044a504cec512443039e15eac186df07e841f896
[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
45 struct lttng_consumer_global_data consumer_data = {
46 .stream_count = 0,
47 .need_update = 1,
48 .type = LTTNG_CONSUMER_UNKNOWN,
49 };
50
51 enum consumer_channel_action {
52 CONSUMER_CHANNEL_ADD,
53 CONSUMER_CHANNEL_DEL,
54 CONSUMER_CHANNEL_QUIT,
55 };
56
57 struct consumer_channel_msg {
58 enum consumer_channel_action action;
59 struct lttng_consumer_channel *chan; /* add */
60 uint64_t key; /* del */
61 };
62
63 /*
64 * Flag to inform the polling thread to quit when all fd hung up. Updated by
65 * the consumer_thread_receive_fds when it notices that all fds has hung up.
66 * Also updated by the signal handler (consumer_should_exit()). Read by the
67 * polling threads.
68 */
69 volatile int consumer_quit;
70
71 /*
72 * Global hash table containing respectively metadata and data streams. The
73 * stream element in this ht should only be updated by the metadata poll thread
74 * for the metadata and the data poll thread for the data.
75 */
76 static struct lttng_ht *metadata_ht;
77 static struct lttng_ht *data_ht;
78
79 /*
80 * Notify a thread lttng pipe to poll back again. This usually means that some
81 * global state has changed so we just send back the thread in a poll wait
82 * call.
83 */
84 static void notify_thread_lttng_pipe(struct lttng_pipe *pipe)
85 {
86 struct lttng_consumer_stream *null_stream = NULL;
87
88 assert(pipe);
89
90 (void) lttng_pipe_write(pipe, &null_stream, sizeof(null_stream));
91 }
92
93 static void notify_channel_pipe(struct lttng_consumer_local_data *ctx,
94 struct lttng_consumer_channel *chan,
95 uint64_t key,
96 enum consumer_channel_action action)
97 {
98 struct consumer_channel_msg msg;
99 int ret;
100
101 memset(&msg, 0, sizeof(msg));
102
103 msg.action = action;
104 msg.chan = chan;
105 do {
106 ret = write(ctx->consumer_channel_pipe[1], &msg, sizeof(msg));
107 } while (ret < 0 && errno == EINTR);
108 }
109
110 void notify_thread_del_channel(struct lttng_consumer_local_data *ctx,
111 uint64_t key)
112 {
113 notify_channel_pipe(ctx, NULL, key, CONSUMER_CHANNEL_DEL);
114 }
115
116 static int read_channel_pipe(struct lttng_consumer_local_data *ctx,
117 struct lttng_consumer_channel **chan,
118 uint64_t *key,
119 enum consumer_channel_action *action)
120 {
121 struct consumer_channel_msg msg;
122 int ret;
123
124 do {
125 ret = read(ctx->consumer_channel_pipe[0], &msg, sizeof(msg));
126 } while (ret < 0 && errno == EINTR);
127 if (ret > 0) {
128 *action = msg.action;
129 *chan = msg.chan;
130 *key = msg.key;
131 }
132 return ret;
133 }
134
135 /*
136 * Find a stream. The consumer_data.lock must be locked during this
137 * call.
138 */
139 static struct lttng_consumer_stream *find_stream(uint64_t key,
140 struct lttng_ht *ht)
141 {
142 struct lttng_ht_iter iter;
143 struct lttng_ht_node_u64 *node;
144 struct lttng_consumer_stream *stream = NULL;
145
146 assert(ht);
147
148 /* -1ULL keys are lookup failures */
149 if (key == (uint64_t) -1ULL) {
150 return NULL;
151 }
152
153 rcu_read_lock();
154
155 lttng_ht_lookup(ht, &key, &iter);
156 node = lttng_ht_iter_get_node_u64(&iter);
157 if (node != NULL) {
158 stream = caa_container_of(node, struct lttng_consumer_stream, node);
159 }
160
161 rcu_read_unlock();
162
163 return stream;
164 }
165
166 static void steal_stream_key(int key, struct lttng_ht *ht)
167 {
168 struct lttng_consumer_stream *stream;
169
170 rcu_read_lock();
171 stream = find_stream(key, ht);
172 if (stream) {
173 stream->key = -1ULL;
174 /*
175 * We don't want the lookup to match, but we still need
176 * to iterate on this stream when iterating over the hash table. Just
177 * change the node key.
178 */
179 stream->node.key = -1ULL;
180 }
181 rcu_read_unlock();
182 }
183
184 /*
185 * Return a channel object for the given key.
186 *
187 * RCU read side lock MUST be acquired before calling this function and
188 * protects the channel ptr.
189 */
190 struct lttng_consumer_channel *consumer_find_channel(uint64_t key)
191 {
192 struct lttng_ht_iter iter;
193 struct lttng_ht_node_u64 *node;
194 struct lttng_consumer_channel *channel = NULL;
195
196 /* -1ULL keys are lookup failures */
197 if (key == (uint64_t) -1ULL) {
198 return NULL;
199 }
200
201 lttng_ht_lookup(consumer_data.channel_ht, &key, &iter);
202 node = lttng_ht_iter_get_node_u64(&iter);
203 if (node != NULL) {
204 channel = caa_container_of(node, struct lttng_consumer_channel, node);
205 }
206
207 return channel;
208 }
209
210 static void free_stream_rcu(struct rcu_head *head)
211 {
212 struct lttng_ht_node_u64 *node =
213 caa_container_of(head, struct lttng_ht_node_u64, head);
214 struct lttng_consumer_stream *stream =
215 caa_container_of(node, struct lttng_consumer_stream, node);
216
217 free(stream);
218 }
219
220 static void free_channel_rcu(struct rcu_head *head)
221 {
222 struct lttng_ht_node_u64 *node =
223 caa_container_of(head, struct lttng_ht_node_u64, head);
224 struct lttng_consumer_channel *channel =
225 caa_container_of(node, struct lttng_consumer_channel, node);
226
227 free(channel);
228 }
229
230 /*
231 * RCU protected relayd socket pair free.
232 */
233 static void free_relayd_rcu(struct rcu_head *head)
234 {
235 struct lttng_ht_node_u64 *node =
236 caa_container_of(head, struct lttng_ht_node_u64, head);
237 struct consumer_relayd_sock_pair *relayd =
238 caa_container_of(node, struct consumer_relayd_sock_pair, node);
239
240 /*
241 * Close all sockets. This is done in the call RCU since we don't want the
242 * socket fds to be reassigned thus potentially creating bad state of the
243 * relayd object.
244 *
245 * We do not have to lock the control socket mutex here since at this stage
246 * there is no one referencing to this relayd object.
247 */
248 (void) relayd_close(&relayd->control_sock);
249 (void) relayd_close(&relayd->data_sock);
250
251 free(relayd);
252 }
253
254 /*
255 * Destroy and free relayd socket pair object.
256 *
257 * This function MUST be called with the consumer_data lock acquired.
258 */
259 static void 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
290 DBG("Consumer delete channel key %" PRIu64, channel->key);
291
292 pthread_mutex_lock(&consumer_data.lock);
293
294 switch (consumer_data.type) {
295 case LTTNG_CONSUMER_KERNEL:
296 break;
297 case LTTNG_CONSUMER32_UST:
298 case LTTNG_CONSUMER64_UST:
299 lttng_ustconsumer_del_channel(channel);
300 break;
301 default:
302 ERR("Unknown consumer_data type");
303 assert(0);
304 goto end;
305 }
306
307 rcu_read_lock();
308 iter.iter.node = &channel->node.node;
309 ret = lttng_ht_del(consumer_data.channel_ht, &iter);
310 assert(!ret);
311 rcu_read_unlock();
312
313 call_rcu(&channel->node.head, free_channel_rcu);
314 end:
315 pthread_mutex_unlock(&consumer_data.lock);
316 }
317
318 /*
319 * Iterate over the relayd hash table and destroy each element. Finally,
320 * destroy the whole hash table.
321 */
322 static void cleanup_relayd_ht(void)
323 {
324 struct lttng_ht_iter iter;
325 struct consumer_relayd_sock_pair *relayd;
326
327 rcu_read_lock();
328
329 cds_lfht_for_each_entry(consumer_data.relayd_ht->ht, &iter.iter, relayd,
330 node.node) {
331 destroy_relayd(relayd);
332 }
333
334 rcu_read_unlock();
335
336 lttng_ht_destroy(consumer_data.relayd_ht);
337 }
338
339 /*
340 * Update the end point status of all streams having the given network sequence
341 * index (relayd index).
342 *
343 * It's atomically set without having the stream mutex locked which is fine
344 * because we handle the write/read race with a pipe wakeup for each thread.
345 */
346 static void update_endpoint_status_by_netidx(int net_seq_idx,
347 enum consumer_endpoint_status status)
348 {
349 struct lttng_ht_iter iter;
350 struct lttng_consumer_stream *stream;
351
352 DBG("Consumer set delete flag on stream by idx %d", net_seq_idx);
353
354 rcu_read_lock();
355
356 /* Let's begin with metadata */
357 cds_lfht_for_each_entry(metadata_ht->ht, &iter.iter, stream, node.node) {
358 if (stream->net_seq_idx == net_seq_idx) {
359 uatomic_set(&stream->endpoint_status, status);
360 DBG("Delete flag set to metadata stream %d", stream->wait_fd);
361 }
362 }
363
364 /* Follow up by the data streams */
365 cds_lfht_for_each_entry(data_ht->ht, &iter.iter, stream, node.node) {
366 if (stream->net_seq_idx == net_seq_idx) {
367 uatomic_set(&stream->endpoint_status, status);
368 DBG("Delete flag set to data stream %d", stream->wait_fd);
369 }
370 }
371 rcu_read_unlock();
372 }
373
374 /*
375 * Cleanup a relayd object by flagging every associated streams for deletion,
376 * destroying the object meaning removing it from the relayd hash table,
377 * closing the sockets and freeing the memory in a RCU call.
378 *
379 * If a local data context is available, notify the threads that the streams'
380 * state have changed.
381 */
382 static void cleanup_relayd(struct consumer_relayd_sock_pair *relayd,
383 struct lttng_consumer_local_data *ctx)
384 {
385 int netidx;
386
387 assert(relayd);
388
389 DBG("Cleaning up relayd sockets");
390
391 /* Save the net sequence index before destroying the object */
392 netidx = relayd->net_seq_idx;
393
394 /*
395 * Delete the relayd from the relayd hash table, close the sockets and free
396 * the object in a RCU call.
397 */
398 destroy_relayd(relayd);
399
400 /* Set inactive endpoint to all streams */
401 update_endpoint_status_by_netidx(netidx, CONSUMER_ENDPOINT_INACTIVE);
402
403 /*
404 * With a local data context, notify the threads that the streams' state
405 * have changed. The write() action on the pipe acts as an "implicit"
406 * memory barrier ordering the updates of the end point status from the
407 * read of this status which happens AFTER receiving this notify.
408 */
409 if (ctx) {
410 notify_thread_lttng_pipe(ctx->consumer_data_pipe);
411 notify_thread_lttng_pipe(ctx->consumer_metadata_pipe);
412 }
413 }
414
415 /*
416 * Flag a relayd socket pair for destruction. Destroy it if the refcount
417 * reaches zero.
418 *
419 * RCU read side lock MUST be aquired before calling this function.
420 */
421 void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair *relayd)
422 {
423 assert(relayd);
424
425 /* Set destroy flag for this object */
426 uatomic_set(&relayd->destroy_flag, 1);
427
428 /* Destroy the relayd if refcount is 0 */
429 if (uatomic_read(&relayd->refcount) == 0) {
430 destroy_relayd(relayd);
431 }
432 }
433
434 /*
435 * Remove a stream from the global list protected by a mutex. This
436 * function is also responsible for freeing its data structures.
437 */
438 void consumer_del_stream(struct lttng_consumer_stream *stream,
439 struct lttng_ht *ht)
440 {
441 int ret;
442 struct lttng_ht_iter iter;
443 struct lttng_consumer_channel *free_chan = NULL;
444 struct consumer_relayd_sock_pair *relayd;
445
446 assert(stream);
447
448 DBG("Consumer del stream %d", stream->wait_fd);
449
450 if (ht == NULL) {
451 /* Means the stream was allocated but not successfully added */
452 goto free_stream_rcu;
453 }
454
455 pthread_mutex_lock(&consumer_data.lock);
456 pthread_mutex_lock(&stream->lock);
457
458 switch (consumer_data.type) {
459 case LTTNG_CONSUMER_KERNEL:
460 if (stream->mmap_base != NULL) {
461 ret = munmap(stream->mmap_base, stream->mmap_len);
462 if (ret != 0) {
463 PERROR("munmap");
464 }
465 }
466 break;
467 case LTTNG_CONSUMER32_UST:
468 case LTTNG_CONSUMER64_UST:
469 lttng_ustconsumer_del_stream(stream);
470 break;
471 default:
472 ERR("Unknown consumer_data type");
473 assert(0);
474 goto end;
475 }
476
477 rcu_read_lock();
478 iter.iter.node = &stream->node.node;
479 ret = lttng_ht_del(ht, &iter);
480 assert(!ret);
481
482 iter.iter.node = &stream->node_channel_id.node;
483 ret = lttng_ht_del(consumer_data.stream_per_chan_id_ht, &iter);
484 assert(!ret);
485
486 iter.iter.node = &stream->node_session_id.node;
487 ret = lttng_ht_del(consumer_data.stream_list_ht, &iter);
488 assert(!ret);
489 rcu_read_unlock();
490
491 assert(consumer_data.stream_count > 0);
492 consumer_data.stream_count--;
493
494 if (stream->out_fd >= 0) {
495 ret = close(stream->out_fd);
496 if (ret) {
497 PERROR("close");
498 }
499 }
500
501 /* Check and cleanup relayd */
502 rcu_read_lock();
503 relayd = consumer_find_relayd(stream->net_seq_idx);
504 if (relayd != NULL) {
505 uatomic_dec(&relayd->refcount);
506 assert(uatomic_read(&relayd->refcount) >= 0);
507
508 /* Closing streams requires to lock the control socket. */
509 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
510 ret = relayd_send_close_stream(&relayd->control_sock,
511 stream->relayd_stream_id,
512 stream->next_net_seq_num - 1);
513 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
514 if (ret < 0) {
515 DBG("Unable to close stream on the relayd. Continuing");
516 /*
517 * Continue here. There is nothing we can do for the relayd.
518 * Chances are that the relayd has closed the socket so we just
519 * continue cleaning up.
520 */
521 }
522
523 /* Both conditions are met, we destroy the relayd. */
524 if (uatomic_read(&relayd->refcount) == 0 &&
525 uatomic_read(&relayd->destroy_flag)) {
526 destroy_relayd(relayd);
527 }
528 }
529 rcu_read_unlock();
530
531 if (!uatomic_sub_return(&stream->chan->refcount, 1)
532 && !uatomic_read(&stream->chan->nb_init_stream_left)) {
533 free_chan = stream->chan;
534 }
535
536 end:
537 consumer_data.need_update = 1;
538 pthread_mutex_unlock(&stream->lock);
539 pthread_mutex_unlock(&consumer_data.lock);
540
541 if (free_chan) {
542 consumer_del_channel(free_chan);
543 }
544
545 free_stream_rcu:
546 call_rcu(&stream->node.head, free_stream_rcu);
547 }
548
549 struct lttng_consumer_stream *consumer_allocate_stream(uint64_t channel_key,
550 uint64_t stream_key,
551 enum lttng_consumer_stream_state state,
552 const char *channel_name,
553 uid_t uid,
554 gid_t gid,
555 int relayd_id,
556 uint64_t session_id,
557 int cpu,
558 int *alloc_ret,
559 enum consumer_channel_type type)
560 {
561 int ret;
562 struct lttng_consumer_stream *stream;
563
564 stream = zmalloc(sizeof(*stream));
565 if (stream == NULL) {
566 PERROR("malloc struct lttng_consumer_stream");
567 ret = -ENOMEM;
568 goto end;
569 }
570
571 rcu_read_lock();
572
573 stream->key = stream_key;
574 stream->out_fd = -1;
575 stream->out_fd_offset = 0;
576 stream->state = state;
577 stream->uid = uid;
578 stream->gid = gid;
579 stream->net_seq_idx = relayd_id;
580 stream->session_id = session_id;
581 pthread_mutex_init(&stream->lock, NULL);
582
583 /* If channel is the metadata, flag this stream as metadata. */
584 if (type == CONSUMER_CHANNEL_TYPE_METADATA) {
585 stream->metadata_flag = 1;
586 /* Metadata is flat out. */
587 strncpy(stream->name, DEFAULT_METADATA_NAME, sizeof(stream->name));
588 } else {
589 /* Format stream name to <channel_name>_<cpu_number> */
590 ret = snprintf(stream->name, sizeof(stream->name), "%s_%d",
591 channel_name, cpu);
592 if (ret < 0) {
593 PERROR("snprintf stream name");
594 goto error;
595 }
596 }
597
598 /* Key is always the wait_fd for streams. */
599 lttng_ht_node_init_u64(&stream->node, stream->key);
600
601 /* Init node per channel id key */
602 lttng_ht_node_init_u64(&stream->node_channel_id, channel_key);
603
604 /* Init session id node with the stream session id */
605 lttng_ht_node_init_u64(&stream->node_session_id, stream->session_id);
606
607 DBG3("Allocated stream %s (key %" PRIu64 ", chan_key %" PRIu64 " relayd_id %" PRIu64 ", session_id %" PRIu64,
608 stream->name, stream->key, channel_key, stream->net_seq_idx, stream->session_id);
609
610 rcu_read_unlock();
611 return stream;
612
613 error:
614 rcu_read_unlock();
615 free(stream);
616 end:
617 if (alloc_ret) {
618 *alloc_ret = ret;
619 }
620 return NULL;
621 }
622
623 /*
624 * Add a stream to the global list protected by a mutex.
625 */
626 static int add_stream(struct lttng_consumer_stream *stream,
627 struct lttng_ht *ht)
628 {
629 int ret = 0;
630 struct consumer_relayd_sock_pair *relayd;
631
632 assert(stream);
633 assert(ht);
634
635 DBG3("Adding consumer stream %" PRIu64, stream->key);
636
637 pthread_mutex_lock(&consumer_data.lock);
638 pthread_mutex_lock(&stream->lock);
639 rcu_read_lock();
640
641 /* Steal stream identifier to avoid having streams with the same key */
642 steal_stream_key(stream->key, ht);
643
644 lttng_ht_add_unique_u64(ht, &stream->node);
645
646 lttng_ht_add_u64(consumer_data.stream_per_chan_id_ht,
647 &stream->node_channel_id);
648
649 /*
650 * Add stream to the stream_list_ht of the consumer data. No need to steal
651 * the key since the HT does not use it and we allow to add redundant keys
652 * into this table.
653 */
654 lttng_ht_add_u64(consumer_data.stream_list_ht, &stream->node_session_id);
655
656 /* Check and cleanup relayd */
657 relayd = consumer_find_relayd(stream->net_seq_idx);
658 if (relayd != NULL) {
659 uatomic_inc(&relayd->refcount);
660 }
661
662 /*
663 * When nb_init_stream_left reaches 0, we don't need to trigger any action
664 * in terms of destroying the associated channel, because the action that
665 * causes the count to become 0 also causes a stream to be added. The
666 * channel deletion will thus be triggered by the following removal of this
667 * stream.
668 */
669 if (uatomic_read(&stream->chan->nb_init_stream_left) > 0) {
670 /* Increment refcount before decrementing nb_init_stream_left */
671 cmm_smp_wmb();
672 uatomic_dec(&stream->chan->nb_init_stream_left);
673 }
674
675 /* Update consumer data once the node is inserted. */
676 consumer_data.stream_count++;
677 consumer_data.need_update = 1;
678
679 rcu_read_unlock();
680 pthread_mutex_unlock(&stream->lock);
681 pthread_mutex_unlock(&consumer_data.lock);
682
683 return ret;
684 }
685
686 /*
687 * Add relayd socket to global consumer data hashtable. RCU read side lock MUST
688 * be acquired before calling this.
689 */
690 static int add_relayd(struct consumer_relayd_sock_pair *relayd)
691 {
692 int ret = 0;
693 struct lttng_ht_node_u64 *node;
694 struct lttng_ht_iter iter;
695
696 assert(relayd);
697
698 lttng_ht_lookup(consumer_data.relayd_ht,
699 &relayd->net_seq_idx, &iter);
700 node = lttng_ht_iter_get_node_u64(&iter);
701 if (node != NULL) {
702 goto end;
703 }
704 lttng_ht_add_unique_u64(consumer_data.relayd_ht, &relayd->node);
705
706 end:
707 return ret;
708 }
709
710 /*
711 * Allocate and return a consumer relayd socket.
712 */
713 struct consumer_relayd_sock_pair *consumer_allocate_relayd_sock_pair(
714 int net_seq_idx)
715 {
716 struct consumer_relayd_sock_pair *obj = NULL;
717
718 /* Negative net sequence index is a failure */
719 if (net_seq_idx < 0) {
720 goto error;
721 }
722
723 obj = zmalloc(sizeof(struct consumer_relayd_sock_pair));
724 if (obj == NULL) {
725 PERROR("zmalloc relayd sock");
726 goto error;
727 }
728
729 obj->net_seq_idx = net_seq_idx;
730 obj->refcount = 0;
731 obj->destroy_flag = 0;
732 obj->control_sock.sock.fd = -1;
733 obj->data_sock.sock.fd = -1;
734 lttng_ht_node_init_u64(&obj->node, obj->net_seq_idx);
735 pthread_mutex_init(&obj->ctrl_sock_mutex, NULL);
736
737 error:
738 return obj;
739 }
740
741 /*
742 * Find a relayd socket pair in the global consumer data.
743 *
744 * Return the object if found else NULL.
745 * RCU read-side lock must be held across this call and while using the
746 * returned object.
747 */
748 struct consumer_relayd_sock_pair *consumer_find_relayd(uint64_t key)
749 {
750 struct lttng_ht_iter iter;
751 struct lttng_ht_node_u64 *node;
752 struct consumer_relayd_sock_pair *relayd = NULL;
753
754 /* Negative keys are lookup failures */
755 if (key == (uint64_t) -1ULL) {
756 goto error;
757 }
758
759 lttng_ht_lookup(consumer_data.relayd_ht, &key,
760 &iter);
761 node = lttng_ht_iter_get_node_u64(&iter);
762 if (node != NULL) {
763 relayd = caa_container_of(node, struct consumer_relayd_sock_pair, node);
764 }
765
766 error:
767 return relayd;
768 }
769
770 /*
771 * Handle stream for relayd transmission if the stream applies for network
772 * streaming where the net sequence index is set.
773 *
774 * Return destination file descriptor or negative value on error.
775 */
776 static int write_relayd_stream_header(struct lttng_consumer_stream *stream,
777 size_t data_size, unsigned long padding,
778 struct consumer_relayd_sock_pair *relayd)
779 {
780 int outfd = -1, ret;
781 struct lttcomm_relayd_data_hdr data_hdr;
782
783 /* Safety net */
784 assert(stream);
785 assert(relayd);
786
787 /* Reset data header */
788 memset(&data_hdr, 0, sizeof(data_hdr));
789
790 if (stream->metadata_flag) {
791 /* Caller MUST acquire the relayd control socket lock */
792 ret = relayd_send_metadata(&relayd->control_sock, data_size);
793 if (ret < 0) {
794 goto error;
795 }
796
797 /* Metadata are always sent on the control socket. */
798 outfd = relayd->control_sock.sock.fd;
799 } else {
800 /* Set header with stream information */
801 data_hdr.stream_id = htobe64(stream->relayd_stream_id);
802 data_hdr.data_size = htobe32(data_size);
803 data_hdr.padding_size = htobe32(padding);
804 /*
805 * Note that net_seq_num below is assigned with the *current* value of
806 * next_net_seq_num and only after that the next_net_seq_num will be
807 * increment. This is why when issuing a command on the relayd using
808 * this next value, 1 should always be substracted in order to compare
809 * the last seen sequence number on the relayd side to the last sent.
810 */
811 data_hdr.net_seq_num = htobe64(stream->next_net_seq_num);
812 /* Other fields are zeroed previously */
813
814 ret = relayd_send_data_hdr(&relayd->data_sock, &data_hdr,
815 sizeof(data_hdr));
816 if (ret < 0) {
817 goto error;
818 }
819
820 ++stream->next_net_seq_num;
821
822 /* Set to go on data socket */
823 outfd = relayd->data_sock.sock.fd;
824 }
825
826 error:
827 return outfd;
828 }
829
830 /*
831 * Allocate and return a new lttng_consumer_channel object using the given key
832 * to initialize the hash table node.
833 *
834 * On error, return NULL.
835 */
836 struct lttng_consumer_channel *consumer_allocate_channel(uint64_t key,
837 uint64_t session_id,
838 const char *pathname,
839 const char *name,
840 uid_t uid,
841 gid_t gid,
842 int relayd_id,
843 enum lttng_event_output output,
844 uint64_t tracefile_size,
845 uint64_t tracefile_count)
846 {
847 struct lttng_consumer_channel *channel;
848
849 channel = zmalloc(sizeof(*channel));
850 if (channel == NULL) {
851 PERROR("malloc struct lttng_consumer_channel");
852 goto end;
853 }
854
855 channel->key = key;
856 channel->refcount = 0;
857 channel->session_id = session_id;
858 channel->uid = uid;
859 channel->gid = gid;
860 channel->relayd_id = relayd_id;
861 channel->output = output;
862 channel->tracefile_size = tracefile_size;
863 channel->tracefile_count = tracefile_count;
864
865 strncpy(channel->pathname, pathname, sizeof(channel->pathname));
866 channel->pathname[sizeof(channel->pathname) - 1] = '\0';
867
868 strncpy(channel->name, name, sizeof(channel->name));
869 channel->name[sizeof(channel->name) - 1] = '\0';
870
871 lttng_ht_node_init_u64(&channel->node, channel->key);
872
873 channel->wait_fd = -1;
874
875 CDS_INIT_LIST_HEAD(&channel->streams.head);
876
877 DBG("Allocated channel (key %" PRIu64 ")", channel->key)
878
879 end:
880 return channel;
881 }
882
883 /*
884 * Add a channel to the global list protected by a mutex.
885 */
886 int consumer_add_channel(struct lttng_consumer_channel *channel,
887 struct lttng_consumer_local_data *ctx)
888 {
889 int ret = 0;
890 struct lttng_ht_node_u64 *node;
891 struct lttng_ht_iter iter;
892
893 pthread_mutex_lock(&consumer_data.lock);
894 rcu_read_lock();
895
896 lttng_ht_lookup(consumer_data.channel_ht, &channel->key, &iter);
897 node = lttng_ht_iter_get_node_u64(&iter);
898 if (node != NULL) {
899 /* Channel already exist. Ignore the insertion */
900 ERR("Consumer add channel key %" PRIu64 " already exists!",
901 channel->key);
902 ret = -1;
903 goto end;
904 }
905
906 lttng_ht_add_unique_u64(consumer_data.channel_ht, &channel->node);
907
908 end:
909 rcu_read_unlock();
910 pthread_mutex_unlock(&consumer_data.lock);
911
912 if (!ret && channel->wait_fd != -1 &&
913 channel->metadata_stream == NULL) {
914 notify_channel_pipe(ctx, channel, -1, CONSUMER_CHANNEL_ADD);
915 }
916 return ret;
917 }
918
919 /*
920 * Allocate the pollfd structure and the local view of the out fds to avoid
921 * doing a lookup in the linked list and concurrency issues when writing is
922 * needed. Called with consumer_data.lock held.
923 *
924 * Returns the number of fds in the structures.
925 */
926 static int update_poll_array(struct lttng_consumer_local_data *ctx,
927 struct pollfd **pollfd, struct lttng_consumer_stream **local_stream,
928 struct lttng_ht *ht)
929 {
930 int i = 0;
931 struct lttng_ht_iter iter;
932 struct lttng_consumer_stream *stream;
933
934 assert(ctx);
935 assert(ht);
936 assert(pollfd);
937 assert(local_stream);
938
939 DBG("Updating poll fd array");
940 rcu_read_lock();
941 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
942 /*
943 * Only active streams with an active end point can be added to the
944 * poll set and local stream storage of the thread.
945 *
946 * There is a potential race here for endpoint_status to be updated
947 * just after the check. However, this is OK since the stream(s) will
948 * be deleted once the thread is notified that the end point state has
949 * changed where this function will be called back again.
950 */
951 if (stream->state != LTTNG_CONSUMER_ACTIVE_STREAM ||
952 stream->endpoint_status == CONSUMER_ENDPOINT_INACTIVE) {
953 continue;
954 }
955 /*
956 * This clobbers way too much the debug output. Uncomment that if you
957 * need it for debugging purposes.
958 *
959 * DBG("Active FD %d", stream->wait_fd);
960 */
961 (*pollfd)[i].fd = stream->wait_fd;
962 (*pollfd)[i].events = POLLIN | POLLPRI;
963 local_stream[i] = stream;
964 i++;
965 }
966 rcu_read_unlock();
967
968 /*
969 * Insert the consumer_data_pipe at the end of the array and don't
970 * increment i so nb_fd is the number of real FD.
971 */
972 (*pollfd)[i].fd = lttng_pipe_get_readfd(ctx->consumer_data_pipe);
973 (*pollfd)[i].events = POLLIN | POLLPRI;
974 return i;
975 }
976
977 /*
978 * Poll on the should_quit pipe and the command socket return -1 on error and
979 * should exit, 0 if data is available on the command socket
980 */
981 int lttng_consumer_poll_socket(struct pollfd *consumer_sockpoll)
982 {
983 int num_rdy;
984
985 restart:
986 num_rdy = poll(consumer_sockpoll, 2, -1);
987 if (num_rdy == -1) {
988 /*
989 * Restart interrupted system call.
990 */
991 if (errno == EINTR) {
992 goto restart;
993 }
994 PERROR("Poll error");
995 goto exit;
996 }
997 if (consumer_sockpoll[0].revents & (POLLIN | POLLPRI)) {
998 DBG("consumer_should_quit wake up");
999 goto exit;
1000 }
1001 return 0;
1002
1003 exit:
1004 return -1;
1005 }
1006
1007 /*
1008 * Set the error socket.
1009 */
1010 void lttng_consumer_set_error_sock(struct lttng_consumer_local_data *ctx,
1011 int sock)
1012 {
1013 ctx->consumer_error_socket = sock;
1014 }
1015
1016 /*
1017 * Set the command socket path.
1018 */
1019 void lttng_consumer_set_command_sock_path(
1020 struct lttng_consumer_local_data *ctx, char *sock)
1021 {
1022 ctx->consumer_command_sock_path = sock;
1023 }
1024
1025 /*
1026 * Send return code to the session daemon.
1027 * If the socket is not defined, we return 0, it is not a fatal error
1028 */
1029 int lttng_consumer_send_error(struct lttng_consumer_local_data *ctx, int cmd)
1030 {
1031 if (ctx->consumer_error_socket > 0) {
1032 return lttcomm_send_unix_sock(ctx->consumer_error_socket, &cmd,
1033 sizeof(enum lttcomm_sessiond_command));
1034 }
1035
1036 return 0;
1037 }
1038
1039 /*
1040 * Close all the tracefiles and stream fds and MUST be called when all
1041 * instances are destroyed i.e. when all threads were joined and are ended.
1042 */
1043 void lttng_consumer_cleanup(void)
1044 {
1045 struct lttng_ht_iter iter;
1046 struct lttng_consumer_channel *channel;
1047
1048 rcu_read_lock();
1049
1050 cds_lfht_for_each_entry(consumer_data.channel_ht->ht, &iter.iter, channel,
1051 node.node) {
1052 consumer_del_channel(channel);
1053 }
1054
1055 rcu_read_unlock();
1056
1057 lttng_ht_destroy(consumer_data.channel_ht);
1058
1059 cleanup_relayd_ht();
1060
1061 lttng_ht_destroy(consumer_data.stream_per_chan_id_ht);
1062
1063 /*
1064 * This HT contains streams that are freed by either the metadata thread or
1065 * the data thread so we do *nothing* on the hash table and simply destroy
1066 * it.
1067 */
1068 lttng_ht_destroy(consumer_data.stream_list_ht);
1069 }
1070
1071 /*
1072 * Called from signal handler.
1073 */
1074 void lttng_consumer_should_exit(struct lttng_consumer_local_data *ctx)
1075 {
1076 int ret;
1077 consumer_quit = 1;
1078 do {
1079 ret = write(ctx->consumer_should_quit[1], "4", 1);
1080 } while (ret < 0 && errno == EINTR);
1081 if (ret < 0 || ret != 1) {
1082 PERROR("write consumer quit");
1083 }
1084
1085 DBG("Consumer flag that it should quit");
1086 }
1087
1088 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream *stream,
1089 off_t orig_offset)
1090 {
1091 int outfd = stream->out_fd;
1092
1093 /*
1094 * This does a blocking write-and-wait on any page that belongs to the
1095 * subbuffer prior to the one we just wrote.
1096 * Don't care about error values, as these are just hints and ways to
1097 * limit the amount of page cache used.
1098 */
1099 if (orig_offset < stream->max_sb_size) {
1100 return;
1101 }
1102 lttng_sync_file_range(outfd, orig_offset - stream->max_sb_size,
1103 stream->max_sb_size,
1104 SYNC_FILE_RANGE_WAIT_BEFORE
1105 | SYNC_FILE_RANGE_WRITE
1106 | SYNC_FILE_RANGE_WAIT_AFTER);
1107 /*
1108 * Give hints to the kernel about how we access the file:
1109 * POSIX_FADV_DONTNEED : we won't re-access data in a near future after
1110 * we write it.
1111 *
1112 * We need to call fadvise again after the file grows because the
1113 * kernel does not seem to apply fadvise to non-existing parts of the
1114 * file.
1115 *
1116 * Call fadvise _after_ having waited for the page writeback to
1117 * complete because the dirty page writeback semantic is not well
1118 * defined. So it can be expected to lead to lower throughput in
1119 * streaming.
1120 */
1121 posix_fadvise(outfd, orig_offset - stream->max_sb_size,
1122 stream->max_sb_size, POSIX_FADV_DONTNEED);
1123 }
1124
1125 /*
1126 * Initialise the necessary environnement :
1127 * - create a new context
1128 * - create the poll_pipe
1129 * - create the should_quit pipe (for signal handler)
1130 * - create the thread pipe (for splice)
1131 *
1132 * Takes a function pointer as argument, this function is called when data is
1133 * available on a buffer. This function is responsible to do the
1134 * kernctl_get_next_subbuf, read the data with mmap or splice depending on the
1135 * buffer configuration and then kernctl_put_next_subbuf at the end.
1136 *
1137 * Returns a pointer to the new context or NULL on error.
1138 */
1139 struct lttng_consumer_local_data *lttng_consumer_create(
1140 enum lttng_consumer_type type,
1141 ssize_t (*buffer_ready)(struct lttng_consumer_stream *stream,
1142 struct lttng_consumer_local_data *ctx),
1143 int (*recv_channel)(struct lttng_consumer_channel *channel),
1144 int (*recv_stream)(struct lttng_consumer_stream *stream),
1145 int (*update_stream)(int stream_key, uint32_t state))
1146 {
1147 int ret;
1148 struct lttng_consumer_local_data *ctx;
1149
1150 assert(consumer_data.type == LTTNG_CONSUMER_UNKNOWN ||
1151 consumer_data.type == type);
1152 consumer_data.type = type;
1153
1154 ctx = zmalloc(sizeof(struct lttng_consumer_local_data));
1155 if (ctx == NULL) {
1156 PERROR("allocating context");
1157 goto error;
1158 }
1159
1160 ctx->consumer_error_socket = -1;
1161 ctx->consumer_metadata_socket = -1;
1162 /* assign the callbacks */
1163 ctx->on_buffer_ready = buffer_ready;
1164 ctx->on_recv_channel = recv_channel;
1165 ctx->on_recv_stream = recv_stream;
1166 ctx->on_update_stream = update_stream;
1167
1168 ctx->consumer_data_pipe = lttng_pipe_open(0);
1169 if (!ctx->consumer_data_pipe) {
1170 goto error_poll_pipe;
1171 }
1172
1173 ret = pipe(ctx->consumer_should_quit);
1174 if (ret < 0) {
1175 PERROR("Error creating recv pipe");
1176 goto error_quit_pipe;
1177 }
1178
1179 ret = pipe(ctx->consumer_thread_pipe);
1180 if (ret < 0) {
1181 PERROR("Error creating thread pipe");
1182 goto error_thread_pipe;
1183 }
1184
1185 ret = pipe(ctx->consumer_channel_pipe);
1186 if (ret < 0) {
1187 PERROR("Error creating channel pipe");
1188 goto error_channel_pipe;
1189 }
1190
1191 ctx->consumer_metadata_pipe = lttng_pipe_open(0);
1192 if (!ctx->consumer_metadata_pipe) {
1193 goto error_metadata_pipe;
1194 }
1195
1196 ret = utils_create_pipe(ctx->consumer_splice_metadata_pipe);
1197 if (ret < 0) {
1198 goto error_splice_pipe;
1199 }
1200
1201 return ctx;
1202
1203 error_splice_pipe:
1204 lttng_pipe_destroy(ctx->consumer_metadata_pipe);
1205 error_metadata_pipe:
1206 utils_close_pipe(ctx->consumer_channel_pipe);
1207 error_channel_pipe:
1208 utils_close_pipe(ctx->consumer_thread_pipe);
1209 error_thread_pipe:
1210 utils_close_pipe(ctx->consumer_should_quit);
1211 error_quit_pipe:
1212 lttng_pipe_destroy(ctx->consumer_data_pipe);
1213 error_poll_pipe:
1214 free(ctx);
1215 error:
1216 return NULL;
1217 }
1218
1219 /*
1220 * Close all fds associated with the instance and free the context.
1221 */
1222 void lttng_consumer_destroy(struct lttng_consumer_local_data *ctx)
1223 {
1224 int ret;
1225
1226 DBG("Consumer destroying it. Closing everything.");
1227
1228 ret = close(ctx->consumer_error_socket);
1229 if (ret) {
1230 PERROR("close");
1231 }
1232 ret = close(ctx->consumer_metadata_socket);
1233 if (ret) {
1234 PERROR("close");
1235 }
1236 utils_close_pipe(ctx->consumer_thread_pipe);
1237 utils_close_pipe(ctx->consumer_channel_pipe);
1238 lttng_pipe_destroy(ctx->consumer_data_pipe);
1239 lttng_pipe_destroy(ctx->consumer_metadata_pipe);
1240 utils_close_pipe(ctx->consumer_should_quit);
1241 utils_close_pipe(ctx->consumer_splice_metadata_pipe);
1242
1243 unlink(ctx->consumer_command_sock_path);
1244 free(ctx);
1245 }
1246
1247 /*
1248 * Write the metadata stream id on the specified file descriptor.
1249 */
1250 static int write_relayd_metadata_id(int fd,
1251 struct lttng_consumer_stream *stream,
1252 struct consumer_relayd_sock_pair *relayd, unsigned long padding)
1253 {
1254 int ret;
1255 struct lttcomm_relayd_metadata_payload hdr;
1256
1257 hdr.stream_id = htobe64(stream->relayd_stream_id);
1258 hdr.padding_size = htobe32(padding);
1259 do {
1260 ret = write(fd, (void *) &hdr, sizeof(hdr));
1261 } while (ret < 0 && errno == EINTR);
1262 if (ret < 0 || ret != sizeof(hdr)) {
1263 /*
1264 * This error means that the fd's end is closed so ignore the perror
1265 * not to clubber the error output since this can happen in a normal
1266 * code path.
1267 */
1268 if (errno != EPIPE) {
1269 PERROR("write metadata stream id");
1270 }
1271 DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno);
1272 /*
1273 * Set ret to a negative value because if ret != sizeof(hdr), we don't
1274 * handle writting the missing part so report that as an error and
1275 * don't lie to the caller.
1276 */
1277 ret = -1;
1278 goto end;
1279 }
1280 DBG("Metadata stream id %" PRIu64 " with padding %lu written before data",
1281 stream->relayd_stream_id, padding);
1282
1283 end:
1284 return ret;
1285 }
1286
1287 /*
1288 * Mmap the ring buffer, read it and write the data to the tracefile. This is a
1289 * core function for writing trace buffers to either the local filesystem or
1290 * the network.
1291 *
1292 * It must be called with the stream lock held.
1293 *
1294 * Careful review MUST be put if any changes occur!
1295 *
1296 * Returns the number of bytes written
1297 */
1298 ssize_t lttng_consumer_on_read_subbuffer_mmap(
1299 struct lttng_consumer_local_data *ctx,
1300 struct lttng_consumer_stream *stream, unsigned long len,
1301 unsigned long padding)
1302 {
1303 unsigned long mmap_offset;
1304 void *mmap_base;
1305 ssize_t ret = 0, written = 0;
1306 off_t orig_offset = stream->out_fd_offset;
1307 /* Default is on the disk */
1308 int outfd = stream->out_fd;
1309 struct consumer_relayd_sock_pair *relayd = NULL;
1310 unsigned int relayd_hang_up = 0;
1311
1312 /* RCU lock for the relayd pointer */
1313 rcu_read_lock();
1314
1315 /* Flag that the current stream if set for network streaming. */
1316 if (stream->net_seq_idx != -1) {
1317 relayd = consumer_find_relayd(stream->net_seq_idx);
1318 if (relayd == NULL) {
1319 goto end;
1320 }
1321 }
1322
1323 /* get the offset inside the fd to mmap */
1324 switch (consumer_data.type) {
1325 case LTTNG_CONSUMER_KERNEL:
1326 mmap_base = stream->mmap_base;
1327 ret = kernctl_get_mmap_read_offset(stream->wait_fd, &mmap_offset);
1328 break;
1329 case LTTNG_CONSUMER32_UST:
1330 case LTTNG_CONSUMER64_UST:
1331 mmap_base = lttng_ustctl_get_mmap_base(stream);
1332 if (!mmap_base) {
1333 ERR("read mmap get mmap base for stream %s", stream->name);
1334 written = -1;
1335 goto end;
1336 }
1337 ret = lttng_ustctl_get_mmap_read_offset(stream, &mmap_offset);
1338
1339 break;
1340 default:
1341 ERR("Unknown consumer_data type");
1342 assert(0);
1343 }
1344 if (ret != 0) {
1345 errno = -ret;
1346 PERROR("tracer ctl get_mmap_read_offset");
1347 written = ret;
1348 goto end;
1349 }
1350
1351 /* Handle stream on the relayd if the output is on the network */
1352 if (relayd) {
1353 unsigned long netlen = len;
1354
1355 /*
1356 * Lock the control socket for the complete duration of the function
1357 * since from this point on we will use the socket.
1358 */
1359 if (stream->metadata_flag) {
1360 /* Metadata requires the control socket. */
1361 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
1362 netlen += sizeof(struct lttcomm_relayd_metadata_payload);
1363 }
1364
1365 ret = write_relayd_stream_header(stream, netlen, padding, relayd);
1366 if (ret >= 0) {
1367 /* Use the returned socket. */
1368 outfd = ret;
1369
1370 /* Write metadata stream id before payload */
1371 if (stream->metadata_flag) {
1372 ret = write_relayd_metadata_id(outfd, stream, relayd, padding);
1373 if (ret < 0) {
1374 written = ret;
1375 /* Socket operation failed. We consider the relayd dead */
1376 if (ret == -EPIPE || ret == -EINVAL) {
1377 relayd_hang_up = 1;
1378 goto write_error;
1379 }
1380 goto end;
1381 }
1382 }
1383 } else {
1384 /* Socket operation failed. We consider the relayd dead */
1385 if (ret == -EPIPE || ret == -EINVAL) {
1386 relayd_hang_up = 1;
1387 goto write_error;
1388 }
1389 /* Else, use the default set before which is the filesystem. */
1390 }
1391 } else {
1392 /* No streaming, we have to set the len with the full padding */
1393 len += padding;
1394
1395 /*
1396 * Check if we need to change the tracefile before writing the packet.
1397 */
1398 if (stream->chan->tracefile_size > 0 &&
1399 (stream->tracefile_size_current + len) >
1400 stream->chan->tracefile_size) {
1401 ret = utils_rotate_stream_file(stream->chan->pathname,
1402 stream->name, stream->chan->tracefile_size,
1403 stream->chan->tracefile_count, stream->uid, stream->gid,
1404 stream->out_fd, &(stream->tracefile_count_current));
1405 if (ret < 0) {
1406 ERR("Rotating output file");
1407 goto end;
1408 }
1409 outfd = stream->out_fd = ret;
1410 /* Reset current size because we just perform a rotation. */
1411 stream->tracefile_size_current = 0;
1412 }
1413 stream->tracefile_size_current += len;
1414 }
1415
1416 while (len > 0) {
1417 do {
1418 ret = write(outfd, mmap_base + mmap_offset, len);
1419 } while (ret < 0 && errno == EINTR);
1420 DBG("Consumer mmap write() ret %zd (len %lu)", ret, len);
1421 if (ret < 0) {
1422 /*
1423 * This is possible if the fd is closed on the other side (outfd)
1424 * or any write problem. It can be verbose a bit for a normal
1425 * execution if for instance the relayd is stopped abruptly. This
1426 * can happen so set this to a DBG statement.
1427 */
1428 DBG("Error in file write mmap");
1429 if (written == 0) {
1430 written = ret;
1431 }
1432 /* Socket operation failed. We consider the relayd dead */
1433 if (errno == EPIPE || errno == EINVAL) {
1434 relayd_hang_up = 1;
1435 goto write_error;
1436 }
1437 goto end;
1438 } else if (ret > len) {
1439 PERROR("Error in file write (ret %zd > len %lu)", ret, len);
1440 written += ret;
1441 goto end;
1442 } else {
1443 len -= ret;
1444 mmap_offset += ret;
1445 }
1446
1447 /* This call is useless on a socket so better save a syscall. */
1448 if (!relayd) {
1449 /* This won't block, but will start writeout asynchronously */
1450 lttng_sync_file_range(outfd, stream->out_fd_offset, ret,
1451 SYNC_FILE_RANGE_WRITE);
1452 stream->out_fd_offset += ret;
1453 }
1454 written += ret;
1455 }
1456 lttng_consumer_sync_trace_file(stream, orig_offset);
1457
1458 write_error:
1459 /*
1460 * This is a special case that the relayd has closed its socket. Let's
1461 * cleanup the relayd object and all associated streams.
1462 */
1463 if (relayd && relayd_hang_up) {
1464 cleanup_relayd(relayd, ctx);
1465 }
1466
1467 end:
1468 /* Unlock only if ctrl socket used */
1469 if (relayd && stream->metadata_flag) {
1470 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
1471 }
1472
1473 rcu_read_unlock();
1474 return written;
1475 }
1476
1477 /*
1478 * Splice the data from the ring buffer to the tracefile.
1479 *
1480 * It must be called with the stream lock held.
1481 *
1482 * Returns the number of bytes spliced.
1483 */
1484 ssize_t lttng_consumer_on_read_subbuffer_splice(
1485 struct lttng_consumer_local_data *ctx,
1486 struct lttng_consumer_stream *stream, unsigned long len,
1487 unsigned long padding)
1488 {
1489 ssize_t ret = 0, written = 0, ret_splice = 0;
1490 loff_t offset = 0;
1491 off_t orig_offset = stream->out_fd_offset;
1492 int fd = stream->wait_fd;
1493 /* Default is on the disk */
1494 int outfd = stream->out_fd;
1495 struct consumer_relayd_sock_pair *relayd = NULL;
1496 int *splice_pipe;
1497 unsigned int relayd_hang_up = 0;
1498
1499 switch (consumer_data.type) {
1500 case LTTNG_CONSUMER_KERNEL:
1501 break;
1502 case LTTNG_CONSUMER32_UST:
1503 case LTTNG_CONSUMER64_UST:
1504 /* Not supported for user space tracing */
1505 return -ENOSYS;
1506 default:
1507 ERR("Unknown consumer_data type");
1508 assert(0);
1509 }
1510
1511 /* RCU lock for the relayd pointer */
1512 rcu_read_lock();
1513
1514 /* Flag that the current stream if set for network streaming. */
1515 if (stream->net_seq_idx != -1) {
1516 relayd = consumer_find_relayd(stream->net_seq_idx);
1517 if (relayd == NULL) {
1518 goto end;
1519 }
1520 }
1521
1522 /*
1523 * Choose right pipe for splice. Metadata and trace data are handled by
1524 * different threads hence the use of two pipes in order not to race or
1525 * corrupt the written data.
1526 */
1527 if (stream->metadata_flag) {
1528 splice_pipe = ctx->consumer_splice_metadata_pipe;
1529 } else {
1530 splice_pipe = ctx->consumer_thread_pipe;
1531 }
1532
1533 /* Write metadata stream id before payload */
1534 if (relayd) {
1535 int total_len = len;
1536
1537 if (stream->metadata_flag) {
1538 /*
1539 * Lock the control socket for the complete duration of the function
1540 * since from this point on we will use the socket.
1541 */
1542 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
1543
1544 ret = write_relayd_metadata_id(splice_pipe[1], stream, relayd,
1545 padding);
1546 if (ret < 0) {
1547 written = ret;
1548 /* Socket operation failed. We consider the relayd dead */
1549 if (ret == -EBADF) {
1550 WARN("Remote relayd disconnected. Stopping");
1551 relayd_hang_up = 1;
1552 goto write_error;
1553 }
1554 goto end;
1555 }
1556
1557 total_len += sizeof(struct lttcomm_relayd_metadata_payload);
1558 }
1559
1560 ret = write_relayd_stream_header(stream, total_len, padding, relayd);
1561 if (ret >= 0) {
1562 /* Use the returned socket. */
1563 outfd = ret;
1564 } else {
1565 /* Socket operation failed. We consider the relayd dead */
1566 if (ret == -EBADF) {
1567 WARN("Remote relayd disconnected. Stopping");
1568 relayd_hang_up = 1;
1569 goto write_error;
1570 }
1571 goto end;
1572 }
1573 } else {
1574 /* No streaming, we have to set the len with the full padding */
1575 len += padding;
1576
1577 /*
1578 * Check if we need to change the tracefile before writing the packet.
1579 */
1580 if (stream->chan->tracefile_size > 0 &&
1581 (stream->tracefile_size_current + len) >
1582 stream->chan->tracefile_size) {
1583 ret = utils_rotate_stream_file(stream->chan->pathname,
1584 stream->name, stream->chan->tracefile_size,
1585 stream->chan->tracefile_count, stream->uid, stream->gid,
1586 stream->out_fd, &(stream->tracefile_count_current));
1587 if (ret < 0) {
1588 ERR("Rotating output file");
1589 goto end;
1590 }
1591 outfd = stream->out_fd = ret;
1592 /* Reset current size because we just perform a rotation. */
1593 stream->tracefile_size_current = 0;
1594 }
1595 stream->tracefile_size_current += len;
1596 }
1597
1598 while (len > 0) {
1599 DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)",
1600 (unsigned long)offset, len, fd, splice_pipe[1]);
1601 ret_splice = splice(fd, &offset, splice_pipe[1], NULL, len,
1602 SPLICE_F_MOVE | SPLICE_F_MORE);
1603 DBG("splice chan to pipe, ret %zd", ret_splice);
1604 if (ret_splice < 0) {
1605 PERROR("Error in relay splice");
1606 if (written == 0) {
1607 written = ret_splice;
1608 }
1609 ret = errno;
1610 goto splice_error;
1611 }
1612
1613 /* Handle stream on the relayd if the output is on the network */
1614 if (relayd) {
1615 if (stream->metadata_flag) {
1616 size_t metadata_payload_size =
1617 sizeof(struct lttcomm_relayd_metadata_payload);
1618
1619 /* Update counter to fit the spliced data */
1620 ret_splice += metadata_payload_size;
1621 len += metadata_payload_size;
1622 /*
1623 * We do this so the return value can match the len passed as
1624 * argument to this function.
1625 */
1626 written -= metadata_payload_size;
1627 }
1628 }
1629
1630 /* Splice data out */
1631 ret_splice = splice(splice_pipe[0], NULL, outfd, NULL,
1632 ret_splice, SPLICE_F_MOVE | SPLICE_F_MORE);
1633 DBG("Consumer splice pipe to file, ret %zd", ret_splice);
1634 if (ret_splice < 0) {
1635 PERROR("Error in file splice");
1636 if (written == 0) {
1637 written = ret_splice;
1638 }
1639 /* Socket operation failed. We consider the relayd dead */
1640 if (errno == EBADF || errno == EPIPE) {
1641 WARN("Remote relayd disconnected. Stopping");
1642 relayd_hang_up = 1;
1643 goto write_error;
1644 }
1645 ret = errno;
1646 goto splice_error;
1647 } else if (ret_splice > len) {
1648 errno = EINVAL;
1649 PERROR("Wrote more data than requested %zd (len: %lu)",
1650 ret_splice, len);
1651 written += ret_splice;
1652 ret = errno;
1653 goto splice_error;
1654 }
1655 len -= ret_splice;
1656
1657 /* This call is useless on a socket so better save a syscall. */
1658 if (!relayd) {
1659 /* This won't block, but will start writeout asynchronously */
1660 lttng_sync_file_range(outfd, stream->out_fd_offset, ret_splice,
1661 SYNC_FILE_RANGE_WRITE);
1662 stream->out_fd_offset += ret_splice;
1663 }
1664 written += ret_splice;
1665 }
1666 lttng_consumer_sync_trace_file(stream, orig_offset);
1667
1668 ret = ret_splice;
1669
1670 goto end;
1671
1672 write_error:
1673 /*
1674 * This is a special case that the relayd has closed its socket. Let's
1675 * cleanup the relayd object and all associated streams.
1676 */
1677 if (relayd && relayd_hang_up) {
1678 cleanup_relayd(relayd, ctx);
1679 /* Skip splice error so the consumer does not fail */
1680 goto end;
1681 }
1682
1683 splice_error:
1684 /* send the appropriate error description to sessiond */
1685 switch (ret) {
1686 case EINVAL:
1687 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_EINVAL);
1688 break;
1689 case ENOMEM:
1690 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_ENOMEM);
1691 break;
1692 case ESPIPE:
1693 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_ESPIPE);
1694 break;
1695 }
1696
1697 end:
1698 if (relayd && stream->metadata_flag) {
1699 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
1700 }
1701
1702 rcu_read_unlock();
1703 return written;
1704 }
1705
1706 /*
1707 * Take a snapshot for a specific fd
1708 *
1709 * Returns 0 on success, < 0 on error
1710 */
1711 int lttng_consumer_take_snapshot(struct lttng_consumer_stream *stream)
1712 {
1713 switch (consumer_data.type) {
1714 case LTTNG_CONSUMER_KERNEL:
1715 return lttng_kconsumer_take_snapshot(stream);
1716 case LTTNG_CONSUMER32_UST:
1717 case LTTNG_CONSUMER64_UST:
1718 return lttng_ustconsumer_take_snapshot(stream);
1719 default:
1720 ERR("Unknown consumer_data type");
1721 assert(0);
1722 return -ENOSYS;
1723 }
1724 }
1725
1726 /*
1727 * Get the produced position
1728 *
1729 * Returns 0 on success, < 0 on error
1730 */
1731 int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream *stream,
1732 unsigned long *pos)
1733 {
1734 switch (consumer_data.type) {
1735 case LTTNG_CONSUMER_KERNEL:
1736 return lttng_kconsumer_get_produced_snapshot(stream, pos);
1737 case LTTNG_CONSUMER32_UST:
1738 case LTTNG_CONSUMER64_UST:
1739 return lttng_ustconsumer_get_produced_snapshot(stream, pos);
1740 default:
1741 ERR("Unknown consumer_data type");
1742 assert(0);
1743 return -ENOSYS;
1744 }
1745 }
1746
1747 int lttng_consumer_recv_cmd(struct lttng_consumer_local_data *ctx,
1748 int sock, struct pollfd *consumer_sockpoll)
1749 {
1750 switch (consumer_data.type) {
1751 case LTTNG_CONSUMER_KERNEL:
1752 return lttng_kconsumer_recv_cmd(ctx, sock, consumer_sockpoll);
1753 case LTTNG_CONSUMER32_UST:
1754 case LTTNG_CONSUMER64_UST:
1755 return lttng_ustconsumer_recv_cmd(ctx, sock, consumer_sockpoll);
1756 default:
1757 ERR("Unknown consumer_data type");
1758 assert(0);
1759 return -ENOSYS;
1760 }
1761 }
1762
1763 /*
1764 * Iterate over all streams of the hashtable and free them properly.
1765 *
1766 * WARNING: *MUST* be used with data stream only.
1767 */
1768 static void destroy_data_stream_ht(struct lttng_ht *ht)
1769 {
1770 struct lttng_ht_iter iter;
1771 struct lttng_consumer_stream *stream;
1772
1773 if (ht == NULL) {
1774 return;
1775 }
1776
1777 rcu_read_lock();
1778 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
1779 /*
1780 * Ignore return value since we are currently cleaning up so any error
1781 * can't be handled.
1782 */
1783 (void) consumer_del_stream(stream, ht);
1784 }
1785 rcu_read_unlock();
1786
1787 lttng_ht_destroy(ht);
1788 }
1789
1790 /*
1791 * Iterate over all streams of the hashtable and free them properly.
1792 *
1793 * XXX: Should not be only for metadata stream or else use an other name.
1794 */
1795 static void destroy_stream_ht(struct lttng_ht *ht)
1796 {
1797 struct lttng_ht_iter iter;
1798 struct lttng_consumer_stream *stream;
1799
1800 if (ht == NULL) {
1801 return;
1802 }
1803
1804 rcu_read_lock();
1805 cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
1806 /*
1807 * Ignore return value since we are currently cleaning up so any error
1808 * can't be handled.
1809 */
1810 (void) consumer_del_metadata_stream(stream, ht);
1811 }
1812 rcu_read_unlock();
1813
1814 lttng_ht_destroy(ht);
1815 }
1816
1817 void lttng_consumer_close_metadata(void)
1818 {
1819 switch (consumer_data.type) {
1820 case LTTNG_CONSUMER_KERNEL:
1821 /*
1822 * The Kernel consumer has a different metadata scheme so we don't
1823 * close anything because the stream will be closed by the session
1824 * daemon.
1825 */
1826 break;
1827 case LTTNG_CONSUMER32_UST:
1828 case LTTNG_CONSUMER64_UST:
1829 /*
1830 * Close all metadata streams. The metadata hash table is passed and
1831 * this call iterates over it by closing all wakeup fd. This is safe
1832 * because at this point we are sure that the metadata producer is
1833 * either dead or blocked.
1834 */
1835 lttng_ustconsumer_close_metadata(metadata_ht);
1836 break;
1837 default:
1838 ERR("Unknown consumer_data type");
1839 assert(0);
1840 }
1841 }
1842
1843 /*
1844 * Clean up a metadata stream and free its memory.
1845 */
1846 void consumer_del_metadata_stream(struct lttng_consumer_stream *stream,
1847 struct lttng_ht *ht)
1848 {
1849 int ret;
1850 struct lttng_ht_iter iter;
1851 struct lttng_consumer_channel *free_chan = NULL;
1852 struct consumer_relayd_sock_pair *relayd;
1853
1854 assert(stream);
1855 /*
1856 * This call should NEVER receive regular stream. It must always be
1857 * metadata stream and this is crucial for data structure synchronization.
1858 */
1859 assert(stream->metadata_flag);
1860
1861 DBG3("Consumer delete metadata stream %d", stream->wait_fd);
1862
1863 if (ht == NULL) {
1864 /* Means the stream was allocated but not successfully added */
1865 goto free_stream_rcu;
1866 }
1867
1868 pthread_mutex_lock(&consumer_data.lock);
1869 pthread_mutex_lock(&stream->lock);
1870
1871 switch (consumer_data.type) {
1872 case LTTNG_CONSUMER_KERNEL:
1873 if (stream->mmap_base != NULL) {
1874 ret = munmap(stream->mmap_base, stream->mmap_len);
1875 if (ret != 0) {
1876 PERROR("munmap metadata stream");
1877 }
1878 }
1879 break;
1880 case LTTNG_CONSUMER32_UST:
1881 case LTTNG_CONSUMER64_UST:
1882 lttng_ustconsumer_del_stream(stream);
1883 break;
1884 default:
1885 ERR("Unknown consumer_data type");
1886 assert(0);
1887 goto end;
1888 }
1889
1890 rcu_read_lock();
1891 iter.iter.node = &stream->node.node;
1892 ret = lttng_ht_del(ht, &iter);
1893 assert(!ret);
1894
1895 iter.iter.node = &stream->node_channel_id.node;
1896 ret = lttng_ht_del(consumer_data.stream_per_chan_id_ht, &iter);
1897 assert(!ret);
1898
1899 iter.iter.node = &stream->node_session_id.node;
1900 ret = lttng_ht_del(consumer_data.stream_list_ht, &iter);
1901 assert(!ret);
1902 rcu_read_unlock();
1903
1904 if (stream->out_fd >= 0) {
1905 ret = close(stream->out_fd);
1906 if (ret) {
1907 PERROR("close");
1908 }
1909 }
1910
1911 /* Check and cleanup relayd */
1912 rcu_read_lock();
1913 relayd = consumer_find_relayd(stream->net_seq_idx);
1914 if (relayd != NULL) {
1915 uatomic_dec(&relayd->refcount);
1916 assert(uatomic_read(&relayd->refcount) >= 0);
1917
1918 /* Closing streams requires to lock the control socket. */
1919 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
1920 ret = relayd_send_close_stream(&relayd->control_sock,
1921 stream->relayd_stream_id, stream->next_net_seq_num - 1);
1922 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
1923 if (ret < 0) {
1924 DBG("Unable to close stream on the relayd. Continuing");
1925 /*
1926 * Continue here. There is nothing we can do for the relayd.
1927 * Chances are that the relayd has closed the socket so we just
1928 * continue cleaning up.
1929 */
1930 }
1931
1932 /* Both conditions are met, we destroy the relayd. */
1933 if (uatomic_read(&relayd->refcount) == 0 &&
1934 uatomic_read(&relayd->destroy_flag)) {
1935 destroy_relayd(relayd);
1936 }
1937 }
1938 rcu_read_unlock();
1939
1940 /* Atomically decrement channel refcount since other threads can use it. */
1941 if (!uatomic_sub_return(&stream->chan->refcount, 1)
1942 && !uatomic_read(&stream->chan->nb_init_stream_left)) {
1943 /* Go for channel deletion! */
1944 free_chan = stream->chan;
1945 }
1946
1947 end:
1948 pthread_mutex_unlock(&stream->lock);
1949 pthread_mutex_unlock(&consumer_data.lock);
1950
1951 if (free_chan) {
1952 consumer_del_channel(free_chan);
1953 }
1954
1955 free_stream_rcu:
1956 call_rcu(&stream->node.head, free_stream_rcu);
1957 }
1958
1959 /*
1960 * Action done with the metadata stream when adding it to the consumer internal
1961 * data structures to handle it.
1962 */
1963 static int add_metadata_stream(struct lttng_consumer_stream *stream,
1964 struct lttng_ht *ht)
1965 {
1966 int ret = 0;
1967 struct consumer_relayd_sock_pair *relayd;
1968 struct lttng_ht_iter iter;
1969 struct lttng_ht_node_u64 *node;
1970
1971 assert(stream);
1972 assert(ht);
1973
1974 DBG3("Adding metadata stream %" PRIu64 " to hash table", stream->key);
1975
1976 pthread_mutex_lock(&consumer_data.lock);
1977 pthread_mutex_lock(&stream->lock);
1978
1979 /*
1980 * From here, refcounts are updated so be _careful_ when returning an error
1981 * after this point.
1982 */
1983
1984 rcu_read_lock();
1985
1986 /*
1987 * Lookup the stream just to make sure it does not exist in our internal
1988 * state. This should NEVER happen.
1989 */
1990 lttng_ht_lookup(ht, &stream->key, &iter);
1991 node = lttng_ht_iter_get_node_u64(&iter);
1992 assert(!node);
1993
1994 /* Find relayd and, if one is found, increment refcount. */
1995 relayd = consumer_find_relayd(stream->net_seq_idx);
1996 if (relayd != NULL) {
1997 uatomic_inc(&relayd->refcount);
1998 }
1999
2000 /* Update channel refcount once added without error(s). */
2001 uatomic_inc(&stream->chan->refcount);
2002
2003 /*
2004 * When nb_init_stream_left reaches 0, we don't need to trigger any action
2005 * in terms of destroying the associated channel, because the action that
2006 * causes the count to become 0 also causes a stream to be added. The
2007 * channel deletion will thus be triggered by the following removal of this
2008 * stream.
2009 */
2010 if (uatomic_read(&stream->chan->nb_init_stream_left) > 0) {
2011 /* Increment refcount before decrementing nb_init_stream_left */
2012 cmm_smp_wmb();
2013 uatomic_dec(&stream->chan->nb_init_stream_left);
2014 }
2015
2016 lttng_ht_add_unique_u64(ht, &stream->node);
2017
2018 lttng_ht_add_unique_u64(consumer_data.stream_per_chan_id_ht,
2019 &stream->node_channel_id);
2020
2021 /*
2022 * Add stream to the stream_list_ht of the consumer data. No need to steal
2023 * the key since the HT does not use it and we allow to add redundant keys
2024 * into this table.
2025 */
2026 lttng_ht_add_u64(consumer_data.stream_list_ht, &stream->node_session_id);
2027
2028 rcu_read_unlock();
2029
2030 pthread_mutex_unlock(&stream->lock);
2031 pthread_mutex_unlock(&consumer_data.lock);
2032 return ret;
2033 }
2034
2035 /*
2036 * Delete data stream that are flagged for deletion (endpoint_status).
2037 */
2038 static void validate_endpoint_status_data_stream(void)
2039 {
2040 struct lttng_ht_iter iter;
2041 struct lttng_consumer_stream *stream;
2042
2043 DBG("Consumer delete flagged data stream");
2044
2045 rcu_read_lock();
2046 cds_lfht_for_each_entry(data_ht->ht, &iter.iter, stream, node.node) {
2047 /* Validate delete flag of the stream */
2048 if (stream->endpoint_status == CONSUMER_ENDPOINT_ACTIVE) {
2049 continue;
2050 }
2051 /* Delete it right now */
2052 consumer_del_stream(stream, data_ht);
2053 }
2054 rcu_read_unlock();
2055 }
2056
2057 /*
2058 * Delete metadata stream that are flagged for deletion (endpoint_status).
2059 */
2060 static void validate_endpoint_status_metadata_stream(
2061 struct lttng_poll_event *pollset)
2062 {
2063 struct lttng_ht_iter iter;
2064 struct lttng_consumer_stream *stream;
2065
2066 DBG("Consumer delete flagged metadata stream");
2067
2068 assert(pollset);
2069
2070 rcu_read_lock();
2071 cds_lfht_for_each_entry(metadata_ht->ht, &iter.iter, stream, node.node) {
2072 /* Validate delete flag of the stream */
2073 if (stream->endpoint_status == CONSUMER_ENDPOINT_ACTIVE) {
2074 continue;
2075 }
2076 /*
2077 * Remove from pollset so the metadata thread can continue without
2078 * blocking on a deleted stream.
2079 */
2080 lttng_poll_del(pollset, stream->wait_fd);
2081
2082 /* Delete it right now */
2083 consumer_del_metadata_stream(stream, metadata_ht);
2084 }
2085 rcu_read_unlock();
2086 }
2087
2088 /*
2089 * Thread polls on metadata file descriptor and write them on disk or on the
2090 * network.
2091 */
2092 void *consumer_thread_metadata_poll(void *data)
2093 {
2094 int ret, i, pollfd;
2095 uint32_t revents, nb_fd;
2096 struct lttng_consumer_stream *stream = NULL;
2097 struct lttng_ht_iter iter;
2098 struct lttng_ht_node_u64 *node;
2099 struct lttng_poll_event events;
2100 struct lttng_consumer_local_data *ctx = data;
2101 ssize_t len;
2102
2103 rcu_register_thread();
2104
2105 metadata_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
2106 if (!metadata_ht) {
2107 /* ENOMEM at this point. Better to bail out. */
2108 goto end_ht;
2109 }
2110
2111 DBG("Thread metadata poll started");
2112
2113 /* Size is set to 1 for the consumer_metadata pipe */
2114 ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC);
2115 if (ret < 0) {
2116 ERR("Poll set creation failed");
2117 goto end_poll;
2118 }
2119
2120 ret = lttng_poll_add(&events,
2121 lttng_pipe_get_readfd(ctx->consumer_metadata_pipe), LPOLLIN);
2122 if (ret < 0) {
2123 goto end;
2124 }
2125
2126 /* Main loop */
2127 DBG("Metadata main loop started");
2128
2129 while (1) {
2130 /* Only the metadata pipe is set */
2131 if (LTTNG_POLL_GETNB(&events) == 0 && consumer_quit == 1) {
2132 goto end;
2133 }
2134
2135 restart:
2136 DBG("Metadata poll wait with %d fd(s)", LTTNG_POLL_GETNB(&events));
2137 ret = lttng_poll_wait(&events, -1);
2138 DBG("Metadata event catched in thread");
2139 if (ret < 0) {
2140 if (errno == EINTR) {
2141 ERR("Poll EINTR catched");
2142 goto restart;
2143 }
2144 goto error;
2145 }
2146
2147 nb_fd = ret;
2148
2149 /* From here, the event is a metadata wait fd */
2150 for (i = 0; i < nb_fd; i++) {
2151 revents = LTTNG_POLL_GETEV(&events, i);
2152 pollfd = LTTNG_POLL_GETFD(&events, i);
2153
2154 /* Just don't waste time if no returned events for the fd */
2155 if (!revents) {
2156 continue;
2157 }
2158
2159 if (pollfd == lttng_pipe_get_readfd(ctx->consumer_metadata_pipe)) {
2160 if (revents & (LPOLLERR | LPOLLHUP )) {
2161 DBG("Metadata thread pipe hung up");
2162 /*
2163 * Remove the pipe from the poll set and continue the loop
2164 * since their might be data to consume.
2165 */
2166 lttng_poll_del(&events,
2167 lttng_pipe_get_readfd(ctx->consumer_metadata_pipe));
2168 lttng_pipe_read_close(ctx->consumer_metadata_pipe);
2169 continue;
2170 } else if (revents & LPOLLIN) {
2171 ssize_t pipe_len;
2172
2173 pipe_len = lttng_pipe_read(ctx->consumer_metadata_pipe,
2174 &stream, sizeof(stream));
2175 if (pipe_len < 0) {
2176 ERR("read metadata stream, ret: %ld", pipe_len);
2177 /*
2178 * Continue here to handle the rest of the streams.
2179 */
2180 continue;
2181 }
2182
2183 /* A NULL stream means that the state has changed. */
2184 if (stream == NULL) {
2185 /* Check for deleted streams. */
2186 validate_endpoint_status_metadata_stream(&events);
2187 goto restart;
2188 }
2189
2190 DBG("Adding metadata stream %d to poll set",
2191 stream->wait_fd);
2192
2193 ret = add_metadata_stream(stream, metadata_ht);
2194 if (ret) {
2195 ERR("Unable to add metadata stream");
2196 /* Stream was not setup properly. Continuing. */
2197 consumer_del_metadata_stream(stream, NULL);
2198 continue;
2199 }
2200
2201 /* Add metadata stream to the global poll events list */
2202 lttng_poll_add(&events, stream->wait_fd,
2203 LPOLLIN | LPOLLPRI);
2204 }
2205
2206 /* Handle other stream */
2207 continue;
2208 }
2209
2210 rcu_read_lock();
2211 {
2212 uint64_t tmp_id = (uint64_t) pollfd;
2213
2214 lttng_ht_lookup(metadata_ht, &tmp_id, &iter);
2215 }
2216 node = lttng_ht_iter_get_node_u64(&iter);
2217 assert(node);
2218
2219 stream = caa_container_of(node, struct lttng_consumer_stream,
2220 node);
2221
2222 /* Check for error event */
2223 if (revents & (LPOLLERR | LPOLLHUP)) {
2224 DBG("Metadata fd %d is hup|err.", pollfd);
2225 if (!stream->hangup_flush_done
2226 && (consumer_data.type == LTTNG_CONSUMER32_UST
2227 || consumer_data.type == LTTNG_CONSUMER64_UST)) {
2228 DBG("Attempting to flush and consume the UST buffers");
2229 lttng_ustconsumer_on_stream_hangup(stream);
2230
2231 /* We just flushed the stream now read it. */
2232 do {
2233 len = ctx->on_buffer_ready(stream, ctx);
2234 /*
2235 * We don't check the return value here since if we get
2236 * a negative len, it means an error occured thus we
2237 * simply remove it from the poll set and free the
2238 * stream.
2239 */
2240 } while (len > 0);
2241 }
2242
2243 lttng_poll_del(&events, stream->wait_fd);
2244 /*
2245 * This call update the channel states, closes file descriptors
2246 * and securely free the stream.
2247 */
2248 consumer_del_metadata_stream(stream, metadata_ht);
2249 } else if (revents & (LPOLLIN | LPOLLPRI)) {
2250 /* Get the data out of the metadata file descriptor */
2251 DBG("Metadata available on fd %d", pollfd);
2252 assert(stream->wait_fd == pollfd);
2253
2254 len = ctx->on_buffer_ready(stream, ctx);
2255 /* It's ok to have an unavailable sub-buffer */
2256 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2257 /* Clean up stream from consumer and free it. */
2258 lttng_poll_del(&events, stream->wait_fd);
2259 consumer_del_metadata_stream(stream, metadata_ht);
2260 } else if (len > 0) {
2261 stream->data_read = 1;
2262 }
2263 }
2264
2265 /* Release RCU lock for the stream looked up */
2266 rcu_read_unlock();
2267 }
2268 }
2269
2270 error:
2271 end:
2272 DBG("Metadata poll thread exiting");
2273
2274 lttng_poll_clean(&events);
2275 end_poll:
2276 destroy_stream_ht(metadata_ht);
2277 end_ht:
2278 rcu_unregister_thread();
2279 return NULL;
2280 }
2281
2282 /*
2283 * This thread polls the fds in the set to consume the data and write
2284 * it to tracefile if necessary.
2285 */
2286 void *consumer_thread_data_poll(void *data)
2287 {
2288 int num_rdy, num_hup, high_prio, ret, i;
2289 struct pollfd *pollfd = NULL;
2290 /* local view of the streams */
2291 struct lttng_consumer_stream **local_stream = NULL, *new_stream = NULL;
2292 /* local view of consumer_data.fds_count */
2293 int nb_fd = 0;
2294 struct lttng_consumer_local_data *ctx = data;
2295 ssize_t len;
2296
2297 rcu_register_thread();
2298
2299 data_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
2300 if (data_ht == NULL) {
2301 /* ENOMEM at this point. Better to bail out. */
2302 goto end;
2303 }
2304
2305 local_stream = zmalloc(sizeof(struct lttng_consumer_stream));
2306
2307 while (1) {
2308 high_prio = 0;
2309 num_hup = 0;
2310
2311 /*
2312 * the fds set has been updated, we need to update our
2313 * local array as well
2314 */
2315 pthread_mutex_lock(&consumer_data.lock);
2316 if (consumer_data.need_update) {
2317 free(pollfd);
2318 pollfd = NULL;
2319
2320 free(local_stream);
2321 local_stream = NULL;
2322
2323 /* allocate for all fds + 1 for the consumer_data_pipe */
2324 pollfd = zmalloc((consumer_data.stream_count + 1) * sizeof(struct pollfd));
2325 if (pollfd == NULL) {
2326 PERROR("pollfd malloc");
2327 pthread_mutex_unlock(&consumer_data.lock);
2328 goto end;
2329 }
2330
2331 /* allocate for all fds + 1 for the consumer_data_pipe */
2332 local_stream = zmalloc((consumer_data.stream_count + 1) *
2333 sizeof(struct lttng_consumer_stream));
2334 if (local_stream == NULL) {
2335 PERROR("local_stream malloc");
2336 pthread_mutex_unlock(&consumer_data.lock);
2337 goto end;
2338 }
2339 ret = update_poll_array(ctx, &pollfd, local_stream,
2340 data_ht);
2341 if (ret < 0) {
2342 ERR("Error in allocating pollfd or local_outfds");
2343 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
2344 pthread_mutex_unlock(&consumer_data.lock);
2345 goto end;
2346 }
2347 nb_fd = ret;
2348 consumer_data.need_update = 0;
2349 }
2350 pthread_mutex_unlock(&consumer_data.lock);
2351
2352 /* No FDs and consumer_quit, consumer_cleanup the thread */
2353 if (nb_fd == 0 && consumer_quit == 1) {
2354 goto end;
2355 }
2356 /* poll on the array of fds */
2357 restart:
2358 DBG("polling on %d fd", nb_fd + 1);
2359 num_rdy = poll(pollfd, nb_fd + 1, -1);
2360 DBG("poll num_rdy : %d", num_rdy);
2361 if (num_rdy == -1) {
2362 /*
2363 * Restart interrupted system call.
2364 */
2365 if (errno == EINTR) {
2366 goto restart;
2367 }
2368 PERROR("Poll error");
2369 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
2370 goto end;
2371 } else if (num_rdy == 0) {
2372 DBG("Polling thread timed out");
2373 goto end;
2374 }
2375
2376 /*
2377 * If the consumer_data_pipe triggered poll go directly to the
2378 * beginning of the loop to update the array. We want to prioritize
2379 * array update over low-priority reads.
2380 */
2381 if (pollfd[nb_fd].revents & (POLLIN | POLLPRI)) {
2382 ssize_t pipe_readlen;
2383
2384 DBG("consumer_data_pipe wake up");
2385 pipe_readlen = lttng_pipe_read(ctx->consumer_data_pipe,
2386 &new_stream, sizeof(new_stream));
2387 if (pipe_readlen < 0) {
2388 ERR("Consumer data pipe ret %ld", pipe_readlen);
2389 /* Continue so we can at least handle the current stream(s). */
2390 continue;
2391 }
2392
2393 /*
2394 * If the stream is NULL, just ignore it. It's also possible that
2395 * the sessiond poll thread changed the consumer_quit state and is
2396 * waking us up to test it.
2397 */
2398 if (new_stream == NULL) {
2399 validate_endpoint_status_data_stream();
2400 continue;
2401 }
2402
2403 ret = add_stream(new_stream, data_ht);
2404 if (ret) {
2405 ERR("Consumer add stream %" PRIu64 " failed. Continuing",
2406 new_stream->key);
2407 /*
2408 * At this point, if the add_stream fails, it is not in the
2409 * hash table thus passing the NULL value here.
2410 */
2411 consumer_del_stream(new_stream, NULL);
2412 }
2413
2414 /* Continue to update the local streams and handle prio ones */
2415 continue;
2416 }
2417
2418 /* Take care of high priority channels first. */
2419 for (i = 0; i < nb_fd; i++) {
2420 if (local_stream[i] == NULL) {
2421 continue;
2422 }
2423 if (pollfd[i].revents & POLLPRI) {
2424 DBG("Urgent read on fd %d", pollfd[i].fd);
2425 high_prio = 1;
2426 len = ctx->on_buffer_ready(local_stream[i], ctx);
2427 /* it's ok to have an unavailable sub-buffer */
2428 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2429 /* Clean the stream and free it. */
2430 consumer_del_stream(local_stream[i], data_ht);
2431 local_stream[i] = NULL;
2432 } else if (len > 0) {
2433 local_stream[i]->data_read = 1;
2434 }
2435 }
2436 }
2437
2438 /*
2439 * If we read high prio channel in this loop, try again
2440 * for more high prio data.
2441 */
2442 if (high_prio) {
2443 continue;
2444 }
2445
2446 /* Take care of low priority channels. */
2447 for (i = 0; i < nb_fd; i++) {
2448 if (local_stream[i] == NULL) {
2449 continue;
2450 }
2451 if ((pollfd[i].revents & POLLIN) ||
2452 local_stream[i]->hangup_flush_done) {
2453 DBG("Normal read on fd %d", pollfd[i].fd);
2454 len = ctx->on_buffer_ready(local_stream[i], ctx);
2455 /* it's ok to have an unavailable sub-buffer */
2456 if (len < 0 && len != -EAGAIN && len != -ENODATA) {
2457 /* Clean the stream and free it. */
2458 consumer_del_stream(local_stream[i], data_ht);
2459 local_stream[i] = NULL;
2460 } else if (len > 0) {
2461 local_stream[i]->data_read = 1;
2462 }
2463 }
2464 }
2465
2466 /* Handle hangup and errors */
2467 for (i = 0; i < nb_fd; i++) {
2468 if (local_stream[i] == NULL) {
2469 continue;
2470 }
2471 if (!local_stream[i]->hangup_flush_done
2472 && (pollfd[i].revents & (POLLHUP | POLLERR | POLLNVAL))
2473 && (consumer_data.type == LTTNG_CONSUMER32_UST
2474 || consumer_data.type == LTTNG_CONSUMER64_UST)) {
2475 DBG("fd %d is hup|err|nval. Attempting flush and read.",
2476 pollfd[i].fd);
2477 lttng_ustconsumer_on_stream_hangup(local_stream[i]);
2478 /* Attempt read again, for the data we just flushed. */
2479 local_stream[i]->data_read = 1;
2480 }
2481 /*
2482 * If the poll flag is HUP/ERR/NVAL and we have
2483 * read no data in this pass, we can remove the
2484 * stream from its hash table.
2485 */
2486 if ((pollfd[i].revents & POLLHUP)) {
2487 DBG("Polling fd %d tells it has hung up.", pollfd[i].fd);
2488 if (!local_stream[i]->data_read) {
2489 consumer_del_stream(local_stream[i], data_ht);
2490 local_stream[i] = NULL;
2491 num_hup++;
2492 }
2493 } else if (pollfd[i].revents & POLLERR) {
2494 ERR("Error returned in polling fd %d.", pollfd[i].fd);
2495 if (!local_stream[i]->data_read) {
2496 consumer_del_stream(local_stream[i], data_ht);
2497 local_stream[i] = NULL;
2498 num_hup++;
2499 }
2500 } else if (pollfd[i].revents & POLLNVAL) {
2501 ERR("Polling fd %d tells fd is not open.", pollfd[i].fd);
2502 if (!local_stream[i]->data_read) {
2503 consumer_del_stream(local_stream[i], data_ht);
2504 local_stream[i] = NULL;
2505 num_hup++;
2506 }
2507 }
2508 if (local_stream[i] != NULL) {
2509 local_stream[i]->data_read = 0;
2510 }
2511 }
2512 }
2513 end:
2514 DBG("polling thread exiting");
2515 free(pollfd);
2516 free(local_stream);
2517
2518 /*
2519 * Close the write side of the pipe so epoll_wait() in
2520 * consumer_thread_metadata_poll can catch it. The thread is monitoring the
2521 * read side of the pipe. If we close them both, epoll_wait strangely does
2522 * not return and could create a endless wait period if the pipe is the
2523 * only tracked fd in the poll set. The thread will take care of closing
2524 * the read side.
2525 */
2526 (void) lttng_pipe_write_close(ctx->consumer_metadata_pipe);
2527
2528 destroy_data_stream_ht(data_ht);
2529
2530 rcu_unregister_thread();
2531 return NULL;
2532 }
2533
2534 /*
2535 * Close wake-up end of each stream belonging to the channel. This will
2536 * allow the poll() on the stream read-side to detect when the
2537 * write-side (application) finally closes them.
2538 */
2539 static
2540 void consumer_close_channel_streams(struct lttng_consumer_channel *channel)
2541 {
2542 struct lttng_ht *ht;
2543 struct lttng_consumer_stream *stream;
2544 struct lttng_ht_iter iter;
2545
2546 ht = consumer_data.stream_per_chan_id_ht;
2547
2548 rcu_read_lock();
2549 cds_lfht_for_each_entry_duplicate(ht->ht,
2550 ht->hash_fct(&channel->key, lttng_ht_seed),
2551 ht->match_fct, &channel->key,
2552 &iter.iter, stream, node_channel_id.node) {
2553 /*
2554 * Protect against teardown with mutex.
2555 */
2556 pthread_mutex_lock(&stream->lock);
2557 if (cds_lfht_is_node_deleted(&stream->node.node)) {
2558 goto next;
2559 }
2560 switch (consumer_data.type) {
2561 case LTTNG_CONSUMER_KERNEL:
2562 break;
2563 case LTTNG_CONSUMER32_UST:
2564 case LTTNG_CONSUMER64_UST:
2565 /*
2566 * Note: a mutex is taken internally within
2567 * liblttng-ust-ctl to protect timer wakeup_fd
2568 * use from concurrent close.
2569 */
2570 lttng_ustconsumer_close_stream_wakeup(stream);
2571 break;
2572 default:
2573 ERR("Unknown consumer_data type");
2574 assert(0);
2575 }
2576 next:
2577 pthread_mutex_unlock(&stream->lock);
2578 }
2579 rcu_read_unlock();
2580 }
2581
2582 static void destroy_channel_ht(struct lttng_ht *ht)
2583 {
2584 struct lttng_ht_iter iter;
2585 struct lttng_consumer_channel *channel;
2586 int ret;
2587
2588 if (ht == NULL) {
2589 return;
2590 }
2591
2592 rcu_read_lock();
2593 cds_lfht_for_each_entry(ht->ht, &iter.iter, channel, wait_fd_node.node) {
2594 ret = lttng_ht_del(ht, &iter);
2595 assert(ret != 0);
2596 }
2597 rcu_read_unlock();
2598
2599 lttng_ht_destroy(ht);
2600 }
2601
2602 /*
2603 * This thread polls the channel fds to detect when they are being
2604 * closed. It closes all related streams if the channel is detected as
2605 * closed. It is currently only used as a shim layer for UST because the
2606 * consumerd needs to keep the per-stream wakeup end of pipes open for
2607 * periodical flush.
2608 */
2609 void *consumer_thread_channel_poll(void *data)
2610 {
2611 int ret, i, pollfd;
2612 uint32_t revents, nb_fd;
2613 struct lttng_consumer_channel *chan = NULL;
2614 struct lttng_ht_iter iter;
2615 struct lttng_ht_node_u64 *node;
2616 struct lttng_poll_event events;
2617 struct lttng_consumer_local_data *ctx = data;
2618 struct lttng_ht *channel_ht;
2619
2620 rcu_register_thread();
2621
2622 channel_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
2623 if (!channel_ht) {
2624 /* ENOMEM at this point. Better to bail out. */
2625 goto end_ht;
2626 }
2627
2628 DBG("Thread channel poll started");
2629
2630 /* Size is set to 1 for the consumer_channel pipe */
2631 ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC);
2632 if (ret < 0) {
2633 ERR("Poll set creation failed");
2634 goto end_poll;
2635 }
2636
2637 ret = lttng_poll_add(&events, ctx->consumer_channel_pipe[0], LPOLLIN);
2638 if (ret < 0) {
2639 goto end;
2640 }
2641
2642 /* Main loop */
2643 DBG("Channel main loop started");
2644
2645 while (1) {
2646 /* Only the channel pipe is set */
2647 if (LTTNG_POLL_GETNB(&events) == 0 && consumer_quit == 1) {
2648 goto end;
2649 }
2650
2651 restart:
2652 DBG("Channel poll wait with %d fd(s)", LTTNG_POLL_GETNB(&events));
2653 ret = lttng_poll_wait(&events, -1);
2654 DBG("Channel event catched in thread");
2655 if (ret < 0) {
2656 if (errno == EINTR) {
2657 ERR("Poll EINTR catched");
2658 goto restart;
2659 }
2660 goto end;
2661 }
2662
2663 nb_fd = ret;
2664
2665 /* From here, the event is a channel wait fd */
2666 for (i = 0; i < nb_fd; i++) {
2667 revents = LTTNG_POLL_GETEV(&events, i);
2668 pollfd = LTTNG_POLL_GETFD(&events, i);
2669
2670 /* Just don't waste time if no returned events for the fd */
2671 if (!revents) {
2672 continue;
2673 }
2674 if (pollfd == ctx->consumer_channel_pipe[0]) {
2675 if (revents & (LPOLLERR | LPOLLHUP)) {
2676 DBG("Channel thread pipe hung up");
2677 /*
2678 * Remove the pipe from the poll set and continue the loop
2679 * since their might be data to consume.
2680 */
2681 lttng_poll_del(&events, ctx->consumer_channel_pipe[0]);
2682 continue;
2683 } else if (revents & LPOLLIN) {
2684 enum consumer_channel_action action;
2685 uint64_t key;
2686
2687 ret = read_channel_pipe(ctx, &chan, &key, &action);
2688 if (ret <= 0) {
2689 ERR("Error reading channel pipe");
2690 continue;
2691 }
2692
2693 switch (action) {
2694 case CONSUMER_CHANNEL_ADD:
2695 DBG("Adding channel %d to poll set",
2696 chan->wait_fd);
2697
2698 lttng_ht_node_init_u64(&chan->wait_fd_node,
2699 chan->wait_fd);
2700 lttng_ht_add_unique_u64(channel_ht,
2701 &chan->wait_fd_node);
2702 /* Add channel to the global poll events list */
2703 lttng_poll_add(&events, chan->wait_fd,
2704 LPOLLIN | LPOLLPRI);
2705 break;
2706 case CONSUMER_CHANNEL_DEL:
2707 {
2708 chan = consumer_find_channel(key);
2709 if (!chan) {
2710 ERR("UST consumer get channel key %" PRIu64 " not found for del channel", key);
2711 break;
2712 }
2713 lttng_poll_del(&events, chan->wait_fd);
2714 ret = lttng_ht_del(channel_ht, &iter);
2715 assert(ret == 0);
2716 consumer_close_channel_streams(chan);
2717
2718 /*
2719 * Release our own refcount. Force channel deletion even if
2720 * streams were not initialized.
2721 */
2722 if (!uatomic_sub_return(&chan->refcount, 1)) {
2723 consumer_del_channel(chan);
2724 }
2725 goto restart;
2726 }
2727 case CONSUMER_CHANNEL_QUIT:
2728 /*
2729 * Remove the pipe from the poll set and continue the loop
2730 * since their might be data to consume.
2731 */
2732 lttng_poll_del(&events, ctx->consumer_channel_pipe[0]);
2733 continue;
2734 default:
2735 ERR("Unknown action");
2736 break;
2737 }
2738 }
2739
2740 /* Handle other stream */
2741 continue;
2742 }
2743
2744 rcu_read_lock();
2745 {
2746 uint64_t tmp_id = (uint64_t) pollfd;
2747
2748 lttng_ht_lookup(channel_ht, &tmp_id, &iter);
2749 }
2750 node = lttng_ht_iter_get_node_u64(&iter);
2751 assert(node);
2752
2753 chan = caa_container_of(node, struct lttng_consumer_channel,
2754 wait_fd_node);
2755
2756 /* Check for error event */
2757 if (revents & (LPOLLERR | LPOLLHUP)) {
2758 DBG("Channel fd %d is hup|err.", pollfd);
2759
2760 lttng_poll_del(&events, chan->wait_fd);
2761 ret = lttng_ht_del(channel_ht, &iter);
2762 assert(ret == 0);
2763 consumer_close_channel_streams(chan);
2764
2765 /* Release our own refcount */
2766 if (!uatomic_sub_return(&chan->refcount, 1)
2767 && !uatomic_read(&chan->nb_init_stream_left)) {
2768 consumer_del_channel(chan);
2769 }
2770 }
2771
2772 /* Release RCU lock for the channel looked up */
2773 rcu_read_unlock();
2774 }
2775 }
2776
2777 end:
2778 lttng_poll_clean(&events);
2779 end_poll:
2780 destroy_channel_ht(channel_ht);
2781 end_ht:
2782 DBG("Channel poll thread exiting");
2783 rcu_unregister_thread();
2784 return NULL;
2785 }
2786
2787 static int set_metadata_socket(struct lttng_consumer_local_data *ctx,
2788 struct pollfd *sockpoll, int client_socket)
2789 {
2790 int ret;
2791
2792 assert(ctx);
2793 assert(sockpoll);
2794
2795 if (lttng_consumer_poll_socket(sockpoll) < 0) {
2796 ret = -1;
2797 goto error;
2798 }
2799 DBG("Metadata connection on client_socket");
2800
2801 /* Blocking call, waiting for transmission */
2802 ctx->consumer_metadata_socket = lttcomm_accept_unix_sock(client_socket);
2803 if (ctx->consumer_metadata_socket < 0) {
2804 WARN("On accept metadata");
2805 ret = -1;
2806 goto error;
2807 }
2808 ret = 0;
2809
2810 error:
2811 return ret;
2812 }
2813
2814 /*
2815 * This thread listens on the consumerd socket and receives the file
2816 * descriptors from the session daemon.
2817 */
2818 void *consumer_thread_sessiond_poll(void *data)
2819 {
2820 int sock = -1, client_socket, ret;
2821 /*
2822 * structure to poll for incoming data on communication socket avoids
2823 * making blocking sockets.
2824 */
2825 struct pollfd consumer_sockpoll[2];
2826 struct lttng_consumer_local_data *ctx = data;
2827
2828 rcu_register_thread();
2829
2830 DBG("Creating command socket %s", ctx->consumer_command_sock_path);
2831 unlink(ctx->consumer_command_sock_path);
2832 client_socket = lttcomm_create_unix_sock(ctx->consumer_command_sock_path);
2833 if (client_socket < 0) {
2834 ERR("Cannot create command socket");
2835 goto end;
2836 }
2837
2838 ret = lttcomm_listen_unix_sock(client_socket);
2839 if (ret < 0) {
2840 goto end;
2841 }
2842
2843 DBG("Sending ready command to lttng-sessiond");
2844 ret = lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY);
2845 /* return < 0 on error, but == 0 is not fatal */
2846 if (ret < 0) {
2847 ERR("Error sending ready command to lttng-sessiond");
2848 goto end;
2849 }
2850
2851 ret = fcntl(client_socket, F_SETFL, O_NONBLOCK);
2852 if (ret < 0) {
2853 PERROR("fcntl O_NONBLOCK");
2854 goto end;
2855 }
2856
2857 /* prepare the FDs to poll : to client socket and the should_quit pipe */
2858 consumer_sockpoll[0].fd = ctx->consumer_should_quit[0];
2859 consumer_sockpoll[0].events = POLLIN | POLLPRI;
2860 consumer_sockpoll[1].fd = client_socket;
2861 consumer_sockpoll[1].events = POLLIN | POLLPRI;
2862
2863 if (lttng_consumer_poll_socket(consumer_sockpoll) < 0) {
2864 goto end;
2865 }
2866 DBG("Connection on client_socket");
2867
2868 /* Blocking call, waiting for transmission */
2869 sock = lttcomm_accept_unix_sock(client_socket);
2870 if (sock < 0) {
2871 WARN("On accept");
2872 goto end;
2873 }
2874 ret = fcntl(sock, F_SETFL, O_NONBLOCK);
2875 if (ret < 0) {
2876 PERROR("fcntl O_NONBLOCK");
2877 goto end;
2878 }
2879
2880 /*
2881 * Setup metadata socket which is the second socket connection on the
2882 * command unix socket.
2883 */
2884 ret = set_metadata_socket(ctx, consumer_sockpoll, client_socket);
2885 if (ret < 0) {
2886 goto end;
2887 }
2888
2889 /* This socket is not useful anymore. */
2890 ret = close(client_socket);
2891 if (ret < 0) {
2892 PERROR("close client_socket");
2893 }
2894 client_socket = -1;
2895
2896 /* update the polling structure to poll on the established socket */
2897 consumer_sockpoll[1].fd = sock;
2898 consumer_sockpoll[1].events = POLLIN | POLLPRI;
2899
2900 while (1) {
2901 if (lttng_consumer_poll_socket(consumer_sockpoll) < 0) {
2902 goto end;
2903 }
2904 DBG("Incoming command on sock");
2905 ret = lttng_consumer_recv_cmd(ctx, sock, consumer_sockpoll);
2906 if (ret == -ENOENT) {
2907 DBG("Received STOP command");
2908 goto end;
2909 }
2910 if (ret <= 0) {
2911 /*
2912 * This could simply be a session daemon quitting. Don't output
2913 * ERR() here.
2914 */
2915 DBG("Communication interrupted on command socket");
2916 goto end;
2917 }
2918 if (consumer_quit) {
2919 DBG("consumer_thread_receive_fds received quit from signal");
2920 goto end;
2921 }
2922 DBG("received command on sock");
2923 }
2924 end:
2925 DBG("Consumer thread sessiond poll exiting");
2926
2927 /*
2928 * Close metadata streams since the producer is the session daemon which
2929 * just died.
2930 *
2931 * NOTE: for now, this only applies to the UST tracer.
2932 */
2933 lttng_consumer_close_metadata();
2934
2935 /*
2936 * when all fds have hung up, the polling thread
2937 * can exit cleanly
2938 */
2939 consumer_quit = 1;
2940
2941 /*
2942 * Notify the data poll thread to poll back again and test the
2943 * consumer_quit state that we just set so to quit gracefully.
2944 */
2945 notify_thread_lttng_pipe(ctx->consumer_data_pipe);
2946
2947 notify_channel_pipe(ctx, NULL, -1, CONSUMER_CHANNEL_QUIT);
2948
2949 /* Cleaning up possibly open sockets. */
2950 if (sock >= 0) {
2951 ret = close(sock);
2952 if (ret < 0) {
2953 PERROR("close sock sessiond poll");
2954 }
2955 }
2956 if (client_socket >= 0) {
2957 ret = close(sock);
2958 if (ret < 0) {
2959 PERROR("close client_socket sessiond poll");
2960 }
2961 }
2962
2963 rcu_unregister_thread();
2964 return NULL;
2965 }
2966
2967 ssize_t lttng_consumer_read_subbuffer(struct lttng_consumer_stream *stream,
2968 struct lttng_consumer_local_data *ctx)
2969 {
2970 ssize_t ret;
2971
2972 pthread_mutex_lock(&stream->lock);
2973
2974 switch (consumer_data.type) {
2975 case LTTNG_CONSUMER_KERNEL:
2976 ret = lttng_kconsumer_read_subbuffer(stream, ctx);
2977 break;
2978 case LTTNG_CONSUMER32_UST:
2979 case LTTNG_CONSUMER64_UST:
2980 ret = lttng_ustconsumer_read_subbuffer(stream, ctx);
2981 break;
2982 default:
2983 ERR("Unknown consumer_data type");
2984 assert(0);
2985 ret = -ENOSYS;
2986 break;
2987 }
2988
2989 pthread_mutex_unlock(&stream->lock);
2990 return ret;
2991 }
2992
2993 int lttng_consumer_on_recv_stream(struct lttng_consumer_stream *stream)
2994 {
2995 switch (consumer_data.type) {
2996 case LTTNG_CONSUMER_KERNEL:
2997 return lttng_kconsumer_on_recv_stream(stream);
2998 case LTTNG_CONSUMER32_UST:
2999 case LTTNG_CONSUMER64_UST:
3000 return lttng_ustconsumer_on_recv_stream(stream);
3001 default:
3002 ERR("Unknown consumer_data type");
3003 assert(0);
3004 return -ENOSYS;
3005 }
3006 }
3007
3008 /*
3009 * Allocate and set consumer data hash tables.
3010 */
3011 void lttng_consumer_init(void)
3012 {
3013 consumer_data.channel_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
3014 consumer_data.relayd_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
3015 consumer_data.stream_list_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
3016 consumer_data.stream_per_chan_id_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
3017 }
3018
3019 /*
3020 * Process the ADD_RELAYD command receive by a consumer.
3021 *
3022 * This will create a relayd socket pair and add it to the relayd hash table.
3023 * The caller MUST acquire a RCU read side lock before calling it.
3024 */
3025 int consumer_add_relayd_socket(int net_seq_idx, int sock_type,
3026 struct lttng_consumer_local_data *ctx, int sock,
3027 struct pollfd *consumer_sockpoll,
3028 struct lttcomm_relayd_sock *relayd_sock, unsigned int sessiond_id)
3029 {
3030 int fd = -1, ret = -1, relayd_created = 0;
3031 enum lttng_error_code ret_code = LTTNG_OK;
3032 struct consumer_relayd_sock_pair *relayd = NULL;
3033
3034 assert(ctx);
3035 assert(relayd_sock);
3036
3037 DBG("Consumer adding relayd socket (idx: %d)", net_seq_idx);
3038
3039 /* First send a status message before receiving the fds. */
3040 ret = consumer_send_status_msg(sock, ret_code);
3041 if (ret < 0) {
3042 /* Somehow, the session daemon is not responding anymore. */
3043 goto error;
3044 }
3045
3046 /* Get relayd reference if exists. */
3047 relayd = consumer_find_relayd(net_seq_idx);
3048 if (relayd == NULL) {
3049 /* Not found. Allocate one. */
3050 relayd = consumer_allocate_relayd_sock_pair(net_seq_idx);
3051 if (relayd == NULL) {
3052 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_OUTFD_ERROR);
3053 ret = -1;
3054 goto error;
3055 }
3056 relayd->sessiond_session_id = (uint64_t) sessiond_id;
3057 relayd_created = 1;
3058 }
3059
3060 /* Poll on consumer socket. */
3061 if (lttng_consumer_poll_socket(consumer_sockpoll) < 0) {
3062 ret = -EINTR;
3063 goto error;
3064 }
3065
3066 /* Get relayd socket from session daemon */
3067 ret = lttcomm_recv_fds_unix_sock(sock, &fd, 1);
3068 if (ret != sizeof(fd)) {
3069 lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_ERROR_RECV_FD);
3070 ret = -1;
3071 fd = -1; /* Just in case it gets set with an invalid value. */
3072 goto error_close;
3073 }
3074
3075 /* We have the fds without error. Send status back. */
3076 ret = consumer_send_status_msg(sock, ret_code);
3077 if (ret < 0) {
3078 /* Somehow, the session daemon is not responding anymore. */
3079 goto error;
3080 }
3081
3082 /* Copy socket information and received FD */
3083 switch (sock_type) {
3084 case LTTNG_STREAM_CONTROL:
3085 /* Copy received lttcomm socket */
3086 lttcomm_copy_sock(&relayd->control_sock.sock, &relayd_sock->sock);
3087 ret = lttcomm_create_sock(&relayd->control_sock.sock);
3088 /* Immediately try to close the created socket if valid. */
3089 if (relayd->control_sock.sock.fd >= 0) {
3090 if (close(relayd->control_sock.sock.fd)) {
3091 PERROR("close relayd control socket");
3092 }
3093 }
3094 /* Handle create_sock error. */
3095 if (ret < 0) {
3096 goto error;
3097 }
3098
3099 /* Assign new file descriptor */
3100 relayd->control_sock.sock.fd = fd;
3101 /* Assign version values. */
3102 relayd->control_sock.major = relayd_sock->major;
3103 relayd->control_sock.minor = relayd_sock->minor;
3104
3105 /*
3106 * Create a session on the relayd and store the returned id. Lock the
3107 * control socket mutex if the relayd was NOT created before.
3108 */
3109 if (!relayd_created) {
3110 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
3111 }
3112 ret = relayd_create_session(&relayd->control_sock,
3113 &relayd->relayd_session_id);
3114 if (!relayd_created) {
3115 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
3116 }
3117 if (ret < 0) {
3118 /*
3119 * Close all sockets of a relayd object. It will be freed if it was
3120 * created at the error code path or else it will be garbage
3121 * collect.
3122 */
3123 (void) relayd_close(&relayd->control_sock);
3124 (void) relayd_close(&relayd->data_sock);
3125 goto error;
3126 }
3127
3128 break;
3129 case LTTNG_STREAM_DATA:
3130 /* Copy received lttcomm socket */
3131 lttcomm_copy_sock(&relayd->data_sock.sock, &relayd_sock->sock);
3132 ret = lttcomm_create_sock(&relayd->data_sock.sock);
3133 /* Immediately try to close the created socket if valid. */
3134 if (relayd->data_sock.sock.fd >= 0) {
3135 if (close(relayd->data_sock.sock.fd)) {
3136 PERROR("close relayd data socket");
3137 }
3138 }
3139 /* Handle create_sock error. */
3140 if (ret < 0) {
3141 goto error;
3142 }
3143
3144 /* Assign new file descriptor */
3145 relayd->data_sock.sock.fd = fd;
3146 /* Assign version values. */
3147 relayd->data_sock.major = relayd_sock->major;
3148 relayd->data_sock.minor = relayd_sock->minor;
3149 break;
3150 default:
3151 ERR("Unknown relayd socket type (%d)", sock_type);
3152 ret = -1;
3153 goto error;
3154 }
3155
3156 DBG("Consumer %s socket created successfully with net idx %" PRIu64 " (fd: %d)",
3157 sock_type == LTTNG_STREAM_CONTROL ? "control" : "data",
3158 relayd->net_seq_idx, fd);
3159
3160 /*
3161 * Add relayd socket pair to consumer data hashtable. If object already
3162 * exists or on error, the function gracefully returns.
3163 */
3164 add_relayd(relayd);
3165
3166 /* All good! */
3167 return 0;
3168
3169 error:
3170 /* Close received socket if valid. */
3171 if (fd >= 0) {
3172 if (close(fd)) {
3173 PERROR("close received socket");
3174 }
3175 }
3176
3177 error_close:
3178 if (relayd_created) {
3179 free(relayd);
3180 }
3181
3182 return ret;
3183 }
3184
3185 /*
3186 * Try to lock the stream mutex.
3187 *
3188 * On success, 1 is returned else 0 indicating that the mutex is NOT lock.
3189 */
3190 static int stream_try_lock(struct lttng_consumer_stream *stream)
3191 {
3192 int ret;
3193
3194 assert(stream);
3195
3196 /*
3197 * Try to lock the stream mutex. On failure, we know that the stream is
3198 * being used else where hence there is data still being extracted.
3199 */
3200 ret = pthread_mutex_trylock(&stream->lock);
3201 if (ret) {
3202 /* For both EBUSY and EINVAL error, the mutex is NOT locked. */
3203 ret = 0;
3204 goto end;
3205 }
3206
3207 ret = 1;
3208
3209 end:
3210 return ret;
3211 }
3212
3213 /*
3214 * Search for a relayd associated to the session id and return the reference.
3215 *
3216 * A rcu read side lock MUST be acquire before calling this function and locked
3217 * until the relayd object is no longer necessary.
3218 */
3219 static struct consumer_relayd_sock_pair *find_relayd_by_session_id(uint64_t id)
3220 {
3221 struct lttng_ht_iter iter;
3222 struct consumer_relayd_sock_pair *relayd = NULL;
3223
3224 /* Iterate over all relayd since they are indexed by net_seq_idx. */
3225 cds_lfht_for_each_entry(consumer_data.relayd_ht->ht, &iter.iter, relayd,
3226 node.node) {
3227 /*
3228 * Check by sessiond id which is unique here where the relayd session
3229 * id might not be when having multiple relayd.
3230 */
3231 if (relayd->sessiond_session_id == id) {
3232 /* Found the relayd. There can be only one per id. */
3233 goto found;
3234 }
3235 }
3236
3237 return NULL;
3238
3239 found:
3240 return relayd;
3241 }
3242
3243 /*
3244 * Check if for a given session id there is still data needed to be extract
3245 * from the buffers.
3246 *
3247 * Return 1 if data is pending or else 0 meaning ready to be read.
3248 */
3249 int consumer_data_pending(uint64_t id)
3250 {
3251 int ret;
3252 struct lttng_ht_iter iter;
3253 struct lttng_ht *ht;
3254 struct lttng_consumer_stream *stream;
3255 struct consumer_relayd_sock_pair *relayd = NULL;
3256 int (*data_pending)(struct lttng_consumer_stream *);
3257
3258 DBG("Consumer data pending command on session id %" PRIu64, id);
3259
3260 rcu_read_lock();
3261 pthread_mutex_lock(&consumer_data.lock);
3262
3263 switch (consumer_data.type) {
3264 case LTTNG_CONSUMER_KERNEL:
3265 data_pending = lttng_kconsumer_data_pending;
3266 break;
3267 case LTTNG_CONSUMER32_UST:
3268 case LTTNG_CONSUMER64_UST:
3269 data_pending = lttng_ustconsumer_data_pending;
3270 break;
3271 default:
3272 ERR("Unknown consumer data type");
3273 assert(0);
3274 }
3275
3276 /* Ease our life a bit */
3277 ht = consumer_data.stream_list_ht;
3278
3279 relayd = find_relayd_by_session_id(id);
3280 if (relayd) {
3281 /* Send init command for data pending. */
3282 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
3283 ret = relayd_begin_data_pending(&relayd->control_sock,
3284 relayd->relayd_session_id);
3285 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
3286 if (ret < 0) {
3287 /* Communication error thus the relayd so no data pending. */
3288 goto data_not_pending;
3289 }
3290 }
3291
3292 cds_lfht_for_each_entry_duplicate(ht->ht,
3293 ht->hash_fct(&id, lttng_ht_seed),
3294 ht->match_fct, &id,
3295 &iter.iter, stream, node_session_id.node) {
3296 /* If this call fails, the stream is being used hence data pending. */
3297 ret = stream_try_lock(stream);
3298 if (!ret) {
3299 goto data_pending;
3300 }
3301
3302 /*
3303 * A removed node from the hash table indicates that the stream has
3304 * been deleted thus having a guarantee that the buffers are closed
3305 * on the consumer side. However, data can still be transmitted
3306 * over the network so don't skip the relayd check.
3307 */
3308 ret = cds_lfht_is_node_deleted(&stream->node.node);
3309 if (!ret) {
3310 /* Check the stream if there is data in the buffers. */
3311 ret = data_pending(stream);
3312 if (ret == 1) {
3313 pthread_mutex_unlock(&stream->lock);
3314 goto data_pending;
3315 }
3316 }
3317
3318 /* Relayd check */
3319 if (relayd) {
3320 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
3321 if (stream->metadata_flag) {
3322 ret = relayd_quiescent_control(&relayd->control_sock,
3323 stream->relayd_stream_id);
3324 } else {
3325 ret = relayd_data_pending(&relayd->control_sock,
3326 stream->relayd_stream_id,
3327 stream->next_net_seq_num - 1);
3328 }
3329 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
3330 if (ret == 1) {
3331 pthread_mutex_unlock(&stream->lock);
3332 goto data_pending;
3333 }
3334 }
3335 pthread_mutex_unlock(&stream->lock);
3336 }
3337
3338 if (relayd) {
3339 unsigned int is_data_inflight = 0;
3340
3341 /* Send init command for data pending. */
3342 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
3343 ret = relayd_end_data_pending(&relayd->control_sock,
3344 relayd->relayd_session_id, &is_data_inflight);
3345 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
3346 if (ret < 0) {
3347 goto data_not_pending;
3348 }
3349 if (is_data_inflight) {
3350 goto data_pending;
3351 }
3352 }
3353
3354 /*
3355 * Finding _no_ node in the hash table and no inflight data means that the
3356 * stream(s) have been removed thus data is guaranteed to be available for
3357 * analysis from the trace files.
3358 */
3359
3360 data_not_pending:
3361 /* Data is available to be read by a viewer. */
3362 pthread_mutex_unlock(&consumer_data.lock);
3363 rcu_read_unlock();
3364 return 0;
3365
3366 data_pending:
3367 /* Data is still being extracted from buffers. */
3368 pthread_mutex_unlock(&consumer_data.lock);
3369 rcu_read_unlock();
3370 return 1;
3371 }
3372
3373 /*
3374 * Send a ret code status message to the sessiond daemon.
3375 *
3376 * Return the sendmsg() return value.
3377 */
3378 int consumer_send_status_msg(int sock, int ret_code)
3379 {
3380 struct lttcomm_consumer_status_msg msg;
3381
3382 msg.ret_code = ret_code;
3383
3384 return lttcomm_send_unix_sock(sock, &msg, sizeof(msg));
3385 }
3386
3387 /*
3388 * Send a channel status message to the sessiond daemon.
3389 *
3390 * Return the sendmsg() return value.
3391 */
3392 int consumer_send_status_channel(int sock,
3393 struct lttng_consumer_channel *channel)
3394 {
3395 struct lttcomm_consumer_status_channel msg;
3396
3397 assert(sock >= 0);
3398
3399 if (!channel) {
3400 msg.ret_code = -LTTNG_ERR_UST_CHAN_FAIL;
3401 } else {
3402 msg.ret_code = LTTNG_OK;
3403 msg.key = channel->key;
3404 msg.stream_count = channel->streams.count;
3405 }
3406
3407 return lttcomm_send_unix_sock(sock, &msg, sizeof(msg));
3408 }
This page took 0.100311 seconds and 4 git commands to generate.