Fix: consumer: unbalanced RCU read-side lock on error
[lttng-tools.git] / src / common / consumer / consumer-stream.c
1 /*
2 * Copyright (C) 2011 Julien Desfossez <julien.desfossez@polymtl.ca>
3 * Copyright (C) 2011 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
4 * Copyright (C) 2013 David Goulet <dgoulet@efficios.com>
5 *
6 * SPDX-License-Identifier: GPL-2.0-only
7 *
8 */
9
10 #define _LGPL_SOURCE
11 #include <assert.h>
12 #include <inttypes.h>
13 #include <sys/mman.h>
14 #include <unistd.h>
15
16 #include <common/common.h>
17 #include <common/index/index.h>
18 #include <common/kernel-consumer/kernel-consumer.h>
19 #include <common/relayd/relayd.h>
20 #include <common/ust-consumer/ust-consumer.h>
21 #include <common/utils.h>
22 #include <common/consumer/consumer.h>
23 #include <common/consumer/consumer-timer.h>
24 #include <common/consumer/metadata-bucket.h>
25 #include <common/kernel-ctl/kernel-ctl.h>
26
27 #include "consumer-stream.h"
28
29 /*
30 * RCU call to free stream. MUST only be used with call_rcu().
31 */
32 static void free_stream_rcu(struct rcu_head *head)
33 {
34 struct lttng_ht_node_u64 *node =
35 caa_container_of(head, struct lttng_ht_node_u64, head);
36 struct lttng_consumer_stream *stream =
37 caa_container_of(node, struct lttng_consumer_stream, node);
38
39 pthread_mutex_destroy(&stream->lock);
40 free(stream);
41 }
42
43 static void consumer_stream_data_lock_all(struct lttng_consumer_stream *stream)
44 {
45 pthread_mutex_lock(&stream->chan->lock);
46 pthread_mutex_lock(&stream->lock);
47 }
48
49 static void consumer_stream_data_unlock_all(struct lttng_consumer_stream *stream)
50 {
51 pthread_mutex_unlock(&stream->lock);
52 pthread_mutex_unlock(&stream->chan->lock);
53 }
54
55 static void consumer_stream_metadata_lock_all(struct lttng_consumer_stream *stream)
56 {
57 consumer_stream_data_lock_all(stream);
58 pthread_mutex_lock(&stream->metadata_rdv_lock);
59 }
60
61 static void consumer_stream_metadata_unlock_all(struct lttng_consumer_stream *stream)
62 {
63 pthread_mutex_unlock(&stream->metadata_rdv_lock);
64 consumer_stream_data_unlock_all(stream);
65 }
66
67 /* Only used for data streams. */
68 static int consumer_stream_update_stats(struct lttng_consumer_stream *stream,
69 const struct stream_subbuffer *subbuf)
70 {
71 int ret = 0;
72 uint64_t sequence_number;
73 const uint64_t discarded_events = subbuf->info.data.events_discarded;
74
75 if (!subbuf->info.data.sequence_number.is_set) {
76 /* Command not supported by the tracer. */
77 sequence_number = -1ULL;
78 stream->sequence_number_unavailable = true;
79 } else {
80 sequence_number = subbuf->info.data.sequence_number.value;
81 }
82
83 /*
84 * Start the sequence when we extract the first packet in case we don't
85 * start at 0 (for example if a consumer is not connected to the
86 * session immediately after the beginning).
87 */
88 if (stream->last_sequence_number == -1ULL) {
89 stream->last_sequence_number = sequence_number;
90 } else if (sequence_number > stream->last_sequence_number) {
91 stream->chan->lost_packets += sequence_number -
92 stream->last_sequence_number - 1;
93 } else {
94 /* seq <= last_sequence_number */
95 ERR("Sequence number inconsistent : prev = %" PRIu64
96 ", current = %" PRIu64,
97 stream->last_sequence_number, sequence_number);
98 ret = -1;
99 goto end;
100 }
101 stream->last_sequence_number = sequence_number;
102
103 if (discarded_events < stream->last_discarded_events) {
104 /*
105 * Overflow has occurred. We assume only one wrap-around
106 * has occurred.
107 */
108 stream->chan->discarded_events +=
109 (1ULL << (CAA_BITS_PER_LONG - 1)) -
110 stream->last_discarded_events +
111 discarded_events;
112 } else {
113 stream->chan->discarded_events += discarded_events -
114 stream->last_discarded_events;
115 }
116 stream->last_discarded_events = discarded_events;
117 ret = 0;
118
119 end:
120 return ret;
121 }
122
123 static
124 void ctf_packet_index_populate(struct ctf_packet_index *index,
125 off_t offset, const struct stream_subbuffer *subbuffer)
126 {
127 *index = (typeof(*index)){
128 .offset = htobe64(offset),
129 .packet_size = htobe64(subbuffer->info.data.packet_size),
130 .content_size = htobe64(subbuffer->info.data.content_size),
131 .timestamp_begin = htobe64(
132 subbuffer->info.data.timestamp_begin),
133 .timestamp_end = htobe64(
134 subbuffer->info.data.timestamp_end),
135 .events_discarded = htobe64(
136 subbuffer->info.data.events_discarded),
137 .stream_id = htobe64(subbuffer->info.data.stream_id),
138 .stream_instance_id = htobe64(
139 subbuffer->info.data.stream_instance_id.is_set ?
140 subbuffer->info.data.stream_instance_id.value : -1ULL),
141 .packet_seq_num = htobe64(
142 subbuffer->info.data.sequence_number.is_set ?
143 subbuffer->info.data.sequence_number.value : -1ULL),
144 };
145 }
146
147 static ssize_t consumer_stream_consume_mmap(
148 struct lttng_consumer_local_data *ctx,
149 struct lttng_consumer_stream *stream,
150 const struct stream_subbuffer *subbuffer)
151 {
152 const unsigned long padding_size =
153 subbuffer->info.data.padded_subbuf_size -
154 subbuffer->info.data.subbuf_size;
155 const ssize_t written_bytes = lttng_consumer_on_read_subbuffer_mmap(
156 stream, &subbuffer->buffer.buffer, padding_size);
157
158 if (stream->net_seq_idx == -1ULL) {
159 /*
160 * When writing on disk, check that only the subbuffer (no
161 * padding) was written to disk.
162 */
163 if (written_bytes != subbuffer->info.data.padded_subbuf_size) {
164 DBG("Failed to write the entire padded subbuffer on disk (written_bytes: %zd, padded subbuffer size %lu)",
165 written_bytes,
166 subbuffer->info.data.padded_subbuf_size);
167 }
168 } else {
169 /*
170 * When streaming over the network, check that the entire
171 * subbuffer including padding was successfully written.
172 */
173 if (written_bytes != subbuffer->info.data.subbuf_size) {
174 DBG("Failed to write only the subbuffer over the network (written_bytes: %zd, subbuffer size %lu)",
175 written_bytes,
176 subbuffer->info.data.subbuf_size);
177 }
178 }
179
180 /*
181 * If `lttng_consumer_on_read_subbuffer_mmap()` returned an error, pass
182 * it along to the caller, else return zero.
183 */
184 if (written_bytes < 0) {
185 ERR("Error reading mmap subbuffer: %zd", written_bytes);
186 }
187
188 return written_bytes;
189 }
190
191 static ssize_t consumer_stream_consume_splice(
192 struct lttng_consumer_local_data *ctx,
193 struct lttng_consumer_stream *stream,
194 const struct stream_subbuffer *subbuffer)
195 {
196 const ssize_t written_bytes = lttng_consumer_on_read_subbuffer_splice(
197 ctx, stream, subbuffer->info.data.padded_subbuf_size, 0);
198
199 if (written_bytes != subbuffer->info.data.padded_subbuf_size) {
200 DBG("Failed to write the entire padded subbuffer (written_bytes: %zd, padded subbuffer size %lu)",
201 written_bytes,
202 subbuffer->info.data.padded_subbuf_size);
203 }
204
205 /*
206 * If `lttng_consumer_on_read_subbuffer_splice()` returned an error,
207 * pass it along to the caller, else return zero.
208 */
209 if (written_bytes < 0) {
210 ERR("Error reading splice subbuffer: %zd", written_bytes);
211 }
212
213 return written_bytes;
214 }
215
216 static int consumer_stream_send_index(
217 struct lttng_consumer_stream *stream,
218 const struct stream_subbuffer *subbuffer,
219 struct lttng_consumer_local_data *ctx)
220 {
221 off_t packet_offset = 0;
222 struct ctf_packet_index index = {};
223
224 /*
225 * This is called after consuming the sub-buffer; substract the
226 * effect this sub-buffer from the offset.
227 */
228 if (stream->net_seq_idx == (uint64_t) -1ULL) {
229 packet_offset = stream->out_fd_offset -
230 subbuffer->info.data.padded_subbuf_size;
231 }
232
233 ctf_packet_index_populate(&index, packet_offset, subbuffer);
234 return consumer_stream_write_index(stream, &index);
235 }
236
237 /*
238 * Actually do the metadata sync using the given metadata stream.
239 *
240 * Return 0 on success else a negative value. ENODATA can be returned also
241 * indicating that there is no metadata available for that stream.
242 */
243 static int do_sync_metadata(struct lttng_consumer_stream *metadata,
244 struct lttng_consumer_local_data *ctx)
245 {
246 int ret;
247 enum sync_metadata_status status;
248
249 assert(metadata);
250 assert(metadata->metadata_flag);
251 assert(ctx);
252
253 /*
254 * In UST, since we have to write the metadata from the cache packet
255 * by packet, we might need to start this procedure multiple times
256 * until all the metadata from the cache has been extracted.
257 */
258 do {
259 /*
260 * Steps :
261 * - Lock the metadata stream
262 * - Check if metadata stream node was deleted before locking.
263 * - if yes, release and return success
264 * - Check if new metadata is ready (flush + snapshot pos)
265 * - If nothing : release and return.
266 * - Lock the metadata_rdv_lock
267 * - Unlock the metadata stream
268 * - cond_wait on metadata_rdv to wait the wakeup from the
269 * metadata thread
270 * - Unlock the metadata_rdv_lock
271 */
272 pthread_mutex_lock(&metadata->lock);
273
274 /*
275 * There is a possibility that we were able to acquire a reference on the
276 * stream from the RCU hash table but between then and now, the node might
277 * have been deleted just before the lock is acquired. Thus, after locking,
278 * we make sure the metadata node has not been deleted which means that the
279 * buffers are closed.
280 *
281 * In that case, there is no need to sync the metadata hence returning a
282 * success return code.
283 */
284 ret = cds_lfht_is_node_deleted(&metadata->node.node);
285 if (ret) {
286 ret = 0;
287 goto end_unlock_mutex;
288 }
289
290 switch (ctx->type) {
291 case LTTNG_CONSUMER_KERNEL:
292 /*
293 * Empty the metadata cache and flush the current stream.
294 */
295 status = lttng_kconsumer_sync_metadata(metadata);
296 break;
297 case LTTNG_CONSUMER32_UST:
298 case LTTNG_CONSUMER64_UST:
299 /*
300 * Ask the sessiond if we have new metadata waiting and update the
301 * consumer metadata cache.
302 */
303 status = lttng_ustconsumer_sync_metadata(ctx, metadata);
304 break;
305 default:
306 abort();
307 }
308
309 switch (status) {
310 case SYNC_METADATA_STATUS_NEW_DATA:
311 break;
312 case SYNC_METADATA_STATUS_NO_DATA:
313 ret = 0;
314 goto end_unlock_mutex;
315 case SYNC_METADATA_STATUS_ERROR:
316 ret = -1;
317 goto end_unlock_mutex;
318 default:
319 abort();
320 }
321
322 /*
323 * At this point, new metadata have been flushed, so we wait on the
324 * rendez-vous point for the metadata thread to wake us up when it
325 * finishes consuming the metadata and continue execution.
326 */
327
328 pthread_mutex_lock(&metadata->metadata_rdv_lock);
329
330 /*
331 * Release metadata stream lock so the metadata thread can process it.
332 */
333 pthread_mutex_unlock(&metadata->lock);
334
335 /*
336 * Wait on the rendez-vous point. Once woken up, it means the metadata was
337 * consumed and thus synchronization is achieved.
338 */
339 pthread_cond_wait(&metadata->metadata_rdv, &metadata->metadata_rdv_lock);
340 pthread_mutex_unlock(&metadata->metadata_rdv_lock);
341 } while (status == SYNC_METADATA_STATUS_NEW_DATA);
342
343 /* Success */
344 return 0;
345
346 end_unlock_mutex:
347 pthread_mutex_unlock(&metadata->lock);
348 return ret;
349 }
350
351 /*
352 * Synchronize the metadata using a given session ID. A successful acquisition
353 * of a metadata stream will trigger a request to the session daemon and a
354 * snapshot so the metadata thread can consume it.
355 *
356 * This function call is a rendez-vous point between the metadata thread and
357 * the data thread.
358 *
359 * Return 0 on success or else a negative value.
360 */
361 int consumer_stream_sync_metadata(struct lttng_consumer_local_data *ctx,
362 uint64_t session_id)
363 {
364 int ret;
365 struct lttng_consumer_stream *stream = NULL;
366 struct lttng_ht_iter iter;
367 struct lttng_ht *ht;
368
369 assert(ctx);
370
371 /* Ease our life a bit. */
372 ht = the_consumer_data.stream_list_ht;
373
374 rcu_read_lock();
375
376 /* Search the metadata associated with the session id of the given stream. */
377
378 cds_lfht_for_each_entry_duplicate(ht->ht,
379 ht->hash_fct(&session_id, lttng_ht_seed), ht->match_fct,
380 &session_id, &iter.iter, stream, node_session_id.node) {
381 if (!stream->metadata_flag) {
382 continue;
383 }
384
385 ret = do_sync_metadata(stream, ctx);
386 if (ret < 0) {
387 goto end;
388 }
389 }
390
391 /*
392 * Force return code to 0 (success) since ret might be ENODATA for instance
393 * which is not an error but rather that we should come back.
394 */
395 ret = 0;
396
397 end:
398 rcu_read_unlock();
399 return ret;
400 }
401
402 static int consumer_stream_sync_metadata_index(
403 struct lttng_consumer_stream *stream,
404 const struct stream_subbuffer *subbuffer,
405 struct lttng_consumer_local_data *ctx)
406 {
407 int ret;
408
409 /* Block until all the metadata is sent. */
410 pthread_mutex_lock(&stream->metadata_timer_lock);
411 assert(!stream->missed_metadata_flush);
412 stream->waiting_on_metadata = true;
413 pthread_mutex_unlock(&stream->metadata_timer_lock);
414
415 ret = consumer_stream_sync_metadata(ctx, stream->session_id);
416
417 pthread_mutex_lock(&stream->metadata_timer_lock);
418 stream->waiting_on_metadata = false;
419 if (stream->missed_metadata_flush) {
420 stream->missed_metadata_flush = false;
421 pthread_mutex_unlock(&stream->metadata_timer_lock);
422 (void) stream->read_subbuffer_ops.send_live_beacon(stream);
423 } else {
424 pthread_mutex_unlock(&stream->metadata_timer_lock);
425 }
426 if (ret < 0) {
427 goto end;
428 }
429
430 ret = consumer_stream_send_index(stream, subbuffer, ctx);
431 end:
432 return ret;
433 }
434
435 /*
436 * Check if the local version of the metadata stream matches with the version
437 * of the metadata stream in the kernel. If it was updated, set the reset flag
438 * on the stream.
439 */
440 static
441 int metadata_stream_check_version(struct lttng_consumer_stream *stream,
442 const struct stream_subbuffer *subbuffer)
443 {
444 if (stream->metadata_version == subbuffer->info.metadata.version) {
445 goto end;
446 }
447
448 DBG("New metadata version detected");
449 consumer_stream_metadata_set_version(stream,
450 subbuffer->info.metadata.version);
451
452 if (stream->read_subbuffer_ops.reset_metadata) {
453 stream->read_subbuffer_ops.reset_metadata(stream);
454 }
455
456 end:
457 return 0;
458 }
459
460 static
461 bool stream_is_rotating_to_null_chunk(
462 const struct lttng_consumer_stream *stream)
463 {
464 bool rotating_to_null_chunk = false;
465
466 if (stream->rotate_position == -1ULL) {
467 /* No rotation ongoing. */
468 goto end;
469 }
470
471 if (stream->trace_chunk == stream->chan->trace_chunk ||
472 !stream->chan->trace_chunk) {
473 rotating_to_null_chunk = true;
474 }
475 end:
476 return rotating_to_null_chunk;
477 }
478
479 enum consumer_stream_open_packet_status consumer_stream_open_packet(
480 struct lttng_consumer_stream *stream)
481 {
482 int ret;
483 enum consumer_stream_open_packet_status status;
484 unsigned long produced_pos_before, produced_pos_after;
485
486 ret = lttng_consumer_sample_snapshot_positions(stream);
487 if (ret < 0) {
488 ERR("Failed to snapshot positions before post-rotation empty packet flush: stream id = %" PRIu64
489 ", channel name = %s, session id = %" PRIu64,
490 stream->key, stream->chan->name,
491 stream->chan->session_id);
492 status = CONSUMER_STREAM_OPEN_PACKET_STATUS_ERROR;
493 goto end;
494 }
495
496 ret = lttng_consumer_get_produced_snapshot(
497 stream, &produced_pos_before);
498 if (ret < 0) {
499 ERR("Failed to read produced position before post-rotation empty packet flush: stream id = %" PRIu64
500 ", channel name = %s, session id = %" PRIu64,
501 stream->key, stream->chan->name,
502 stream->chan->session_id);
503 status = CONSUMER_STREAM_OPEN_PACKET_STATUS_ERROR;
504 goto end;
505 }
506
507 ret = consumer_stream_flush_buffer(stream, 0);
508 if (ret) {
509 ERR("Failed to flush an empty packet at rotation point: stream id = %" PRIu64
510 ", channel name = %s, session id = %" PRIu64,
511 stream->key, stream->chan->name,
512 stream->chan->session_id);
513 status = CONSUMER_STREAM_OPEN_PACKET_STATUS_ERROR;
514 goto end;
515 }
516
517 ret = lttng_consumer_sample_snapshot_positions(stream);
518 if (ret < 0) {
519 ERR("Failed to snapshot positions after post-rotation empty packet flush: stream id = %" PRIu64
520 ", channel name = %s, session id = %" PRIu64,
521 stream->key, stream->chan->name,
522 stream->chan->session_id);
523 status = CONSUMER_STREAM_OPEN_PACKET_STATUS_ERROR;
524 goto end;
525 }
526
527 ret = lttng_consumer_get_produced_snapshot(stream, &produced_pos_after);
528 if (ret < 0) {
529 ERR("Failed to read produced position after post-rotation empty packet flush: stream id = %" PRIu64
530 ", channel name = %s, session id = %" PRIu64,
531 stream->key, stream->chan->name,
532 stream->chan->session_id);
533 status = CONSUMER_STREAM_OPEN_PACKET_STATUS_ERROR;
534 goto end;
535 }
536
537 /*
538 * Determine if the flush had an effect by comparing the produced
539 * positons before and after the flush.
540 */
541 status = produced_pos_before != produced_pos_after ?
542 CONSUMER_STREAM_OPEN_PACKET_STATUS_OPENED :
543 CONSUMER_STREAM_OPEN_PACKET_STATUS_NO_SPACE;
544 if (status == CONSUMER_STREAM_OPEN_PACKET_STATUS_OPENED) {
545 stream->opened_packet_in_current_trace_chunk = true;
546 }
547
548 end:
549 return status;
550 }
551
552 /*
553 * An attempt to open a new packet is performed after a rotation completes to
554 * get a begin timestamp as close as possible to the rotation point.
555 *
556 * However, that initial attempt at opening a packet can fail due to a full
557 * ring-buffer. In that case, a second attempt is performed after consuming
558 * a packet since that will have freed enough space in the ring-buffer.
559 */
560 static
561 int post_consume_open_new_packet(struct lttng_consumer_stream *stream,
562 const struct stream_subbuffer *subbuffer,
563 struct lttng_consumer_local_data *ctx)
564 {
565 int ret = 0;
566
567 if (!stream->opened_packet_in_current_trace_chunk &&
568 stream->trace_chunk &&
569 !stream_is_rotating_to_null_chunk(stream)) {
570 const enum consumer_stream_open_packet_status status =
571 consumer_stream_open_packet(stream);
572
573 switch (status) {
574 case CONSUMER_STREAM_OPEN_PACKET_STATUS_OPENED:
575 DBG("Opened a packet after consuming a packet rotation: stream id = %" PRIu64
576 ", channel name = %s, session id = %" PRIu64,
577 stream->key, stream->chan->name,
578 stream->chan->session_id);
579 stream->opened_packet_in_current_trace_chunk = true;
580 break;
581 case CONSUMER_STREAM_OPEN_PACKET_STATUS_NO_SPACE:
582 /*
583 * Can't open a packet as there is no space left.
584 * This means that new events were produced, resulting
585 * in a packet being opened, which is what we want
586 * anyhow.
587 */
588 DBG("No space left to open a packet after consuming a packet: stream id = %" PRIu64
589 ", channel name = %s, session id = %" PRIu64,
590 stream->key, stream->chan->name,
591 stream->chan->session_id);
592 stream->opened_packet_in_current_trace_chunk = true;
593 break;
594 case CONSUMER_STREAM_OPEN_PACKET_STATUS_ERROR:
595 /* Logged by callee. */
596 ret = -1;
597 goto end;
598 default:
599 abort();
600 }
601
602 stream->opened_packet_in_current_trace_chunk = true;
603 }
604
605 end:
606 return ret;
607 }
608
609 struct lttng_consumer_stream *consumer_stream_create(
610 struct lttng_consumer_channel *channel,
611 uint64_t channel_key,
612 uint64_t stream_key,
613 const char *channel_name,
614 uint64_t relayd_id,
615 uint64_t session_id,
616 struct lttng_trace_chunk *trace_chunk,
617 int cpu,
618 int *alloc_ret,
619 enum consumer_channel_type type,
620 unsigned int monitor)
621 {
622 int ret;
623 struct lttng_consumer_stream *stream;
624
625 stream = zmalloc(sizeof(*stream));
626 if (stream == NULL) {
627 PERROR("malloc struct lttng_consumer_stream");
628 ret = -ENOMEM;
629 goto end;
630 }
631
632 rcu_read_lock();
633
634 if (trace_chunk && !lttng_trace_chunk_get(trace_chunk)) {
635 ERR("Failed to acquire trace chunk reference during the creation of a stream");
636 ret = -1;
637 goto error;
638 }
639
640 stream->chan = channel;
641 stream->key = stream_key;
642 stream->trace_chunk = trace_chunk;
643 stream->out_fd = -1;
644 stream->out_fd_offset = 0;
645 stream->output_written = 0;
646 stream->net_seq_idx = relayd_id;
647 stream->session_id = session_id;
648 stream->monitor = monitor;
649 stream->endpoint_status = CONSUMER_ENDPOINT_ACTIVE;
650 stream->index_file = NULL;
651 stream->last_sequence_number = -1ULL;
652 stream->rotate_position = -1ULL;
653 /* Buffer is created with an open packet. */
654 stream->opened_packet_in_current_trace_chunk = true;
655 pthread_mutex_init(&stream->lock, NULL);
656 pthread_mutex_init(&stream->metadata_timer_lock, NULL);
657
658 /* If channel is the metadata, flag this stream as metadata. */
659 if (type == CONSUMER_CHANNEL_TYPE_METADATA) {
660 stream->metadata_flag = 1;
661 /* Metadata is flat out. */
662 strncpy(stream->name, DEFAULT_METADATA_NAME, sizeof(stream->name));
663 /* Live rendez-vous point. */
664 pthread_cond_init(&stream->metadata_rdv, NULL);
665 pthread_mutex_init(&stream->metadata_rdv_lock, NULL);
666 } else {
667 /* Format stream name to <channel_name>_<cpu_number> */
668 ret = snprintf(stream->name, sizeof(stream->name), "%s_%d",
669 channel_name, cpu);
670 if (ret < 0) {
671 PERROR("snprintf stream name");
672 goto error;
673 }
674 }
675
676 switch (channel->output) {
677 case CONSUMER_CHANNEL_SPLICE:
678 stream->output = LTTNG_EVENT_SPLICE;
679 ret = utils_create_pipe(stream->splice_pipe);
680 if (ret < 0) {
681 goto error;
682 }
683 break;
684 case CONSUMER_CHANNEL_MMAP:
685 stream->output = LTTNG_EVENT_MMAP;
686 break;
687 default:
688 abort();
689 }
690
691 /* Key is always the wait_fd for streams. */
692 lttng_ht_node_init_u64(&stream->node, stream->key);
693
694 /* Init node per channel id key */
695 lttng_ht_node_init_u64(&stream->node_channel_id, channel_key);
696
697 /* Init session id node with the stream session id */
698 lttng_ht_node_init_u64(&stream->node_session_id, stream->session_id);
699
700 DBG3("Allocated stream %s (key %" PRIu64 ", chan_key %" PRIu64
701 " relayd_id %" PRIu64 ", session_id %" PRIu64,
702 stream->name, stream->key, channel_key,
703 stream->net_seq_idx, stream->session_id);
704
705 rcu_read_unlock();
706
707 lttng_dynamic_array_init(&stream->read_subbuffer_ops.post_consume_cbs,
708 sizeof(post_consume_cb), NULL);
709
710 if (type == CONSUMER_CHANNEL_TYPE_METADATA) {
711 stream->read_subbuffer_ops.lock =
712 consumer_stream_metadata_lock_all;
713 stream->read_subbuffer_ops.unlock =
714 consumer_stream_metadata_unlock_all;
715 stream->read_subbuffer_ops.pre_consume_subbuffer =
716 metadata_stream_check_version;
717 } else {
718 const post_consume_cb post_consume_index_op = channel->is_live ?
719 consumer_stream_sync_metadata_index :
720 consumer_stream_send_index;
721
722 ret = lttng_dynamic_array_add_element(
723 &stream->read_subbuffer_ops.post_consume_cbs,
724 &post_consume_index_op);
725 if (ret) {
726 PERROR("Failed to add `send index` callback to stream's post consumption callbacks");
727 goto error;
728 }
729
730 ret = lttng_dynamic_array_add_element(
731 &stream->read_subbuffer_ops.post_consume_cbs,
732 &(post_consume_cb) { post_consume_open_new_packet });
733 if (ret) {
734 PERROR("Failed to add `open new packet` callback to stream's post consumption callbacks");
735 goto error;
736 }
737
738 stream->read_subbuffer_ops.lock = consumer_stream_data_lock_all;
739 stream->read_subbuffer_ops.unlock =
740 consumer_stream_data_unlock_all;
741 stream->read_subbuffer_ops.pre_consume_subbuffer =
742 consumer_stream_update_stats;
743 }
744
745 if (channel->output == CONSUMER_CHANNEL_MMAP) {
746 stream->read_subbuffer_ops.consume_subbuffer =
747 consumer_stream_consume_mmap;
748 } else {
749 stream->read_subbuffer_ops.consume_subbuffer =
750 consumer_stream_consume_splice;
751 }
752
753 return stream;
754
755 error:
756 rcu_read_unlock();
757 lttng_trace_chunk_put(stream->trace_chunk);
758 lttng_dynamic_array_reset(&stream->read_subbuffer_ops.post_consume_cbs);
759 free(stream);
760 end:
761 if (alloc_ret) {
762 *alloc_ret = ret;
763 }
764 return NULL;
765 }
766
767 /*
768 * Close stream on the relayd side. This call can destroy a relayd if the
769 * conditions are met.
770 *
771 * A RCU read side lock MUST be acquired if the relayd object was looked up in
772 * a hash table before calling this.
773 */
774 void consumer_stream_relayd_close(struct lttng_consumer_stream *stream,
775 struct consumer_relayd_sock_pair *relayd)
776 {
777 int ret;
778
779 assert(stream);
780 assert(relayd);
781
782 if (stream->sent_to_relayd) {
783 uatomic_dec(&relayd->refcount);
784 assert(uatomic_read(&relayd->refcount) >= 0);
785 }
786
787 /* Closing streams requires to lock the control socket. */
788 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
789 ret = relayd_send_close_stream(&relayd->control_sock,
790 stream->relayd_stream_id,
791 stream->next_net_seq_num - 1);
792 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
793 if (ret < 0) {
794 ERR("Relayd send close stream failed. Cleaning up relayd %" PRIu64 ".", relayd->net_seq_idx);
795 lttng_consumer_cleanup_relayd(relayd);
796 }
797
798 /* Both conditions are met, we destroy the relayd. */
799 if (uatomic_read(&relayd->refcount) == 0 &&
800 uatomic_read(&relayd->destroy_flag)) {
801 consumer_destroy_relayd(relayd);
802 }
803 stream->net_seq_idx = (uint64_t) -1ULL;
804 stream->sent_to_relayd = 0;
805 }
806
807 /*
808 * Close stream's file descriptors and, if needed, close stream also on the
809 * relayd side.
810 *
811 * The consumer data lock MUST be acquired.
812 * The stream lock MUST be acquired.
813 */
814 void consumer_stream_close(struct lttng_consumer_stream *stream)
815 {
816 int ret;
817 struct consumer_relayd_sock_pair *relayd;
818
819 assert(stream);
820
821 switch (the_consumer_data.type) {
822 case LTTNG_CONSUMER_KERNEL:
823 if (stream->mmap_base != NULL) {
824 ret = munmap(stream->mmap_base, stream->mmap_len);
825 if (ret != 0) {
826 PERROR("munmap");
827 }
828 }
829
830 if (stream->wait_fd >= 0) {
831 ret = close(stream->wait_fd);
832 if (ret) {
833 PERROR("close");
834 }
835 stream->wait_fd = -1;
836 }
837 if (stream->chan->output == CONSUMER_CHANNEL_SPLICE) {
838 utils_close_pipe(stream->splice_pipe);
839 }
840 break;
841 case LTTNG_CONSUMER32_UST:
842 case LTTNG_CONSUMER64_UST:
843 {
844 /*
845 * Special case for the metadata since the wait fd is an internal pipe
846 * polled in the metadata thread.
847 */
848 if (stream->metadata_flag && stream->chan->monitor) {
849 int rpipe = stream->ust_metadata_poll_pipe[0];
850
851 /*
852 * This will stop the channel timer if one and close the write side
853 * of the metadata poll pipe.
854 */
855 lttng_ustconsumer_close_metadata(stream->chan);
856 if (rpipe >= 0) {
857 ret = close(rpipe);
858 if (ret < 0) {
859 PERROR("closing metadata pipe read side");
860 }
861 stream->ust_metadata_poll_pipe[0] = -1;
862 }
863 }
864 break;
865 }
866 default:
867 ERR("Unknown consumer_data type");
868 assert(0);
869 }
870
871 /* Close output fd. Could be a socket or local file at this point. */
872 if (stream->out_fd >= 0) {
873 ret = close(stream->out_fd);
874 if (ret) {
875 PERROR("close");
876 }
877 stream->out_fd = -1;
878 }
879
880 if (stream->index_file) {
881 lttng_index_file_put(stream->index_file);
882 stream->index_file = NULL;
883 }
884
885 lttng_trace_chunk_put(stream->trace_chunk);
886 stream->trace_chunk = NULL;
887
888 /* Check and cleanup relayd if needed. */
889 rcu_read_lock();
890 relayd = consumer_find_relayd(stream->net_seq_idx);
891 if (relayd != NULL) {
892 consumer_stream_relayd_close(stream, relayd);
893 }
894 rcu_read_unlock();
895 }
896
897 /*
898 * Delete the stream from all possible hash tables.
899 *
900 * The consumer data lock MUST be acquired.
901 * The stream lock MUST be acquired.
902 */
903 void consumer_stream_delete(struct lttng_consumer_stream *stream,
904 struct lttng_ht *ht)
905 {
906 int ret;
907 struct lttng_ht_iter iter;
908
909 assert(stream);
910 /* Should NEVER be called not in monitor mode. */
911 assert(stream->chan->monitor);
912
913 rcu_read_lock();
914
915 if (ht) {
916 iter.iter.node = &stream->node.node;
917 ret = lttng_ht_del(ht, &iter);
918 assert(!ret);
919 }
920
921 /* Delete from stream per channel ID hash table. */
922 iter.iter.node = &stream->node_channel_id.node;
923 /*
924 * The returned value is of no importance. Even if the node is NOT in the
925 * hash table, we continue since we may have been called by a code path
926 * that did not add the stream to a (all) hash table. Same goes for the
927 * next call ht del call.
928 */
929 (void) lttng_ht_del(the_consumer_data.stream_per_chan_id_ht, &iter);
930
931 /* Delete from the global stream list. */
932 iter.iter.node = &stream->node_session_id.node;
933 /* See the previous ht del on why we ignore the returned value. */
934 (void) lttng_ht_del(the_consumer_data.stream_list_ht, &iter);
935
936 rcu_read_unlock();
937
938 if (!stream->metadata_flag) {
939 /* Decrement the stream count of the global consumer data. */
940 assert(the_consumer_data.stream_count > 0);
941 the_consumer_data.stream_count--;
942 }
943 }
944
945 /*
946 * Free the given stream within a RCU call.
947 */
948 void consumer_stream_free(struct lttng_consumer_stream *stream)
949 {
950 assert(stream);
951
952 metadata_bucket_destroy(stream->metadata_bucket);
953 call_rcu(&stream->node.head, free_stream_rcu);
954 }
955
956 /*
957 * Destroy the stream's buffers of the tracer.
958 */
959 void consumer_stream_destroy_buffers(struct lttng_consumer_stream *stream)
960 {
961 assert(stream);
962
963 switch (the_consumer_data.type) {
964 case LTTNG_CONSUMER_KERNEL:
965 break;
966 case LTTNG_CONSUMER32_UST:
967 case LTTNG_CONSUMER64_UST:
968 lttng_ustconsumer_del_stream(stream);
969 break;
970 default:
971 ERR("Unknown consumer_data type");
972 assert(0);
973 }
974 }
975
976 /*
977 * Destroy and close a already created stream.
978 */
979 static void destroy_close_stream(struct lttng_consumer_stream *stream)
980 {
981 assert(stream);
982
983 DBG("Consumer stream destroy monitored key: %" PRIu64, stream->key);
984
985 /* Destroy tracer buffers of the stream. */
986 consumer_stream_destroy_buffers(stream);
987 /* Close down everything including the relayd if one. */
988 consumer_stream_close(stream);
989 }
990
991 /*
992 * Decrement the stream's channel refcount and if down to 0, return the channel
993 * pointer so it can be destroyed by the caller or NULL if not.
994 */
995 static struct lttng_consumer_channel *unref_channel(
996 struct lttng_consumer_stream *stream)
997 {
998 struct lttng_consumer_channel *free_chan = NULL;
999
1000 assert(stream);
1001 assert(stream->chan);
1002
1003 /* Update refcount of channel and see if we need to destroy it. */
1004 if (!uatomic_sub_return(&stream->chan->refcount, 1)
1005 && !uatomic_read(&stream->chan->nb_init_stream_left)) {
1006 free_chan = stream->chan;
1007 }
1008
1009 return free_chan;
1010 }
1011
1012 /*
1013 * Destroy a stream completely. This will delete, close and free the stream.
1014 * Once return, the stream is NO longer usable. Its channel may get destroyed
1015 * if conditions are met for a monitored stream.
1016 *
1017 * This MUST be called WITHOUT the consumer data and stream lock acquired if
1018 * the stream is in _monitor_ mode else it does not matter.
1019 */
1020 void consumer_stream_destroy(struct lttng_consumer_stream *stream,
1021 struct lttng_ht *ht)
1022 {
1023 assert(stream);
1024
1025 /* Stream is in monitor mode. */
1026 if (stream->monitor) {
1027 struct lttng_consumer_channel *free_chan = NULL;
1028
1029 /*
1030 * This means that the stream was successfully removed from the streams
1031 * list of the channel and sent to the right thread managing this
1032 * stream thus being globally visible.
1033 */
1034 if (stream->globally_visible) {
1035 pthread_mutex_lock(&the_consumer_data.lock);
1036 pthread_mutex_lock(&stream->chan->lock);
1037 pthread_mutex_lock(&stream->lock);
1038 /* Remove every reference of the stream in the consumer. */
1039 consumer_stream_delete(stream, ht);
1040
1041 destroy_close_stream(stream);
1042
1043 /* Update channel's refcount of the stream. */
1044 free_chan = unref_channel(stream);
1045
1046 /* Indicates that the consumer data state MUST be updated after this. */
1047 the_consumer_data.need_update = 1;
1048
1049 pthread_mutex_unlock(&stream->lock);
1050 pthread_mutex_unlock(&stream->chan->lock);
1051 pthread_mutex_unlock(&the_consumer_data.lock);
1052 } else {
1053 /*
1054 * If the stream is not visible globally, this needs to be done
1055 * outside of the consumer data lock section.
1056 */
1057 free_chan = unref_channel(stream);
1058 }
1059
1060 if (free_chan) {
1061 consumer_del_channel(free_chan);
1062 }
1063 } else {
1064 destroy_close_stream(stream);
1065 }
1066
1067 /* Free stream within a RCU call. */
1068 lttng_trace_chunk_put(stream->trace_chunk);
1069 stream->trace_chunk = NULL;
1070 lttng_dynamic_array_reset(&stream->read_subbuffer_ops.post_consume_cbs);
1071 consumer_stream_free(stream);
1072 }
1073
1074 /*
1075 * Write index of a specific stream either on the relayd or local disk.
1076 *
1077 * Return 0 on success or else a negative value.
1078 */
1079 int consumer_stream_write_index(struct lttng_consumer_stream *stream,
1080 struct ctf_packet_index *element)
1081 {
1082 int ret;
1083
1084 assert(stream);
1085 assert(element);
1086
1087 rcu_read_lock();
1088 if (stream->net_seq_idx != (uint64_t) -1ULL) {
1089 struct consumer_relayd_sock_pair *relayd;
1090 relayd = consumer_find_relayd(stream->net_seq_idx);
1091 if (relayd) {
1092 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
1093 ret = relayd_send_index(&relayd->control_sock, element,
1094 stream->relayd_stream_id, stream->next_net_seq_num - 1);
1095 if (ret < 0) {
1096 /*
1097 * Communication error with lttng-relayd,
1098 * perform cleanup now
1099 */
1100 ERR("Relayd send index failed. Cleaning up relayd %" PRIu64 ".", relayd->net_seq_idx);
1101 lttng_consumer_cleanup_relayd(relayd);
1102 ret = -1;
1103 }
1104 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
1105 } else {
1106 ERR("Stream %" PRIu64 " relayd ID %" PRIu64 " unknown. Can't write index.",
1107 stream->key, stream->net_seq_idx);
1108 ret = -1;
1109 }
1110 } else {
1111 if (lttng_index_file_write(stream->index_file, element)) {
1112 ret = -1;
1113 } else {
1114 ret = 0;
1115 }
1116 }
1117 if (ret < 0) {
1118 goto error;
1119 }
1120
1121 error:
1122 rcu_read_unlock();
1123 return ret;
1124 }
1125
1126 int consumer_stream_create_output_files(struct lttng_consumer_stream *stream,
1127 bool create_index)
1128 {
1129 int ret;
1130 enum lttng_trace_chunk_status chunk_status;
1131 const int flags = O_WRONLY | O_CREAT | O_TRUNC;
1132 const mode_t mode = S_IRUSR | S_IWUSR | S_IRGRP | S_IWGRP;
1133 char stream_path[LTTNG_PATH_MAX];
1134
1135 ASSERT_LOCKED(stream->lock);
1136 assert(stream->trace_chunk);
1137
1138 ret = utils_stream_file_path(stream->chan->pathname, stream->name,
1139 stream->chan->tracefile_size,
1140 stream->tracefile_count_current, NULL,
1141 stream_path, sizeof(stream_path));
1142 if (ret < 0) {
1143 goto end;
1144 }
1145
1146 if (stream->out_fd >= 0) {
1147 ret = close(stream->out_fd);
1148 if (ret < 0) {
1149 PERROR("Failed to close stream file \"%s\"",
1150 stream->name);
1151 goto end;
1152 }
1153 stream->out_fd = -1;
1154 }
1155
1156 DBG("Opening stream output file \"%s\"", stream_path);
1157 chunk_status = lttng_trace_chunk_open_file(stream->trace_chunk, stream_path,
1158 flags, mode, &stream->out_fd, false);
1159 if (chunk_status != LTTNG_TRACE_CHUNK_STATUS_OK) {
1160 ERR("Failed to open stream file \"%s\"", stream->name);
1161 ret = -1;
1162 goto end;
1163 }
1164
1165 if (!stream->metadata_flag && (create_index || stream->index_file)) {
1166 if (stream->index_file) {
1167 lttng_index_file_put(stream->index_file);
1168 }
1169 chunk_status = lttng_index_file_create_from_trace_chunk(
1170 stream->trace_chunk,
1171 stream->chan->pathname,
1172 stream->name,
1173 stream->chan->tracefile_size,
1174 stream->tracefile_count_current,
1175 CTF_INDEX_MAJOR, CTF_INDEX_MINOR,
1176 false, &stream->index_file);
1177 if (chunk_status != LTTNG_TRACE_CHUNK_STATUS_OK) {
1178 ret = -1;
1179 goto end;
1180 }
1181 }
1182
1183 /* Reset current size because we just perform a rotation. */
1184 stream->tracefile_size_current = 0;
1185 stream->out_fd_offset = 0;
1186 end:
1187 return ret;
1188 }
1189
1190 int consumer_stream_rotate_output_files(struct lttng_consumer_stream *stream)
1191 {
1192 int ret;
1193
1194 stream->tracefile_count_current++;
1195 if (stream->chan->tracefile_count > 0) {
1196 stream->tracefile_count_current %=
1197 stream->chan->tracefile_count;
1198 }
1199
1200 DBG("Rotating output files of stream \"%s\"", stream->name);
1201 ret = consumer_stream_create_output_files(stream, true);
1202 if (ret) {
1203 goto end;
1204 }
1205
1206 end:
1207 return ret;
1208 }
1209
1210 bool consumer_stream_is_deleted(struct lttng_consumer_stream *stream)
1211 {
1212 /*
1213 * This function does not take a const stream since
1214 * cds_lfht_is_node_deleted was not const before liburcu 0.12.
1215 */
1216 assert(stream);
1217 return cds_lfht_is_node_deleted(&stream->node.node);
1218 }
1219
1220 static ssize_t metadata_bucket_flush(
1221 const struct stream_subbuffer *buffer, void *data)
1222 {
1223 ssize_t ret;
1224 struct lttng_consumer_stream *stream = data;
1225
1226 ret = consumer_stream_consume_mmap(NULL, stream, buffer);
1227 if (ret < 0) {
1228 goto end;
1229 }
1230 end:
1231 return ret;
1232 }
1233
1234 static ssize_t metadata_bucket_consume(
1235 struct lttng_consumer_local_data *unused,
1236 struct lttng_consumer_stream *stream,
1237 const struct stream_subbuffer *subbuffer)
1238 {
1239 ssize_t ret;
1240 enum metadata_bucket_status status;
1241
1242 status = metadata_bucket_fill(stream->metadata_bucket, subbuffer);
1243 switch (status) {
1244 case METADATA_BUCKET_STATUS_OK:
1245 /* Return consumed size. */
1246 ret = subbuffer->buffer.buffer.size;
1247 break;
1248 default:
1249 ret = -1;
1250 }
1251
1252 return ret;
1253 }
1254
1255 int consumer_stream_enable_metadata_bucketization(
1256 struct lttng_consumer_stream *stream)
1257 {
1258 int ret = 0;
1259
1260 assert(stream->metadata_flag);
1261 assert(!stream->metadata_bucket);
1262 assert(stream->chan->output == CONSUMER_CHANNEL_MMAP);
1263
1264 stream->metadata_bucket = metadata_bucket_create(
1265 metadata_bucket_flush, stream);
1266 if (!stream->metadata_bucket) {
1267 ret = -1;
1268 goto end;
1269 }
1270
1271 stream->read_subbuffer_ops.consume_subbuffer = metadata_bucket_consume;
1272 end:
1273 return ret;
1274 }
1275
1276 void consumer_stream_metadata_set_version(
1277 struct lttng_consumer_stream *stream, uint64_t new_version)
1278 {
1279 assert(new_version > stream->metadata_version);
1280 stream->metadata_version = new_version;
1281 stream->reset_metadata_flag = 1;
1282
1283 if (stream->metadata_bucket) {
1284 metadata_bucket_reset(stream->metadata_bucket);
1285 }
1286 }
1287
1288 int consumer_stream_flush_buffer(struct lttng_consumer_stream *stream,
1289 bool producer_active)
1290 {
1291 int ret = 0;
1292
1293 switch (the_consumer_data.type) {
1294 case LTTNG_CONSUMER_KERNEL:
1295 if (producer_active) {
1296 ret = kernctl_buffer_flush(stream->wait_fd);
1297 if (ret < 0) {
1298 ERR("Failed to flush kernel stream");
1299 goto end;
1300 }
1301 } else {
1302 ret = kernctl_buffer_flush_empty(stream->wait_fd);
1303 if (ret < 0) {
1304 /*
1305 * Doing a buffer flush which does not take into
1306 * account empty packets. This is not perfect,
1307 * but required as a fall-back when
1308 * "flush_empty" is not implemented by
1309 * lttng-modules.
1310 */
1311 ret = kernctl_buffer_flush(stream->wait_fd);
1312 if (ret < 0) {
1313 ERR("Failed to flush kernel stream");
1314 goto end;
1315 }
1316 }
1317 }
1318 break;
1319 case LTTNG_CONSUMER32_UST:
1320 case LTTNG_CONSUMER64_UST:
1321 ret = lttng_ustconsumer_flush_buffer(stream, (int) producer_active);
1322 break;
1323 default:
1324 ERR("Unknown consumer_data type");
1325 abort();
1326 }
1327
1328 end:
1329 return ret;
1330 }
This page took 0.088068 seconds and 5 git commands to generate.