto end;
2363 }
2364
2365 /*
2366 * If the consumer_data_pipe triggered poll go directly to the
2367 * beginning of the loop to update the array. We want to prioritize
2368 * array update over low-priority reads.
2369 */
2370 if (pollfd[nb_fd].revents & (POLLIN | POLLPRI)) {
2371 ssize_t pipe_readlen;
2372
2373 DBG("consumer_data_pipe wake up");
2374 pipe_readlen = lttng_pipe_read(ctx->consumer_data_pipe,
2375 &new_stream, sizeof(new_stream));
2376 if (pipe_readlen < 0) {
2377 ERR("Consumer data pipe ret %ld", pipe_readlen);
2378 /* Continue so we can at least handle the current stream(s). */
2379 continue;
2380 }
2381
2382 /*
2383 * If the stream is NULL, just ignore it. It's also possible that
2384 * the sessiond poll thread changed the consumer_quit state and is
2385 * waking us up to test it.
2386 */
2387 if (new_stream == NULL) {
2388 validate_endpoint_status_data_stream();
2389 continue;
2390 }
2391
2392 ret = add_stream(new_stream, data_ht);
2393 if (ret) {
2394 ERR("Consumer add stream %" PRIu64 " failed. Continuing",
2395 new_stream->key);
2396 /*
2397 * At this point, if the add_stream fails, it is not in the
2398 * hash table thus passing the NULL value here.
2399 */
2400 consumer_del_stream(new_stream, NULL);
2401 }
2402
2403 /* Continue to update the local streams and handle prio ones */
2404 continue;
2405 }
2406
2407 /* Take care of high priority channels first. */
2408 for (i = 0; i < nb_fd; i++) {
2409 if (local_stream[i] == NULL) {
2410 continue;
2411 }
2412 if (pollfd[i].revents & POLLPRI) {
2413 DBG("Urgent read on fd %d", pollfd[i].fd);
2414 high_prio = 1;
2415 len = ctx->on_buffer_ready(local_stream[i], ctx);
2416 /* it's ok to have an unavailable sub-buffer */
2417 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2418 /* Clean the stream and free it. */
2419 consumer_del_stream(local_stream[i], data_ht);
2420 local_stream[i] = NULL;
2421 } else if (len > 0) {
2422 local_stream[i]->data_read = 1;
2423 }
2424 }
2425 }
2426
2427 /*
2428 * If we read high prio channel in this loop, try again
2429 * for more high prio data.
2430 */
2431 if (high_prio) {
2432 continue;
2433 }
2434
2435 /* Take care of low priority channels. */
2436 for (i = 0; i < nb_fd; i++) {
2437 if (local_stream[i] == NULL) {
2438 continue;
2439 }
2440 if ((pollfd[i].revents & POLLIN) ||
2441 local_stream[i]->hangup_flush_done) {
2442 DBG("Normal read on fd %d", pollfd[i].fd);
2443 len = ctx->on_buffer_ready(local_stream[i], ctx);
2444 /* it's ok to have an unavailable sub-buffer */
2445 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2446 /* Clean the stream and free it. */
2447 consumer_del_stream(local_stream[i], data_ht);
2448 local_stream[i] = NULL;
2449 } else if (len > 0) {
2450 local_stream[i]->data_read = 1;
2451 }
2452 }
2453 }
2454
2455 /* Handle hangup and errors */
2456 for (i = 0; i < nb_fd; i++) {
2457 if (local_stream[i] == NULL) {
2458 continue;
2459 }
2460 if (!local_stream[i]->hangup_flush_done
2461 && (pollfd[i].revents & (POLLHUP | POLLERR | POLLNVAL))
2462 && (consumer_data.type == LTTNG_CONSUMER32_UST
2463 || consumer_data.type == LTTNG_CONSUMER64_UST)) {
2464 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2465 pollfd[i].fd);
2466 lttng_ustconsumer_on_stream_hangup(local_stream[i]);
2467 /* Attempt read again, for the data we just flushed. */
2468 local_stream[i]->data_read = 1;
2469 }
2470 /*
2471 * If the poll flag is HUP/ERR/NVAL and we have
2472 * read no data in this pass, we can remove the
2473 * stream from its hash table.
2474 */
2475 if ((pollfd[i].revents & POLLHUP)) {
2476 DBG("Polling fd %d tells it has hung up.", 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 } else if (pollfd[i].revents & POLLERR) {
2483 ERR("Error returned in polling fd %d.", pollfd[i].fd);
2484 if (!local_stream[i]->data_read) {
2485 consumer_del_stream(local_stream[i], data_ht);
2486 local_stream[i] = NULL;
2487 num_hup++;
2488 }
2489 } else if (pollfd[i].revents & POLLNVAL) {
2490 ERR("Polling fd %d tells fd is not open.", pollfd[i].fd);
2491 if (!local_stream[i]->data_read) {
2492 consumer_del_stream(local_stream[i], data_ht);
2493 local_stream[i] = NULL;
2494 num_hup++;
2495 }
2496 }
2497 if (local_stream[i] != NULL) {
2498 local_stream[i]->data_read = 0;
2499 }
2500 }
2501 }
2502 end:
2503 DBG("polling thread exiting");
2504 free(pollfd);
2505 free(local_stream);
2506
2507 /*
2508 * Close the write side of the pipe so epoll_wait() in
2509 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2510 * read side of the pipe. If we close them both, epoll_wait strangely does
2511 * not return and could create a endless wait period if the pipe is the
2512 * only tracked fd in the poll set. The thread will take care of closing
2513 * the read side.
2514 */
2515 (void) lttng_pipe_write_close(ctx->consumer_metadata_pipe);
2516
2517 destroy_data_stream_ht(data_ht);
2518
2519 rcu_unregister_thread();
2520 return NULL;
2521 }
2522
2523 /*
2524 * Close wake-up end of each stream belonging to the channel. This will
2525 * allow the poll() on the stream read-side to detect when the
2526 * write-side (application) finally closes them.
2527 */
2528 static
2529 void consumer_close_channel_streams(struct lttng_consumer_channel *channel)
2530 {
2531 struct lttng_ht *ht;
2532 struct lttng_consumer_stream *stream;
2533 struct lttng_ht_iter iter;
2534
2535 ht = consumer_data.stream_per_chan_id_ht;
2536
2537 rcu_read_lock();
2538 cds_lfht_for_each_entry_duplicate(ht->ht,
2539 ht->hash_fct(&channel->key, lttng_ht_seed),
2540 ht->match_fct, &channel->key,
2541 &iter.iter, stream, node_channel_id.node) {
2542 /*
2543 * Protect against teardown with mutex.
2544 */
2545 pthread_mutex_lock(&stream->lock);
2546 if (cds_lfht_is_node_deleted(&stream->node.node)) {
2547 goto next;
2548 }
2549 switch (consumer_data.type) {
2550 case LTTNG_CONSUMER_KERNEL:
2551 break;
2552 case LTTNG_CONSUMER32_UST:
2553 case LTTNG_CONSUMER64_UST:
2554 /*
2555 * Note: a mutex is taken internally within
2556 * liblttng-ust-ctl to protect timer wakeup_fd
2557 * use from concurrent close.
2558 */
2559 lttng_ustconsumer_close_stream_wakeup(stream);
2560 break;
2561 default:
2562 ERR("Unknown consumer_data type");
2563 assert(0);
2564 }
2565 next:
2566 pthread_mutex_unlock(&stream->lock);
2567 }
2568 rcu_read_unlock();
2569 }
2570
2571 static void destroy_channel_ht(struct lttng_ht *ht)
2572 {
2573 struct lttng_ht_iter iter;
2574 struct lttng_consumer_channel *channel;
2575 int ret;
2576
2577 if (ht == NULL) {
2578 return;
2579 }
2580
2581 rcu_read_lock();
2582 cds_lfht_for_each_entry(ht->ht, &iter.iter, channel, wait_fd_node.node) {
2583 ret = lttng_ht_del(ht, &iter);
2584 assert(ret != 0);
2585 }
2586 rcu_read_unlock();
2587
2588 lttng_ht_destroy(ht);
2589 }
2590
2591 /*
2592 * This thread polls the channel fds to detect when they are being
2593 * closed. It closes all related streams if the channel is detected as
2594 * closed. It is currently only used as a shim layer for UST because the
2595 * consumerd needs to keep the per-stream wakeup end of pipes open for
2596 * periodical flush.
2597 */
2598 void *consumer_thread_channel_poll(void *data)
2599 {
2600 int ret, i, pollfd;
2601 uint32_t revents, nb_fd;
2602 struct lttng_consumer_channel *chan = NULL;
2603 struct lttng_ht_iter iter;
2604 struct lttng_ht_node_u64 *node;
2605 struct lttng_poll_event events;
2606 struct lttng_consumer_local_data *ctx = data;
2607 struct lttng_ht *channel_ht;
2608
2609 rcu_register_thread();
2610
2611 channel_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
2612 if (!channel_ht) {
2613 /* ENOMEM at this point. Better to bail out. */
2614 goto end_ht;
2615 }
2616
2617 DBG("Thread channel poll started");
2618
2619 /* Size is set to 1 for the consumer_channel pipe */
2620 ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC);
2621 if (ret < 0) {
2622 ERR("Poll set creation failed");
2623 goto end_poll;
2624 }
2625
2626 ret = lttng_poll_add(&events, ctx->consumer_channel_pipe[0], LPOLLIN);
2627 if (ret < 0) {
2628 goto end;
2629 }
2630
2631 /* Main loop */
2632 DBG("Channel main loop started");
2633
2634 while (1) {
2635 /* Only the channel pipe is set */
2636 if (LTTNG_POLL_GETNB(&events) == 0 && consumer_quit == 1) {
2637 goto end;
2638 }
2639
2640 restart:
2641 DBG("Channel poll wait with %d fd(s)", LTTNG_POLL_GETNB(&events));
2642 ret = lttng_poll_wait(&events, -1);
2643 DBG("Channel event catched in thread");
2644 if (ret < 0) {
2645 if (errno == EINTR) {
2646 ERR("Poll EINTR catched");
2647 goto restart;
2648 }
2649 goto end;
2650 }
2651
2652 nb_fd = ret;
2653
2654 /* From here, the event is a channel wait fd */
2655 for (i = 0; i < nb_fd; i++) {
2656 revents = LTTNG_POLL_GETEV(&events, i);
2657 pollfd = LTTNG_POLL_GETFD(&events, i);
2658
2659 /* Just don't waste time if no returned events for the fd */
2660 if (!revents) {
2661 continue;
2662 }
2663 if (pollfd == ctx->consumer_channel_pipe[0]) {
2664 if (revents & (LPOLLERR | LPOLLHUP)) {
2665 DBG("Channel thread pipe hung up");
2666 /*
2667 * Remove the pipe from the poll set and continue the loop
2668 * since their might be data to consume.
2669 */
2670 lttng_poll_del(&events, ctx->consumer_channel_pipe[0]);
2671 continue;
2672 } else if (revents & LPOLLIN) {
2673 enum consumer_channel_action action;
2674 uint64_t key;
2675
2676 ret = read_channel_pipe(ctx, &chan, &key, &action);
2677 if (ret <= 0) {
2678 ERR("Error reading channel pipe");
2679 continue;
2680 }
2681
2682 switch (action) {
2683 case CONSUMER_CHANNEL_ADD:
2684 DBG("Adding channel %d to poll set",
2685 chan->wait_fd);
2686
2687 lttng_ht_node_init_u64(&chan->wait_fd_node,
2688 chan->wait_fd);
2689 rcu_read_lock();
2690 lttng_ht_add_unique_u64(channel_ht,
2691 &chan->wait_fd_node);
2692 rcu_read_unlock();
2693 /* Add channel to the global poll events list */
2694 lttng_poll_add(&events, chan->wait_fd,
2695 LPOLLIN | LPOLLPRI);
2696 break;
2697 case CONSUMER_CHANNEL_DEL:
2698 {
2699 struct lttng_consumer_stream *stream, *stmp;
2700
2701 rcu_read_lock();
2702 chan = consumer_find_channel(key);
2703 if (!chan) {
2704 rcu_read_unlock();
2705 ERR("UST consumer get channel key %" PRIu64 " not found for del channel", key);
2706 break;
2707 }
2708 lttng_poll_del(&events, chan->wait_fd);
2709 iter.iter.node = &chan->wait_fd_node.node;
2710 ret = lttng_ht_del(channel_ht, &iter);
2711 assert(ret == 0);
2712 consumer_close_channel_streams(chan);
2713
2714 switch (consumer_data.type) {
2715 case LTTNG_CONSUMER_KERNEL:
2716 break;
2717 case LTTNG_CONSUMER32_UST:
2718 case LTTNG_CONSUMER64_UST:
2719 /* Delete streams that might have been left in the stream list. */
2720 cds_list_for_each_entry_safe(stream, stmp, &chan->streams.head,
2721 send_node) {
2722 cds_list_del(&stream->send_node);
2723 lttng_ustconsumer_del_stream(stream);
2724 uatomic_sub(&stream->chan->refcount, 1);
2725 assert(&chan->refcount);
2726 free(stream);
2727 }
2728 break;
2729 default:
2730 ERR("Unknown consumer_data type");
2731 assert(0);
2732 }
2733
2734 /*
2735 * Release our own refcount. Force channel deletion even if
2736 * streams were not initialized.
2737 */
2738 if (!uatomic_sub_return(&chan->refcount, 1)) {
2739 consumer_del_channel(chan);
2740 }
2741 rcu_read_unlock();
2742 goto restart;
2743 }
2744 case CONSUMER_CHANNEL_QUIT:
2745 /*
2746 * Remove the pipe from the poll set and continue the loop
2747 * since their might be data to consume.
2748 */
2749 lttng_poll_del(&events, ctx->consumer_channel_pipe[0]);
2750 continue;
2751 default:
2752 ERR("Unknown action");
2753 break;
2754 }
2755 }
2756
2757 /* Handle other stream */
2758 continue;
2759 }
2760
2761 rcu_read_lock();
2762 {
2763 uint64_t tmp_id = (uint64_t) pollfd;
2764
2765 lttng_ht_lookup(channel_ht, &tmp_id, &iter);
2766 }
2767 node = lttng_ht_iter_get_node_u64(&iter);
2768 assert(node);
2769
2770 chan = caa_container_of(node, struct lttng_consumer_channel,
2771 wait_fd_node);
2772
2773 /* Check for error event */
2774 if (revents & (LPOLLERR | LPOLLHUP)) {
2775 DBG("Channel fd %d is hup|err.", pollfd);
2776
2777 lttng_poll_del(&events, chan->wait_fd);
2778 ret = lttng_ht_del(channel_ht, &iter);
2779 assert(ret == 0);
2780 consumer_close_channel_streams(chan);
2781
2782 /* Release our own refcount */
2783 if (!uatomic_sub_return(&chan->refcount, 1)
2784 && !uatomic_read(&chan->nb_init_stream_left)) {
2785 consumer_del_channel(chan);
2786 }
2787 }
2788
2789 /* Release RCU lock for the channel looked up */
2790 rcu_read_unlock();
2791 }
2792 }
2793
2794 end:
2795 lttng_poll_clean(&events);
2796 end_poll:
2797 destroy_channel_ht(channel_ht);
2798 end_ht:
2799 DBG("Channel poll thread exiting");
2800 rcu_unregister_thread();
2801 return NULL;
2802 }
2803
2804 static int set_metadata_socket(struct lttng_consumer_local_data *ctx,
2805 struct pollfd *sockpoll, int client_socket)
2806 {
2807 int ret;
2808
2809 assert(ctx);
2810 assert(sockpoll);
2811
2812 if (lttng_consumer_poll_socket(sockpoll) < 0) {
2813 ret = -1;
2814 goto error;
2815 }
2816 DBG("Metadata connection on client_socket");
2817
2818 /* Blocking call, waiting for transmission */
2819 ctx->consumer_metadata_socket = lttcomm_accept_unix_sock(client_socket);
2820 if (ctx->consumer_metadata_socket < 0) {
2821 WARN("On accept metadata");
2822 ret = -1;
2823 goto error;
2824 }
2825 ret = 0;
2826
2827 error:
2828 return ret;
2829 }
2830
2831 /*
2832 * This thread listens on the consumerd socket and receives the file
2833 * descriptors from the session daemon.
2834 */
2835 void *consumer_thread_sessiond_poll(void *data)
2836 {
2837 int sock = -1, client_socket, ret;
2838 /*
2839 * structure to poll for incoming data on communication socket avoids
2840 * making blocking sockets.
2841 */
2842 struct pollfd consumer_sockpoll[2];
2843 struct lttng_consumer_local_data *ctx = data;
2844
2845 rcu_register_thread();
2846
2847 DBG("Creating command socket %s", ctx->consumer_command_sock_path);
2848 unlink(ctx->consumer_command_sock_path);
2849 client_socket = lttcomm_create_unix_sock(ctx->consumer_command_sock_path);
2850 if (client_socket < 0) {
2851 ERR("Cannot create command socket");
2852 goto end;
2853 }
2854
2855 ret = lttcomm_listen_unix_sock(client_socket);
2856 if (ret < 0) {
2857 goto end;
2858 }
2859
2860 DBG("Sending ready command to lttng-sessiond");
2861 ret = lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY);
2862 /* return < 0 on error, but == 0 is not fatal */
2863 if (ret < 0) {
2864 ERR("Error sending ready command to lttng-sessiond");
2865 goto end;
2866 }
2867
2868 /* prepare the FDs to poll : to client socket and the should_quit pipe */
2869 consumer_sockpoll[0].fd = ctx->consumer_should_quit[0];
2870 consumer_sockpoll[0].events = POLLIN | POLLPRI;
2871 consumer_sockpoll[1].fd = client_socket;
2872 consumer_sockpoll[1].events = POLLIN | POLLPRI;
2873
2874 if (lttng_consumer_poll_socket(consumer_sockpoll) < 0) {
2875 goto end;
2876 }
2877 DBG("Connection on client_socket");
2878
2879 /* Blocking call, waiting for transmission */
2880 sock = lttcomm_accept_unix_sock(client_socket);
2881 if (sock < 0) {
2882 WARN("On accept");
2883 goto end;
2884 }
2885
2886 /*
2887 * Setup metadata socket which is the second socket connection on the
2888 * command unix socket.
2889 */
2890 ret = set_metadata_socket(ctx, consumer_sockpoll, client_socket);
2891 if (ret < 0) {
2892 goto end;
2893 }
2894
2895 /* This socket is not useful anymore. */
2896 ret = close(client_socket);
2897 if (ret < 0) {
2898 PERROR("close client_socket");
2899 }
2900 client_socket = -1;
2901
2902 /* update the polling structure to poll on the established socket */
2903 consumer_sockpoll[1].fd = sock;
2904 consumer_sockpoll[1].events = POLLIN | POLLPRI;
2905
2906 while (1) {
2907 if (lttng_consumer_poll_socket(consumer_sockpoll) < 0) {
2908 goto end;
2909 }
2910 DBG("Incoming command on sock");
2911 ret = lttng_consumer_recv_cmd(ctx, sock, consumer_sockpoll);
2912 if (ret == -ENOENT) {
2913 DBG("Received STOP command");
2914 goto end;
2915 }
2916 if (ret <= 0) {
2917 /*
2918 * This could simply be a session daemon quitting. Don't output
2919 * ERR() here.
2920 */
2921 DBG("Communication interrupted on command socket");
2922 goto end;
2923 }
2924 if (consumer_quit) {
2925 DBG("consumer_thread_receive_fds received quit from signal");
2926 goto end;
2927 }
2928 DBG("received command on sock");
2929 }
2930 end:
2931 DBG("Consumer thread sessiond poll exiting");
2932
2933 /*
2934 * Close metadata streams since the producer is the session daemon which
2935 * just died.
2936 *
2937 * NOTE: for now, this only applies to the UST tracer.
2938 */
2939 lttng_consumer_close_metadata();
2940
2941 /*
2942 * when all fds have hung up, the polling thread
2943 * can exit cleanly
2944 */
2945 consumer_quit = 1;
2946
2947 /*
2948 * Notify the data poll thread to poll back again and test the
2949 * consumer_quit state that we just set so to quit gracefully.
2950 */
2951 notify_thread_lttng_pipe(ctx->consumer_data_pipe);
2952
2953 notify_channel_pipe(ctx, NULL, -1, CONSUMER_CHANNEL_QUIT);
2954
2955 /* Cleaning up possibly open sockets. */
2956 if (sock >= 0) {
2957 ret = close(sock);
2958 if (ret < 0) {
2959 PERROR("close sock sessiond poll");
2960 }
2961 }
2962 if (client_socket >= 0) {
2963 ret = close(client_socket);
2964 if (ret < 0) {
2965 PERROR("close client_socket sessiond poll");
2966 }
2967 }
2968
2969 rcu_unregister_thread();
2970 return NULL;
2971 }
2972
2973 ssize_t lttng_consumer_read_subbuffer(struct lttng_consumer_stream *stream,
2974 struct lttng_consumer_local_data *ctx)
2975 {
2976 ssize_t ret;
2977
2978 pthread_mutex_lock(&stream->lock);
2979
2980 switch (consumer_data.type) {
2981 case LTTNG_CONSUMER_KERNEL:
2982 ret = lttng_kconsumer_read_subbuffer(stream, ctx);
2983 break;
2984 case LTTNG_CONSUMER32_UST:
2985 case LTTNG_CONSUMER64_UST:
2986 ret = lttng_ustconsumer_read_subbuffer(stream, ctx);
2987 break;
2988 default:
2989 ERR("Unknown consumer_data type");
2990 assert(0);
2991 ret = -ENOSYS;
2992 break;
2993 }
2994
2995 pthread_mutex_unlock(&stream->lock);
2996 return ret;
2997 }
2998
2999 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream *stream)
3000 {
3001 switch (consumer_data.type) {
3002 case LTTNG_CONSUMER_KERNEL:
3003 return lttng_kconsumer_on_recv_stream(stream);
3004 case LTTNG_CONSUMER32_UST:
3005 case LTTNG_CONSUMER64_UST:
3006 return lttng_ustconsumer_on_recv_stream(stream);
3007 default:
3008 ERR("Unknown consumer_data type");
3009 assert(0);
3010 return -ENOSYS;
3011 }
3012 }
3013
3014 /*
3015 * Allocate and set consumer data hash tables.
3016 */
3017 void lttng_consumer_init(void)
3018 {
3019 consumer_data.channel_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
3020 consumer_data.relayd_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
3021 consumer_data.stream_list_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
3022 consumer_data.stream_per_chan_id_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
3023 }
3024
3025 /*
3026 * Process the ADD_RELAYD command receive by a consumer.
3027 *
3028 * This will create a relayd socket pair and add it to the relayd hash table.
3029 * The caller MUST acquire a RCU read side lock before calling it.
3030 */
3031 int consumer_add_relayd_socket(uint64_t net_seq_idx, int sock_type,
3032 struct lttng_consumer_local_data *ctx, int sock,
3033 struct pollfd *consumer_sockpoll,
3034 struct lttcomm_relayd_sock *relayd_sock, unsigned int sessiond_id)
3035 {
3036 int fd = -1, ret = -1, relayd_created = 0;
3037 enum lttng_error_code ret_code = LTTNG_OK;
3038 struct consumer_relayd_sock_pair *relayd = NULL;
3039
3040 assert(ctx);
3041 assert(relayd_sock);
3042
3043 DBG("Consumer adding relayd socket (idx: %" PRIu64 ")", net_seq_idx);
3044
3045 /* Get relayd reference if exists. */
3046 relayd = consumer_find_relayd(net_seq_idx);
3047 if (relayd == NULL) {
3048 assert(sock_type == LTTNG_STREAM_CONTROL);
3049 /* Not found. Allocate one. */
3050 relayd = consumer_allocate_relayd_sock_pair(net_seq_idx);
3051 if (relayd == NULL) {
3052 ret = -ENOMEM;
3053 ret_code = LTTCOMM_CONSUMERD_ENOMEM;
3054 goto error;
3055 } else {
3056 relayd->sessiond_session_id = (uint64_t) sessiond_id;
3057 relayd_created = 1;
3058 }
3059
3060 /*
3061 * This code path MUST continue to the consumer send status message to
3062 * we can notify the session daemon and continue our work without
3063 * killing everything.
3064 */
3065 } else {
3066 /*
3067 * relayd key should never be found for control socket.
3068 */
3069 assert(sock_type != LTTNG_STREAM_CONTROL);
3070 }
3071
3072 /* First send a status message before receiving the fds. */
3073 ret = consumer_send_status_msg(sock, LTTNG_OK);
3074 if (ret < 0) {
3075 /* Somehow, the session daemon is not responding anymore. */
3076 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_FATAL);
3077 goto error_nosignal;
3078 }
3079
3080 /* Poll on consumer socket. */
3081 if (lttng_consumer_poll_socket(consumer_sockpoll) < 0) {
3082 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
3083 ret = -EINTR;
3084 goto error_nosignal;
3085 }
3086
3087 /* Get relayd socket from session daemon */
3088 ret = lttcomm_recv_fds_unix_sock(sock, &fd, 1);
3089 if (ret != sizeof(fd)) {
3090 ret = -1;
3091 fd = -1; /* Just in case it gets set with an invalid value. */
3092
3093 /*
3094 * Failing to receive FDs might indicate a major problem such as
3095 * reaching a fd limit during the receive where the kernel returns a
3096 * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we
3097 * don't take any chances and stop everything.
3098 *
3099 * XXX: Feature request #558 will fix that and avoid this possible
3100 * issue when reaching the fd limit.
3101 */
3102 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_ERROR_RECV_FD);
3103 ret_code = LTTCOMM_CONSUMERD_ERROR_RECV_FD;
3104 goto error;
3105 }
3106
3107 /* Copy socket information and received FD */
3108 switch (sock_type) {
3109 case LTTNG_STREAM_CONTROL:
3110 /* Copy received lttcomm socket */
3111 lttcomm_copy_sock(&relayd->control_sock.sock, &relayd_sock->sock);
3112 ret = lttcomm_create_sock(&relayd->control_sock.sock);
3113 /* Handle create_sock error. */
3114 if (ret < 0) {
3115 ret_code = LTTCOMM_CONSUMERD_ENOMEM;
3116 goto error;
3117 }
3118 /*
3119 * Close the socket created internally by
3120 * lttcomm_create_sock, so we can replace it by the one
3121 * received from sessiond.
3122 */
3123 if (close(relayd->control_sock.sock.fd)) {
3124 PERROR("close");
3125 }
3126
3127 /* Assign new file descriptor */
3128 relayd->control_sock.sock.fd = fd;
3129 fd = -1; /* For error path */
3130 /* Assign version values. */
3131 relayd->control_sock.major = relayd_sock->major;
3132 relayd->control_sock.minor = relayd_sock->minor;
3133
3134 /*
3135 * Create a session on the relayd and store the returned id. Lock the
3136 * control socket mutex if the relayd was NOT created before.
3137 */
3138 if (!relayd_created) {
3139 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
3140 }
3141 ret = relayd_create_session(&relayd->control_sock,
3142 &relayd->relayd_session_id);
3143 if (!relayd_created) {
3144 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
3145 }
3146 if (ret < 0) {
3147 /*
3148 * Close all sockets of a relayd object. It will be freed if it was
3149 * created at the error code path or else it will be garbage
3150 * collect.
3151 */
3152 (void) relayd_close(&relayd->control_sock);
3153 (void) relayd_close(&relayd->data_sock);
3154 ret_code = LTTCOMM_CONSUMERD_RELAYD_FAIL;
3155 goto error;
3156 }
3157
3158 break;
3159 case LTTNG_STREAM_DATA:
3160 /* Copy received lttcomm socket */
3161 lttcomm_copy_sock(&relayd->data_sock.sock, &relayd_sock->sock);
3162 ret = lttcomm_create_sock(&relayd->data_sock.sock);
3163 /* Handle create_sock error. */
3164 if (ret < 0) {
3165 ret_code = LTTCOMM_CONSUMERD_ENOMEM;
3166 goto error;
3167 }
3168 /*
3169 * Close the socket created internally by
3170 * lttcomm_create_sock, so we can replace it by the one
3171 * received from sessiond.
3172 */
3173 if (close(relayd->data_sock.sock.fd)) {
3174 PERROR("close");
3175 }
3176
3177 /* Assign new file descriptor */
3178 relayd->data_sock.sock.fd = fd;
3179 fd = -1; /* for eventual error paths */
3180 /* Assign version values. */
3181 relayd->data_sock.major = relayd_sock->major;
3182 relayd->data_sock.minor = relayd_sock->minor;
3183 break;
3184 default:
3185 ERR("Unknown relayd socket type (%d)", sock_type);
3186 ret = -1;
3187 ret_code = LTTCOMM_CONSUMERD_FATAL;
3188 goto error;
3189 }
3190
3191 DBG("Consumer %s socket created successfully with net idx %" PRIu64 " (fd: %d)",
3192 sock_type == LTTNG_STREAM_CONTROL ? "control" : "data",
3193 relayd->net_seq_idx, fd);
3194
3195 /* We successfully added the socket. Send status back. */
3196 ret = consumer_send_status_msg(sock, ret_code);
3197 if (ret < 0) {
3198 /* Somehow, the session daemon is not responding anymore. */
3199 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_FATAL);
3200 goto error_nosignal;
3201 }
3202
3203 /*
3204 * Add relayd socket pair to consumer data hashtable. If object already
3205 * exists or on error, the function gracefully returns.
3206 */
3207 add_relayd(relayd);
3208
3209 /* All good! */
3210 return 0;
3211
3212 error:
3213 if (consumer_send_status_msg(sock, ret_code) < 0) {
3214 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_FATAL);
3215 }
3216
3217 error_nosignal:
3218 /* Close received socket if valid. */
3219 if (fd >= 0) {
3220 if (close(fd)) {
3221 PERROR("close received socket");
3222 }
3223 }
3224
3225 if (relayd_created) {
3226 free(relayd);
3227 }
3228
3229 return ret;
3230 }
3231
3232 /*
3233 * Try to lock the stream mutex.
3234 *
3235 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
3236 */
3237 static int stream_try_lock(struct lttng_consumer_stream *stream)
3238 {
3239 int ret;
3240
3241 assert(stream);
3242
3243 /*
3244 * Try to lock the stream mutex. On failure, we know that the stream is
3245 * being used else where hence there is data still being extracted.
3246 */
3247 ret = pthread_mutex_trylock(&stream->lock);
3248 if (ret) {
3249 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
3250 ret = 0;
3251 goto end;
3252 }
3253
3254 ret = 1;
3255
3256 end:
3257 return ret;
3258 }
3259
3260 /*
3261 * Search for a relayd associated to the session id and return the reference.
3262 *
3263 * A rcu read side lock MUST be acquire before calling this function and locked
3264 * until the relayd object is no longer necessary.
3265 */
3266 static struct consumer_relayd_sock_pair *find_relayd_by_session_id(uint64_t id)
3267 {
3268 struct lttng_ht_iter iter;
3269 struct consumer_relayd_sock_pair *relayd = NULL;
3270
3271 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3272 cds_lfht_for_each_entry(consumer_data.relayd_ht->ht, &iter.iter, relayd,
3273 node.node) {
3274 /*
3275 * Check by sessiond id which is unique here where the relayd session
3276 * id might not be when having multiple relayd.
3277 */
3278 if (relayd->sessiond_session_id == id) {
3279 /* Found the relayd. There can be only one per id. */
3280 goto found;
3281 }
3282 }
3283
3284 return NULL;
3285
3286 found:
3287 return relayd;
3288 }
3289
3290 /*
3291 * Check if for a given session id there is still data needed to be extract
3292 * from the buffers.
3293 *
3294 * Return 1 if data is pending or else 0 meaning ready to be read.
3295 */
3296 int consumer_data_pending(uint64_t id)
3297 {
3298 int ret;
3299 struct lttng_ht_iter iter;
3300 struct lttng_ht *ht;
3301 struct lttng_consumer_stream *stream;
3302 struct consumer_relayd_sock_pair *relayd = NULL;
3303 int (*data_pending)(struct lttng_consumer_stream *);
3304
3305 DBG("Consumer data pending command on session id %" PRIu64, id);
3306
3307 rcu_read_lock();
3308 pthread_mutex_lock(&consumer_data.lock);
3309
3310 switch (consumer_data.type) {
3311 case LTTNG_CONSUMER_KERNEL:
3312 data_pending = lttng_kconsumer_data_pending;
3313 break;
3314 case LTTNG_CONSUMER32_UST:
3315 case LTTNG_CONSUMER64_UST:
3316 data_pending = lttng_ustconsumer_data_pending;
3317 break;
3318 default:
3319 ERR("Unknown consumer data type");
3320 assert(0);
3321 }
3322
3323 /* Ease our life a bit */
3324 ht = consumer_data.stream_list_ht;
3325
3326 relayd = find_relayd_by_session_id(id);
3327 if (relayd) {
3328 /* Send init command for data pending. */
3329 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
3330 ret = relayd_begin_data_pending(&relayd->control_sock,
3331 relayd->relayd_session_id);
3332 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
3333 if (ret < 0) {
3334 /* Communication error thus the relayd so no data pending. */
3335 goto data_not_pending;
3336 }
3337 }
3338
3339 cds_lfht_for_each_entry_duplicate(ht->ht,
3340 ht->hash_fct(&id, lttng_ht_seed),
3341 ht->match_fct, &id,
3342 &iter.iter, stream, node_session_id.node) {
3343 /* If this call fails, the stream is being used hence data pending. */
3344 ret = stream_try_lock(stream);
3345 if (!ret) {
3346 goto data_pending;
3347 }
3348
3349 /*
3350 * A removed node from the hash table indicates that the stream has
3351 * been deleted thus having a guarantee that the buffers are closed
3352 * on the consumer side. However, data can still be transmitted
3353 * over the network so don't skip the relayd check.
3354 */
3355 ret = cds_lfht_is_node_deleted(&stream->node.node);
3356 if (!ret) {
3357 /* Check the stream if there is data in the buffers. */
3358 ret = data_pending(stream);
3359 if (ret == 1) {
3360 pthread_mutex_unlock(&stream->lock);
3361 goto data_pending;
3362 }
3363 }
3364
3365 /* Relayd check */
3366 if (relayd) {
3367 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
3368 if (stream->metadata_flag) {
3369 ret = relayd_quiescent_control(&relayd->control_sock,
3370 stream->relayd_stream_id);
3371 } else {
3372 ret = relayd_data_pending(&relayd->control_sock,
3373 stream->relayd_stream_id,
3374 stream->next_net_seq_num - 1);
3375 }
3376 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
3377 if (ret == 1) {
3378 pthread_mutex_unlock(&stream->lock);
3379 goto data_pending;
3380 }
3381 }
3382 pthread_mutex_unlock(&stream->lock);
3383 }
3384
3385 if (relayd) {
3386 unsigned int is_data_inflight = 0;
3387
3388 /* Send init command for data pending. */
3389 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
3390 ret = relayd_end_data_pending(&relayd->control_sock,
3391 relayd->relayd_session_id, &is_data_inflight);
3392 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
3393 if (ret < 0) {
3394 goto data_not_pending;
3395 }
3396 if (is_data_inflight) {
3397 goto data_pending;
3398 }
3399 }
3400
3401 /*
3402 * Finding _no_ node in the hash table and no inflight data means that the
3403 * stream(s) have been removed thus data is guaranteed to be available for
3404 * analysis from the trace files.
3405 */
3406
3407 data_not_pending:
3408 /* Data is available to be read by a viewer. */
3409 pthread_mutex_unlock(&consumer_data.lock);
3410 rcu_read_unlock();
3411 return 0;
3412
3413 data_pending:
3414 /* Data is still being extracted from buffers. */
3415 pthread_mutex_unlock(&consumer_data.lock);
3416 rcu_read_unlock();
3417 return 1;
3418 }
3419
3420 /*
3421 * Send a ret code status message to the sessiond daemon.
3422 *
3423 * Return the sendmsg() return value.
3424 */
3425 int consumer_send_status_msg(int sock, int ret_code)
3426 {
3427 struct lttcomm_consumer_status_msg msg;
3428
3429 msg.ret_code = ret_code;
3430
3431 return lttcomm_send_unix_sock(sock, &msg, sizeof(msg));
3432 }
3433
3434 /*
3435 * Send a channel status message to the sessiond daemon.
3436 *
3437 * Return the sendmsg() return value.
3438 */
3439 int consumer_send_status_channel(int sock,
3440 struct lttng_consumer_channel *channel)
3441 {
3442 struct lttcomm_consumer_status_channel msg;
3443
3444 assert(sock >= 0);
3445
3446 if (!channel) {
3447 msg.ret_code = -LTTNG_ERR_UST_CHAN_FAIL;
3448 } else {
3449 msg.ret_code = LTTNG_OK;
3450 msg.key = channel->key;
3451 msg.stream_count = channel->streams.count;
3452 }
3453
3454 return lttcomm_send_unix_sock(sock, &msg, sizeof(msg));
3455 }
This page took 0.120691 seconds and 4 git commands to generate.