Fix: fail on relayd lookup when finding a relayd is expected
[lttng-tools.git] / src / common / consumer / consumer-stream.c
1 /*
2 * Copyright (C) 2011 - Julien Desfossez <julien.desfossez@polymtl.ca>
3 * Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
4 * Copyright (C) 2013 - David Goulet <dgoulet@efficios.com>
5 *
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License, version 2 only, as
8 * 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 with
16 * this program; if not, write to the Free Software Foundation, Inc., 51
17 * Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
18 */
19
20 #define _LGPL_SOURCE
21 #include <assert.h>
22 #include <inttypes.h>
23 #include <sys/mman.h>
24 #include <unistd.h>
25
26 #include <common/common.h>
27 #include <common/index/index.h>
28 #include <common/kernel-consumer/kernel-consumer.h>
29 #include <common/relayd/relayd.h>
30 #include <common/ust-consumer/ust-consumer.h>
31 #include <common/utils.h>
32
33 #include "consumer-stream.h"
34
35 /*
36 * RCU call to free stream. MUST only be used with call_rcu().
37 */
38 static void free_stream_rcu(struct rcu_head *head)
39 {
40 struct lttng_ht_node_u64 *node =
41 caa_container_of(head, struct lttng_ht_node_u64, head);
42 struct lttng_consumer_stream *stream =
43 caa_container_of(node, struct lttng_consumer_stream, node);
44
45 pthread_mutex_destroy(&stream->lock);
46 free(stream);
47 }
48
49 /*
50 * Close stream on the relayd side. This call can destroy a relayd if the
51 * conditions are met.
52 *
53 * A RCU read side lock MUST be acquired if the relayd object was looked up in
54 * a hash table before calling this.
55 */
56 void consumer_stream_relayd_close(struct lttng_consumer_stream *stream,
57 struct consumer_relayd_sock_pair *relayd)
58 {
59 int ret;
60
61 assert(stream);
62 assert(relayd);
63
64 if (stream->sent_to_relayd) {
65 uatomic_dec(&relayd->refcount);
66 assert(uatomic_read(&relayd->refcount) >= 0);
67 }
68
69 /* Closing streams requires to lock the control socket. */
70 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
71 ret = relayd_send_close_stream(&relayd->control_sock,
72 stream->relayd_stream_id,
73 stream->next_net_seq_num - 1);
74 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
75 if (ret < 0) {
76 DBG("Unable to close stream on the relayd. Continuing");
77 /*
78 * Continue here. There is nothing we can do for the relayd.
79 * Chances are that the relayd has closed the socket so we just
80 * continue cleaning up.
81 */
82 }
83
84 /* Both conditions are met, we destroy the relayd. */
85 if (uatomic_read(&relayd->refcount) == 0 &&
86 uatomic_read(&relayd->destroy_flag)) {
87 consumer_destroy_relayd(relayd);
88 }
89 stream->net_seq_idx = (uint64_t) -1ULL;
90 stream->sent_to_relayd = 0;
91 }
92
93 /*
94 * Close stream's file descriptors and, if needed, close stream also on the
95 * relayd side.
96 *
97 * The consumer data lock MUST be acquired.
98 * The stream lock MUST be acquired.
99 */
100 void consumer_stream_close(struct lttng_consumer_stream *stream)
101 {
102 int ret;
103 struct consumer_relayd_sock_pair *relayd;
104
105 assert(stream);
106
107 switch (consumer_data.type) {
108 case LTTNG_CONSUMER_KERNEL:
109 if (stream->mmap_base != NULL) {
110 ret = munmap(stream->mmap_base, stream->mmap_len);
111 if (ret != 0) {
112 PERROR("munmap");
113 }
114 }
115
116 if (stream->wait_fd >= 0) {
117 ret = close(stream->wait_fd);
118 if (ret) {
119 PERROR("close");
120 }
121 stream->wait_fd = -1;
122 }
123 if (stream->chan->output == CONSUMER_CHANNEL_SPLICE) {
124 utils_close_pipe(stream->splice_pipe);
125 }
126 break;
127 case LTTNG_CONSUMER32_UST:
128 case LTTNG_CONSUMER64_UST:
129 {
130 /*
131 * Special case for the metadata since the wait fd is an internal pipe
132 * polled in the metadata thread.
133 */
134 if (stream->metadata_flag && stream->chan->monitor) {
135 int rpipe = stream->ust_metadata_poll_pipe[0];
136
137 /*
138 * This will stop the channel timer if one and close the write side
139 * of the metadata poll pipe.
140 */
141 lttng_ustconsumer_close_metadata(stream->chan);
142 if (rpipe >= 0) {
143 ret = close(rpipe);
144 if (ret < 0) {
145 PERROR("closing metadata pipe read side");
146 }
147 stream->ust_metadata_poll_pipe[0] = -1;
148 }
149 }
150 break;
151 }
152 default:
153 ERR("Unknown consumer_data type");
154 assert(0);
155 }
156
157 /* Close output fd. Could be a socket or local file at this point. */
158 if (stream->out_fd >= 0) {
159 ret = close(stream->out_fd);
160 if (ret) {
161 PERROR("close");
162 }
163 stream->out_fd = -1;
164 }
165
166 if (stream->index_file) {
167 lttng_index_file_put(stream->index_file);
168 stream->index_file = NULL;
169 }
170
171 /* Check and cleanup relayd if needed. */
172 rcu_read_lock();
173 relayd = consumer_find_relayd(stream->net_seq_idx);
174 if (relayd != NULL) {
175 consumer_stream_relayd_close(stream, relayd);
176 }
177 rcu_read_unlock();
178 }
179
180 /*
181 * Delete the stream from all possible hash tables.
182 *
183 * The consumer data lock MUST be acquired.
184 * The stream lock MUST be acquired.
185 */
186 void consumer_stream_delete(struct lttng_consumer_stream *stream,
187 struct lttng_ht *ht)
188 {
189 int ret;
190 struct lttng_ht_iter iter;
191
192 assert(stream);
193 /* Should NEVER be called not in monitor mode. */
194 assert(stream->chan->monitor);
195
196 rcu_read_lock();
197
198 if (ht) {
199 iter.iter.node = &stream->node.node;
200 ret = lttng_ht_del(ht, &iter);
201 assert(!ret);
202 }
203
204 /* Delete from stream per channel ID hash table. */
205 iter.iter.node = &stream->node_channel_id.node;
206 /*
207 * The returned value is of no importance. Even if the node is NOT in the
208 * hash table, we continue since we may have been called by a code path
209 * that did not add the stream to a (all) hash table. Same goes for the
210 * next call ht del call.
211 */
212 (void) lttng_ht_del(consumer_data.stream_per_chan_id_ht, &iter);
213
214 /* Delete from the global stream list. */
215 iter.iter.node = &stream->node_session_id.node;
216 /* See the previous ht del on why we ignore the returned value. */
217 (void) lttng_ht_del(consumer_data.stream_list_ht, &iter);
218
219 rcu_read_unlock();
220
221 if (!stream->metadata_flag) {
222 /* Decrement the stream count of the global consumer data. */
223 assert(consumer_data.stream_count > 0);
224 consumer_data.stream_count--;
225 }
226 }
227
228 /*
229 * Free the given stream within a RCU call.
230 */
231 void consumer_stream_free(struct lttng_consumer_stream *stream)
232 {
233 assert(stream);
234
235 call_rcu(&stream->node.head, free_stream_rcu);
236 }
237
238 /*
239 * Destroy the stream's buffers of the tracer.
240 */
241 void consumer_stream_destroy_buffers(struct lttng_consumer_stream *stream)
242 {
243 assert(stream);
244
245 switch (consumer_data.type) {
246 case LTTNG_CONSUMER_KERNEL:
247 break;
248 case LTTNG_CONSUMER32_UST:
249 case LTTNG_CONSUMER64_UST:
250 lttng_ustconsumer_del_stream(stream);
251 break;
252 default:
253 ERR("Unknown consumer_data type");
254 assert(0);
255 }
256 }
257
258 /*
259 * Destroy and close a already created stream.
260 */
261 static void destroy_close_stream(struct lttng_consumer_stream *stream)
262 {
263 assert(stream);
264
265 DBG("Consumer stream destroy monitored key: %" PRIu64, stream->key);
266
267 /* Destroy tracer buffers of the stream. */
268 consumer_stream_destroy_buffers(stream);
269 /* Close down everything including the relayd if one. */
270 consumer_stream_close(stream);
271 }
272
273 /*
274 * Decrement the stream's channel refcount and if down to 0, return the channel
275 * pointer so it can be destroyed by the caller or NULL if not.
276 */
277 static struct lttng_consumer_channel *unref_channel(
278 struct lttng_consumer_stream *stream)
279 {
280 struct lttng_consumer_channel *free_chan = NULL;
281
282 assert(stream);
283 assert(stream->chan);
284
285 /* Update refcount of channel and see if we need to destroy it. */
286 if (!uatomic_sub_return(&stream->chan->refcount, 1)
287 && !uatomic_read(&stream->chan->nb_init_stream_left)) {
288 free_chan = stream->chan;
289 }
290
291 return free_chan;
292 }
293
294 /*
295 * Destroy a stream completely. This will delete, close and free the stream.
296 * Once return, the stream is NO longer usable. Its channel may get destroyed
297 * if conditions are met for a monitored stream.
298 *
299 * This MUST be called WITHOUT the consumer data and stream lock acquired if
300 * the stream is in _monitor_ mode else it does not matter.
301 */
302 void consumer_stream_destroy(struct lttng_consumer_stream *stream,
303 struct lttng_ht *ht)
304 {
305 assert(stream);
306
307 /* Stream is in monitor mode. */
308 if (stream->monitor) {
309 struct lttng_consumer_channel *free_chan = NULL;
310
311 /*
312 * This means that the stream was successfully removed from the streams
313 * list of the channel and sent to the right thread managing this
314 * stream thus being globally visible.
315 */
316 if (stream->globally_visible) {
317 pthread_mutex_lock(&consumer_data.lock);
318 pthread_mutex_lock(&stream->chan->lock);
319 pthread_mutex_lock(&stream->lock);
320 /* Remove every reference of the stream in the consumer. */
321 consumer_stream_delete(stream, ht);
322
323 destroy_close_stream(stream);
324
325 /* Update channel's refcount of the stream. */
326 free_chan = unref_channel(stream);
327
328 /* Indicates that the consumer data state MUST be updated after this. */
329 consumer_data.need_update = 1;
330
331 pthread_mutex_unlock(&stream->lock);
332 pthread_mutex_unlock(&stream->chan->lock);
333 pthread_mutex_unlock(&consumer_data.lock);
334 } else {
335 /*
336 * If the stream is not visible globally, this needs to be done
337 * outside of the consumer data lock section.
338 */
339 free_chan = unref_channel(stream);
340 }
341
342 if (free_chan) {
343 consumer_del_channel(free_chan);
344 }
345 } else {
346 destroy_close_stream(stream);
347 }
348
349 /* Free stream within a RCU call. */
350 consumer_stream_free(stream);
351 }
352
353 /*
354 * Write index of a specific stream either on the relayd or local disk.
355 *
356 * Return 0 on success or else a negative value.
357 */
358 int consumer_stream_write_index(struct lttng_consumer_stream *stream,
359 struct ctf_packet_index *element)
360 {
361 int ret;
362
363 assert(stream);
364 assert(element);
365
366 rcu_read_lock();
367 if (stream->net_seq_idx != (uint64_t) -1ULL) {
368 struct consumer_relayd_sock_pair *relayd;
369 relayd = consumer_find_relayd(stream->net_seq_idx);
370 if (relayd) {
371 pthread_mutex_lock(&relayd->ctrl_sock_mutex);
372 ret = relayd_send_index(&relayd->control_sock, element,
373 stream->relayd_stream_id, stream->next_net_seq_num - 1);
374 pthread_mutex_unlock(&relayd->ctrl_sock_mutex);
375 } else {
376 ERR("Stream %" PRIu64 " relayd ID %" PRIu64 " unknown. Can't write index.",
377 stream->key, stream->net_seq_idx);
378 ret = -1;
379 }
380 } else {
381 if (lttng_index_file_write(stream->index_file, element)) {
382 ret = -1;
383 } else {
384 ret = 0;
385 }
386 }
387 if (ret < 0) {
388 goto error;
389 }
390
391 error:
392 rcu_read_unlock();
393 return ret;
394 }
395
396 /*
397 * Actually do the metadata sync using the given metadata stream.
398 *
399 * Return 0 on success else a negative value. ENODATA can be returned also
400 * indicating that there is no metadata available for that stream.
401 */
402 static int do_sync_metadata(struct lttng_consumer_stream *metadata,
403 struct lttng_consumer_local_data *ctx)
404 {
405 int ret;
406
407 assert(metadata);
408 assert(metadata->metadata_flag);
409 assert(ctx);
410
411 /*
412 * In UST, since we have to write the metadata from the cache packet
413 * by packet, we might need to start this procedure multiple times
414 * until all the metadata from the cache has been extracted.
415 */
416 do {
417 /*
418 * Steps :
419 * - Lock the metadata stream
420 * - Check if metadata stream node was deleted before locking.
421 * - if yes, release and return success
422 * - Check if new metadata is ready (flush + snapshot pos)
423 * - If nothing : release and return.
424 * - Lock the metadata_rdv_lock
425 * - Unlock the metadata stream
426 * - cond_wait on metadata_rdv to wait the wakeup from the
427 * metadata thread
428 * - Unlock the metadata_rdv_lock
429 */
430 pthread_mutex_lock(&metadata->lock);
431
432 /*
433 * There is a possibility that we were able to acquire a reference on the
434 * stream from the RCU hash table but between then and now, the node might
435 * have been deleted just before the lock is acquired. Thus, after locking,
436 * we make sure the metadata node has not been deleted which means that the
437 * buffers are closed.
438 *
439 * In that case, there is no need to sync the metadata hence returning a
440 * success return code.
441 */
442 ret = cds_lfht_is_node_deleted(&metadata->node.node);
443 if (ret) {
444 ret = 0;
445 goto end_unlock_mutex;
446 }
447
448 switch (ctx->type) {
449 case LTTNG_CONSUMER_KERNEL:
450 /*
451 * Empty the metadata cache and flush the current stream.
452 */
453 ret = lttng_kconsumer_sync_metadata(metadata);
454 break;
455 case LTTNG_CONSUMER32_UST:
456 case LTTNG_CONSUMER64_UST:
457 /*
458 * Ask the sessiond if we have new metadata waiting and update the
459 * consumer metadata cache.
460 */
461 ret = lttng_ustconsumer_sync_metadata(ctx, metadata);
462 break;
463 default:
464 assert(0);
465 ret = -1;
466 break;
467 }
468 /*
469 * Error or no new metadata, we exit here.
470 */
471 if (ret <= 0 || ret == ENODATA) {
472 goto end_unlock_mutex;
473 }
474
475 /*
476 * At this point, new metadata have been flushed, so we wait on the
477 * rendez-vous point for the metadata thread to wake us up when it
478 * finishes consuming the metadata and continue execution.
479 */
480
481 pthread_mutex_lock(&metadata->metadata_rdv_lock);
482
483 /*
484 * Release metadata stream lock so the metadata thread can process it.
485 */
486 pthread_mutex_unlock(&metadata->lock);
487
488 /*
489 * Wait on the rendez-vous point. Once woken up, it means the metadata was
490 * consumed and thus synchronization is achieved.
491 */
492 pthread_cond_wait(&metadata->metadata_rdv, &metadata->metadata_rdv_lock);
493 pthread_mutex_unlock(&metadata->metadata_rdv_lock);
494 } while (ret == EAGAIN);
495
496 /* Success */
497 return 0;
498
499 end_unlock_mutex:
500 pthread_mutex_unlock(&metadata->lock);
501 return ret;
502 }
503
504 /*
505 * Synchronize the metadata using a given session ID. A successful acquisition
506 * of a metadata stream will trigger a request to the session daemon and a
507 * snapshot so the metadata thread can consume it.
508 *
509 * This function call is a rendez-vous point between the metadata thread and
510 * the data thread.
511 *
512 * Return 0 on success or else a negative value.
513 */
514 int consumer_stream_sync_metadata(struct lttng_consumer_local_data *ctx,
515 uint64_t session_id)
516 {
517 int ret;
518 struct lttng_consumer_stream *stream = NULL;
519 struct lttng_ht_iter iter;
520 struct lttng_ht *ht;
521
522 assert(ctx);
523
524 /* Ease our life a bit. */
525 ht = consumer_data.stream_list_ht;
526
527 rcu_read_lock();
528
529 /* Search the metadata associated with the session id of the given stream. */
530
531 cds_lfht_for_each_entry_duplicate(ht->ht,
532 ht->hash_fct(&session_id, lttng_ht_seed), ht->match_fct,
533 &session_id, &iter.iter, stream, node_session_id.node) {
534 if (!stream->metadata_flag) {
535 continue;
536 }
537
538 ret = do_sync_metadata(stream, ctx);
539 if (ret < 0) {
540 goto end;
541 }
542 }
543
544 /*
545 * Force return code to 0 (success) since ret might be ENODATA for instance
546 * which is not an error but rather that we should come back.
547 */
548 ret = 0;
549
550 end:
551 rcu_read_unlock();
552 return ret;
553 }
This page took 0.053619 seconds and 6 git commands to generate.