Tests: Add test to check shared-memory FD leaks after relayd dies
[lttng-tools.git] / src / bin / lttng-relayd / stream.cpp
1 /*
2 * Copyright (C) 2013 Julien Desfossez <jdesfossez@efficios.com>
3 * Copyright (C) 2013 David Goulet <dgoulet@efficios.com>
4 * Copyright (C) 2015 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
5 * Copyright (C) 2019 Jérémie Galarneau <jeremie.galarneau@efficios.com>
6 *
7 * SPDX-License-Identifier: GPL-2.0-only
8 *
9 */
10
11 #define _LGPL_SOURCE
12 #include "index.hpp"
13 #include "lttng-relayd.hpp"
14 #include "stream.hpp"
15 #include "viewer-stream.hpp"
16
17 #include <common/common.hpp>
18 #include <common/defaults.hpp>
19 #include <common/fs-handle.hpp>
20 #include <common/sessiond-comm/relayd.hpp>
21 #include <common/urcu.hpp>
22 #include <common/utils.hpp>
23
24 #include <algorithm>
25 #include <fcntl.h>
26 #include <sys/stat.h>
27 #include <sys/types.h>
28 #include <urcu/rculist.h>
29
30 #define FILE_IO_STACK_BUFFER_SIZE 65536
31
32 /* Should be called with RCU read-side lock held. */
33 bool stream_get(struct relay_stream *stream)
34 {
35 ASSERT_RCU_READ_LOCKED();
36
37 return urcu_ref_get_unless_zero(&stream->ref);
38 }
39
40 /*
41 * Get stream from stream id from the streams hash table. Return stream
42 * if found else NULL. A stream reference is taken when a stream is
43 * returned. stream_put() must be called on that stream.
44 */
45 struct relay_stream *stream_get_by_id(uint64_t stream_id)
46 {
47 struct lttng_ht_node_u64 *node;
48 struct lttng_ht_iter iter;
49 struct relay_stream *stream = nullptr;
50
51 lttng::urcu::read_lock_guard read_lock;
52 lttng_ht_lookup(relay_streams_ht, &stream_id, &iter);
53 node = lttng_ht_iter_get_node_u64(&iter);
54 if (!node) {
55 DBG("Relay stream %" PRIu64 " not found", stream_id);
56 goto end;
57 }
58 stream = lttng::utils::container_of(node, &relay_stream::node);
59 if (!stream_get(stream)) {
60 stream = nullptr;
61 }
62 end:
63 return stream;
64 }
65
66 static void stream_complete_rotation(struct relay_stream *stream)
67 {
68 DBG("Rotation completed for stream %" PRIu64, stream->stream_handle);
69 if (stream->ongoing_rotation.value.next_trace_chunk) {
70 tracefile_array_reset(stream->tfa);
71 tracefile_array_commit_seq(stream->tfa, stream->index_received_seqcount);
72 }
73 lttng_trace_chunk_put(stream->trace_chunk);
74 stream->trace_chunk = stream->ongoing_rotation.value.next_trace_chunk;
75 stream->ongoing_rotation = LTTNG_OPTIONAL_INIT_UNSET;
76 stream->completed_rotation_count++;
77 }
78
79 static int stream_create_data_output_file_from_trace_chunk(struct relay_stream *stream,
80 struct lttng_trace_chunk *trace_chunk,
81 bool force_unlink,
82 struct fs_handle **out_file)
83 {
84 int ret;
85 char stream_path[LTTNG_PATH_MAX];
86 enum lttng_trace_chunk_status status;
87 const int flags = O_RDWR | O_CREAT | O_TRUNC;
88 const mode_t mode = S_IRUSR | S_IWUSR | S_IRGRP | S_IWGRP;
89
90 ASSERT_LOCKED(stream->lock);
91
92 ret = utils_stream_file_path(stream->path_name,
93 stream->channel_name,
94 stream->tracefile_size,
95 stream->tracefile_current_index,
96 nullptr,
97 stream_path,
98 sizeof(stream_path));
99 if (ret < 0) {
100 goto end;
101 }
102
103 if (stream->tracefile_wrapped_around || force_unlink) {
104 /*
105 * The on-disk ring-buffer has wrapped around.
106 * Newly created stream files will replace existing files. Since
107 * live clients may be consuming existing files, the file about
108 * to be replaced is unlinked in order to not overwrite its
109 * content.
110 */
111 status = (lttng_trace_chunk_status) lttng_trace_chunk_unlink_file(trace_chunk,
112 stream_path);
113 if (status != LTTNG_TRACE_CHUNK_STATUS_OK) {
114 PERROR("Failed to unlink stream file \"%s\" during trace file rotation",
115 stream_path);
116 /*
117 * Don't abort if the file doesn't exist, it is
118 * unexpected, but should not be a fatal error.
119 */
120 if (errno != ENOENT) {
121 ret = -1;
122 goto end;
123 }
124 }
125 }
126
127 status = lttng_trace_chunk_open_fs_handle(
128 trace_chunk, stream_path, flags, mode, out_file, false);
129 if (status != LTTNG_TRACE_CHUNK_STATUS_OK) {
130 ERR("Failed to open stream file \"%s\"", stream->channel_name);
131 ret = -1;
132 goto end;
133 }
134 end:
135 return ret;
136 }
137
138 static int stream_rotate_data_file(struct relay_stream *stream)
139 {
140 int ret = 0;
141
142 DBG("Rotating stream %" PRIu64 " data file with size %" PRIu64,
143 stream->stream_handle,
144 stream->tracefile_size_current);
145
146 if (stream->file) {
147 fs_handle_close(stream->file);
148 stream->file = nullptr;
149 }
150
151 stream->tracefile_wrapped_around = false;
152 stream->tracefile_current_index = 0;
153
154 if (stream->ongoing_rotation.value.next_trace_chunk) {
155 enum lttng_trace_chunk_status chunk_status;
156
157 chunk_status = lttng_trace_chunk_create_subdirectory(
158 stream->ongoing_rotation.value.next_trace_chunk, stream->path_name);
159 if (chunk_status != LTTNG_TRACE_CHUNK_STATUS_OK) {
160 ret = -1;
161 goto end;
162 }
163
164 /* Rotate the data file. */
165 ret = stream_create_data_output_file_from_trace_chunk(
166 stream,
167 stream->ongoing_rotation.value.next_trace_chunk,
168 false,
169 &stream->file);
170 if (ret < 0) {
171 ERR("Failed to rotate stream data file");
172 goto end;
173 }
174 }
175 DBG("%s: reset tracefile_size_current for stream %" PRIu64 " was %" PRIu64,
176 __func__,
177 stream->stream_handle,
178 stream->tracefile_size_current);
179 stream->tracefile_size_current = 0;
180 stream->pos_after_last_complete_data_index = 0;
181 stream->ongoing_rotation.value.data_rotated = true;
182
183 if (stream->ongoing_rotation.value.index_rotated) {
184 /* Rotation completed; reset its state. */
185 stream_complete_rotation(stream);
186 }
187 end:
188 return ret;
189 }
190
191 /*
192 * If too much data has been written in a tracefile before we received the
193 * rotation command, we have to move the excess data to the new tracefile and
194 * perform the rotation. This can happen because the control and data
195 * connections are separate, the indexes as well as the commands arrive from
196 * the control connection and we have no control over the order so we could be
197 * in a situation where too much data has been received on the data connection
198 * before the rotation command on the control connection arrives.
199 */
200 static int rotate_truncate_stream(struct relay_stream *stream)
201 {
202 int ret;
203 off_t lseek_ret, previous_stream_copy_origin;
204 uint64_t copy_bytes_left, misplaced_data_size;
205 bool acquired_reference;
206 struct fs_handle *previous_stream_file = nullptr;
207 struct lttng_trace_chunk *previous_chunk = nullptr;
208
209 if (!LTTNG_OPTIONAL_GET(stream->ongoing_rotation).next_trace_chunk) {
210 ERR("Protocol error encoutered in %s(): stream rotation "
211 "sequence number is before the current sequence number "
212 "and the next trace chunk is unset. Honoring this "
213 "rotation command would result in data loss",
214 __FUNCTION__);
215 ret = -1;
216 goto end;
217 }
218
219 ASSERT_LOCKED(stream->lock);
220 /*
221 * Acquire a reference to the current trace chunk to ensure
222 * it is not reclaimed when `stream_rotate_data_file` is called.
223 * Failing to do so would violate the contract of the trace
224 * chunk API as an active file descriptor would outlive the
225 * trace chunk.
226 */
227 acquired_reference = lttng_trace_chunk_get(stream->trace_chunk);
228 LTTNG_ASSERT(acquired_reference);
229 previous_chunk = stream->trace_chunk;
230
231 /*
232 * Steal the stream's reference to its stream_fd. A new
233 * stream_fd will be created when the rotation completes and
234 * the orinal stream_fd will be used to copy the "extra" data
235 * to the new file.
236 */
237 LTTNG_ASSERT(stream->file);
238 previous_stream_file = stream->file;
239 stream->file = nullptr;
240
241 LTTNG_ASSERT(!stream->is_metadata);
242 LTTNG_ASSERT(stream->tracefile_size_current > stream->pos_after_last_complete_data_index);
243 misplaced_data_size =
244 stream->tracefile_size_current - stream->pos_after_last_complete_data_index;
245 copy_bytes_left = misplaced_data_size;
246 previous_stream_copy_origin = stream->pos_after_last_complete_data_index;
247
248 ret = stream_rotate_data_file(stream);
249 if (ret) {
250 goto end;
251 }
252
253 LTTNG_ASSERT(stream->file);
254 /*
255 * Seek the current tracefile to the position at which the rotation
256 * should have occurred.
257 */
258 lseek_ret = fs_handle_seek(previous_stream_file, previous_stream_copy_origin, SEEK_SET);
259 if (lseek_ret < 0) {
260 PERROR("Failed to seek to offset %" PRIu64
261 " while copying extra data received before a stream rotation",
262 (uint64_t) previous_stream_copy_origin);
263 ret = -1;
264 goto end;
265 }
266
267 /* Move data from the old file to the new file. */
268 while (copy_bytes_left) {
269 ssize_t io_ret;
270 char copy_buffer[FILE_IO_STACK_BUFFER_SIZE];
271 const off_t copy_size_this_pass =
272 std::min<uint64_t>(copy_bytes_left, sizeof(copy_buffer));
273
274 io_ret = fs_handle_read(previous_stream_file, copy_buffer, copy_size_this_pass);
275 if (io_ret < (ssize_t) copy_size_this_pass) {
276 if (io_ret == -1) {
277 PERROR("Failed to read %" PRIu64
278 " bytes from previous stream file in %s(), returned %zi: stream id = %" PRIu64,
279 copy_size_this_pass,
280 __FUNCTION__,
281 io_ret,
282 stream->stream_handle);
283 } else {
284 ERR("Failed to read %" PRIu64
285 " bytes from previous stream file in %s(), returned %zi: stream id = %" PRIu64,
286 copy_size_this_pass,
287 __FUNCTION__,
288 io_ret,
289 stream->stream_handle);
290 }
291 ret = -1;
292 goto end;
293 }
294
295 io_ret = fs_handle_write(stream->file, copy_buffer, copy_size_this_pass);
296 if (io_ret < (ssize_t) copy_size_this_pass) {
297 if (io_ret == -1) {
298 PERROR("Failed to write %" PRIu64
299 " bytes from previous stream file in %s(), returned %zi: stream id = %" PRIu64,
300 copy_size_this_pass,
301 __FUNCTION__,
302 io_ret,
303 stream->stream_handle);
304 } else {
305 ERR("Failed to write %" PRIu64
306 " bytes from previous stream file in %s(), returned %zi: stream id = %" PRIu64,
307 copy_size_this_pass,
308 __FUNCTION__,
309 io_ret,
310 stream->stream_handle);
311 }
312 ret = -1;
313 goto end;
314 }
315 copy_bytes_left -= copy_size_this_pass;
316 }
317
318 /* Truncate the file to get rid of the excess data. */
319 ret = fs_handle_truncate(previous_stream_file, previous_stream_copy_origin);
320 if (ret) {
321 PERROR("Failed to truncate current stream file to offset %" PRIu64,
322 previous_stream_copy_origin);
323 goto end;
324 }
325
326 /*
327 * Update the offset and FD of all the eventual indexes created by the
328 * data connection before the rotation command arrived.
329 */
330 ret = relay_index_switch_all_files(stream);
331 if (ret < 0) {
332 ERR("Failed to rotate index file");
333 goto end;
334 }
335
336 stream->tracefile_size_current = misplaced_data_size;
337 /* Index and data contents are back in sync. */
338 stream->pos_after_last_complete_data_index = 0;
339 ret = 0;
340 end:
341 lttng_trace_chunk_put(previous_chunk);
342 return ret;
343 }
344
345 /*
346 * Check if a stream's data file (as opposed to index) should be rotated
347 * (for session rotation).
348 * Must be called with the stream lock held.
349 *
350 * Return 0 on success, a negative value on error.
351 */
352 static int try_rotate_stream_data(struct relay_stream *stream)
353 {
354 int ret = 0;
355
356 if (caa_likely(!stream->ongoing_rotation.is_set)) {
357 /* No rotation expected. */
358 goto end;
359 }
360
361 if (stream->ongoing_rotation.value.data_rotated) {
362 /* Rotation of the data file has already occurred. */
363 goto end;
364 }
365
366 DBG("%s: Stream %" PRIu64 " (rotate_at_index_packet_seq_num = %" PRIu64
367 ", rotate_at_prev_data_net_seq = %" PRIu64 ", prev_data_seq = %" PRIu64 ")",
368 __func__,
369 stream->stream_handle,
370 stream->ongoing_rotation.value.packet_seq_num,
371 stream->ongoing_rotation.value.prev_data_net_seq,
372 stream->prev_data_seq);
373
374 if (stream->prev_data_seq == -1ULL ||
375 stream->ongoing_rotation.value.prev_data_net_seq == -1ULL ||
376 stream->prev_data_seq < stream->ongoing_rotation.value.prev_data_net_seq) {
377 /*
378 * The next packet that will be written is not part of the next
379 * chunk yet.
380 */
381 DBG("Stream %" PRIu64 " data not yet ready for rotation "
382 "(rotate_at_index_packet_seq_num = %" PRIu64
383 ", rotate_at_prev_data_net_seq = %" PRIu64 ", prev_data_seq = %" PRIu64 ")",
384 stream->stream_handle,
385 stream->ongoing_rotation.value.packet_seq_num,
386 stream->ongoing_rotation.value.prev_data_net_seq,
387 stream->prev_data_seq);
388 goto end;
389 } else if (stream->prev_data_seq > stream->ongoing_rotation.value.prev_data_net_seq) {
390 /*
391 * prev_data_seq is checked here since indexes and rotation
392 * commands are serialized with respect to each other.
393 */
394 DBG("Rotation after too much data has been written in tracefile "
395 "for stream %" PRIu64 ", need to truncate before "
396 "rotating",
397 stream->stream_handle);
398 ret = rotate_truncate_stream(stream);
399 if (ret) {
400 ERR("Failed to truncate stream");
401 goto end;
402 }
403 } else {
404 ret = stream_rotate_data_file(stream);
405 }
406
407 end:
408 return ret;
409 }
410
411 /*
412 * Close the current index file if it is open, and create a new one.
413 *
414 * Return 0 on success, -1 on error.
415 */
416 static int create_index_file(struct relay_stream *stream, struct lttng_trace_chunk *chunk)
417 {
418 int ret;
419 uint32_t major, minor;
420 char *index_subpath = nullptr;
421 enum lttng_trace_chunk_status status;
422
423 ASSERT_LOCKED(stream->lock);
424
425 /* Put ref on previous index_file. */
426 if (stream->index_file) {
427 lttng_index_file_put(stream->index_file);
428 stream->index_file = nullptr;
429 }
430 major = stream->trace->session->major;
431 minor = stream->trace->session->minor;
432
433 if (!chunk) {
434 ret = 0;
435 goto end;
436 }
437 ret = asprintf(&index_subpath, "%s/%s", stream->path_name, DEFAULT_INDEX_DIR);
438 if (ret < 0) {
439 goto end;
440 }
441
442 status = lttng_trace_chunk_create_subdirectory(chunk, index_subpath);
443 free(index_subpath);
444 if (status != LTTNG_TRACE_CHUNK_STATUS_OK) {
445 ret = -1;
446 goto end;
447 }
448 status = lttng_index_file_create_from_trace_chunk(chunk,
449 stream->path_name,
450 stream->channel_name,
451 stream->tracefile_size,
452 stream->tracefile_current_index,
453 lttng_to_index_major(major, minor),
454 lttng_to_index_minor(major, minor),
455 true,
456 &stream->index_file);
457 if (status != LTTNG_TRACE_CHUNK_STATUS_OK) {
458 ret = -1;
459 goto end;
460 }
461
462 ret = 0;
463
464 end:
465 return ret;
466 }
467
468 /*
469 * Check if a stream's index file should be rotated (for session rotation).
470 * Must be called with the stream lock held.
471 *
472 * Return 0 on success, a negative value on error.
473 */
474 static int try_rotate_stream_index(struct relay_stream *stream)
475 {
476 int ret = 0;
477
478 if (!stream->ongoing_rotation.is_set) {
479 /* No rotation expected. */
480 goto end;
481 }
482
483 if (stream->ongoing_rotation.value.index_rotated) {
484 /* Rotation of the index has already occurred. */
485 goto end;
486 }
487
488 DBG("%s: Stream %" PRIu64 " (rotate_at_packet_seq_num = %" PRIu64
489 ", received_packet_seq_num = "
490 "(value = %" PRIu64 ", is_set = %" PRIu8 "))",
491 __func__,
492 stream->stream_handle,
493 stream->ongoing_rotation.value.packet_seq_num,
494 stream->received_packet_seq_num.value,
495 stream->received_packet_seq_num.is_set);
496
497 if (!stream->received_packet_seq_num.is_set ||
498 LTTNG_OPTIONAL_GET(stream->received_packet_seq_num) + 1 <
499 stream->ongoing_rotation.value.packet_seq_num) {
500 DBG("Stream %" PRIu64 " index not yet ready for rotation "
501 "(rotate_at_packet_seq_num = %" PRIu64 ", received_packet_seq_num = "
502 "(value = %" PRIu64 ", is_set = %" PRIu8 "))",
503 stream->stream_handle,
504 stream->ongoing_rotation.value.packet_seq_num,
505 stream->received_packet_seq_num.value,
506 stream->received_packet_seq_num.is_set);
507 goto end;
508 } else {
509 /*
510 * The next index belongs to the new trace chunk; rotate.
511 * In overwrite mode, the packet seq num may jump over the
512 * rotation position.
513 */
514 LTTNG_ASSERT(LTTNG_OPTIONAL_GET(stream->received_packet_seq_num) + 1 >=
515 stream->ongoing_rotation.value.packet_seq_num);
516 DBG("Rotating stream %" PRIu64 " index file", stream->stream_handle);
517 if (stream->index_file) {
518 lttng_index_file_put(stream->index_file);
519 stream->index_file = nullptr;
520 }
521 stream->ongoing_rotation.value.index_rotated = true;
522
523 /*
524 * Set the rotation pivot position for the data, now that we have the
525 * net_seq_num matching the packet_seq_num index pivot position.
526 */
527 stream->ongoing_rotation.value.prev_data_net_seq = stream->prev_index_seq;
528 if (stream->ongoing_rotation.value.data_rotated &&
529 stream->ongoing_rotation.value.index_rotated) {
530 /* Rotation completed; reset its state. */
531 DBG("Rotation completed for stream %" PRIu64, stream->stream_handle);
532 stream_complete_rotation(stream);
533 }
534 }
535
536 end:
537 return ret;
538 }
539
540 static int stream_set_trace_chunk(struct relay_stream *stream, struct lttng_trace_chunk *chunk)
541 {
542 int ret = 0;
543 enum lttng_trace_chunk_status status;
544 bool acquired_reference;
545
546 status = lttng_trace_chunk_create_subdirectory(chunk, stream->path_name);
547 if (status != LTTNG_TRACE_CHUNK_STATUS_OK) {
548 ret = -1;
549 goto end;
550 }
551
552 lttng_trace_chunk_put(stream->trace_chunk);
553 acquired_reference = lttng_trace_chunk_get(chunk);
554 LTTNG_ASSERT(acquired_reference);
555 stream->trace_chunk = chunk;
556
557 if (stream->file) {
558 fs_handle_close(stream->file);
559 stream->file = nullptr;
560 }
561 ret = stream_create_data_output_file_from_trace_chunk(stream, chunk, false, &stream->file);
562 end:
563 return ret;
564 }
565
566 /*
567 * We keep ownership of path_name and channel_name.
568 */
569 struct relay_stream *stream_create(struct ctf_trace *trace,
570 uint64_t stream_handle,
571 char *path_name,
572 char *channel_name,
573 uint64_t tracefile_size,
574 uint64_t tracefile_count)
575 {
576 int ret;
577 struct relay_stream *stream = nullptr;
578 struct relay_session *session = trace->session;
579 bool acquired_reference = false;
580 struct lttng_trace_chunk *current_trace_chunk;
581
582 stream = zmalloc<relay_stream>();
583 if (stream == nullptr) {
584 PERROR("relay stream zmalloc");
585 goto error_no_alloc;
586 }
587
588 stream->stream_handle = stream_handle;
589 stream->prev_data_seq = -1ULL;
590 stream->prev_index_seq = -1ULL;
591 stream->last_net_seq_num = -1ULL;
592 stream->ctf_stream_id = -1ULL;
593 stream->tracefile_size = tracefile_size;
594 stream->tracefile_count = tracefile_count;
595 stream->path_name = path_name;
596 stream->channel_name = channel_name;
597 stream->beacon_ts_end = -1ULL;
598 lttng_ht_node_init_u64(&stream->node, stream->stream_handle);
599 pthread_mutex_init(&stream->lock, nullptr);
600 urcu_ref_init(&stream->ref);
601 ctf_trace_get(trace);
602 stream->trace = trace;
603
604 pthread_mutex_lock(&trace->session->lock);
605 current_trace_chunk = trace->session->current_trace_chunk;
606 if (current_trace_chunk) {
607 acquired_reference = lttng_trace_chunk_get(current_trace_chunk);
608 }
609 pthread_mutex_unlock(&trace->session->lock);
610 if (!acquired_reference) {
611 ERR("Cannot create stream for channel \"%s\" as a reference to the session's current trace chunk could not be acquired",
612 channel_name);
613 ret = -1;
614 goto end;
615 }
616
617 stream->indexes_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
618 if (!stream->indexes_ht) {
619 ERR("Cannot created indexes_ht");
620 ret = -1;
621 goto end;
622 }
623
624 pthread_mutex_lock(&stream->lock);
625 ret = stream_set_trace_chunk(stream, current_trace_chunk);
626 pthread_mutex_unlock(&stream->lock);
627 if (ret) {
628 ERR("Failed to set the current trace chunk of session \"%s\" on newly created stream of channel \"%s\"",
629 trace->session->session_name,
630 stream->channel_name);
631 ret = -1;
632 goto end;
633 }
634 stream->tfa = tracefile_array_create(stream->tracefile_count);
635 if (!stream->tfa) {
636 ret = -1;
637 goto end;
638 }
639
640 stream->is_metadata = !strcmp(stream->channel_name, DEFAULT_METADATA_NAME);
641 stream->in_recv_list = true;
642
643 /*
644 * Add the stream in the recv list of the session. Once the end stream
645 * message is received, all session streams are published.
646 */
647 pthread_mutex_lock(&session->recv_list_lock);
648 cds_list_add_rcu(&stream->recv_node, &session->recv_list);
649 session->stream_count++;
650 pthread_mutex_unlock(&session->recv_list_lock);
651
652 /*
653 * Both in the ctf_trace object and the global stream ht since the data
654 * side of the relayd does not have the concept of session.
655 */
656 lttng_ht_add_unique_u64(relay_streams_ht, &stream->node);
657 stream->in_stream_ht = true;
658
659 DBG("Relay new stream added %s with ID %" PRIu64,
660 stream->channel_name,
661 stream->stream_handle);
662 ret = 0;
663
664 end:
665 if (ret) {
666 if (stream->file) {
667 fs_handle_close(stream->file);
668 stream->file = nullptr;
669 }
670 stream_put(stream);
671 stream = nullptr;
672 }
673 if (acquired_reference) {
674 lttng_trace_chunk_put(current_trace_chunk);
675 }
676 return stream;
677
678 error_no_alloc:
679 /*
680 * path_name and channel_name need to be freed explicitly here
681 * because we cannot rely on stream_put().
682 */
683 free(path_name);
684 free(channel_name);
685 return nullptr;
686 }
687
688 /*
689 * Called with the session lock held.
690 */
691 void stream_publish(struct relay_stream *stream)
692 {
693 struct relay_session *session;
694
695 pthread_mutex_lock(&stream->lock);
696 if (stream->published) {
697 goto unlock;
698 }
699
700 session = stream->trace->session;
701
702 pthread_mutex_lock(&session->recv_list_lock);
703 if (stream->in_recv_list) {
704 cds_list_del_rcu(&stream->recv_node);
705 stream->in_recv_list = false;
706 }
707 pthread_mutex_unlock(&session->recv_list_lock);
708
709 pthread_mutex_lock(&stream->trace->stream_list_lock);
710 cds_list_add_rcu(&stream->stream_node, &stream->trace->stream_list);
711 pthread_mutex_unlock(&stream->trace->stream_list_lock);
712
713 stream->published = true;
714 unlock:
715 pthread_mutex_unlock(&stream->lock);
716 }
717
718 /*
719 * Stream must be protected by holding the stream lock or by virtue of being
720 * called from stream_destroy.
721 */
722 static void stream_unpublish(struct relay_stream *stream)
723 {
724 if (stream->in_stream_ht) {
725 struct lttng_ht_iter iter;
726 int ret;
727
728 iter.iter.node = &stream->node.node;
729 ret = lttng_ht_del(relay_streams_ht, &iter);
730 LTTNG_ASSERT(!ret);
731 stream->in_stream_ht = false;
732 }
733 if (stream->published) {
734 pthread_mutex_lock(&stream->trace->stream_list_lock);
735 cds_list_del_rcu(&stream->stream_node);
736 pthread_mutex_unlock(&stream->trace->stream_list_lock);
737 stream->published = false;
738 }
739 }
740
741 static void stream_destroy(struct relay_stream *stream)
742 {
743 if (stream->indexes_ht) {
744 /*
745 * Calling lttng_ht_destroy in call_rcu worker thread so
746 * we don't hold the RCU read-side lock while calling
747 * it.
748 */
749 lttng_ht_destroy(stream->indexes_ht);
750 }
751 if (stream->tfa) {
752 tracefile_array_destroy(stream->tfa);
753 }
754 free(stream->path_name);
755 free(stream->channel_name);
756 free(stream);
757 }
758
759 static void stream_destroy_rcu(struct rcu_head *rcu_head)
760 {
761 struct relay_stream *stream = lttng::utils::container_of(rcu_head, &relay_stream::rcu_node);
762
763 stream_destroy(stream);
764 }
765
766 /*
767 * No need to take stream->lock since this is only called on the final
768 * stream_put which ensures that a single thread may act on the stream.
769 */
770 static void stream_release(struct urcu_ref *ref)
771 {
772 struct relay_stream *stream = lttng::utils::container_of(ref, &relay_stream::ref);
773 struct relay_session *session;
774
775 session = stream->trace->session;
776
777 DBG("Releasing stream id %" PRIu64, stream->stream_handle);
778
779 pthread_mutex_lock(&session->recv_list_lock);
780 session->stream_count--;
781 if (stream->in_recv_list) {
782 cds_list_del_rcu(&stream->recv_node);
783 stream->in_recv_list = false;
784 }
785 pthread_mutex_unlock(&session->recv_list_lock);
786
787 stream_unpublish(stream);
788
789 if (stream->file) {
790 fs_handle_close(stream->file);
791 stream->file = nullptr;
792 }
793 if (stream->index_file) {
794 lttng_index_file_put(stream->index_file);
795 stream->index_file = nullptr;
796 }
797 if (stream->trace) {
798 ctf_trace_put(stream->trace);
799 stream->trace = nullptr;
800 }
801 stream_complete_rotation(stream);
802 lttng_trace_chunk_put(stream->trace_chunk);
803 stream->trace_chunk = nullptr;
804
805 call_rcu(&stream->rcu_node, stream_destroy_rcu);
806 }
807
808 void stream_put(struct relay_stream *stream)
809 {
810 lttng::urcu::read_lock_guard read_lock;
811 LTTNG_ASSERT(stream->ref.refcount != 0);
812 /*
813 * Wait until we have processed all the stream packets before
814 * actually putting our last stream reference.
815 */
816 urcu_ref_put(&stream->ref, stream_release);
817 }
818
819 int stream_set_pending_rotation(struct relay_stream *stream,
820 struct lttng_trace_chunk *next_trace_chunk,
821 uint64_t rotation_sequence_number)
822 {
823 int ret = 0;
824 const struct relay_stream_rotation rotation = {
825 .data_rotated = false,
826 .index_rotated = false,
827 .packet_seq_num = rotation_sequence_number,
828 .prev_data_net_seq = -1ULL,
829 .next_trace_chunk = next_trace_chunk,
830 };
831
832 if (stream->ongoing_rotation.is_set) {
833 ERR("Attempted to set a pending rotation on a stream already being rotated (protocol error)");
834 ret = -1;
835 goto end;
836 }
837
838 if (next_trace_chunk) {
839 const bool reference_acquired = lttng_trace_chunk_get(next_trace_chunk);
840
841 LTTNG_ASSERT(reference_acquired);
842 }
843 LTTNG_OPTIONAL_SET(&stream->ongoing_rotation, rotation);
844
845 DBG("Setting pending rotation: stream_id = %" PRIu64
846 ", rotate_at_packet_seq_num = %" PRIu64,
847 stream->stream_handle,
848 rotation_sequence_number);
849 if (stream->is_metadata) {
850 /*
851 * A metadata stream has no index; consider it already rotated.
852 */
853 stream->ongoing_rotation.value.index_rotated = true;
854 if (next_trace_chunk) {
855 /*
856 * The metadata will be received again in the new chunk.
857 */
858 stream->metadata_received = 0;
859 }
860 ret = stream_rotate_data_file(stream);
861 } else {
862 ret = try_rotate_stream_index(stream);
863 if (ret < 0) {
864 goto end;
865 }
866
867 ret = try_rotate_stream_data(stream);
868 if (ret < 0) {
869 goto end;
870 }
871 }
872 end:
873 return ret;
874 }
875
876 void try_stream_close(struct relay_stream *stream)
877 {
878 bool session_aborted;
879 struct relay_session *session = stream->trace->session;
880
881 DBG("Trying to close stream %" PRIu64, stream->stream_handle);
882
883 pthread_mutex_lock(&session->lock);
884 session_aborted = session->aborted;
885 pthread_mutex_unlock(&session->lock);
886
887 pthread_mutex_lock(&stream->lock);
888 /*
889 * Can be called concurently by connection close and reception of last
890 * pending data.
891 */
892 if (stream->closed) {
893 pthread_mutex_unlock(&stream->lock);
894 DBG("closing stream %" PRIu64 " aborted since it is already marked as closed",
895 stream->stream_handle);
896 return;
897 }
898
899 stream->close_requested = true;
900
901 if (stream->last_net_seq_num == -1ULL) {
902 /*
903 * Handle connection close without explicit stream close
904 * command.
905 *
906 * We can be clever about indexes partially received in
907 * cases where we received the data socket part, but not
908 * the control socket part: since we're currently closing
909 * the stream on behalf of the control socket, we *know*
910 * there won't be any more control information for this
911 * socket. Therefore, we can destroy all indexes for
912 * which we have received only the file descriptor (from
913 * data socket). This takes care of consumerd crashes
914 * between sending the data and control information for
915 * a packet. Since those are sent in that order, we take
916 * care of consumerd crashes.
917 */
918 DBG("relay_index_close_partial_fd");
919 relay_index_close_partial_fd(stream);
920 /*
921 * Use the highest net_seq_num we currently have pending
922 * As end of stream indicator. Leave last_net_seq_num
923 * at -1ULL if we cannot find any index.
924 */
925 stream->last_net_seq_num = relay_index_find_last(stream);
926 DBG("Updating stream->last_net_seq_num to %" PRIu64, stream->last_net_seq_num);
927 /* Fall-through into the next check. */
928 }
929
930 if (stream->last_net_seq_num != -1ULL &&
931 ((int64_t) (stream->prev_data_seq - stream->last_net_seq_num)) < 0 &&
932 !session_aborted) {
933 /*
934 * Don't close since we still have data pending. This
935 * handles cases where an explicit close command has
936 * been received for this stream, and cases where the
937 * connection has been closed, and we are awaiting for
938 * index information from the data socket. It is
939 * therefore expected that all the index fd information
940 * we need has already been received on the control
941 * socket. Matching index information from data socket
942 * should be Expected Soon(TM).
943 *
944 * TODO: We should implement a timer to garbage collect
945 * streams after a timeout to be resilient against a
946 * consumerd implementation that would not match this
947 * expected behavior.
948 */
949 pthread_mutex_unlock(&stream->lock);
950 DBG("closing stream %" PRIu64 " aborted since it still has data pending",
951 stream->stream_handle);
952 return;
953 }
954 /*
955 * We received all the indexes we can expect.
956 */
957 stream_unpublish(stream);
958 stream->closed = true;
959 /* Relay indexes are only used by the "consumer/sessiond" end. */
960 relay_index_close_all(stream);
961
962 /*
963 * If we are closed by an application exiting (per-pid buffers),
964 * we need to put our reference on the stream trace chunk right
965 * away, because otherwise still holding the reference on the
966 * trace chunk could allow a viewer stream (which holds a reference
967 * to the stream) to postpone destroy waiting for the chunk to cease
968 * to exist endlessly until the viewer is detached.
969 */
970
971 /* Put stream fd before put chunk. */
972 if (stream->file) {
973 fs_handle_close(stream->file);
974 stream->file = nullptr;
975 }
976 if (stream->index_file) {
977 lttng_index_file_put(stream->index_file);
978 stream->index_file = nullptr;
979 }
980 lttng_trace_chunk_put(stream->trace_chunk);
981 stream->trace_chunk = nullptr;
982 pthread_mutex_unlock(&stream->lock);
983 DBG("Succeeded in closing stream %" PRIu64, stream->stream_handle);
984 stream_put(stream);
985 }
986
987 int stream_init_packet(struct relay_stream *stream, size_t packet_size, bool *file_rotated)
988 {
989 int ret = 0;
990
991 ASSERT_LOCKED(stream->lock);
992
993 if (!stream->file || !stream->trace_chunk) {
994 ERR("Protocol error: received a packet for a stream that doesn't have a current trace chunk: stream_id = %" PRIu64
995 ", channel_name = %s",
996 stream->stream_handle,
997 stream->channel_name);
998 ret = -1;
999 goto end;
1000 }
1001
1002 if (caa_likely(stream->tracefile_size == 0)) {
1003 /* No size limit set; nothing to check. */
1004 goto end;
1005 }
1006
1007 /*
1008 * Check if writing the new packet would exceed the maximal file size.
1009 */
1010 if (caa_unlikely((stream->tracefile_size_current + packet_size) > stream->tracefile_size)) {
1011 const uint64_t new_file_index =
1012 (stream->tracefile_current_index + 1) % stream->tracefile_count;
1013
1014 if (new_file_index < stream->tracefile_current_index) {
1015 stream->tracefile_wrapped_around = true;
1016 }
1017 DBG("New stream packet causes stream file rotation: stream_id = %" PRIu64
1018 ", current_file_size = %" PRIu64
1019 ", packet_size = %zu, current_file_index = %" PRIu64
1020 " new_file_index = %" PRIu64,
1021 stream->stream_handle,
1022 stream->tracefile_size_current,
1023 packet_size,
1024 stream->tracefile_current_index,
1025 new_file_index);
1026 tracefile_array_file_rotate(stream->tfa, TRACEFILE_ROTATE_WRITE);
1027 stream->tracefile_current_index = new_file_index;
1028
1029 if (stream->file) {
1030 fs_handle_close(stream->file);
1031 stream->file = nullptr;
1032 }
1033 ret = stream_create_data_output_file_from_trace_chunk(
1034 stream, stream->trace_chunk, false, &stream->file);
1035 if (ret) {
1036 ERR("Failed to perform trace file rotation of stream %" PRIu64,
1037 stream->stream_handle);
1038 goto end;
1039 }
1040
1041 /*
1042 * Reset current size because we just performed a stream
1043 * rotation.
1044 */
1045 DBG("%s: reset tracefile_size_current for stream %" PRIu64 " was %" PRIu64,
1046 __func__,
1047 stream->stream_handle,
1048 stream->tracefile_size_current);
1049 stream->tracefile_size_current = 0;
1050 *file_rotated = true;
1051 } else {
1052 *file_rotated = false;
1053 }
1054 end:
1055 return ret;
1056 }
1057
1058 /* Note that the packet is not necessarily complete. */
1059 int stream_write(struct relay_stream *stream,
1060 const struct lttng_buffer_view *packet,
1061 size_t padding_len)
1062 {
1063 int ret = 0;
1064 ssize_t write_ret;
1065 size_t padding_to_write = padding_len;
1066 char padding_buffer[FILE_IO_STACK_BUFFER_SIZE];
1067
1068 ASSERT_LOCKED(stream->lock);
1069 memset(padding_buffer, 0, std::min(sizeof(padding_buffer), padding_to_write));
1070
1071 if (!stream->file || !stream->trace_chunk) {
1072 ERR("Protocol error: received a packet for a stream that doesn't have a current trace chunk: stream_id = %" PRIu64
1073 ", channel_name = %s",
1074 stream->stream_handle,
1075 stream->channel_name);
1076 ret = -1;
1077 goto end;
1078 }
1079 if (packet) {
1080 write_ret = fs_handle_write(stream->file, packet->data, packet->size);
1081 if (write_ret != packet->size) {
1082 PERROR("Failed to write to stream file of %sstream %" PRIu64,
1083 stream->is_metadata ? "metadata " : "",
1084 stream->stream_handle);
1085 ret = -1;
1086 goto end;
1087 }
1088 }
1089
1090 while (padding_to_write > 0) {
1091 const size_t padding_to_write_this_pass =
1092 std::min(padding_to_write, sizeof(padding_buffer));
1093
1094 write_ret =
1095 fs_handle_write(stream->file, padding_buffer, padding_to_write_this_pass);
1096 if (write_ret != padding_to_write_this_pass) {
1097 PERROR("Failed to write padding to file of %sstream %" PRIu64,
1098 stream->is_metadata ? "metadata " : "",
1099 stream->stream_handle);
1100 ret = -1;
1101 goto end;
1102 }
1103 padding_to_write -= padding_to_write_this_pass;
1104 }
1105
1106 if (stream->is_metadata) {
1107 size_t recv_len;
1108
1109 recv_len = packet ? packet->size : 0;
1110 recv_len += padding_len;
1111 stream->metadata_received += recv_len;
1112 if (recv_len) {
1113 stream->no_new_metadata_notified = false;
1114 }
1115 }
1116
1117 DBG("Wrote to %sstream %" PRIu64 ": data_length = %zu, padding_length = %zu",
1118 stream->is_metadata ? "metadata " : "",
1119 stream->stream_handle,
1120 packet ? packet->size : (size_t) 0,
1121 padding_len);
1122 end:
1123 return ret;
1124 }
1125
1126 /*
1127 * Update index after receiving a packet for a data stream.
1128 *
1129 * Called with the stream lock held.
1130 *
1131 * Return 0 on success else a negative value.
1132 */
1133 int stream_update_index(struct relay_stream *stream,
1134 uint64_t net_seq_num,
1135 bool rotate_index,
1136 bool *flushed,
1137 uint64_t total_size)
1138 {
1139 int ret = 0;
1140 uint64_t data_offset;
1141 struct relay_index *index;
1142
1143 LTTNG_ASSERT(stream->trace_chunk);
1144 ASSERT_LOCKED(stream->lock);
1145 /* Get data offset because we are about to update the index. */
1146 data_offset = htobe64(stream->tracefile_size_current);
1147
1148 DBG("handle_index_data: stream %" PRIu64 " net_seq_num %" PRIu64 " data offset %" PRIu64,
1149 stream->stream_handle,
1150 net_seq_num,
1151 stream->tracefile_size_current);
1152
1153 /*
1154 * Lookup for an existing index for that stream id/sequence
1155 * number. If it exists, the control thread has already received the
1156 * data for it, thus we need to write it to disk.
1157 */
1158 index = relay_index_get_by_id_or_create(stream, net_seq_num);
1159 if (!index) {
1160 ret = -1;
1161 goto end;
1162 }
1163
1164 if (rotate_index || !stream->index_file) {
1165 ret = create_index_file(stream, stream->trace_chunk);
1166 if (ret) {
1167 ERR("Failed to create index file for stream %" PRIu64,
1168 stream->stream_handle);
1169 /* Put self-ref for this index due to error. */
1170 relay_index_put(index);
1171 index = nullptr;
1172 goto end;
1173 }
1174 }
1175
1176 if (relay_index_set_file(index, stream->index_file, data_offset)) {
1177 ret = -1;
1178 /* Put self-ref for this index due to error. */
1179 relay_index_put(index);
1180 index = nullptr;
1181 goto end;
1182 }
1183
1184 ret = relay_index_try_flush(index);
1185 if (ret == 0) {
1186 tracefile_array_file_rotate(stream->tfa, TRACEFILE_ROTATE_READ);
1187 tracefile_array_commit_seq(stream->tfa, stream->index_received_seqcount);
1188 stream->index_received_seqcount++;
1189 LTTNG_OPTIONAL_SET(&stream->received_packet_seq_num,
1190 be64toh(index->index_data.packet_seq_num));
1191 *flushed = true;
1192 } else if (ret > 0) {
1193 index->total_size = total_size;
1194 /* No flush. */
1195 ret = 0;
1196 } else {
1197 /*
1198 * ret < 0
1199 *
1200 * relay_index_try_flush is responsible for the self-reference
1201 * put of the index object on error.
1202 */
1203 ERR("relay_index_try_flush error %d", ret);
1204 ret = -1;
1205 }
1206 end:
1207 return ret;
1208 }
1209
1210 int stream_complete_packet(struct relay_stream *stream,
1211 size_t packet_total_size,
1212 uint64_t sequence_number,
1213 bool index_flushed)
1214 {
1215 int ret = 0;
1216
1217 ASSERT_LOCKED(stream->lock);
1218
1219 stream->tracefile_size_current += packet_total_size;
1220 if (index_flushed) {
1221 stream->pos_after_last_complete_data_index = stream->tracefile_size_current;
1222 stream->prev_index_seq = sequence_number;
1223 ret = try_rotate_stream_index(stream);
1224 if (ret < 0) {
1225 goto end;
1226 }
1227 }
1228
1229 stream->prev_data_seq = sequence_number;
1230 ret = try_rotate_stream_data(stream);
1231
1232 end:
1233 return ret;
1234 }
1235
1236 int stream_add_index(struct relay_stream *stream, const struct lttcomm_relayd_index *index_info)
1237 {
1238 int ret = 0;
1239 struct relay_index *index;
1240
1241 ASSERT_LOCKED(stream->lock);
1242
1243 DBG("stream_add_index for stream %" PRIu64, stream->stream_handle);
1244
1245 /* Live beacon handling */
1246 if (index_info->packet_size == 0) {
1247 DBG("Received live beacon for stream %" PRIu64, stream->stream_handle);
1248
1249 /*
1250 * Only flag a stream inactive when it has already
1251 * received data and no indexes are in flight.
1252 */
1253 if (stream->index_received_seqcount > 0 && stream->indexes_in_flight == 0) {
1254 stream->beacon_ts_end = index_info->timestamp_end;
1255 }
1256 ret = 0;
1257 goto end;
1258 } else {
1259 stream->beacon_ts_end = -1ULL;
1260 }
1261
1262 if (stream->ctf_stream_id == -1ULL) {
1263 stream->ctf_stream_id = index_info->stream_id;
1264 }
1265
1266 index = relay_index_get_by_id_or_create(stream, index_info->net_seq_num);
1267 if (!index) {
1268 ret = -1;
1269 ERR("Failed to get or create index %" PRIu64, index_info->net_seq_num);
1270 goto end;
1271 }
1272 if (relay_index_set_control_data(index, index_info, stream->trace->session->minor)) {
1273 ERR("set_index_control_data error");
1274 relay_index_put(index);
1275 ret = -1;
1276 goto end;
1277 }
1278 ret = relay_index_try_flush(index);
1279 if (ret == 0) {
1280 tracefile_array_file_rotate(stream->tfa, TRACEFILE_ROTATE_READ);
1281 tracefile_array_commit_seq(stream->tfa, stream->index_received_seqcount);
1282 stream->index_received_seqcount++;
1283 stream->pos_after_last_complete_data_index += index->total_size;
1284 stream->prev_index_seq = index_info->net_seq_num;
1285 LTTNG_OPTIONAL_SET(&stream->received_packet_seq_num, index_info->packet_seq_num);
1286
1287 ret = try_rotate_stream_index(stream);
1288 if (ret < 0) {
1289 goto end;
1290 }
1291 ret = try_rotate_stream_data(stream);
1292 if (ret < 0) {
1293 goto end;
1294 }
1295 } else if (ret > 0) {
1296 /* no flush. */
1297 ret = 0;
1298 } else {
1299 /*
1300 * ret < 0
1301 *
1302 * relay_index_try_flush is responsible for the self-reference
1303 * put of the index object on error.
1304 */
1305 ERR("relay_index_try_flush error %d", ret);
1306 ret = -1;
1307 }
1308 end:
1309 return ret;
1310 }
1311
1312 static void print_stream_indexes(struct relay_stream *stream)
1313 {
1314 struct lttng_ht_iter iter;
1315 struct relay_index *index;
1316
1317 {
1318 lttng::urcu::read_lock_guard read_lock;
1319
1320 cds_lfht_for_each_entry (stream->indexes_ht->ht, &iter.iter, index, index_n.node) {
1321 DBG("index %p net_seq_num %" PRIu64 " refcount %ld"
1322 " stream %" PRIu64 " trace %" PRIu64 " session %" PRIu64,
1323 index,
1324 index->index_n.key,
1325 stream->ref.refcount,
1326 index->stream->stream_handle,
1327 index->stream->trace->id,
1328 index->stream->trace->session->id);
1329 }
1330 }
1331 }
1332
1333 int stream_reset_file(struct relay_stream *stream)
1334 {
1335 ASSERT_LOCKED(stream->lock);
1336
1337 if (stream->file) {
1338 int ret;
1339
1340 ret = fs_handle_close(stream->file);
1341 if (ret) {
1342 ERR("Failed to close stream file handle: channel name = \"%s\", id = %" PRIu64,
1343 stream->channel_name,
1344 stream->stream_handle);
1345 }
1346 stream->file = nullptr;
1347 }
1348
1349 DBG("%s: reset tracefile_size_current for stream %" PRIu64 " was %" PRIu64,
1350 __func__,
1351 stream->stream_handle,
1352 stream->tracefile_size_current);
1353 stream->tracefile_size_current = 0;
1354 stream->prev_data_seq = 0;
1355 stream->prev_index_seq = 0;
1356 /* Note that this does not reset the tracefile array. */
1357 stream->tracefile_current_index = 0;
1358 stream->pos_after_last_complete_data_index = 0;
1359
1360 return stream_create_data_output_file_from_trace_chunk(
1361 stream, stream->trace_chunk, true, &stream->file);
1362 }
1363
1364 void print_relay_streams()
1365 {
1366 struct lttng_ht_iter iter;
1367 struct relay_stream *stream;
1368
1369 if (!relay_streams_ht) {
1370 return;
1371 }
1372
1373 {
1374 lttng::urcu::read_lock_guard read_lock;
1375
1376 cds_lfht_for_each_entry (relay_streams_ht->ht, &iter.iter, stream, node.node) {
1377 if (!stream_get(stream)) {
1378 continue;
1379 }
1380
1381 DBG("stream %p refcount %ld stream %" PRIu64 " trace %" PRIu64
1382 " session %" PRIu64,
1383 stream,
1384 stream->ref.refcount,
1385 stream->stream_handle,
1386 stream->trace->id,
1387 stream->trace->session->id);
1388 print_stream_indexes(stream);
1389 stream_put(stream);
1390 }
1391 }
1392 }
This page took 0.09242 seconds and 4 git commands to generate.