projects / lttng-tools.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
Clean code base from redundant verification
[lttng-tools.git] / src / bin / lttng-relayd / main.c
CommitLineData
b8aa1682
JD		 1/*
2 * Copyright (C) 2012 - Julien Desfossez <jdesfossez@efficios.com>
3 * David Goulet <dgoulet@efficios.com>
cd60b05a 4 * 2013 - Jérémie Galarneau <jeremie.galarneau@efficios.com>
7591bab1 5 * 2015 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
b8aa1682
JD		 6 *
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License, version 2 only,
9 * as published by the Free Software Foundation.
10 *
11 * This program is distributed in the hope that it will be useful, but WITHOUT
12 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
14 * more details.
15 *
16 * You should have received a copy of the GNU General Public License along
17 * with this program; if not, write to the Free Software Foundation, Inc.,
18 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
19 */
20
6c1c0768 21#define _LGPL_SOURCE
b8aa1682
JD		 22#include <getopt.h>
23#include <grp.h>
24#include <limits.h>
25#include <pthread.h>
26#include <signal.h>
27#include <stdio.h>
28#include <stdlib.h>
29#include <string.h>
30#include <sys/mman.h>
31#include <sys/mount.h>
32#include <sys/resource.h>
33#include <sys/socket.h>
34#include <sys/stat.h>
35#include <sys/types.h>
36#include <sys/wait.h>
173af62f 37#include <inttypes.h>
b8aa1682
JD		 38#include <urcu/futex.h>
39#include <urcu/uatomic.h>
40#include <unistd.h>
41#include <fcntl.h>
b8aa1682
JD		 42
43#include <lttng/lttng.h>
44#include <common/common.h>
45#include <common/compat/poll.h>
46#include <common/compat/socket.h>
f263b7fd 47#include <common/compat/endian.h>
e8fa9fb0 48#include <common/compat/getenv.h>
b8aa1682 49#include <common/defaults.h>
3fd27398 50#include <common/daemonize.h>
b8aa1682
JD		 51#include <common/futex.h>
52#include <common/sessiond-comm/sessiond-comm.h>
53#include <common/sessiond-comm/inet.h>
b8aa1682
JD		 54#include <common/sessiond-comm/relayd.h>
55#include <common/uri.h>
a02de639 56#include <common/utils.h>
d3ecc550 57#include <common/align.h>
f40ef1d5 58#include <common/config/session-config.h>
5312a3ed
JG		 59#include <common/dynamic-buffer.h>
60#include <common/buffer-view.h>
7591bab1 61#include <urcu/rculist.h>
b8aa1682 62
0f907de1 63#include "cmd.h"
d3e2ba59 64#include "ctf-trace.h"
1c20f0e2 65#include "index.h"
0f907de1 66#include "utils.h"
b8aa1682 67#include "lttng-relayd.h"
d3e2ba59 68#include "live.h"
55706a7d 69#include "health-relayd.h"
9b5e0863 70#include "testpoint.h"
2f8f53af 71#include "viewer-stream.h"
2a174661
DG		 72#include "session.h"
73#include "stream.h"
58eb9381 74#include "connection.h"
a44ca2ca 75#include "tracefile-array.h"
f056029c 76#include "tcp_keep_alive.h"
b8aa1682 77
4fc83d94
PP		 78static const char *help_msg =
79#ifdef LTTNG_EMBED_HELP
80#include <lttng-relayd.8.h>
81#else
82NULL
83#endif
84;
85
5569b118
JG		 86enum relay_connection_status {
87 RELAY_CONNECTION_STATUS_OK,
a9577b76 88 /* An error occurred while processing an event on the connection. */
5569b118
JG		 89 RELAY_CONNECTION_STATUS_ERROR,
90 /* Connection closed/shutdown cleanly. */
91 RELAY_CONNECTION_STATUS_CLOSED,
92};
93
b8aa1682 94/* command line options */
0f907de1 95char *opt_output_path;
b5218ffb 96static int opt_daemon, opt_background;
3fd27398
MD		 97
98/*
99 * We need to wait for listener and live listener threads, as well as
100 * health check thread, before being ready to signal readiness.
101 */
102#define NR_LTTNG_RELAY_READY 3
103static int lttng_relay_ready = NR_LTTNG_RELAY_READY;
0848dba7
MD		 104
105/* Size of receive buffer. */
106#define RECV_DATA_BUFFER_SIZE 65536
d3ecc550 107#define FILE_COPY_BUFFER_SIZE 65536
0848dba7 108
3fd27398
MD		 109static int recv_child_signal; /* Set to 1 when a SIGUSR1 signal is received. */
110static pid_t child_ppid; /* Internal parent PID use with daemonize. */
111
095a4ae5
MD		 112static struct lttng_uri *control_uri;
113static struct lttng_uri *data_uri;
d3e2ba59 114static struct lttng_uri *live_uri;
b8aa1682
JD		 115
116const char *progname;
b8aa1682 117
65931c8b 118const char *tracing_group_name = DEFAULT_TRACING_GROUP;
cd60b05a
JG		 119static int tracing_group_name_override;
120
121const char * const config_section_name = "relayd";
65931c8b 122
b8aa1682
JD		 123/*
124 * Quit pipe for all threads. This permits a single cancellation point
125 * for all threads when receiving an event on the pipe.
126 */
0b242f62 127int thread_quit_pipe[2] = { -1, -1 };
b8aa1682
JD		 128
129/*
130 * This pipe is used to inform the worker thread that a command is queued and
131 * ready to be processed.
132 */
58eb9381 133static int relay_conn_pipe[2] = { -1, -1 };
b8aa1682 134
26c9d55e 135/* Shared between threads */
b8aa1682
JD		 136static int dispatch_thread_exit;
137
138static pthread_t listener_thread;
139static pthread_t dispatcher_thread;
140static pthread_t worker_thread;
65931c8b 141static pthread_t health_thread;
b8aa1682 142
7591bab1
MD		 143/*
144 * last_relay_stream_id_lock protects last_relay_stream_id increment
145 * atomicity on 32-bit architectures.
146 */
147static pthread_mutex_t last_relay_stream_id_lock = PTHREAD_MUTEX_INITIALIZER;
095a4ae5 148static uint64_t last_relay_stream_id;
b8aa1682
JD		 149
150/*
151 * Relay command queue.
152 *
153 * The relay_thread_listener and relay_thread_dispatcher communicate with this
154 * queue.
155 */
58eb9381 156static struct relay_conn_queue relay_conn_queue;
b8aa1682 157
d3e2ba59
JD		 158/* Global relay stream hash table. */
159struct lttng_ht *relay_streams_ht;
160
92c6ca54
DG		 161/* Global relay viewer stream hash table. */
162struct lttng_ht *viewer_streams_ht;
163
7591bab1
MD		 164/* Global relay sessions hash table. */
165struct lttng_ht *sessions_ht;
0a6518b0 166
55706a7d 167/* Relayd health monitoring */
eea7556c 168struct health_app *health_relayd;
55706a7d 169
cd60b05a
JG		 170static struct option long_options[] = {
171 { "control-port", 1, 0, 'C', },
172 { "data-port", 1, 0, 'D', },
8d5c808e 173 { "live-port", 1, 0, 'L', },
cd60b05a 174 { "daemonize", 0, 0, 'd', },
b5218ffb 175 { "background", 0, 0, 'b', },
cd60b05a
JG		 176 { "group", 1, 0, 'g', },
177 { "help", 0, 0, 'h', },
178 { "output", 1, 0, 'o', },
179 { "verbose", 0, 0, 'v', },
180 { "config", 1, 0, 'f' },
3a904098 181 { "version", 0, 0, 'V' },
cd60b05a
JG		 182 { NULL, 0, 0, 0, },
183};
184
3a904098 185static const char *config_ignore_options[] = { "help", "config", "version" };
cd60b05a 186
cd60b05a
JG		 187/*
188 * Take an option from the getopt output and set it in the right variable to be
189 * used later.
190 *
191 * Return 0 on success else a negative value.
192 */
7591bab1 193static int set_option(int opt, const char *arg, const char *optname)
b8aa1682 194{
cd60b05a
JG		 195 int ret;
196
197 switch (opt) {
198 case 0:
199 fprintf(stderr, "option %s", optname);
200 if (arg) {
201 fprintf(stderr, " with arg %s\n", arg);
202 }
203 break;
204 case 'C':
e8fa9fb0
MD		 205 if (lttng_is_setuid_setgid()) {
206 WARN("Getting '%s' argument from setuid/setgid binary refused for security reasons.",
207 "-C, --control-port");
208 } else {
209 ret = uri_parse(arg, &control_uri);
210 if (ret < 0) {
211 ERR("Invalid control URI specified");
212 goto end;
213 }
214 if (control_uri->port == 0) {
215 control_uri->port = DEFAULT_NETWORK_CONTROL_PORT;
216 }
cd60b05a
JG		 217 }
218 break;
219 case 'D':
e8fa9fb0
MD		 220 if (lttng_is_setuid_setgid()) {
221 WARN("Getting '%s' argument from setuid/setgid binary refused for security reasons.",
222 "-D, -data-port");
223 } else {
224 ret = uri_parse(arg, &data_uri);
225 if (ret < 0) {
226 ERR("Invalid data URI specified");
227 goto end;
228 }
229 if (data_uri->port == 0) {
230 data_uri->port = DEFAULT_NETWORK_DATA_PORT;
231 }
cd60b05a
JG		 232 }
233 break;
8d5c808e 234 case 'L':
e8fa9fb0
MD		 235 if (lttng_is_setuid_setgid()) {
236 WARN("Getting '%s' argument from setuid/setgid binary refused for security reasons.",
237 "-L, -live-port");
238 } else {
239 ret = uri_parse(arg, &live_uri);
240 if (ret < 0) {
241 ERR("Invalid live URI specified");
242 goto end;
243 }
244 if (live_uri->port == 0) {
245 live_uri->port = DEFAULT_NETWORK_VIEWER_PORT;
246 }
8d5c808e
AM		 247 }
248 break;
cd60b05a
JG		 249 case 'd':
250 opt_daemon = 1;
251 break;
b5218ffb
MD		 252 case 'b':
253 opt_background = 1;
254 break;
cd60b05a 255 case 'g':
e8fa9fb0
MD		 256 if (lttng_is_setuid_setgid()) {
257 WARN("Getting '%s' argument from setuid/setgid binary refused for security reasons.",
258 "-g, --group");
259 } else {
260 tracing_group_name = strdup(arg);
261 if (tracing_group_name == NULL) {
262 ret = -errno;
263 PERROR("strdup");
264 goto end;
265 }
266 tracing_group_name_override = 1;
330a40bb 267 }
cd60b05a
JG		 268 break;
269 case 'h':
4fc83d94 270 ret = utils_show_help(8, "lttng-relayd", help_msg);
655b5cc1 271 if (ret) {
4fc83d94 272 ERR("Cannot show --help for `lttng-relayd`");
655b5cc1
PP		 273 perror("exec");
274 }
cd60b05a 275 exit(EXIT_FAILURE);
3a904098
AW		 276 case 'V':
277 fprintf(stdout, "%s\n", VERSION);
278 exit(EXIT_SUCCESS);
cd60b05a 279 case 'o':
e8fa9fb0
MD		 280 if (lttng_is_setuid_setgid()) {
281 WARN("Getting '%s' argument from setuid/setgid binary refused for security reasons.",
282 "-o, --output");
283 } else {
284 ret = asprintf(&opt_output_path, "%s", arg);
285 if (ret < 0) {
286 ret = -errno;
287 PERROR("asprintf opt_output_path");
288 goto end;
289 }
cd60b05a
JG		 290 }
291 break;
292 case 'v':
293 /* Verbose level can increase using multiple -v */
294 if (arg) {
295 lttng_opt_verbose = config_parse_value(arg);
296 } else {
849e5b7b
DG		 297 /* Only 3 level of verbosity (-vvv). */
298 if (lttng_opt_verbose < 3) {
299 lttng_opt_verbose += 1;
300 }
cd60b05a
JG		 301 }
302 break;
303 default:
304 /* Unknown option or other error.
305 * Error is printed by getopt, just return */
306 ret = -1;
307 goto end;
308 }
309
310 /* All good. */
311 ret = 0;
312
313end:
314 return ret;
315}
316
317/*
318 * config_entry_handler_cb used to handle options read from a config file.
f40ef1d5 319 * See config_entry_handler_cb comment in common/config/session-config.h for the
cd60b05a
JG		 320 * return value conventions.
321 */
7591bab1 322static int config_entry_handler(const struct config_entry *entry, void *unused)
cd60b05a
JG		 323{
324 int ret = 0, i;
325
326 if (!entry || !entry->name || !entry->value) {
327 ret = -EINVAL;
328 goto end;
329 }
330
331 /* Check if the option is to be ignored */
332 for (i = 0; i < sizeof(config_ignore_options) / sizeof(char *); i++) {
333 if (!strcmp(entry->name, config_ignore_options[i])) {
334 goto end;
335 }
336 }
337
338 for (i = 0; i < (sizeof(long_options) / sizeof(struct option)) - 1; i++) {
339 /* Ignore if entry name is not fully matched. */
340 if (strcmp(entry->name, long_options[i].name)) {
341 continue;
342 }
343
344 /*
7591bab1
MD		 345 * If the option takes no argument on the command line,
346 * we have to check if the value is "true". We support
347 * non-zero numeric values, true, on and yes.
cd60b05a
JG		 348 */
349 if (!long_options[i].has_arg) {
350 ret = config_parse_value(entry->value);
351 if (ret <= 0) {
352 if (ret) {
353 WARN("Invalid configuration value \"%s\" for option %s",
354 entry->value, entry->name);
355 }
356 /* False, skip boolean config option. */
357 goto end;
358 }
359 }
360
361 ret = set_option(long_options[i].val, entry->value, entry->name);
362 goto end;
363 }
364
365 WARN("Unrecognized option \"%s\" in daemon configuration file.",
366 entry->name);
367
368end:
369 return ret;
370}
371
7591bab1 372static int set_options(int argc, char **argv)
cd60b05a 373{
178a0557 374 int c, ret = 0, option_index = 0, retval = 0;
cd60b05a
JG		 375 int orig_optopt = optopt, orig_optind = optind;
376 char *default_address, *optstring;
377 const char *config_path = NULL;
378
379 optstring = utils_generate_optstring(long_options,
380 sizeof(long_options) / sizeof(struct option));
381 if (!optstring) {
178a0557 382 retval = -ENOMEM;
cd60b05a
JG		 383 goto exit;
384 }
385
386 /* Check for the --config option */
387
388 while ((c = getopt_long(argc, argv, optstring, long_options,
389 &option_index)) != -1) {
390 if (c == '?') {
178a0557 391 retval = -EINVAL;
cd60b05a
JG		 392 goto exit;
393 } else if (c != 'f') {
394 continue;
395 }
396
e8fa9fb0
MD		 397 if (lttng_is_setuid_setgid()) {
398 WARN("Getting '%s' argument from setuid/setgid binary refused for security reasons.",
399 "-f, --config");
400 } else {
401 config_path = utils_expand_path(optarg);
402 if (!config_path) {
403 ERR("Failed to resolve path: %s", optarg);
404 }
cd60b05a
JG		 405 }
406 }
407
408 ret = config_get_section_entries(config_path, config_section_name,
409 config_entry_handler, NULL);
410 if (ret) {
411 if (ret > 0) {
412 ERR("Invalid configuration option at line %i", ret);
cd60b05a 413 }
178a0557 414 retval = -1;
cd60b05a
JG		 415 goto exit;
416 }
b8aa1682 417
cd60b05a
JG		 418 /* Reset getopt's global state */
419 optopt = orig_optopt;
420 optind = orig_optind;
b8aa1682 421 while (1) {
cd60b05a 422 c = getopt_long(argc, argv, optstring, long_options, &option_index);
b8aa1682
JD		 423 if (c == -1) {
424 break;
425 }
426
cd60b05a
JG		 427 ret = set_option(c, optarg, long_options[option_index].name);
428 if (ret < 0) {
178a0557 429 retval = -1;
b8aa1682
JD		 430 goto exit;
431 }
432 }
433
434 /* assign default values */
435 if (control_uri == NULL) {
fa91dc52
MD		 436 ret = asprintf(&default_address,
437 "tcp://" DEFAULT_NETWORK_CONTROL_BIND_ADDRESS ":%d",
438 DEFAULT_NETWORK_CONTROL_PORT);
b8aa1682
JD		 439 if (ret < 0) {
440 PERROR("asprintf default data address");
178a0557 441 retval = -1;
b8aa1682
JD		 442 goto exit;
443 }
444
445 ret = uri_parse(default_address, &control_uri);
446 free(default_address);
447 if (ret < 0) {
448 ERR("Invalid control URI specified");
178a0557 449 retval = -1;
b8aa1682
JD		 450 goto exit;
451 }
452 }
453 if (data_uri == NULL) {
fa91dc52
MD		 454 ret = asprintf(&default_address,
455 "tcp://" DEFAULT_NETWORK_DATA_BIND_ADDRESS ":%d",
456 DEFAULT_NETWORK_DATA_PORT);
b8aa1682
JD		 457 if (ret < 0) {
458 PERROR("asprintf default data address");
178a0557 459 retval = -1;
b8aa1682
JD		 460 goto exit;
461 }
462
463 ret = uri_parse(default_address, &data_uri);
464 free(default_address);
465 if (ret < 0) {
466 ERR("Invalid data URI specified");
178a0557 467 retval = -1;
b8aa1682
JD		 468 goto exit;
469 }
470 }
d3e2ba59 471 if (live_uri == NULL) {
fa91dc52
MD		 472 ret = asprintf(&default_address,
473 "tcp://" DEFAULT_NETWORK_VIEWER_BIND_ADDRESS ":%d",
474 DEFAULT_NETWORK_VIEWER_PORT);
d3e2ba59
JD		 475 if (ret < 0) {
476 PERROR("asprintf default viewer control address");
178a0557 477 retval = -1;
d3e2ba59
JD		 478 goto exit;
479 }
480
481 ret = uri_parse(default_address, &live_uri);
482 free(default_address);
483 if (ret < 0) {
484 ERR("Invalid viewer control URI specified");
178a0557 485 retval = -1;
d3e2ba59
JD		 486 goto exit;
487 }
488 }
b8aa1682
JD		 489
490exit:
cd60b05a 491 free(optstring);
178a0557 492 return retval;
b8aa1682
JD		 493}
494
7591bab1
MD		 495static void print_global_objects(void)
496{
497 rcu_register_thread();
498
499 print_viewer_streams();
500 print_relay_streams();
501 print_sessions();
502
503 rcu_unregister_thread();
504}
505
b8aa1682
JD		 506/*
507 * Cleanup the daemon
508 */
7591bab1 509static void relayd_cleanup(void)
b8aa1682 510{
7591bab1
MD		 511 print_global_objects();
512
b8aa1682
JD		 513 DBG("Cleaning up");
514
178a0557
MD		 515 if (viewer_streams_ht)
516 lttng_ht_destroy(viewer_streams_ht);
517 if (relay_streams_ht)
518 lttng_ht_destroy(relay_streams_ht);
7591bab1
MD		 519 if (sessions_ht)
520 lttng_ht_destroy(sessions_ht);
178a0557 521
095a4ae5
MD		 522 /* free the dynamically allocated opt_output_path */
523 free(opt_output_path);
524
a02de639
CB		 525 /* Close thread quit pipes */
526 utils_close_pipe(thread_quit_pipe);
527
710c1f73
DG		 528 uri_free(control_uri);
529 uri_free(data_uri);
8d5c808e 530 /* Live URI is freed in the live thread. */
cd60b05a
JG		 531
532 if (tracing_group_name_override) {
533 free((void *) tracing_group_name);
534 }
b8aa1682
JD		 535}
536
537/*
538 * Write to writable pipe used to notify a thread.
539 */
7591bab1 540static int notify_thread_pipe(int wpipe)
b8aa1682 541{
6cd525e8 542 ssize_t ret;
b8aa1682 543
6cd525e8
MD		 544 ret = lttng_write(wpipe, "!", 1);
545 if (ret < 1) {
b8aa1682 546 PERROR("write poll pipe");
b4aacfdc 547 goto end;
b8aa1682 548 }
b4aacfdc
MD		 549 ret = 0;
550end:
b8aa1682
JD		 551 return ret;
552}
553
7591bab1 554static int notify_health_quit_pipe(int *pipe)
65931c8b 555{
6cd525e8 556 ssize_t ret;
65931c8b 557
6cd525e8
MD		 558 ret = lttng_write(pipe[1], "4", 1);
559 if (ret < 1) {
65931c8b 560 PERROR("write relay health quit");
b4aacfdc 561 goto end;
65931c8b 562 }
b4aacfdc
MD		 563 ret = 0;
564end:
565 return ret;
65931c8b
MD		 566}
567
b8aa1682 568/*
b4aacfdc 569 * Stop all relayd and relayd-live threads.
b8aa1682 570 */
b4aacfdc 571int lttng_relay_stop_threads(void)
b8aa1682 572{
b4aacfdc 573 int retval = 0;
b8aa1682
JD		 574
575 /* Stopping all threads */
576 DBG("Terminating all threads");
b4aacfdc 577 if (notify_thread_pipe(thread_quit_pipe[1])) {
b8aa1682 578 ERR("write error on thread quit pipe");
b4aacfdc 579 retval = -1;
b8aa1682
JD		 580 }
581
b4aacfdc
MD		 582 if (notify_health_quit_pipe(health_quit_pipe)) {
583 ERR("write error on health quit pipe");
584 }
65931c8b 585
b8aa1682 586 /* Dispatch thread */
26c9d55e 587 CMM_STORE_SHARED(dispatch_thread_exit, 1);
58eb9381 588 futex_nto1_wake(&relay_conn_queue.futex);
178a0557 589
b4aacfdc 590 if (relayd_live_stop()) {
178a0557 591 ERR("Error stopping live threads");
b4aacfdc 592 retval = -1;
178a0557 593 }
b4aacfdc 594 return retval;
b8aa1682
JD		 595}
596
597/*
598 * Signal handler for the daemon
599 *
600 * Simply stop all worker threads, leaving main() return gracefully after
601 * joining all threads and calling cleanup().
602 */
7591bab1 603static void sighandler(int sig)
b8aa1682
JD		 604{
605 switch (sig) {
b8aa1682
JD		 606 case SIGINT:
607 DBG("SIGINT caught");
b4aacfdc
MD		 608 if (lttng_relay_stop_threads()) {
609 ERR("Error stopping threads");
610 }
b8aa1682
JD		 611 break;
612 case SIGTERM:
613 DBG("SIGTERM caught");
b4aacfdc
MD		 614 if (lttng_relay_stop_threads()) {
615 ERR("Error stopping threads");
616 }
b8aa1682 617 break;
3fd27398
MD		 618 case SIGUSR1:
619 CMM_STORE_SHARED(recv_child_signal, 1);
620 break;
b8aa1682
JD		 621 default:
622 break;
623 }
624}
625
626/*
627 * Setup signal handler for :
628 * SIGINT, SIGTERM, SIGPIPE
629 */
7591bab1 630static int set_signal_handler(void)
b8aa1682
JD		 631{
632 int ret = 0;
633 struct sigaction sa;
634 sigset_t sigset;
635
636 if ((ret = sigemptyset(&sigset)) < 0) {
637 PERROR("sigemptyset");
638 return ret;
639 }
640
b8aa1682
JD		 641 sa.sa_mask = sigset;
642 sa.sa_flags = 0;
0072e5e2
MD		 643
644 sa.sa_handler = sighandler;
b8aa1682
JD		 645 if ((ret = sigaction(SIGTERM, &sa, NULL)) < 0) {
646 PERROR("sigaction");
647 return ret;
648 }
649
650 if ((ret = sigaction(SIGINT, &sa, NULL)) < 0) {
651 PERROR("sigaction");
652 return ret;
653 }
654
0072e5e2 655 if ((ret = sigaction(SIGUSR1, &sa, NULL)) < 0) {
b8aa1682
JD		 656 PERROR("sigaction");
657 return ret;
658 }
659
0072e5e2
MD		 660 sa.sa_handler = SIG_IGN;
661 if ((ret = sigaction(SIGPIPE, &sa, NULL)) < 0) {
3fd27398
MD		 662 PERROR("sigaction");
663 return ret;
664 }
665
666 DBG("Signal handler set for SIGTERM, SIGUSR1, SIGPIPE and SIGINT");
b8aa1682
JD		 667
668 return ret;
669}
670
3fd27398
MD		 671void lttng_relay_notify_ready(void)
672{
673 /* Notify the parent of the fork() process that we are ready. */
674 if (opt_daemon || opt_background) {
675 if (uatomic_sub_return(&lttng_relay_ready, 1) == 0) {
676 kill(child_ppid, SIGUSR1);
677 }
678 }
679}
680
b8aa1682
JD		 681/*
682 * Init thread quit pipe.
683 *
684 * Return -1 on error or 0 if all pipes are created.
685 */
7591bab1 686static int init_thread_quit_pipe(void)
b8aa1682 687{
a02de639 688 int ret;
b8aa1682 689
a02de639 690 ret = utils_create_pipe_cloexec(thread_quit_pipe);
b8aa1682 691
b8aa1682
JD		 692 return ret;
693}
694
695/*
696 * Create a poll set with O_CLOEXEC and add the thread quit pipe to the set.
697 */
7591bab1 698static int create_thread_poll_set(struct lttng_poll_event *events, int size)
b8aa1682
JD		 699{
700 int ret;
701
702 if (events == NULL || size == 0) {
703 ret = -1;
704 goto error;
705 }
706
707 ret = lttng_poll_create(events, size, LTTNG_CLOEXEC);
708 if (ret < 0) {
709 goto error;
710 }
711
712 /* Add quit pipe */
c7759e6a 713 ret = lttng_poll_add(events, thread_quit_pipe[0], LPOLLIN | LPOLLERR);
b8aa1682
JD		 714 if (ret < 0) {
715 goto error;
716 }
717
718 return 0;
719
720error:
721 return ret;
722}
723
724/*
725 * Check if the thread quit pipe was triggered.
726 *
727 * Return 1 if it was triggered else 0;
728 */
7591bab1 729static int check_thread_quit_pipe(int fd, uint32_t events)
b8aa1682
JD		 730{
731 if (fd == thread_quit_pipe[0] && (events & LPOLLIN)) {
732 return 1;
733 }
734
735 return 0;
736}
737
738/*
739 * Create and init socket from uri.
740 */
7591bab1 741static struct lttcomm_sock *relay_socket_create(struct lttng_uri *uri)
b8aa1682
JD		 742{
743 int ret;
744 struct lttcomm_sock *sock = NULL;
745
746 sock = lttcomm_alloc_sock_from_uri(uri);
747 if (sock == NULL) {
748 ERR("Allocating socket");
749 goto error;
750 }
751
752 ret = lttcomm_create_sock(sock);
753 if (ret < 0) {
754 goto error;
755 }
756 DBG("Listening on sock %d", sock->fd);
757
758 ret = sock->ops->bind(sock);
759 if (ret < 0) {
2288467f 760 PERROR("Failed to bind socket");
b8aa1682
JD		 761 goto error;
762 }
763
764 ret = sock->ops->listen(sock, -1);
765 if (ret < 0) {
766 goto error;
767
768 }
769
770 return sock;
771
772error:
773 if (sock) {
774 lttcomm_destroy_sock(sock);
775 }
776 return NULL;
777}
778
779/*
780 * This thread manages the listening for new connections on the network
781 */
7591bab1 782static void *relay_thread_listener(void *data)
b8aa1682 783{
095a4ae5 784 int i, ret, pollfd, err = -1;
b8aa1682
JD		 785 uint32_t revents, nb_fd;
786 struct lttng_poll_event events;
787 struct lttcomm_sock *control_sock, *data_sock;
788
b8aa1682
JD		 789 DBG("[thread] Relay listener started");
790
55706a7d
MD		 791 health_register(health_relayd, HEALTH_RELAYD_TYPE_LISTENER);
792
f385ae0a
MD		 793 health_code_update();
794
7591bab1 795 control_sock = relay_socket_create(control_uri);
b8aa1682 796 if (!control_sock) {
095a4ae5 797 goto error_sock_control;
b8aa1682
JD		 798 }
799
7591bab1 800 data_sock = relay_socket_create(data_uri);
b8aa1682 801 if (!data_sock) {
095a4ae5 802 goto error_sock_relay;
b8aa1682
JD		 803 }
804
805 /*
7591bab1
MD		 806 * Pass 3 as size here for the thread quit pipe, control and
807 * data socket.
b8aa1682
JD		 808 */
809 ret = create_thread_poll_set(&events, 3);
810 if (ret < 0) {
811 goto error_create_poll;
812 }
813
814 /* Add the control socket */
815 ret = lttng_poll_add(&events, control_sock->fd, LPOLLIN | LPOLLRDHUP);
816 if (ret < 0) {
817 goto error_poll_add;
818 }
819
820 /* Add the data socket */
821 ret = lttng_poll_add(&events, data_sock->fd, LPOLLIN | LPOLLRDHUP);
822 if (ret < 0) {
823 goto error_poll_add;
824 }
825
3fd27398
MD		 826 lttng_relay_notify_ready();
827
9b5e0863
MD		 828 if (testpoint(relayd_thread_listener)) {
829 goto error_testpoint;
830 }
831
b8aa1682 832 while (1) {
f385ae0a
MD		 833 health_code_update();
834
b8aa1682
JD		 835 DBG("Listener accepting connections");
836
b8aa1682 837restart:
f385ae0a 838 health_poll_entry();
b8aa1682 839 ret = lttng_poll_wait(&events, -1);
f385ae0a 840 health_poll_exit();
b8aa1682
JD		 841 if (ret < 0) {
842 /*
843 * Restart interrupted system call.
844 */
845 if (errno == EINTR) {
846 goto restart;
847 }
848 goto error;
849 }
850
0d9c5d77
DG		 851 nb_fd = ret;
852
b8aa1682
JD		 853 DBG("Relay new connection received");
854 for (i = 0; i < nb_fd; i++) {
f385ae0a
MD		 855 health_code_update();
856
b8aa1682
JD		 857 /* Fetch once the poll data */
858 revents = LTTNG_POLL_GETEV(&events, i);
859 pollfd = LTTNG_POLL_GETFD(&events, i);
860
861 /* Thread quit pipe has been closed. Killing thread. */
862 ret = check_thread_quit_pipe(pollfd, revents);
863 if (ret) {
095a4ae5
MD		 864 err = 0;
865 goto exit;
b8aa1682
JD		 866 }
867
03e43155 868 if (revents & LPOLLIN) {
4b7f17b2 869 /*
7591bab1
MD		 870 * A new connection is requested, therefore a
871 * sessiond/consumerd connection is allocated in
872 * this thread, enqueued to a global queue and
873 * dequeued (and freed) in the worker thread.
4b7f17b2 874 */
58eb9381
DG		 875 int val = 1;
876 struct relay_connection *new_conn;
4b7f17b2 877 struct lttcomm_sock *newsock;
7591bab1 878 enum connection_type type;
b8aa1682
JD		 879
880 if (pollfd == data_sock->fd) {
7591bab1 881 type = RELAY_DATA;
b8aa1682 882 newsock = data_sock->ops->accept(data_sock);
58eb9381
DG		 883 DBG("Relay data connection accepted, socket %d",
884 newsock->fd);
4b7f17b2
MD		 885 } else {
886 assert(pollfd == control_sock->fd);
7591bab1 887 type = RELAY_CONTROL;
b8aa1682 888 newsock = control_sock->ops->accept(control_sock);
58eb9381
DG		 889 DBG("Relay control connection accepted, socket %d",
890 newsock->fd);
b8aa1682 891 }
58eb9381
DG		 892 if (!newsock) {
893 PERROR("accepting sock");
58eb9381
DG		 894 goto error;
895 }
896
897 ret = setsockopt(newsock->fd, SOL_SOCKET, SO_REUSEADDR, &val,
898 sizeof(val));
b8aa1682
JD		 899 if (ret < 0) {
900 PERROR("setsockopt inet");
4b7f17b2 901 lttcomm_destroy_sock(newsock);
b8aa1682
JD		 902 goto error;
903 }
f056029c
JR		 904
905 ret = socket_apply_keep_alive_config(newsock->fd);
906 if (ret < 0) {
907 ERR("Failed to apply TCP keep-alive configuration on socket (%i)",
908 newsock->fd);
909 lttcomm_destroy_sock(newsock);
910 goto error;
911 }
912
7591bab1
MD		 913 new_conn = connection_create(newsock, type);
914 if (!new_conn) {
915 lttcomm_destroy_sock(newsock);
916 goto error;
917 }
58eb9381
DG		 918
919 /* Enqueue request for the dispatcher thread. */
8bdee6e2
SM		 920 cds_wfcq_enqueue(&relay_conn_queue.head, &relay_conn_queue.tail,
921 &new_conn->qnode);
b8aa1682
JD		 922
923 /*
7591bab1
MD		 924 * Wake the dispatch queue futex.
925 * Implicit memory barrier with the
926 * exchange in cds_wfcq_enqueue.
b8aa1682 927 */
58eb9381 928 futex_nto1_wake(&relay_conn_queue.futex);
03e43155
MD		 929 } else if (revents & (LPOLLERR | LPOLLHUP | LPOLLRDHUP)) {
930 ERR("socket poll error");
931 goto error;
932 } else {
933 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
934 goto error;
b8aa1682
JD		 935 }
936 }
937 }
938
095a4ae5 939exit:
b8aa1682
JD		 940error:
941error_poll_add:
9b5e0863 942error_testpoint:
b8aa1682
JD		 943 lttng_poll_clean(&events);
944error_create_poll:
095a4ae5
MD		 945 if (data_sock->fd >= 0) {
946 ret = data_sock->ops->close(data_sock);
b8aa1682
JD		 947 if (ret) {
948 PERROR("close");
949 }
b8aa1682 950 }
095a4ae5
MD		 951 lttcomm_destroy_sock(data_sock);
952error_sock_relay:
953 if (control_sock->fd >= 0) {
954 ret = control_sock->ops->close(control_sock);
b8aa1682
JD		 955 if (ret) {
956 PERROR("close");
957 }
b8aa1682 958 }
095a4ae5
MD		 959 lttcomm_destroy_sock(control_sock);
960error_sock_control:
961 if (err) {
f385ae0a
MD		 962 health_error();
963 ERR("Health error occurred in %s", __func__);
095a4ae5 964 }
55706a7d 965 health_unregister(health_relayd);
b8aa1682 966 DBG("Relay listener thread cleanup complete");
b4aacfdc 967 lttng_relay_stop_threads();
b8aa1682
JD		 968 return NULL;
969}
970
971/*
972 * This thread manages the dispatching of the requests to worker threads
973 */
7591bab1 974static void *relay_thread_dispatcher(void *data)
b8aa1682 975{
6cd525e8
MD		 976 int err = -1;
977 ssize_t ret;
8bdee6e2 978 struct cds_wfcq_node *node;
58eb9381 979 struct relay_connection *new_conn = NULL;
b8aa1682
JD		 980
981 DBG("[thread] Relay dispatcher started");
982
55706a7d
MD		 983 health_register(health_relayd, HEALTH_RELAYD_TYPE_DISPATCHER);
984
9b5e0863
MD		 985 if (testpoint(relayd_thread_dispatcher)) {
986 goto error_testpoint;
987 }
988
f385ae0a
MD		 989 health_code_update();
990
0ed3b1a8 991 for (;;) {
f385ae0a
MD		 992 health_code_update();
993
b8aa1682 994 /* Atomically prepare the queue futex */
58eb9381 995 futex_nto1_prepare(&relay_conn_queue.futex);
b8aa1682 996
0ed3b1a8
MD		 997 if (CMM_LOAD_SHARED(dispatch_thread_exit)) {
998 break;
999 }
1000
b8aa1682 1001 do {
f385ae0a
MD		 1002 health_code_update();
1003
b8aa1682 1004 /* Dequeue commands */
8bdee6e2
SM		 1005 node = cds_wfcq_dequeue_blocking(&relay_conn_queue.head,
1006 &relay_conn_queue.tail);
b8aa1682
JD		 1007 if (node == NULL) {
1008 DBG("Woken up but nothing in the relay command queue");
1009 /* Continue thread execution */
1010 break;
1011 }
58eb9381 1012 new_conn = caa_container_of(node, struct relay_connection, qnode);
b8aa1682 1013
58eb9381 1014 DBG("Dispatching request waiting on sock %d", new_conn->sock->fd);
b8aa1682
JD		 1015
1016 /*
7591bab1
MD		 1017 * Inform worker thread of the new request. This
1018 * call is blocking so we can be assured that
1019 * the data will be read at some point in time
1020 * or wait to the end of the world :)
b8aa1682 1021 */
58eb9381
DG		 1022 ret = lttng_write(relay_conn_pipe[1], &new_conn, sizeof(new_conn));
1023 if (ret < 0) {
1024 PERROR("write connection pipe");
7591bab1 1025 connection_put(new_conn);
b8aa1682
JD		 1026 goto error;
1027 }
1028 } while (node != NULL);
1029
1030 /* Futex wait on queue. Blocking call on futex() */
f385ae0a 1031 health_poll_entry();
58eb9381 1032 futex_nto1_wait(&relay_conn_queue.futex);
f385ae0a 1033 health_poll_exit();
b8aa1682
JD		 1034 }
1035
f385ae0a
MD		 1036 /* Normal exit, no error */
1037 err = 0;
1038
b8aa1682 1039error:
9b5e0863 1040error_testpoint:
f385ae0a
MD		 1041 if (err) {
1042 health_error();
1043 ERR("Health error occurred in %s", __func__);
1044 }
55706a7d 1045 health_unregister(health_relayd);
b8aa1682 1046 DBG("Dispatch thread dying");
b4aacfdc 1047 lttng_relay_stop_threads();
b8aa1682
JD		 1048 return NULL;
1049}
1050
b8aa1682 1051/*
7591bab1 1052 * Set index data from the control port to a given index object.
b8aa1682 1053 */
7591bab1 1054static int set_index_control_data(struct relay_index *index,
234cd636
JD		 1055 struct lttcomm_relayd_index *data,
1056 struct relay_connection *conn)
1c20f0e2 1057{
7591bab1 1058 struct ctf_packet_index index_data;
1c20f0e2
JD		 1059
1060 /*
5312a3ed 1061 * The index on disk is encoded in big endian.
1c20f0e2 1062 */
5312a3ed
JG		 1063 index_data.packet_size = htobe64(data->packet_size);
1064 index_data.content_size = htobe64(data->content_size);
1065 index_data.timestamp_begin = htobe64(data->timestamp_begin);
1066 index_data.timestamp_end = htobe64(data->timestamp_end);
1067 index_data.events_discarded = htobe64(data->events_discarded);
1068 index_data.stream_id = htobe64(data->stream_id);
234cd636
JD		 1069
1070 if (conn->minor >= 8) {
5312a3ed
JG		 1071 index->index_data.stream_instance_id = htobe64(data->stream_instance_id);
1072 index->index_data.packet_seq_num = htobe64(data->packet_seq_num);
234cd636
JD		 1073 }
1074
7591bab1 1075 return relay_index_set_data(index, &index_data);
1c20f0e2
JD		 1076}
1077
298a25ca
JG		 1078static bool session_streams_have_index(const struct relay_session *session)
1079{
1080 return session->minor >= 4 && !session->snapshot;
1081}
1082
c5b6f4f0
DG		 1083/*
1084 * Handle the RELAYD_CREATE_SESSION command.
1085 *
1086 * On success, send back the session id or else return a negative value.
1087 */
5312a3ed
JG		 1088static int relay_create_session(const struct lttcomm_relayd_hdr *recv_hdr,
1089 struct relay_connection *conn,
1090 const struct lttng_buffer_view *payload)
c5b6f4f0 1091{
5312a3ed
JG		 1092 int ret = 0;
1093 ssize_t send_ret;
4c6885d2 1094 struct relay_session *session = NULL;
c5b6f4f0 1095 struct lttcomm_relayd_status_session reply;
36d2e35d 1096 char session_name[LTTNG_NAME_MAX];
9cf3eb20 1097 char hostname[LTTNG_HOST_NAME_MAX];
7591bab1
MD		 1098 uint32_t live_timer = 0;
1099 bool snapshot = false;
c5b6f4f0 1100
36d2e35d 1101 memset(session_name, 0, LTTNG_NAME_MAX);
9cf3eb20 1102 memset(hostname, 0, LTTNG_HOST_NAME_MAX);
c5b6f4f0
DG		 1103
1104 memset(&reply, 0, sizeof(reply));
1105
f86f6389
JR		 1106 if (conn->minor < 4) {
1107 /* From 2.1 to 2.3 */
1108 ret = 0;
1109 } else if (conn->minor >= 4 && conn->minor < 11) {
1110 /* From 2.4 to 2.10 */
5312a3ed 1111 ret = cmd_create_session_2_4(payload, session_name,
7591bab1 1112 hostname, &live_timer, &snapshot);
f86f6389
JR		 1113 } else {
1114 /* From 2.11 to ... */
1115 ret = cmd_create_session_2_11(payload, session_name,
1116 hostname, &live_timer, &snapshot);
7591bab1 1117 }
f86f6389 1118
7591bab1
MD		 1119 if (ret < 0) {
1120 goto send_reply;
d3e2ba59
JD		 1121 }
1122
7591bab1
MD		 1123 session = session_create(session_name, hostname, live_timer,
1124 snapshot, conn->major, conn->minor);
1125 if (!session) {
1126 ret = -1;
1127 goto send_reply;
1128 }
1129 assert(!conn->session);
1130 conn->session = session;
c5b6f4f0
DG		 1131 DBG("Created session %" PRIu64, session->id);
1132
7591bab1
MD		 1133 reply.session_id = htobe64(session->id);
1134
1135send_reply:
c5b6f4f0
DG		 1136 if (ret < 0) {
1137 reply.ret_code = htobe32(LTTNG_ERR_FATAL);
1138 } else {
1139 reply.ret_code = htobe32(LTTNG_OK);
1140 }
1141
58eb9381 1142 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply, sizeof(reply), 0);
5312a3ed
JG		 1143 if (send_ret < (ssize_t) sizeof(reply)) {
1144 ERR("Failed to send \"create session\" command reply (ret = %zd)",
1145 send_ret);
1146 ret = -1;
c5b6f4f0 1147 }
4c6885d2
JG		 1148 if (ret < 0 && session) {
1149 session_put(session);
1150 }
c5b6f4f0
DG		 1151 return ret;
1152}
1153
a4baae1b
JD		 1154/*
1155 * When we have received all the streams and the metadata for a channel,
1156 * we make them visible to the viewer threads.
1157 */
7591bab1 1158static void publish_connection_local_streams(struct relay_connection *conn)
a4baae1b 1159{
7591bab1
MD		 1160 struct relay_stream *stream;
1161 struct relay_session *session = conn->session;
a4baae1b 1162
7591bab1
MD		 1163 /*
1164 * We publish all streams belonging to a session atomically wrt
1165 * session lock.
1166 */
1167 pthread_mutex_lock(&session->lock);
1168 rcu_read_lock();
1169 cds_list_for_each_entry_rcu(stream, &session->recv_list,
1170 recv_node) {
1171 stream_publish(stream);
a4baae1b 1172 }
7591bab1 1173 rcu_read_unlock();
a4baae1b 1174
7591bab1
MD		 1175 /*
1176 * Inform the viewer that there are new streams in the session.
1177 */
1178 if (session->viewer_attached) {
1179 uatomic_set(&session->new_streams, 1);
1180 }
1181 pthread_mutex_unlock(&session->lock);
a4baae1b
JD		 1182}
1183
b8aa1682
JD		 1184/*
1185 * relay_add_stream: allocate a new stream for a session
1186 */
5312a3ed
JG		 1187static int relay_add_stream(const struct lttcomm_relayd_hdr *recv_hdr,
1188 struct relay_connection *conn,
1189 const struct lttng_buffer_view *payload)
b8aa1682 1190{
7591bab1
MD		 1191 int ret;
1192 ssize_t send_ret;
58eb9381 1193 struct relay_session *session = conn->session;
b8aa1682
JD		 1194 struct relay_stream *stream = NULL;
1195 struct lttcomm_relayd_status_stream reply;
4030a636 1196 struct ctf_trace *trace = NULL;
7591bab1
MD		 1197 uint64_t stream_handle = -1ULL;
1198 char *path_name = NULL, *channel_name = NULL;
1199 uint64_t tracefile_size = 0, tracefile_count = 0;
81164b6b 1200 struct relay_stream_chunk_id stream_chunk_id = { 0 };
b8aa1682 1201
5312a3ed 1202 if (!session || !conn->version_check_done) {
b8aa1682
JD		 1203 ERR("Trying to add a stream before version check");
1204 ret = -1;
1205 goto end_no_session;
1206 }
1207
2f21a469
JR		 1208 if (session->minor == 1) {
1209 /* For 2.1 */
5312a3ed 1210 ret = cmd_recv_stream_2_1(payload, &path_name,
7591bab1 1211 &channel_name);
2f21a469
JR		 1212 } else if (session->minor > 1 && session->minor < 11) {
1213 /* From 2.2 to 2.10 */
5312a3ed 1214 ret = cmd_recv_stream_2_2(payload, &path_name,
7591bab1 1215 &channel_name, &tracefile_size, &tracefile_count);
2f21a469
JR		 1216 } else {
1217 /* From 2.11 to ... */
1218 ret = cmd_recv_stream_2_11(payload, &path_name,
0b50e4b3
JG		 1219 &channel_name, &tracefile_size, &tracefile_count,
1220 &stream_chunk_id.value);
1221 stream_chunk_id.is_set = true;
0f907de1 1222 }
2f21a469 1223
0f907de1 1224 if (ret < 0) {
7591bab1 1225 goto send_reply;
b8aa1682
JD		 1226 }
1227
7591bab1 1228 trace = ctf_trace_get_by_path_or_create(session, path_name);
2a174661 1229 if (!trace) {
7591bab1 1230 goto send_reply;
2a174661 1231 }
7591bab1 1232 /* This stream here has one reference on the trace. */
2a174661 1233
7591bab1
MD		 1234 pthread_mutex_lock(&last_relay_stream_id_lock);
1235 stream_handle = ++last_relay_stream_id;
1236 pthread_mutex_unlock(&last_relay_stream_id_lock);
d3e2ba59 1237
7591bab1
MD		 1238 /* We pass ownership of path_name and channel_name. */
1239 stream = stream_create(trace, stream_handle, path_name,
81164b6b
JG		 1240 channel_name, tracefile_size, tracefile_count,
1241 &stream_chunk_id);
7591bab1
MD		 1242 path_name = NULL;
1243 channel_name = NULL;
a4baae1b 1244
2a174661 1245 /*
7591bab1
MD		 1246 * Streams are the owners of their trace. Reference to trace is
1247 * kept within stream_create().
2a174661 1248 */
7591bab1 1249 ctf_trace_put(trace);
d3e2ba59 1250
7591bab1 1251send_reply:
53efb85a 1252 memset(&reply, 0, sizeof(reply));
7591bab1
MD		 1253 reply.handle = htobe64(stream_handle);
1254 if (!stream) {
f73fabfd 1255 reply.ret_code = htobe32(LTTNG_ERR_UNK);
b8aa1682 1256 } else {
f73fabfd 1257 reply.ret_code = htobe32(LTTNG_OK);
b8aa1682 1258 }
5af40280 1259
58eb9381 1260 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply,
b8aa1682 1261 sizeof(struct lttcomm_relayd_status_stream), 0);
5312a3ed
JG		 1262 if (send_ret < (ssize_t) sizeof(reply)) {
1263 ERR("Failed to send \"add stream\" command reply (ret = %zd)",
1264 send_ret);
1265 ret = -1;
b8aa1682
JD		 1266 }
1267
1268end_no_session:
7591bab1
MD		 1269 free(path_name);
1270 free(channel_name);
0f907de1 1271 return ret;
b8aa1682
JD		 1272}
1273
173af62f
DG		 1274/*
1275 * relay_close_stream: close a specific stream
1276 */
5312a3ed
JG		 1277static int relay_close_stream(const struct lttcomm_relayd_hdr *recv_hdr,
1278 struct relay_connection *conn,
1279 const struct lttng_buffer_view *payload)
173af62f 1280{
5312a3ed
JG		 1281 int ret;
1282 ssize_t send_ret;
58eb9381 1283 struct relay_session *session = conn->session;
173af62f
DG		 1284 struct lttcomm_relayd_close_stream stream_info;
1285 struct lttcomm_relayd_generic_reply reply;
1286 struct relay_stream *stream;
173af62f
DG		 1287
1288 DBG("Close stream received");
1289
5312a3ed 1290 if (!session || !conn->version_check_done) {
173af62f
DG		 1291 ERR("Trying to close a stream before version check");
1292 ret = -1;
1293 goto end_no_session;
1294 }
1295
5312a3ed
JG		 1296 if (payload->size < sizeof(stream_info)) {
1297 ERR("Unexpected payload size in \"relay_close_stream\": expected >= %zu bytes, got %zu bytes",
1298 sizeof(stream_info), payload->size);
173af62f
DG		 1299 ret = -1;
1300 goto end_no_session;
1301 }
5312a3ed
JG		 1302 memcpy(&stream_info, payload->data, sizeof(stream_info));
1303 stream_info.stream_id = be64toh(stream_info.stream_id);
1304 stream_info.last_net_seq_num = be64toh(stream_info.last_net_seq_num);
173af62f 1305
5312a3ed 1306 stream = stream_get_by_id(stream_info.stream_id);
173af62f
DG		 1307 if (!stream) {
1308 ret = -1;
7591bab1 1309 goto end;
173af62f 1310 }
77f7bd85
MD		 1311
1312 /*
1313 * Set last_net_seq_num before the close flag. Required by data
1314 * pending check.
1315 */
7591bab1 1316 pthread_mutex_lock(&stream->lock);
5312a3ed 1317 stream->last_net_seq_num = stream_info.last_net_seq_num;
77f7bd85
MD		 1318 pthread_mutex_unlock(&stream->lock);
1319
bda7c7b9
JG		 1320 /*
1321 * This is one of the conditions which may trigger a stream close
1322 * with the others being:
1323 * 1) A close command is received for a stream
1324 * 2) The control connection owning the stream is closed
1325 * 3) We have received all of the stream's data _after_ a close
1326 * request.
1327 */
1328 try_stream_close(stream);
7591bab1
MD		 1329 if (stream->is_metadata) {
1330 struct relay_viewer_stream *vstream;
173af62f 1331
7591bab1
MD		 1332 vstream = viewer_stream_get_by_id(stream->stream_handle);
1333 if (vstream) {
1334 if (vstream->metadata_sent == stream->metadata_received) {
1335 /*
1336 * Since all the metadata has been sent to the
1337 * viewer and that we have a request to close
1338 * its stream, we can safely teardown the
1339 * corresponding metadata viewer stream.
1340 */
1341 viewer_stream_put(vstream);
1342 }
1343 /* Put local reference. */
1344 viewer_stream_put(vstream);
1345 }
1346 }
7591bab1 1347 stream_put(stream);
5312a3ed 1348 ret = 0;
173af62f 1349
7591bab1 1350end:
53efb85a 1351 memset(&reply, 0, sizeof(reply));
173af62f 1352 if (ret < 0) {
f73fabfd 1353 reply.ret_code = htobe32(LTTNG_ERR_UNK);
173af62f 1354 } else {
f73fabfd 1355 reply.ret_code = htobe32(LTTNG_OK);
173af62f 1356 }
58eb9381 1357 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply,
173af62f 1358 sizeof(struct lttcomm_relayd_generic_reply), 0);
5312a3ed
JG		 1359 if (send_ret < (ssize_t) sizeof(reply)) {
1360 ERR("Failed to send \"close stream\" command reply (ret = %zd)",
1361 send_ret);
1362 ret = -1;
173af62f
DG		 1363 }
1364
1365end_no_session:
1366 return ret;
1367}
1368
93ec662e
JD		 1369/*
1370 * relay_reset_metadata: reset a metadata stream
1371 */
1372static
5312a3ed
JG		 1373int relay_reset_metadata(const struct lttcomm_relayd_hdr *recv_hdr,
1374 struct relay_connection *conn,
1375 const struct lttng_buffer_view *payload)
93ec662e 1376{
5312a3ed
JG		 1377 int ret;
1378 ssize_t send_ret;
93ec662e
JD		 1379 struct relay_session *session = conn->session;
1380 struct lttcomm_relayd_reset_metadata stream_info;
1381 struct lttcomm_relayd_generic_reply reply;
1382 struct relay_stream *stream;
1383
1384 DBG("Reset metadata received");
1385
5312a3ed 1386 if (!session || !conn->version_check_done) {
93ec662e
JD		 1387 ERR("Trying to reset a metadata stream before version check");
1388 ret = -1;
1389 goto end_no_session;
1390 }
1391
5312a3ed
JG		 1392 if (payload->size < sizeof(stream_info)) {
1393 ERR("Unexpected payload size in \"relay_reset_metadata\": expected >= %zu bytes, got %zu bytes",
1394 sizeof(stream_info), payload->size);
93ec662e
JD		 1395 ret = -1;
1396 goto end_no_session;
1397 }
5312a3ed
JG		 1398 memcpy(&stream_info, payload->data, sizeof(stream_info));
1399 stream_info.stream_id = be64toh(stream_info.stream_id);
1400 stream_info.version = be64toh(stream_info.version);
1401
1402 DBG("Update metadata to version %" PRIu64, stream_info.version);
93ec662e
JD		 1403
1404 /* Unsupported for live sessions for now. */
1405 if (session->live_timer != 0) {
1406 ret = -1;
1407 goto end;
1408 }
1409
5312a3ed 1410 stream = stream_get_by_id(stream_info.stream_id);
93ec662e
JD		 1411 if (!stream) {
1412 ret = -1;
1413 goto end;
1414 }
1415 pthread_mutex_lock(&stream->lock);
1416 if (!stream->is_metadata) {
1417 ret = -1;
1418 goto end_unlock;
1419 }
1420
1421 ret = utils_rotate_stream_file(stream->path_name, stream->channel_name,
1422 0, 0, -1, -1, stream->stream_fd->fd, NULL,
1423 &stream->stream_fd->fd);
1424 if (ret < 0) {
1425 ERR("Failed to rotate metadata file %s of channel %s",
1426 stream->path_name, stream->channel_name);
1427 goto end_unlock;
1428 }
1429
1430end_unlock:
1431 pthread_mutex_unlock(&stream->lock);
1432 stream_put(stream);
1433
1434end:
1435 memset(&reply, 0, sizeof(reply));
1436 if (ret < 0) {
1437 reply.ret_code = htobe32(LTTNG_ERR_UNK);
1438 } else {
1439 reply.ret_code = htobe32(LTTNG_OK);
1440 }
1441 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply,
1442 sizeof(struct lttcomm_relayd_generic_reply), 0);
5312a3ed
JG		 1443 if (send_ret < (ssize_t) sizeof(reply)) {
1444 ERR("Failed to send \"reset metadata\" command reply (ret = %zd)",
1445 send_ret);
1446 ret = -1;
93ec662e
JD		 1447 }
1448
1449end_no_session:
1450 return ret;
1451}
1452
b8aa1682
JD		 1453/*
1454 * relay_unknown_command: send -1 if received unknown command
1455 */
7591bab1 1456static void relay_unknown_command(struct relay_connection *conn)
b8aa1682
JD		 1457{
1458 struct lttcomm_relayd_generic_reply reply;
5312a3ed 1459 ssize_t send_ret;
b8aa1682 1460
53efb85a 1461 memset(&reply, 0, sizeof(reply));
f73fabfd 1462 reply.ret_code = htobe32(LTTNG_ERR_UNK);
5312a3ed
JG		 1463 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply, sizeof(reply), 0);
1464 if (send_ret < sizeof(reply)) {
1465 ERR("Failed to send \"unknown command\" command reply (ret = %zd)", send_ret);
b8aa1682
JD		 1466 }
1467}
1468
1469/*
1470 * relay_start: send an acknowledgment to the client to tell if we are
1471 * ready to receive data. We are ready if a session is established.
1472 */
5312a3ed
JG		 1473static int relay_start(const struct lttcomm_relayd_hdr *recv_hdr,
1474 struct relay_connection *conn,
1475 const struct lttng_buffer_view *payload)
b8aa1682 1476{
5312a3ed
JG		 1477 int ret = 0;
1478 ssize_t send_ret;
b8aa1682 1479 struct lttcomm_relayd_generic_reply reply;
58eb9381 1480 struct relay_session *session = conn->session;
b8aa1682
JD		 1481
1482 if (!session) {
1483 DBG("Trying to start the streaming without a session established");
f73fabfd 1484 ret = htobe32(LTTNG_ERR_UNK);
b8aa1682
JD		 1485 }
1486
53efb85a 1487 memset(&reply, 0, sizeof(reply));
5312a3ed
JG		 1488 reply.ret_code = htobe32(LTTNG_OK);
1489 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply,
1490 sizeof(reply), 0);
1491 if (send_ret < (ssize_t) sizeof(reply)) {
1492 ERR("Failed to send \"relay_start\" command reply (ret = %zd)",
1493 send_ret);
1494 ret = -1;
b8aa1682
JD		 1495 }
1496
1497 return ret;
1498}
1499
1d4dfdef
DG		 1500/*
1501 * Append padding to the file pointed by the file descriptor fd.
1502 */
1503static int write_padding_to_file(int fd, uint32_t size)
1504{
6cd525e8 1505 ssize_t ret = 0;
1d4dfdef
DG		 1506 char *zeros;
1507
1508 if (size == 0) {
1509 goto end;
1510 }
1511
1512 zeros = zmalloc(size);
1513 if (zeros == NULL) {
1514 PERROR("zmalloc zeros for padding");
1515 ret = -1;
1516 goto end;
1517 }
1518
6cd525e8
MD		 1519 ret = lttng_write(fd, zeros, size);
1520 if (ret < size) {
1d4dfdef
DG		 1521 PERROR("write padding to file");
1522 }
1523
e986c7a1
DG		 1524 free(zeros);
1525
1d4dfdef
DG		 1526end:
1527 return ret;
1528}
1529
d3ecc550
JD		 1530/*
1531 * Close the current index file if it is open, and create a new one.
1532 *
1533 * Return 0 on success, -1 on error.
1534 */
1535static
cb523e02
JG		 1536int create_rotate_index_file(struct relay_stream *stream,
1537 const char *stream_path)
d3ecc550
JD		 1538{
1539 int ret;
1540 uint32_t major, minor;
1541
1542 /* Put ref on previous index_file. */
1543 if (stream->index_file) {
1544 lttng_index_file_put(stream->index_file);
1545 stream->index_file = NULL;
1546 }
1547 major = stream->trace->session->major;
1548 minor = stream->trace->session->minor;
cb523e02 1549 stream->index_file = lttng_index_file_create(stream_path,
d3ecc550
JD		 1550 stream->channel_name,
1551 -1, -1, stream->tracefile_size,
1552 tracefile_array_get_file_index_head(stream->tfa),
1553 lttng_to_index_major(major, minor),
1554 lttng_to_index_minor(major, minor));
1555 if (!stream->index_file) {
1556 ret = -1;
1557 goto end;
1558 }
1559
1560 ret = 0;
1561
1562end:
1563 return ret;
1564}
1565
1566static
c6db3843 1567int do_rotate_stream_data(struct relay_stream *stream)
d3ecc550
JD		 1568{
1569 int ret;
1570
c6db3843
JG		 1571 DBG("Rotating stream %" PRIu64 " data file",
1572 stream->stream_handle);
d3ecc550
JD		 1573 /* Perform the stream rotation. */
1574 ret = utils_rotate_stream_file(stream->path_name,
1575 stream->channel_name, stream->tracefile_size,
1576 stream->tracefile_count, -1,
1577 -1, stream->stream_fd->fd,
1578 NULL, &stream->stream_fd->fd);
1579 if (ret < 0) {
1580 ERR("Rotating stream output file");
1581 goto end;
1582 }
1583 stream->tracefile_size_current = 0;
d3ecc550 1584 stream->pos_after_last_complete_data_index = 0;
c6db3843 1585 stream->data_rotated = true;
d3ecc550 1586
c6db3843
JG		 1587 if (stream->data_rotated && stream->index_rotated) {
1588 /* Rotation completed; reset its state. */
1589 DBG("Rotation completed for stream %" PRIu64,
1590 stream->stream_handle);
1591 stream->rotate_at_seq_num = -1ULL;
1592 stream->data_rotated = false;
1593 stream->index_rotated = false;
1594 }
d3ecc550
JD		 1595end:
1596 return ret;
1597}
1598
1599/*
1600 * If too much data has been written in a tracefile before we received the
1601 * rotation command, we have to move the excess data to the new tracefile and
1602 * perform the rotation. This can happen because the control and data
1603 * connections are separate, the indexes as well as the commands arrive from
1604 * the control connection and we have no control over the order so we could be
1605 * in a situation where too much data has been received on the data connection
c6db3843 1606 * before the rotation command on the control connection arrives.
d3ecc550
JD		 1607 */
1608static
1609int rotate_truncate_stream(struct relay_stream *stream)
1610{
1611 int ret, new_fd;
a5df8828 1612 off_t lseek_ret;
d3ecc550
JD		 1613 uint64_t diff, pos = 0;
1614 char buf[FILE_COPY_BUFFER_SIZE];
1615
1616 assert(!stream->is_metadata);
1617
1618 assert(stream->tracefile_size_current >
1619 stream->pos_after_last_complete_data_index);
1620 diff = stream->tracefile_size_current -
1621 stream->pos_after_last_complete_data_index;
1622
1623 /* Create the new tracefile. */
1624 new_fd = utils_create_stream_file(stream->path_name,
1625 stream->channel_name,
1626 stream->tracefile_size, stream->tracefile_count,
1627 /* uid */ -1, /* gid */ -1, /* suffix */ NULL);
1628 if (new_fd < 0) {
1629 ERR("Failed to create new stream file at path %s for channel %s",
1630 stream->path_name, stream->channel_name);
1631 ret = -1;
1632 goto end;
1633 }
1634
1635 /*
1636 * Rewind the current tracefile to the position at which the rotation
a9577b76 1637 * should have occurred.
d3ecc550 1638 */
a5df8828 1639 lseek_ret = lseek(stream->stream_fd->fd,
d3ecc550 1640 stream->pos_after_last_complete_data_index, SEEK_SET);
a5df8828 1641 if (lseek_ret < 0) {
d3ecc550 1642 PERROR("seek truncate stream");
a5df8828 1643 ret = -1;
d3ecc550
JD		 1644 goto end;
1645 }
1646
1647 /* Move data from the old file to the new file. */
1648 while (pos < diff) {
1649 uint64_t count, bytes_left;
1650 ssize_t io_ret;
1651
1652 bytes_left = diff - pos;
1653 count = bytes_left > sizeof(buf) ? sizeof(buf) : bytes_left;
1654 assert(count <= SIZE_MAX);
1655
1656 io_ret = lttng_read(stream->stream_fd->fd, buf, count);
1657 if (io_ret < (ssize_t) count) {
1658 char error_string[256];
1659
1660 snprintf(error_string, sizeof(error_string),
1661 "Failed to read %" PRIu64 " bytes from fd %i in rotate_truncate_stream(), returned %zi",
1662 count, stream->stream_fd->fd, io_ret);
1663 if (io_ret == -1) {
1664 PERROR("%s", error_string);
1665 } else {
1666 ERR("%s", error_string);
1667 }
1668 ret = -1;
1669 goto end;
1670 }
1671
1672 io_ret = lttng_write(new_fd, buf, count);
1673 if (io_ret < (ssize_t) count) {
1674 char error_string[256];
1675
1676 snprintf(error_string, sizeof(error_string),
1677 "Failed to write %" PRIu64 " bytes from fd %i in rotate_truncate_stream(), returned %zi",
1678 count, new_fd, io_ret);
1679 if (io_ret == -1) {
1680 PERROR("%s", error_string);
1681 } else {
1682 ERR("%s", error_string);
1683 }
1684 ret = -1;
1685 goto end;
1686 }
1687
1688 pos += count;
1689 }
1690
1691 /* Truncate the file to get rid of the excess data. */
1692 ret = ftruncate(stream->stream_fd->fd,
1693 stream->pos_after_last_complete_data_index);
1694 if (ret) {
1695 PERROR("ftruncate");
1696 goto end;
1697 }
1698
1699 ret = close(stream->stream_fd->fd);
1700 if (ret < 0) {
1701 PERROR("Closing tracefile");
1702 goto end;
1703 }
1704
d3ecc550
JD		 1705 /*
1706 * Update the offset and FD of all the eventual indexes created by the
1707 * data connection before the rotation command arrived.
1708 */
1709 ret = relay_index_switch_all_files(stream);
1710 if (ret < 0) {
1711 ERR("Failed to rotate index file");
1712 goto end;
1713 }
1714
1715 stream->stream_fd->fd = new_fd;
1716 stream->tracefile_size_current = diff;
1717 stream->pos_after_last_complete_data_index = 0;
1718 stream->rotate_at_seq_num = -1ULL;
1719
1720 ret = 0;
1721
1722end:
1723 return ret;
1724}
1725
1726/*
c6db3843 1727 * Check if a stream's index file should be rotated (for session rotation).
d3ecc550
JD		 1728 * Must be called with the stream lock held.
1729 *
1730 * Return 0 on success, a negative value on error.
1731 */
1732static
c6db3843 1733int try_rotate_stream_index(struct relay_stream *stream)
d3ecc550
JD		 1734{
1735 int ret = 0;
1736
d3ecc550 1737 if (stream->rotate_at_seq_num == -1ULL) {
c6db3843 1738 /* No rotation expected. */
d3ecc550
JD		 1739 goto end;
1740 }
1741
c6db3843
JG		 1742 if (stream->index_rotated) {
1743 /* Rotation of the index has already occurred. */
1744 goto end;
1745 }
1746
1747 if (stream->prev_index_seq == -1ULL ||
1748 stream->prev_index_seq < stream->rotate_at_seq_num) {
1749 DBG("Stream %" PRIu64 " index not yet ready for rotation (rotate_at_seq_num = %" PRIu64 ", prev_index_seq = %" PRIu64 ")",
1750 stream->stream_handle,
1751 stream->rotate_at_seq_num,
1752 stream->prev_index_seq);
1753 goto end;
1754 } else if (stream->prev_index_seq != stream->rotate_at_seq_num) {
1755 /*
1756 * Unexpected, protocol error/bug.
1757 * It could mean that we received a rotation position
1758 * that is in the past.
1759 */
1760 ERR("Stream %" PRIu64 " index is in an inconsistent state (rotate_at_seq_num = %" PRIu64 ", prev_data_seq = %" PRIu64 ", prev_index_seq = %" PRIu64 ")",
c64c70f5
JG		 1761 stream->stream_handle,
1762 stream->rotate_at_seq_num,
a8f9f353 1763 stream->prev_data_seq,
c64c70f5 1764 stream->prev_index_seq);
c6db3843
JG		 1765 ret = -1;
1766 goto end;
1767 } else {
1768 DBG("Rotating stream %" PRIu64 " index file",
1769 stream->stream_handle);
1770 ret = create_rotate_index_file(stream, stream->path_name);
1771 stream->index_rotated = true;
1772
1773 if (stream->data_rotated && stream->index_rotated) {
1774 /* Rotation completed; reset its state. */
1775 DBG("Rotation completed for stream %" PRIu64,
1776 stream->stream_handle);
1777 stream->rotate_at_seq_num = -1ULL;
1778 stream->data_rotated = false;
1779 stream->index_rotated = false;
1780 }
1781 }
1782
1783end:
1784 return ret;
1785}
1786
1787/*
1788 * Check if a stream's data file (as opposed to index) should be rotated
1789 * (for session rotation).
1790 * Must be called with the stream lock held.
1791 *
1792 * Return 0 on success, a negative value on error.
1793 */
1794static
1795int try_rotate_stream_data(struct relay_stream *stream)
1796{
1797 int ret = 0;
1798
1799 if (stream->rotate_at_seq_num == -1ULL) {
1800 /* No rotation expected. */
1801 goto end;
1802 }
1803
1804 if (stream->data_rotated) {
1805 /* Rotation of the data file has already occurred. */
1806 goto end;
1807 }
1808
1809 if (stream->prev_data_seq == -1ULL ||
1810 stream->prev_data_seq < stream->rotate_at_seq_num) {
1811 DBG("Stream %" PRIu64 " not yet ready for rotation (rotate_at_seq_num = %" PRIu64 ", prev_data_seq = %" PRIu64 ")",
1812 stream->stream_handle,
1813 stream->rotate_at_seq_num,
1814 stream->prev_data_seq);
d3ecc550 1815 goto end;
a8f9f353 1816 } else if (stream->prev_data_seq > stream->rotate_at_seq_num) {
c64c70f5 1817 /*
a8f9f353 1818 * prev_data_seq is checked here since indexes and rotation
c64c70f5
JG		 1819 * commands are serialized with respect to each other.
1820 */
d3ecc550
JD		 1821 DBG("Rotation after too much data has been written in tracefile "
1822 "for stream %" PRIu64 ", need to truncate before "
1823 "rotating", stream->stream_handle);
1824 ret = rotate_truncate_stream(stream);
1825 if (ret) {
1826 ERR("Failed to truncate stream");
1827 goto end;
1828 }
c6db3843
JG		 1829 } else if (stream->prev_data_seq != stream->rotate_at_seq_num) {
1830 /*
1831 * Unexpected, protocol error/bug.
1832 * It could mean that we received a rotation position
1833 * that is in the past.
1834 */
1835 ERR("Stream %" PRIu64 " data is in an inconsistent state (rotate_at_seq_num = %" PRIu64 ", prev_data_seq = %" PRIu64 ")",
c64c70f5
JG		 1836 stream->stream_handle,
1837 stream->rotate_at_seq_num,
c6db3843
JG		 1838 stream->prev_data_seq);
1839 ret = -1;
1840 goto end;
1841 } else {
1842 ret = do_rotate_stream_data(stream);
d3ecc550
JD		 1843 }
1844
1845end:
1846 return ret;
1847}
1848
b8aa1682 1849/*
7591bab1 1850 * relay_recv_metadata: receive the metadata for the session.
b8aa1682 1851 */
5312a3ed
JG		 1852static int relay_recv_metadata(const struct lttcomm_relayd_hdr *recv_hdr,
1853 struct relay_connection *conn,
1854 const struct lttng_buffer_view *payload)
b8aa1682 1855{
32d1569c 1856 int ret = 0;
6cd525e8 1857 ssize_t size_ret;
58eb9381 1858 struct relay_session *session = conn->session;
5312a3ed 1859 struct lttcomm_relayd_metadata_payload metadata_payload_header;
b8aa1682 1860 struct relay_stream *metadata_stream;
5312a3ed 1861 uint64_t metadata_payload_size;
b8aa1682
JD		 1862
1863 if (!session) {
1864 ERR("Metadata sent before version check");
1865 ret = -1;
1866 goto end;
1867 }
1868
5312a3ed 1869 if (recv_hdr->data_size < sizeof(struct lttcomm_relayd_metadata_payload)) {
f6416125
MD		 1870 ERR("Incorrect data size");
1871 ret = -1;
1872 goto end;
1873 }
5312a3ed
JG		 1874 metadata_payload_size = recv_hdr->data_size -
1875 sizeof(struct lttcomm_relayd_metadata_payload);
f6416125 1876
5312a3ed
JG		 1877 memcpy(&metadata_payload_header, payload->data,
1878 sizeof(metadata_payload_header));
1879 metadata_payload_header.stream_id = be64toh(
1880 metadata_payload_header.stream_id);
1881 metadata_payload_header.padding_size = be32toh(
1882 metadata_payload_header.padding_size);
9d1bbf21 1883
5312a3ed 1884 metadata_stream = stream_get_by_id(metadata_payload_header.stream_id);
b8aa1682
JD		 1885 if (!metadata_stream) {
1886 ret = -1;
7591bab1 1887 goto end;
b8aa1682
JD		 1888 }
1889
7591bab1
MD		 1890 pthread_mutex_lock(&metadata_stream->lock);
1891
5312a3ed
JG		 1892 size_ret = lttng_write(metadata_stream->stream_fd->fd,
1893 payload->data + sizeof(metadata_payload_header),
1894 metadata_payload_size);
1895 if (size_ret < metadata_payload_size) {
b8aa1682
JD		 1896 ERR("Relay error writing metadata on file");
1897 ret = -1;
7591bab1 1898 goto end_put;
b8aa1682 1899 }
1d4dfdef 1900
32d1569c 1901 size_ret = write_padding_to_file(metadata_stream->stream_fd->fd,
5312a3ed 1902 metadata_payload_header.padding_size);
715e6fb1 1903 if (size_ret < (int64_t) metadata_payload_header.padding_size) {
5312a3ed 1904 ret = -1;
7591bab1 1905 goto end_put;
1d4dfdef 1906 }
2a174661 1907
7591bab1 1908 metadata_stream->metadata_received +=
5312a3ed 1909 metadata_payload_size + metadata_payload_header.padding_size;
7591bab1
MD		 1910 DBG2("Relay metadata written. Updated metadata_received %" PRIu64,
1911 metadata_stream->metadata_received);
1d4dfdef 1912
c6db3843 1913 ret = try_rotate_stream_data(metadata_stream);
d3ecc550
JD		 1914 if (ret < 0) {
1915 goto end_put;
1916 }
1917
7591bab1
MD		 1918end_put:
1919 pthread_mutex_unlock(&metadata_stream->lock);
1920 stream_put(metadata_stream);
b8aa1682
JD		 1921end:
1922 return ret;
1923}
1924
1925/*
1926 * relay_send_version: send relayd version number
1927 */
5312a3ed
JG		 1928static int relay_send_version(const struct lttcomm_relayd_hdr *recv_hdr,
1929 struct relay_connection *conn,
1930 const struct lttng_buffer_view *payload)
b8aa1682 1931{
7f51dcba 1932 int ret;
5312a3ed 1933 ssize_t send_ret;
092b6259 1934 struct lttcomm_relayd_version reply, msg;
87cb6359 1935 bool compatible = true;
b8aa1682 1936
5312a3ed 1937 conn->version_check_done = true;
b8aa1682 1938
092b6259 1939 /* Get version from the other side. */
5312a3ed
JG		 1940 if (payload->size < sizeof(msg)) {
1941 ERR("Unexpected payload size in \"relay_send_version\": expected >= %zu bytes, got %zu bytes",
1942 sizeof(msg), payload->size);
092b6259 1943 ret = -1;
092b6259
DG		 1944 goto end;
1945 }
1946
5312a3ed
JG		 1947 memcpy(&msg, payload->data, sizeof(msg));
1948 msg.major = be32toh(msg.major);
1949 msg.minor = be32toh(msg.minor);
1950
53efb85a 1951 memset(&reply, 0, sizeof(reply));
d83a952c
MD		 1952 reply.major = RELAYD_VERSION_COMM_MAJOR;
1953 reply.minor = RELAYD_VERSION_COMM_MINOR;
d4519fa3
JD		 1954
1955 /* Major versions must be the same */
5312a3ed 1956 if (reply.major != msg.major) {
6151a90f 1957 DBG("Incompatible major versions (%u vs %u), deleting session",
5312a3ed 1958 reply.major, msg.major);
87cb6359 1959 compatible = false;
d4519fa3
JD		 1960 }
1961
58eb9381 1962 conn->major = reply.major;
0f907de1 1963 /* We adapt to the lowest compatible version */
5312a3ed 1964 if (reply.minor <= msg.minor) {
58eb9381 1965 conn->minor = reply.minor;
0f907de1 1966 } else {
5312a3ed 1967 conn->minor = msg.minor;
0f907de1
JD		 1968 }
1969
6151a90f
JD		 1970 reply.major = htobe32(reply.major);
1971 reply.minor = htobe32(reply.minor);
5312a3ed
JG		 1972 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply,
1973 sizeof(reply), 0);
1974 if (send_ret < (ssize_t) sizeof(reply)) {
1975 ERR("Failed to send \"send version\" command reply (ret = %zd)",
1976 send_ret);
1977 ret = -1;
1978 goto end;
1979 } else {
1980 ret = 0;
6151a90f
JD		 1981 }
1982
87cb6359
JD		 1983 if (!compatible) {
1984 ret = -1;
1985 goto end;
1986 }
1987
58eb9381
DG		 1988 DBG("Version check done using protocol %u.%u", conn->major,
1989 conn->minor);
b8aa1682
JD		 1990
1991end:
1992 return ret;
1993}
1994
c8f59ee5 1995/*
6d805429 1996 * Check for data pending for a given stream id from the session daemon.
c8f59ee5 1997 */
5312a3ed
JG		 1998static int relay_data_pending(const struct lttcomm_relayd_hdr *recv_hdr,
1999 struct relay_connection *conn,
2000 const struct lttng_buffer_view *payload)
c8f59ee5 2001{
58eb9381 2002 struct relay_session *session = conn->session;
6d805429 2003 struct lttcomm_relayd_data_pending msg;
c8f59ee5
DG		 2004 struct lttcomm_relayd_generic_reply reply;
2005 struct relay_stream *stream;
5312a3ed 2006 ssize_t send_ret;
c8f59ee5 2007 int ret;
298a25ca 2008 uint64_t stream_seq;
c8f59ee5 2009
6d805429 2010 DBG("Data pending command received");
c8f59ee5 2011
5312a3ed 2012 if (!session || !conn->version_check_done) {
c8f59ee5
DG		 2013 ERR("Trying to check for data before version check");
2014 ret = -1;
2015 goto end_no_session;
2016 }
2017
5312a3ed
JG		 2018 if (payload->size < sizeof(msg)) {
2019 ERR("Unexpected payload size in \"relay_data_pending\": expected >= %zu bytes, got %zu bytes",
2020 sizeof(msg), payload->size);
c8f59ee5
DG		 2021 ret = -1;
2022 goto end_no_session;
2023 }
5312a3ed
JG		 2024 memcpy(&msg, payload->data, sizeof(msg));
2025 msg.stream_id = be64toh(msg.stream_id);
2026 msg.last_net_seq_num = be64toh(msg.last_net_seq_num);
c8f59ee5 2027
5312a3ed 2028 stream = stream_get_by_id(msg.stream_id);
de91f48a 2029 if (stream == NULL) {
c8f59ee5 2030 ret = -1;
7591bab1 2031 goto end;
c8f59ee5
DG		 2032 }
2033
7591bab1
MD		 2034 pthread_mutex_lock(&stream->lock);
2035
298a25ca
JG		 2036 if (session_streams_have_index(session)) {
2037 /*
2038 * Ensure that both the index and stream data have been
2039 * flushed up to the requested point.
2040 */
a8f9f353 2041 stream_seq = min(stream->prev_data_seq, stream->prev_index_seq);
298a25ca 2042 } else {
a8f9f353 2043 stream_seq = stream->prev_data_seq;
298a25ca 2044 }
a8f9f353 2045 DBG("Data pending for stream id %" PRIu64 ": prev_data_seq %" PRIu64
298a25ca
JG		 2046 ", prev_index_seq %" PRIu64
2047 ", and last_seq %" PRIu64, msg.stream_id,
a8f9f353 2048 stream->prev_data_seq, stream->prev_index_seq,
298a25ca 2049 msg.last_net_seq_num);
c8f59ee5 2050
33832e64 2051 /* Avoid wrapping issue */
298a25ca 2052 if (((int64_t) (stream_seq - msg.last_net_seq_num)) >= 0) {
6d805429 2053 /* Data has in fact been written and is NOT pending */
c8f59ee5 2054 ret = 0;
6d805429
DG		 2055 } else {
2056 /* Data still being streamed thus pending */
2057 ret = 1;
c8f59ee5
DG		 2058 }
2059
7591bab1
MD		 2060 stream->data_pending_check_done = true;
2061 pthread_mutex_unlock(&stream->lock);
f7079f67 2062
7591bab1
MD		 2063 stream_put(stream);
2064end:
c8f59ee5 2065
53efb85a 2066 memset(&reply, 0, sizeof(reply));
c8f59ee5 2067 reply.ret_code = htobe32(ret);
5312a3ed
JG		 2068 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply, sizeof(reply), 0);
2069 if (send_ret < (ssize_t) sizeof(reply)) {
2070 ERR("Failed to send \"data pending\" command reply (ret = %zd)",
2071 send_ret);
2072 ret = -1;
c8f59ee5
DG		 2073 }
2074
2075end_no_session:
2076 return ret;
2077}
2078
2079/*
2080 * Wait for the control socket to reach a quiescent state.
2081 *
7591bab1
MD		 2082 * Note that for now, when receiving this command from the session
2083 * daemon, this means that every subsequent commands or data received on
2084 * the control socket has been handled. So, this is why we simply return
2085 * OK here.
c8f59ee5 2086 */
5312a3ed
JG		 2087static int relay_quiescent_control(const struct lttcomm_relayd_hdr *recv_hdr,
2088 struct relay_connection *conn,
2089 const struct lttng_buffer_view *payload)
c8f59ee5
DG		 2090{
2091 int ret;
5312a3ed 2092 ssize_t send_ret;
ad7051c0 2093 struct relay_stream *stream;
ad7051c0 2094 struct lttcomm_relayd_quiescent_control msg;
c8f59ee5
DG		 2095 struct lttcomm_relayd_generic_reply reply;
2096
2097 DBG("Checking quiescent state on control socket");
2098
5312a3ed 2099 if (!conn->session || !conn->version_check_done) {
ad7051c0
DG		 2100 ERR("Trying to check for data before version check");
2101 ret = -1;
2102 goto end_no_session;
2103 }
2104
5312a3ed
JG		 2105 if (payload->size < sizeof(msg)) {
2106 ERR("Unexpected payload size in \"relay_quiescent_control\": expected >= %zu bytes, got %zu bytes",
2107 sizeof(msg), payload->size);
ad7051c0
DG		 2108 ret = -1;
2109 goto end_no_session;
2110 }
5312a3ed
JG		 2111 memcpy(&msg, payload->data, sizeof(msg));
2112 msg.stream_id = be64toh(msg.stream_id);
ad7051c0 2113
5312a3ed 2114 stream = stream_get_by_id(msg.stream_id);
7591bab1
MD		 2115 if (!stream) {
2116 goto reply;
2117 }
2118 pthread_mutex_lock(&stream->lock);
2119 stream->data_pending_check_done = true;
2120 pthread_mutex_unlock(&stream->lock);
5312a3ed
JG		 2121
2122 DBG("Relay quiescent control pending flag set to %" PRIu64, msg.stream_id);
7591bab1
MD		 2123 stream_put(stream);
2124reply:
53efb85a 2125 memset(&reply, 0, sizeof(reply));
c8f59ee5 2126 reply.ret_code = htobe32(LTTNG_OK);
5312a3ed
JG		 2127 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply, sizeof(reply), 0);
2128 if (send_ret < (ssize_t) sizeof(reply)) {
2129 ERR("Failed to send \"quiescent control\" command reply (ret = %zd)",
2130 send_ret);
2131 ret = -1;
2132 } else {
2133 ret = 0;
c8f59ee5
DG		 2134 }
2135
ad7051c0 2136end_no_session:
c8f59ee5
DG		 2137 return ret;
2138}
2139
f7079f67 2140/*
7591bab1
MD		 2141 * Initialize a data pending command. This means that a consumer is about
2142 * to ask for data pending for each stream it holds. Simply iterate over
2143 * all streams of a session and set the data_pending_check_done flag.
f7079f67
DG		 2144 *
2145 * This command returns to the client a LTTNG_OK code.
2146 */
5312a3ed
JG		 2147static int relay_begin_data_pending(const struct lttcomm_relayd_hdr *recv_hdr,
2148 struct relay_connection *conn,
2149 const struct lttng_buffer_view *payload)
f7079f67
DG		 2150{
2151 int ret;
5312a3ed 2152 ssize_t send_ret;
f7079f67
DG		 2153 struct lttng_ht_iter iter;
2154 struct lttcomm_relayd_begin_data_pending msg;
2155 struct lttcomm_relayd_generic_reply reply;
2156 struct relay_stream *stream;
f7079f67
DG		 2157
2158 assert(recv_hdr);
58eb9381 2159 assert(conn);
f7079f67
DG		 2160
2161 DBG("Init streams for data pending");
2162
5312a3ed 2163 if (!conn->session || !conn->version_check_done) {
f7079f67
DG		 2164 ERR("Trying to check for data before version check");
2165 ret = -1;
2166 goto end_no_session;
2167 }
2168
5312a3ed
JG		 2169 if (payload->size < sizeof(msg)) {
2170 ERR("Unexpected payload size in \"relay_begin_data_pending\": expected >= %zu bytes, got %zu bytes",
2171 sizeof(msg), payload->size);
f7079f67
DG		 2172 ret = -1;
2173 goto end_no_session;
2174 }
5312a3ed
JG		 2175 memcpy(&msg, payload->data, sizeof(msg));
2176 msg.session_id = be64toh(msg.session_id);
f7079f67
DG		 2177
2178 /*
7591bab1
MD		 2179 * Iterate over all streams to set the begin data pending flag.
2180 * For now, the streams are indexed by stream handle so we have
2181 * to iterate over all streams to find the one associated with
2182 * the right session_id.
f7079f67
DG		 2183 */
2184 rcu_read_lock();
d3e2ba59 2185 cds_lfht_for_each_entry(relay_streams_ht->ht, &iter.iter, stream,
2a174661 2186 node.node) {
7591bab1
MD		 2187 if (!stream_get(stream)) {
2188 continue;
2189 }
5312a3ed 2190 if (stream->trace->session->id == msg.session_id) {
7591bab1
MD		 2191 pthread_mutex_lock(&stream->lock);
2192 stream->data_pending_check_done = false;
2193 pthread_mutex_unlock(&stream->lock);
f7079f67
DG		 2194 DBG("Set begin data pending flag to stream %" PRIu64,
2195 stream->stream_handle);
2196 }
7591bab1 2197 stream_put(stream);
f7079f67
DG		 2198 }
2199 rcu_read_unlock();
2200
53efb85a 2201 memset(&reply, 0, sizeof(reply));
f7079f67
DG		 2202 /* All good, send back reply. */
2203 reply.ret_code = htobe32(LTTNG_OK);
2204
5312a3ed
JG		 2205 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply, sizeof(reply), 0);
2206 if (send_ret < (ssize_t) sizeof(reply)) {
2207 ERR("Failed to send \"begin data pending\" command reply (ret = %zd)",
2208 send_ret);
2209 ret = -1;
2210 } else {
2211 ret = 0;
f7079f67
DG		 2212 }
2213
2214end_no_session:
2215 return ret;
2216}
2217
2218/*
7591bab1
MD		 2219 * End data pending command. This will check, for a given session id, if
2220 * each stream associated with it has its data_pending_check_done flag
2221 * set. If not, this means that the client lost track of the stream but
2222 * the data is still being streamed on our side. In this case, we inform
2223 * the client that data is in flight.
f7079f67
DG		 2224 *
2225 * Return to the client if there is data in flight or not with a ret_code.
2226 */
5312a3ed
JG		 2227static int relay_end_data_pending(const struct lttcomm_relayd_hdr *recv_hdr,
2228 struct relay_connection *conn,
2229 const struct lttng_buffer_view *payload)
f7079f67
DG		 2230{
2231 int ret;
5312a3ed 2232 ssize_t send_ret;
f7079f67
DG		 2233 struct lttng_ht_iter iter;
2234 struct lttcomm_relayd_end_data_pending msg;
2235 struct lttcomm_relayd_generic_reply reply;
2236 struct relay_stream *stream;
f7079f67
DG		 2237 uint32_t is_data_inflight = 0;
2238
f7079f67
DG		 2239 DBG("End data pending command");
2240
5312a3ed 2241 if (!conn->session || !conn->version_check_done) {
f7079f67
DG		 2242 ERR("Trying to check for data before version check");
2243 ret = -1;
2244 goto end_no_session;
2245 }
2246
5312a3ed
JG		 2247 if (payload->size < sizeof(msg)) {
2248 ERR("Unexpected payload size in \"relay_end_data_pending\": expected >= %zu bytes, got %zu bytes",
2249 sizeof(msg), payload->size);
f7079f67
DG		 2250 ret = -1;
2251 goto end_no_session;
2252 }
5312a3ed
JG		 2253 memcpy(&msg, payload->data, sizeof(msg));
2254 msg.session_id = be64toh(msg.session_id);
f7079f67 2255
7591bab1
MD		 2256 /*
2257 * Iterate over all streams to see if the begin data pending
2258 * flag is set.
2259 */
f7079f67 2260 rcu_read_lock();
d3e2ba59 2261 cds_lfht_for_each_entry(relay_streams_ht->ht, &iter.iter, stream,
2a174661 2262 node.node) {
7591bab1
MD		 2263 if (!stream_get(stream)) {
2264 continue;
2265 }
5312a3ed 2266 if (stream->trace->session->id != msg.session_id) {
7591bab1
MD		 2267 stream_put(stream);
2268 continue;
2269 }
2270 pthread_mutex_lock(&stream->lock);
2271 if (!stream->data_pending_check_done) {
298a25ca
JG		 2272 uint64_t stream_seq;
2273
2274 if (session_streams_have_index(conn->session)) {
2275 /*
2276 * Ensure that both the index and stream data have been
2277 * flushed up to the requested point.
2278 */
a8f9f353 2279 stream_seq = min(stream->prev_data_seq, stream->prev_index_seq);
298a25ca 2280 } else {
a8f9f353 2281 stream_seq = stream->prev_data_seq;
298a25ca
JG		 2282 }
2283 if (!stream->closed || !(((int64_t) (stream_seq - stream->last_net_seq_num)) >= 0)) {
7591bab1
MD		 2284 is_data_inflight = 1;
2285 DBG("Data is still in flight for stream %" PRIu64,
2286 stream->stream_handle);
2287 pthread_mutex_unlock(&stream->lock);
2288 stream_put(stream);
2289 break;
2290 }
f7079f67 2291 }
7591bab1
MD		 2292 pthread_mutex_unlock(&stream->lock);
2293 stream_put(stream);
f7079f67
DG		 2294 }
2295 rcu_read_unlock();
2296
53efb85a 2297 memset(&reply, 0, sizeof(reply));
f7079f67
DG		 2298 /* All good, send back reply. */
2299 reply.ret_code = htobe32(is_data_inflight);
2300
5312a3ed
JG		 2301 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply, sizeof(reply), 0);
2302 if (send_ret < (ssize_t) sizeof(reply)) {
2303 ERR("Failed to send \"end data pending\" command reply (ret = %zd)",
2304 send_ret);
2305 ret = -1;
2306 } else {
2307 ret = 0;
f7079f67
DG		 2308 }
2309
2310end_no_session:
2311 return ret;
2312}
2313
1c20f0e2
JD		 2314/*
2315 * Receive an index for a specific stream.
2316 *
2317 * Return 0 on success else a negative value.
2318 */
5312a3ed
JG		 2319static int relay_recv_index(const struct lttcomm_relayd_hdr *recv_hdr,
2320 struct relay_connection *conn,
2321 const struct lttng_buffer_view *payload)
1c20f0e2 2322{
5312a3ed
JG		 2323 int ret;
2324 ssize_t send_ret;
58eb9381 2325 struct relay_session *session = conn->session;
1c20f0e2 2326 struct lttcomm_relayd_index index_info;
7591bab1 2327 struct relay_index *index;
1c20f0e2
JD		 2328 struct lttcomm_relayd_generic_reply reply;
2329 struct relay_stream *stream;
f8f3885c 2330 size_t msg_len;
1c20f0e2 2331
58eb9381 2332 assert(conn);
1c20f0e2
JD		 2333
2334 DBG("Relay receiving index");
2335
5312a3ed 2336 if (!session || !conn->version_check_done) {
1c20f0e2
JD		 2337 ERR("Trying to close a stream before version check");
2338 ret = -1;
2339 goto end_no_session;
2340 }
2341
f8f3885c
MD		 2342 msg_len = lttcomm_relayd_index_len(
2343 lttng_to_index_major(conn->major, conn->minor),
2344 lttng_to_index_minor(conn->major, conn->minor));
5312a3ed
JG		 2345 if (payload->size < msg_len) {
2346 ERR("Unexpected payload size in \"relay_recv_index\": expected >= %zu bytes, got %zu bytes",
2347 msg_len, payload->size);
1c20f0e2
JD		 2348 ret = -1;
2349 goto end_no_session;
2350 }
5312a3ed
JG		 2351 memcpy(&index_info, payload->data, msg_len);
2352 index_info.relay_stream_id = be64toh(index_info.relay_stream_id);
2353 index_info.net_seq_num = be64toh(index_info.net_seq_num);
2354 index_info.packet_size = be64toh(index_info.packet_size);
2355 index_info.content_size = be64toh(index_info.content_size);
2356 index_info.timestamp_begin = be64toh(index_info.timestamp_begin);
2357 index_info.timestamp_end = be64toh(index_info.timestamp_end);
2358 index_info.events_discarded = be64toh(index_info.events_discarded);
2359 index_info.stream_id = be64toh(index_info.stream_id);
81df238b
JR		 2360
2361 if (conn->minor >= 8) {
2362 index_info.stream_instance_id =
2363 be64toh(index_info.stream_instance_id);
2364 index_info.packet_seq_num = be64toh(index_info.packet_seq_num);
2365 }
5312a3ed
JG		 2366
2367 stream = stream_get_by_id(index_info.relay_stream_id);
1c20f0e2 2368 if (!stream) {
7591bab1 2369 ERR("stream_get_by_id not found");
1c20f0e2 2370 ret = -1;
7591bab1 2371 goto end;
1c20f0e2 2372 }
7591bab1 2373 pthread_mutex_lock(&stream->lock);
1c20f0e2 2374
d3e2ba59
JD		 2375 /* Live beacon handling */
2376 if (index_info.packet_size == 0) {
7591bab1
MD		 2377 DBG("Received live beacon for stream %" PRIu64,
2378 stream->stream_handle);
d3e2ba59
JD		 2379
2380 /*
7591bab1
MD		 2381 * Only flag a stream inactive when it has already
2382 * received data and no indexes are in flight.
d3e2ba59 2383 */
a44ca2ca 2384 if (stream->index_received_seqcount > 0
7591bab1 2385 && stream->indexes_in_flight == 0) {
5312a3ed 2386 stream->beacon_ts_end = index_info.timestamp_end;
d3e2ba59
JD		 2387 }
2388 ret = 0;
7591bab1 2389 goto end_stream_put;
d3e2ba59
JD		 2390 } else {
2391 stream->beacon_ts_end = -1ULL;
2392 }
2393
528f2ffa 2394 if (stream->ctf_stream_id == -1ULL) {
5312a3ed 2395 stream->ctf_stream_id = index_info.stream_id;
528f2ffa 2396 }
5312a3ed 2397 index = relay_index_get_by_id_or_create(stream, index_info.net_seq_num);
7591bab1
MD		 2398 if (!index) {
2399 ret = -1;
2400 ERR("relay_index_get_by_id_or_create index NULL");
2401 goto end_stream_put;
1c20f0e2 2402 }
234cd636 2403 if (set_index_control_data(index, &index_info, conn)) {
7591bab1
MD		 2404 ERR("set_index_control_data error");
2405 relay_index_put(index);
2406 ret = -1;
2407 goto end_stream_put;
2408 }
2409 ret = relay_index_try_flush(index);
2410 if (ret == 0) {
a44ca2ca
MD		 2411 tracefile_array_commit_seq(stream->tfa);
2412 stream->index_received_seqcount++;
d3ecc550 2413 stream->pos_after_last_complete_data_index += index->total_size;
7a45c7e6 2414 stream->prev_index_seq = index_info.net_seq_num;
c6db3843
JG		 2415
2416 ret = try_rotate_stream_index(stream);
2417 if (ret < 0) {
2418 goto end_stream_put;
2419 }
7591bab1
MD		 2420 } else if (ret > 0) {
2421 /* no flush. */
2422 ret = 0;
2423 } else {
245f8bec
JR		 2424 /*
2425 * ret < 0
2426 *
2427 * relay_index_try_flush is responsible for the self-reference
2428 * put of the index object on error.
2429 */
7591bab1 2430 ERR("relay_index_try_flush error %d", ret);
7591bab1 2431 ret = -1;
1c20f0e2
JD		 2432 }
2433
7591bab1
MD		 2434end_stream_put:
2435 pthread_mutex_unlock(&stream->lock);
2436 stream_put(stream);
2437
2438end:
1c20f0e2 2439
53efb85a 2440 memset(&reply, 0, sizeof(reply));
1c20f0e2
JD		 2441 if (ret < 0) {
2442 reply.ret_code = htobe32(LTTNG_ERR_UNK);
2443 } else {
2444 reply.ret_code = htobe32(LTTNG_OK);
2445 }
58eb9381 2446 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply, sizeof(reply), 0);
5312a3ed
JG		 2447 if (send_ret < (ssize_t) sizeof(reply)) {
2448 ERR("Failed to send \"recv index\" command reply (ret = %zd)", send_ret);
2449 ret = -1;
1c20f0e2
JD		 2450 }
2451
2452end_no_session:
2453 return ret;
2454}
2455
a4baae1b
JD		 2456/*
2457 * Receive the streams_sent message.
2458 *
2459 * Return 0 on success else a negative value.
2460 */
5312a3ed
JG		 2461static int relay_streams_sent(const struct lttcomm_relayd_hdr *recv_hdr,
2462 struct relay_connection *conn,
2463 const struct lttng_buffer_view *payload)
a4baae1b 2464{
5312a3ed
JG		 2465 int ret;
2466 ssize_t send_ret;
a4baae1b
JD		 2467 struct lttcomm_relayd_generic_reply reply;
2468
58eb9381 2469 assert(conn);
a4baae1b
JD		 2470
2471 DBG("Relay receiving streams_sent");
2472
5312a3ed 2473 if (!conn->session || !conn->version_check_done) {
a4baae1b
JD		 2474 ERR("Trying to close a stream before version check");
2475 ret = -1;
2476 goto end_no_session;
2477 }
2478
2479 /*
7591bab1
MD		 2480 * Publish every pending stream in the connection recv list which are
2481 * now ready to be used by the viewer.
4a9daf17 2482 */
7591bab1 2483 publish_connection_local_streams(conn);
4a9daf17 2484
53efb85a 2485 memset(&reply, 0, sizeof(reply));
a4baae1b 2486 reply.ret_code = htobe32(LTTNG_OK);
58eb9381 2487 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply, sizeof(reply), 0);
5312a3ed
JG		 2488 if (send_ret < (ssize_t) sizeof(reply)) {
2489 ERR("Failed to send \"streams sent\" command reply (ret = %zd)",
2490 send_ret);
2491 ret = -1;
a4baae1b
JD		 2492 } else {
2493 /* Success. */
2494 ret = 0;
2495 }
2496
2497end_no_session:
2498 return ret;
2499}
2500
d3ecc550 2501/*
5312a3ed 2502 * relay_rotate_session_stream: rotate a stream to a new tracefile for the session
d3ecc550
JD		 2503 * rotation feature (not the tracefile rotation feature).
2504 */
5312a3ed
JG		 2505static int relay_rotate_session_stream(const struct lttcomm_relayd_hdr *recv_hdr,
2506 struct relay_connection *conn,
2507 const struct lttng_buffer_view *payload)
d3ecc550 2508{
5312a3ed
JG		 2509 int ret;
2510 ssize_t send_ret;
d3ecc550
JD		 2511 struct relay_session *session = conn->session;
2512 struct lttcomm_relayd_rotate_stream stream_info;
2513 struct lttcomm_relayd_generic_reply reply;
2514 struct relay_stream *stream;
5312a3ed
JG		 2515 size_t header_len;
2516 size_t path_len;
2517 struct lttng_buffer_view new_path_view;
d3ecc550
JD		 2518
2519 DBG("Rotate stream received");
2520
2521 if (!session || !conn->version_check_done) {
2522 ERR("Trying to rotate a stream before version check");
2523 ret = -1;
2524 goto end_no_reply;
2525 }
2526
2527 if (session->major == 2 && session->minor < 11) {
2528 ERR("Unsupported feature before 2.11");
2529 ret = -1;
2530 goto end_no_reply;
2531 }
2532
5312a3ed 2533 header_len = sizeof(struct lttcomm_relayd_rotate_stream);
d3ecc550 2534
5312a3ed
JG		 2535 if (payload->size < header_len) {
2536 ERR("Unexpected payload size in \"relay_rotate_session_stream\": expected >= %zu bytes, got %zu bytes",
2537 header_len, payload->size);
d3ecc550
JD		 2538 ret = -1;
2539 goto end_no_reply;
2540 }
2541
5312a3ed
JG		 2542 memcpy(&stream_info, payload->data, header_len);
2543
2544 /* Convert to host */
2545 stream_info.pathname_length = be32toh(stream_info.pathname_length);
2546 stream_info.stream_id = be64toh(stream_info.stream_id);
2547 stream_info.new_chunk_id = be64toh(stream_info.new_chunk_id);
2548 stream_info.rotate_at_seq_num = be64toh(stream_info.rotate_at_seq_num);
d3ecc550 2549
5312a3ed
JG		 2550 path_len = stream_info.pathname_length;
2551 if (payload->size < header_len + path_len) {
2552 ERR("Unexpected payload size in \"relay_rotate_session_stream\" including path: expected >= %zu bytes, got %zu bytes",
2553 header_len + path_len, payload->size);
2554 ret = -1;
2555 goto end_no_reply;
2556 }
2557
d3ecc550 2558 /* Ensure it fits in local filename length. */
5312a3ed 2559 if (path_len >= LTTNG_PATH_MAX) {
d3ecc550
JD		 2560 ret = -ENAMETOOLONG;
2561 ERR("Length of relay_rotate_session_stream command's path name (%zu bytes) exceeds the maximal allowed length of %i bytes",
5312a3ed 2562 path_len, LTTNG_PATH_MAX);
d3ecc550
JD		 2563 goto end;
2564 }
2565
5312a3ed
JG		 2566 new_path_view = lttng_buffer_view_from_view(payload, header_len,
2567 stream_info.pathname_length);
d3ecc550 2568
5312a3ed
JG		 2569 stream = stream_get_by_id(stream_info.stream_id);
2570 if (!stream) {
d3ecc550 2571 ret = -1;
5312a3ed 2572 goto end;
d3ecc550
JD		 2573 }
2574
2575 pthread_mutex_lock(&stream->lock);
2576
2577 /*
2578 * Update the trace path (just the folder, the stream name does not
2579 * change).
2580 */
cb523e02
JG		 2581 free(stream->prev_path_name);
2582 stream->prev_path_name = stream->path_name;
5312a3ed 2583 stream->path_name = create_output_path(new_path_view.data);
d3ecc550
JD		 2584 if (!stream->path_name) {
2585 ERR("Failed to create a new output path");
5312a3ed 2586 ret = -1;
d3ecc550
JD		 2587 goto end_stream_unlock;
2588 }
2589 ret = utils_mkdir_recursive(stream->path_name, S_IRWXU | S_IRWXG,
2590 -1, -1);
2591 if (ret < 0) {
2592 ERR("relay creating output directory");
5312a3ed 2593 ret = -1;
d3ecc550
JD		 2594 goto end_stream_unlock;
2595 }
5312a3ed 2596
81164b6b
JG		 2597 assert(stream->current_chunk_id.is_set);
2598 stream->current_chunk_id.value = stream_info.new_chunk_id;
d3ecc550
JD		 2599
2600 if (stream->is_metadata) {
c6db3843
JG		 2601 /*
2602 * Metadata streams have no index; consider its rotation
2603 * complete.
2604 */
2605 stream->index_rotated = true;
d3ecc550
JD		 2606 /*
2607 * The metadata stream is sent only over the control connection
2608 * so we know we have all the data to perform the stream
2609 * rotation.
2610 */
c6db3843 2611 ret = do_rotate_stream_data(stream);
d3ecc550 2612 } else {
5312a3ed 2613 stream->rotate_at_seq_num = stream_info.rotate_at_seq_num;
c6db3843
JG		 2614 ret = try_rotate_stream_data(stream);
2615 if (ret < 0) {
2616 goto end_stream_unlock;
2617 }
2618
2619 ret = try_rotate_stream_index(stream);
2620 if (ret < 0) {
2621 goto end_stream_unlock;
2622 }
d3ecc550
JD		 2623 }
2624
2625end_stream_unlock:
2626 pthread_mutex_unlock(&stream->lock);
2627 stream_put(stream);
2628end:
2629 memset(&reply, 0, sizeof(reply));
2630 if (ret < 0) {
2631 reply.ret_code = htobe32(LTTNG_ERR_UNK);
2632 } else {
2633 reply.ret_code = htobe32(LTTNG_OK);
2634 }
2635 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply,
2636 sizeof(struct lttcomm_relayd_generic_reply), 0);
5312a3ed
JG		 2637 if (send_ret < (ssize_t) sizeof(reply)) {
2638 ERR("Failed to send \"rotate session stream\" command reply (ret = %zd)",
2639 send_ret);
2640 ret = -1;
d3ecc550
JD		 2641 }
2642
2643end_no_reply:
d3ecc550
JD		 2644 return ret;
2645}
2646
a1ae2ea5
JD		 2647/*
2648 * relay_mkdir: Create a folder on the disk.
2649 */
5312a3ed
JG		 2650static int relay_mkdir(const struct lttcomm_relayd_hdr *recv_hdr,
2651 struct relay_connection *conn,
2652 const struct lttng_buffer_view *payload)
a1ae2ea5
JD		 2653{
2654 int ret;
a1ae2ea5
JD		 2655 struct relay_session *session = conn->session;
2656 struct lttcomm_relayd_mkdir path_info_header;
a1ae2ea5
JD		 2657 struct lttcomm_relayd_generic_reply reply;
2658 char *path = NULL;
5312a3ed
JG		 2659 size_t header_len;
2660 ssize_t send_ret;
2661 struct lttng_buffer_view path_view;
a1ae2ea5
JD		 2662
2663 if (!session || !conn->version_check_done) {
00fb02ac 2664 ERR("Trying to create a directory before version check");
a1ae2ea5
JD		 2665 ret = -1;
2666 goto end_no_session;
2667 }
2668
2669 if (session->major == 2 && session->minor < 11) {
2670 /*
2671 * This client is not supposed to use this command since
2672 * it predates its introduction.
2673 */
2674 ERR("relay_mkdir command is unsupported before LTTng 2.11");
2675 ret = -1;
2676 goto end_no_session;
2677 }
2678
5312a3ed
JG		 2679 header_len = sizeof(path_info_header);
2680 if (payload->size < header_len) {
2681 ERR("Unexpected payload size in \"relay_mkdir\": expected >= %zu bytes, got %zu bytes",
2682 header_len, payload->size);
a1ae2ea5
JD		 2683 ret = -1;
2684 goto end_no_session;
2685 }
2686
5312a3ed
JG		 2687 memcpy(&path_info_header, payload->data, header_len);
2688
a1ae2ea5
JD		 2689 path_info_header.length = be32toh(path_info_header.length);
2690
5312a3ed
JG		 2691 if (payload->size < header_len + path_info_header.length) {
2692 ERR("Unexpected payload size in \"relay_mkdir\" including path: expected >= %zu bytes, got %zu bytes",
2693 header_len + path_info_header.length, payload->size);
2694 ret = -1;
2695 goto end_no_session;
2696 }
2697
a1ae2ea5
JD		 2698 /* Ensure that it fits in local path length. */
2699 if (path_info_header.length >= LTTNG_PATH_MAX) {
2700 ret = -ENAMETOOLONG;
2701 ERR("Path name argument of mkdir command (%" PRIu32 " bytes) exceeds the maximal length allowed (%d bytes)",
2702 path_info_header.length, LTTNG_PATH_MAX);
2703 goto end;
2704 }
2705
5312a3ed
JG		 2706 path_view = lttng_buffer_view_from_view(payload, header_len,
2707 path_info_header.length);
a1ae2ea5 2708
5312a3ed 2709 path = create_output_path(path_view.data);
a1ae2ea5
JD		 2710 if (!path) {
2711 ERR("Failed to create output path");
2712 ret = -1;
2713 goto end;
2714 }
2715
116ffe56 2716 DBG("MKDIR command has path \"%s\", changed to \"%s\"", path_view.data, path);
a1ae2ea5
JD		 2717 ret = utils_mkdir_recursive(path, S_IRWXU | S_IRWXG, -1, -1);
2718 if (ret < 0) {
2719 ERR("relay creating output directory");
2720 goto end;
2721 }
2722
2723 ret = 0;
2724
2725end:
2726 memset(&reply, 0, sizeof(reply));
2727 if (ret < 0) {
2728 reply.ret_code = htobe32(LTTNG_ERR_UNK);
2729 } else {
2730 reply.ret_code = htobe32(LTTNG_OK);
2731 }
5312a3ed
JG		 2732 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply, sizeof(reply), 0);
2733 if (send_ret < (ssize_t) sizeof(reply)) {
2734 ERR("Failed to send \"mkdir\" command reply (ret = %zd)", send_ret);
a1ae2ea5
JD		 2735 ret = -1;
2736 }
2737
2738end_no_session:
2739 free(path);
a1ae2ea5
JD		 2740 return ret;
2741}
2742
00fb02ac
JD		 2743static int validate_rotate_rename_path_length(const char *path_type,
2744 uint32_t path_length)
2745{
2746 int ret = 0;
2747
2748 if (path_length > LTTNG_PATH_MAX) {
2749 ret = -ENAMETOOLONG;
2750 ERR("rotate rename \"%s\" path name length (%" PRIu32 " bytes) exceeds the allowed size of %i bytes",
2751 path_type, path_length, LTTNG_PATH_MAX);
2752 } else if (path_length == 0) {
2753 ret = -EINVAL;
2754 ERR("rotate rename \"%s\" path name has an illegal length of 0", path_type);
2755 }
2756 return ret;
2757}
2758
2759/*
2760 * relay_rotate_rename: rename the trace folder after a rotation is
2761 * completed. We are not closing any fd here, just moving the folder, so it
2762 * works even if data is still in-flight.
2763 */
5312a3ed
JG		 2764static int relay_rotate_rename(const struct lttcomm_relayd_hdr *recv_hdr,
2765 struct relay_connection *conn,
2766 const struct lttng_buffer_view *payload)
00fb02ac
JD		 2767{
2768 int ret;
5312a3ed 2769 ssize_t send_ret;
00fb02ac
JD		 2770 struct relay_session *session = conn->session;
2771 struct lttcomm_relayd_generic_reply reply;
2772 struct lttcomm_relayd_rotate_rename header;
5312a3ed 2773 size_t header_len;
00fb02ac 2774 size_t received_paths_size;
00fb02ac 2775 char *complete_old_path = NULL, *complete_new_path = NULL;
5312a3ed
JG		 2776 struct lttng_buffer_view old_path_view;
2777 struct lttng_buffer_view new_path_view;
00fb02ac
JD		 2778
2779 if (!session || !conn->version_check_done) {
2780 ERR("Trying to rename a trace folder before version check");
2781 ret = -1;
2782 goto end_no_reply;
2783 }
2784
2785 if (session->major == 2 && session->minor < 11) {
2786 ERR("relay_rotate_rename command is unsupported before LTTng 2.11");
2787 ret = -1;
2788 goto end_no_reply;
2789 }
2790
5312a3ed
JG		 2791 header_len = sizeof(header);
2792 if (payload->size < header_len) {
2793 ERR("Unexpected payload size in \"relay_rotate_rename\": expected >= %zu bytes, got %zu bytes",
2794 header_len, payload->size);
00fb02ac
JD		 2795 ret = -1;
2796 goto end_no_reply;
2797 }
2798
5312a3ed
JG		 2799 memcpy(&header, payload->data, header_len);
2800
00fb02ac
JD		 2801 header.old_path_length = be32toh(header.old_path_length);
2802 header.new_path_length = be32toh(header.new_path_length);
2803 received_paths_size = header.old_path_length + header.new_path_length;
2804
5312a3ed
JG		 2805 if (payload->size < header_len + received_paths_size) {
2806 ERR("Unexpected payload size in \"relay_rotate_rename\" including paths: expected >= %zu bytes, got %zu bytes",
2807 header_len, payload->size);
2808 ret = -1;
2809 goto end_no_reply;
2810 }
2811
00fb02ac
JD		 2812 /* Ensure the paths don't exceed their allowed size. */
2813 ret = validate_rotate_rename_path_length("old", header.old_path_length);
2814 if (ret) {
2815 goto end;
2816 }
2817 ret = validate_rotate_rename_path_length("new", header.new_path_length);
2818 if (ret) {
2819 goto end;
2820 }
2821
5312a3ed
JG		 2822 old_path_view = lttng_buffer_view_from_view(payload, header_len,
2823 header.old_path_length);
2824 new_path_view = lttng_buffer_view_from_view(payload,
2825 header_len + header.old_path_length,
2826 header.new_path_length);
00fb02ac
JD		 2827
2828 /* Validate that both paths received are NULL terminated. */
5312a3ed 2829 if (old_path_view.data[old_path_view.size - 1] != '\0') {
00fb02ac
JD		 2830 ERR("relay_rotate_rename command's \"old\" path is invalid (not NULL terminated)");
2831 ret = -1;
2832 goto end;
2833 }
5312a3ed 2834 if (new_path_view.data[new_path_view.size - 1] != '\0') {
00fb02ac
JD		 2835 ERR("relay_rotate_rename command's \"new\" path is invalid (not NULL terminated)");
2836 ret = -1;
2837 goto end;
2838 }
2839
116ffe56
JG		 2840 DBG("ROTATE_RENAME command has argument old path = \"%s\", new_path = \"%s\"",
2841 old_path_view.data, new_path_view.data);
5312a3ed 2842 complete_old_path = create_output_path(old_path_view.data);
00fb02ac
JD		 2843 if (!complete_old_path) {
2844 ERR("Failed to build old output path in rotate_rename command");
2845 ret = -1;
2846 goto end;
2847 }
2848
5312a3ed 2849 complete_new_path = create_output_path(new_path_view.data);
00fb02ac
JD		 2850 if (!complete_new_path) {
2851 ERR("Failed to build new output path in rotate_rename command");
2852 ret = -1;
2853 goto end;
2854 }
116ffe56
JG		 2855 DBG("Expanded ROTATE_RENAME arguments to old path = \"%s\", new_path = \"%s\"",
2856 complete_old_path, complete_new_path);
00fb02ac
JD		 2857
2858 ret = utils_mkdir_recursive(complete_new_path, S_IRWXU | S_IRWXG,
2859 -1, -1);
2860 if (ret < 0) {
2861 ERR("Failed to mkdir() rotate_rename's \"new\" output directory at \"%s\"",
2862 complete_new_path);
2863 goto end;
2864 }
2865
2866 /*
2867 * If a domain has not yet created its channel, the domain-specific
2868 * folder might not exist, but this is not an error.
2869 */
2870 ret = rename(complete_old_path, complete_new_path);
2871 if (ret < 0 && errno != ENOENT) {
2872 PERROR("Renaming chunk in rotate_rename command from \"%s\" to \"%s\"",
2873 complete_old_path, complete_new_path);
2874 goto end;
2875 }
2876 ret = 0;
2877
2878end:
2879 memset(&reply, 0, sizeof(reply));
2880 if (ret < 0) {
2881 reply.ret_code = htobe32(LTTNG_ERR_UNK);
2882 } else {
2883 reply.ret_code = htobe32(LTTNG_OK);
2884 }
5312a3ed
JG		 2885 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply,
2886 sizeof(reply), 0);
2887 if (send_ret < sizeof(reply)) {
2888 ERR("Failed to send \"rotate rename\" command reply (ret = %zd)",
2889 send_ret);
00fb02ac
JD		 2890 ret = -1;
2891 }
2892
2893end_no_reply:
00fb02ac
JD		 2894 free(complete_old_path);
2895 free(complete_new_path);
2896 return ret;
2897}
2898
e2acc8d2
JD		 2899/*
2900 * Check if all the streams in the session have completed the last rotation.
2901 * The chunk_id value is used to distinguish the cases where a stream was
2902 * closed on the consumerd before the rotation started but it still active on
2903 * the relayd, and the case where a stream appeared on the consumerd/relayd
2904 * after the last rotation started (in that case, it is already writing in the
2905 * new chunk folder).
2906 */
2907static
5312a3ed
JG		 2908int relay_rotate_pending(const struct lttcomm_relayd_hdr *recv_hdr,
2909 struct relay_connection *conn,
2910 const struct lttng_buffer_view *payload)
e2acc8d2
JD		 2911{
2912 struct relay_session *session = conn->session;
2913 struct lttcomm_relayd_rotate_pending msg;
d88744a4 2914 struct lttcomm_relayd_rotate_pending_reply reply;
e2acc8d2
JD		 2915 struct lttng_ht_iter iter;
2916 struct relay_stream *stream;
4f5fb4c3 2917 int ret = 0;
5312a3ed 2918 ssize_t send_ret;
e2acc8d2
JD		 2919 uint64_t chunk_id;
2920 bool rotate_pending = false;
2921
2922 DBG("Rotate pending command received");
2923
2924 if (!session || !conn->version_check_done) {
2925 ERR("Trying to check for data before version check");
2926 ret = -1;
2927 goto end_no_reply;
2928 }
2929
2930 if (session->major == 2 && session->minor < 11) {
2931 ERR("Unsupported feature before 2.11");
2932 ret = -1;
2933 goto end_no_reply;
2934 }
2935
5312a3ed
JG		 2936 if (payload->size < sizeof(msg)) {
2937 ERR("Unexpected payload size in \"relay_rotate_pending\": expected >= %zu bytes, got %zu bytes",
2938 sizeof(msg), payload->size);
e2acc8d2
JD		 2939 ret = -1;
2940 goto end_no_reply;
2941 }
2942
5312a3ed
JG		 2943 memcpy(&msg, payload->data, sizeof(msg));
2944
e2acc8d2 2945 chunk_id = be64toh(msg.chunk_id);
5312a3ed 2946
ce6877c9
JG		 2947 DBG("Evaluating rotate pending for session \"%s\" and chunk id %" PRIu64,
2948 session->session_name, chunk_id);
e2acc8d2
JD		 2949
2950 /*
2951 * Iterate over all the streams in the session and check if they are
2952 * still waiting for data to perform their rotation.
2953 */
2954 rcu_read_lock();
2955 cds_lfht_for_each_entry(relay_streams_ht->ht, &iter.iter, stream,
2956 node.node) {
2957 if (!stream_get(stream)) {
2958 continue;
2959 }
2960 if (stream->trace->session != session) {
2961 stream_put(stream);
2962 continue;
2963 }
2964 pthread_mutex_lock(&stream->lock);
2965 if (stream->rotate_at_seq_num != -1ULL) {
2966 /* We have not yet performed the rotation. */
2967 rotate_pending = true;
2968 DBG("Stream %" PRIu64 " is still rotating",
2969 stream->stream_handle);
2280aad1 2970 } else if (stream->current_chunk_id.value <= chunk_id) {
e2acc8d2
JD		 2971 /*
2972 * Stream closed on the consumer but still active on the
2973 * relay.
2974 */
2975 rotate_pending = true;
2976 DBG("Stream %" PRIu64 " did not exist on the consumer "
2977 "when the last rotation started, but is"
2978 "still waiting for data before getting"
2979 "closed",
2980 stream->stream_handle);
2981 }
2982 pthread_mutex_unlock(&stream->lock);
2983 stream_put(stream);
2984 if (rotate_pending) {
2985 goto send_reply;
2986 }
2987 }
2988
2989send_reply:
2990 rcu_read_unlock();
2991 memset(&reply, 0, sizeof(reply));
d88744a4
JD		 2992 reply.generic.ret_code = htobe32((uint32_t) LTTNG_OK);
2993 reply.is_pending = (uint8_t) !!rotate_pending;
5312a3ed 2994 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply,
e2acc8d2 2995 sizeof(reply), 0);
5312a3ed
JG		 2996 if (send_ret < (ssize_t) sizeof(reply)) {
2997 ERR("Failed to send \"rotate pending\" command reply (ret = %zd)",
2998 send_ret);
e2acc8d2
JD		 2999 ret = -1;
3000 }
3001
3002end_no_reply:
3003 return ret;
3004}
3005
5312a3ed
JG		 3006#define DBG_CMD(cmd_name, conn) \
3007 DBG3("Processing \"%s\" command for socket %i", cmd_name, conn->sock->fd);
3008
3009static int relay_process_control_command(struct relay_connection *conn,
3010 const struct lttcomm_relayd_hdr *header,
3011 const struct lttng_buffer_view *payload)
b8aa1682
JD		 3012{
3013 int ret = 0;
3014
5312a3ed 3015 switch (header->cmd) {
b8aa1682 3016 case RELAYD_CREATE_SESSION:
5312a3ed
JG		 3017 DBG_CMD("RELAYD_CREATE_SESSION", conn);
3018 ret = relay_create_session(header, conn, payload);
b8aa1682 3019 break;
b8aa1682 3020 case RELAYD_ADD_STREAM:
5312a3ed
JG		 3021 DBG_CMD("RELAYD_ADD_STREAM", conn);
3022 ret = relay_add_stream(header, conn, payload);
b8aa1682
JD		 3023 break;
3024 case RELAYD_START_DATA:
5312a3ed
JG		 3025 DBG_CMD("RELAYD_START_DATA", conn);
3026 ret = relay_start(header, conn, payload);
b8aa1682
JD		 3027 break;
3028 case RELAYD_SEND_METADATA:
5312a3ed
JG		 3029 DBG_CMD("RELAYD_SEND_METADATA", conn);
3030 ret = relay_recv_metadata(header, conn, payload);
b8aa1682
JD		 3031 break;
3032 case RELAYD_VERSION:
5312a3ed
JG		 3033 DBG_CMD("RELAYD_VERSION", conn);
3034 ret = relay_send_version(header, conn, payload);
b8aa1682 3035 break;
173af62f 3036 case RELAYD_CLOSE_STREAM:
5312a3ed
JG		 3037 DBG_CMD("RELAYD_CLOSE_STREAM", conn);
3038 ret = relay_close_stream(header, conn, payload);
173af62f 3039 break;
6d805429 3040 case RELAYD_DATA_PENDING:
5312a3ed
JG		 3041 DBG_CMD("RELAYD_DATA_PENDING", conn);
3042 ret = relay_data_pending(header, conn, payload);
c8f59ee5
DG		 3043 break;
3044 case RELAYD_QUIESCENT_CONTROL:
5312a3ed
JG		 3045 DBG_CMD("RELAYD_QUIESCENT_CONTROL", conn);
3046 ret = relay_quiescent_control(header, conn, payload);
c8f59ee5 3047 break;
f7079f67 3048 case RELAYD_BEGIN_DATA_PENDING:
5312a3ed
JG		 3049 DBG_CMD("RELAYD_BEGIN_DATA_PENDING", conn);
3050 ret = relay_begin_data_pending(header, conn, payload);
f7079f67
DG		 3051 break;
3052 case RELAYD_END_DATA_PENDING:
5312a3ed
JG		 3053 DBG_CMD("RELAYD_END_DATA_PENDING", conn);
3054 ret = relay_end_data_pending(header, conn, payload);
f7079f67 3055 break;
1c20f0e2 3056 case RELAYD_SEND_INDEX:
5312a3ed
JG		 3057 DBG_CMD("RELAYD_SEND_INDEX", conn);
3058 ret = relay_recv_index(header, conn, payload);
1c20f0e2 3059 break;
a4baae1b 3060 case RELAYD_STREAMS_SENT:
5312a3ed
JG		 3061 DBG_CMD("RELAYD_STREAMS_SENT", conn);
3062 ret = relay_streams_sent(header, conn, payload);
a4baae1b 3063 break;
93ec662e 3064 case RELAYD_RESET_METADATA:
5312a3ed
JG		 3065 DBG_CMD("RELAYD_RESET_METADATA", conn);
3066 ret = relay_reset_metadata(header, conn, payload);
93ec662e 3067 break;
d3ecc550 3068 case RELAYD_ROTATE_STREAM:
5312a3ed
JG		 3069 DBG_CMD("RELAYD_ROTATE_STREAM", conn);
3070 ret = relay_rotate_session_stream(header, conn, payload);
d3ecc550 3071 break;
00fb02ac 3072 case RELAYD_ROTATE_RENAME:
5312a3ed
JG		 3073 DBG_CMD("RELAYD_ROTATE_RENAME", conn);
3074 ret = relay_rotate_rename(header, conn, payload);
00fb02ac 3075 break;
e2acc8d2 3076 case RELAYD_ROTATE_PENDING:
5312a3ed
JG		 3077 DBG_CMD("RELAYD_ROTATE_PENDING", conn);
3078 ret = relay_rotate_pending(header, conn, payload);
e2acc8d2 3079 break;
a1ae2ea5 3080 case RELAYD_MKDIR:
5312a3ed
JG		 3081 DBG_CMD("RELAYD_MKDIR", conn);
3082 ret = relay_mkdir(header, conn, payload);
a1ae2ea5 3083 break;
b8aa1682
JD		 3084 case RELAYD_UPDATE_SYNC_INFO:
3085 default:
5312a3ed 3086 ERR("Received unknown command (%u)", header->cmd);
58eb9381 3087 relay_unknown_command(conn);
b8aa1682
JD		 3088 ret = -1;
3089 goto end;
3090 }
3091
3092end:
3093 return ret;
3094}
3095
5569b118
JG		 3096static enum relay_connection_status relay_process_control_receive_payload(
3097 struct relay_connection *conn)
5312a3ed
JG		 3098{
3099 int ret = 0;
5569b118 3100 enum relay_connection_status status = RELAY_CONNECTION_STATUS_OK;
5312a3ed
JG		 3101 struct lttng_dynamic_buffer *reception_buffer =
3102 &conn->protocol.ctrl.reception_buffer;
3103 struct ctrl_connection_state_receive_payload *state =
3104 &conn->protocol.ctrl.state.receive_payload;
3105 struct lttng_buffer_view payload_view;
3106
3107 if (state->left_to_receive == 0) {
3108 /* Short-circuit for payload-less commands. */
3109 goto reception_complete;
3110 }
3111
3112 ret = conn->sock->ops->recvmsg(conn->sock,
3113 reception_buffer->data + state->received,
3114 state->left_to_receive, MSG_DONTWAIT);
3115 if (ret < 0) {
5569b118
JG		 3116 if (errno != EAGAIN && errno != EWOULDBLOCK) {
3117 PERROR("Unable to receive command payload on sock %d",
3118 conn->sock->fd);
3119 status = RELAY_CONNECTION_STATUS_ERROR;
3120 }
5312a3ed
JG		 3121 goto end;
3122 } else if (ret == 0) {
3123 DBG("Socket %d performed an orderly shutdown (received EOF)", conn->sock->fd);
5569b118 3124 status = RELAY_CONNECTION_STATUS_CLOSED;
5312a3ed
JG		 3125 goto end;
3126 }
3127
3128 assert(ret > 0);
3129 assert(ret <= state->left_to_receive);
3130
3131 state->left_to_receive -= ret;
3132 state->received += ret;
3133
3134 if (state->left_to_receive > 0) {
3135 /*
3136 * Can't transition to the protocol's next state, wait to
3137 * receive the rest of the header.
3138 */
3139 DBG3("Partial reception of control connection protocol payload (received %" PRIu64 " bytes, %" PRIu64 " bytes left to receive, fd = %i)",
3140 state->received, state->left_to_receive,
3141 conn->sock->fd);
5312a3ed
JG		 3142 goto end;
3143 }
3144
3145reception_complete:
3146 DBG("Done receiving control command payload: fd = %i, payload size = %" PRIu64 " bytes",
3147 conn->sock->fd, state->received);
3148 /*
3149 * The payload required to process the command has been received.
3150 * A view to the reception buffer is forwarded to the various
3151 * commands and the state of the control is reset on success.
3152 *
3153 * Commands are responsible for sending their reply to the peer.
3154 */
3155 payload_view = lttng_buffer_view_from_dynamic_buffer(reception_buffer,
3156 0, -1);
3157 ret = relay_process_control_command(conn,
3158 &state->header, &payload_view);
3159 if (ret < 0) {
5569b118 3160 status = RELAY_CONNECTION_STATUS_ERROR;
5312a3ed
JG		 3161 goto end;
3162 }
3163
3164 ret = connection_reset_protocol_state(conn);
5569b118
JG		 3165 if (ret) {
3166 status = RELAY_CONNECTION_STATUS_ERROR;
3167 }
5312a3ed 3168end:
5569b118 3169 return status;
5312a3ed
JG		 3170}
3171
5569b118
JG		 3172static enum relay_connection_status relay_process_control_receive_header(
3173 struct relay_connection *conn)
5312a3ed
JG		 3174{
3175 int ret = 0;
5569b118 3176 enum relay_connection_status status = RELAY_CONNECTION_STATUS_OK;
5312a3ed
JG		 3177 struct lttcomm_relayd_hdr header;
3178 struct lttng_dynamic_buffer *reception_buffer =
3179 &conn->protocol.ctrl.reception_buffer;
3180 struct ctrl_connection_state_receive_header *state =
3181 &conn->protocol.ctrl.state.receive_header;
3182
3183 assert(state->left_to_receive != 0);
3184
3185 ret = conn->sock->ops->recvmsg(conn->sock,
3186 reception_buffer->data + state->received,
3187 state->left_to_receive, MSG_DONTWAIT);
3188 if (ret < 0) {
5569b118
JG		 3189 if (errno != EAGAIN && errno != EWOULDBLOCK) {
3190 PERROR("Unable to receive control command header on sock %d",
3191 conn->sock->fd);
3192 status = RELAY_CONNECTION_STATUS_ERROR;
3193 }
5312a3ed
JG		 3194 goto end;
3195 } else if (ret == 0) {
3196 DBG("Socket %d performed an orderly shutdown (received EOF)", conn->sock->fd);
5569b118 3197 status = RELAY_CONNECTION_STATUS_CLOSED;
5312a3ed
JG		 3198 goto end;
3199 }
3200
3201 assert(ret > 0);
3202 assert(ret <= state->left_to_receive);
3203
3204 state->left_to_receive -= ret;
3205 state->received += ret;
3206
3207 if (state->left_to_receive > 0) {
3208 /*
3209 * Can't transition to the protocol's next state, wait to
3210 * receive the rest of the header.
3211 */
3212 DBG3("Partial reception of control connection protocol header (received %" PRIu64 " bytes, %" PRIu64 " bytes left to receive, fd = %i)",
3213 state->received, state->left_to_receive,
3214 conn->sock->fd);
5312a3ed
JG		 3215 goto end;
3216 }
3217
3218 /* Transition to next state: receiving the command's payload. */
3219 conn->protocol.ctrl.state_id =
3220 CTRL_CONNECTION_STATE_RECEIVE_PAYLOAD;
3221 memcpy(&header, reception_buffer->data, sizeof(header));
3222 header.circuit_id = be64toh(header.circuit_id);
3223 header.data_size = be64toh(header.data_size);
3224 header.cmd = be32toh(header.cmd);
3225 header.cmd_version = be32toh(header.cmd_version);
3226 memcpy(&conn->protocol.ctrl.state.receive_payload.header,
3227 &header, sizeof(header));
3228
3229 DBG("Done receiving control command header: fd = %i, cmd = %" PRIu32 ", cmd_version = %" PRIu32 ", payload size = %" PRIu64 " bytes",
3230 conn->sock->fd, header.cmd, header.cmd_version,
3231 header.data_size);
3232
715e6fb1 3233 if (header.data_size > DEFAULT_NETWORK_RELAYD_CTRL_MAX_PAYLOAD_SIZE) {
5312a3ed
JG		 3234 ERR("Command header indicates a payload (%" PRIu64 " bytes) that exceeds the maximal payload size allowed on a control connection.",
3235 header.data_size);
5569b118 3236 status = RELAY_CONNECTION_STATUS_ERROR;
5312a3ed
JG		 3237 goto end;
3238 }
3239
3240 conn->protocol.ctrl.state.receive_payload.left_to_receive =
3241 header.data_size;
3242 conn->protocol.ctrl.state.receive_payload.received = 0;
3243 ret = lttng_dynamic_buffer_set_size(reception_buffer,
3244 header.data_size);
3245 if (ret) {
5569b118 3246 status = RELAY_CONNECTION_STATUS_ERROR;
5312a3ed
JG		 3247 goto end;
3248 }
3249
3250 if (header.data_size == 0) {
3251 /*
3252 * Manually invoke the next state as the poll loop
3253 * will not wake-up to allow us to proceed further.
3254 */
5569b118 3255 status = relay_process_control_receive_payload(conn);
5312a3ed
JG		 3256 }
3257end:
5569b118 3258 return status;
5312a3ed
JG		 3259}
3260
3261/*
3262 * Process the commands received on the control socket
3263 */
5569b118
JG		 3264static enum relay_connection_status relay_process_control(
3265 struct relay_connection *conn)
5312a3ed 3266{
5569b118 3267 enum relay_connection_status status;
5312a3ed
JG		 3268
3269 switch (conn->protocol.ctrl.state_id) {
3270 case CTRL_CONNECTION_STATE_RECEIVE_HEADER:
5569b118 3271 status = relay_process_control_receive_header(conn);
5312a3ed
JG		 3272 break;
3273 case CTRL_CONNECTION_STATE_RECEIVE_PAYLOAD:
5569b118 3274 status = relay_process_control_receive_payload(conn);
5312a3ed
JG		 3275 break;
3276 default:
3277 ERR("Unknown control connection protocol state encountered.");
3278 abort();
3279 }
3280
5569b118 3281 return status;
5312a3ed
JG		 3282}
3283
7d2f7452
DG		 3284/*
3285 * Handle index for a data stream.
3286 *
7591bab1 3287 * Called with the stream lock held.
7d2f7452
DG		 3288 *
3289 * Return 0 on success else a negative value.
3290 */
3291static int handle_index_data(struct relay_stream *stream, uint64_t net_seq_num,
5312a3ed 3292 bool rotate_index, bool *flushed, uint64_t total_size)
7d2f7452 3293{
7591bab1
MD		 3294 int ret = 0;
3295 uint64_t data_offset;
3296 struct relay_index *index;
7d2f7452 3297
7d2f7452
DG		 3298 /* Get data offset because we are about to update the index. */
3299 data_offset = htobe64(stream->tracefile_size_current);
3300
65d66b19
MD		 3301 DBG("handle_index_data: stream %" PRIu64 " net_seq_num %" PRIu64 " data offset %" PRIu64,
3302 stream->stream_handle, net_seq_num, stream->tracefile_size_current);
7591bab1 3303
7d2f7452 3304 /*
7591bab1
MD		 3305 * Lookup for an existing index for that stream id/sequence
3306 * number. If it exists, the control thread has already received the
3307 * data for it, thus we need to write it to disk.
7d2f7452 3308 */
7591bab1 3309 index = relay_index_get_by_id_or_create(stream, net_seq_num);
7d2f7452 3310 if (!index) {
7591bab1
MD		 3311 ret = -1;
3312 goto end;
7d2f7452
DG		 3313 }
3314
f8f3885c 3315 if (rotate_index || !stream->index_file) {
cb523e02
JG		 3316 const char *stream_path;
3317
3318 /*
3319 * The data connection creates the stream's first index file.
3320 *
3321 * This can happen _after_ a ROTATE_STREAM command. In
3322 * other words, the data of the first packet of this stream
3323 * can be received after a ROTATE_STREAM command.
3324 *
3325 * The ROTATE_STREAM command changes the stream's path_name
3326 * to point to the "next" chunk. If a rotation is pending for
3327 * this stream, as indicated by "rotate_at_seq_num != -1ULL",
3328 * it means that we are still receiving data that belongs in the
3329 * stream's former path.
3330 *
3331 * In this very specific case, we must ensure that the index
3332 * file is created in the streams's former path,
3333 * "prev_path_name".
3334 *
3335 * All other rotations beyond the first one are not affected
3336 * by this problem since the actual rotation operation creates
3337 * the new chunk's index file.
3338 */
3339 stream_path = stream->rotate_at_seq_num == -1ULL ?
3340 stream->path_name:
3341 stream->prev_path_name;
3342
3343 ret = create_rotate_index_file(stream, stream_path);
d3ecc550
JD		 3344 if (ret < 0) {
3345 ERR("Failed to rotate index");
7591bab1
MD		 3346 /* Put self-ref for this index due to error. */
3347 relay_index_put(index);
f8f3885c 3348 index = NULL;
7591bab1 3349 goto end;
7d2f7452 3350 }
7d2f7452
DG		 3351 }
3352
f8f3885c 3353 if (relay_index_set_file(index, stream->index_file, data_offset)) {
7591bab1
MD		 3354 ret = -1;
3355 /* Put self-ref for this index due to error. */
3356 relay_index_put(index);
f8f3885c 3357 index = NULL;
7591bab1 3358 goto end;
7d2f7452
DG		 3359 }
3360
7591bab1
MD		 3361 ret = relay_index_try_flush(index);
3362 if (ret == 0) {
a44ca2ca
MD		 3363 tracefile_array_commit_seq(stream->tfa);
3364 stream->index_received_seqcount++;
d3ecc550 3365 *flushed = true;
7591bab1 3366 } else if (ret > 0) {
d3ecc550 3367 index->total_size = total_size;
7591bab1
MD		 3368 /* No flush. */
3369 ret = 0;
3370 } else {
245f8bec
JR		 3371 /*
3372 * ret < 0
3373 *
3374 * relay_index_try_flush is responsible for the self-reference
3375 * put of the index object on error.
3376 */
3377 ERR("relay_index_try_flush error %d", ret);
7591bab1
MD		 3378 ret = -1;
3379 }
3380end:
7d2f7452
DG		 3381 return ret;
3382}
3383
5569b118
JG		 3384static enum relay_connection_status relay_process_data_receive_header(
3385 struct relay_connection *conn)
b8aa1682 3386{
5312a3ed 3387 int ret;
5569b118 3388 enum relay_connection_status status = RELAY_CONNECTION_STATUS_OK;
5312a3ed
JG		 3389 struct data_connection_state_receive_header *state =
3390 &conn->protocol.data.state.receive_header;
3391 struct lttcomm_relayd_data_hdr header;
b8aa1682 3392 struct relay_stream *stream;
5312a3ed
JG		 3393
3394 assert(state->left_to_receive != 0);
3395
3396 ret = conn->sock->ops->recvmsg(conn->sock,
3397 state->header_reception_buffer + state->received,
3398 state->left_to_receive, MSG_DONTWAIT);
3399 if (ret < 0) {
5569b118
JG		 3400 if (errno != EAGAIN && errno != EWOULDBLOCK) {
3401 PERROR("Unable to receive data header on sock %d", conn->sock->fd);
3402 status = RELAY_CONNECTION_STATUS_ERROR;
3403 }
5312a3ed
JG		 3404 goto end;
3405 } else if (ret == 0) {
3406 /* Orderly shutdown. Not necessary to print an error. */
3407 DBG("Socket %d performed an orderly shutdown (received EOF)", conn->sock->fd);
5569b118 3408 status = RELAY_CONNECTION_STATUS_CLOSED;
b8aa1682
JD		 3409 goto end;
3410 }
3411
5312a3ed
JG		 3412 assert(ret > 0);
3413 assert(ret <= state->left_to_receive);
3414
3415 state->left_to_receive -= ret;
3416 state->received += ret;
3417
3418 if (state->left_to_receive > 0) {
3419 /*
3420 * Can't transition to the protocol's next state, wait to
3421 * receive the rest of the header.
3422 */
3423 DBG3("Partial reception of data connection header (received %" PRIu64 " bytes, %" PRIu64 " bytes left to receive, fd = %i)",
3424 state->received, state->left_to_receive,
3425 conn->sock->fd);
7591bab1 3426 goto end;
b8aa1682 3427 }
b8aa1682 3428
5312a3ed
JG		 3429 /* Transition to next state: receiving the payload. */
3430 conn->protocol.data.state_id = DATA_CONNECTION_STATE_RECEIVE_PAYLOAD;
173af62f 3431
5312a3ed
JG		 3432 memcpy(&header, state->header_reception_buffer, sizeof(header));
3433 header.circuit_id = be64toh(header.circuit_id);
3434 header.stream_id = be64toh(header.stream_id);
3435 header.data_size = be32toh(header.data_size);
3436 header.net_seq_num = be64toh(header.net_seq_num);
3437 header.padding_size = be32toh(header.padding_size);
3438 memcpy(&conn->protocol.data.state.receive_payload.header, &header, sizeof(header));
3439
3440 conn->protocol.data.state.receive_payload.left_to_receive =
3441 header.data_size;
3442 conn->protocol.data.state.receive_payload.received = 0;
3443 conn->protocol.data.state.receive_payload.rotate_index = false;
3444
3445 DBG("Received data connection header on fd %i: circuit_id = %" PRIu64 ", stream_id = %" PRIu64 ", data_size = %" PRIu32 ", net_seq_num = %" PRIu64 ", padding_size = %" PRIu32,
3446 conn->sock->fd, header.circuit_id,
3447 header.stream_id, header.data_size,
3448 header.net_seq_num, header.padding_size);
3449
3450 stream = stream_get_by_id(header.stream_id);
3451 if (!stream) {
3452 DBG("relay_process_data_receive_payload: Cannot find stream %" PRIu64,
3453 header.stream_id);
5569b118
JG		 3454 /* Protocol error. */
3455 status = RELAY_CONNECTION_STATUS_ERROR;
5312a3ed
JG		 3456 goto end;
3457 }
b8aa1682 3458
7591bab1
MD		 3459 pthread_mutex_lock(&stream->lock);
3460
1c20f0e2 3461 /* Check if a rotation is needed. */
0f907de1 3462 if (stream->tracefile_size > 0 &&
5312a3ed 3463 (stream->tracefile_size_current + header.data_size) >
0f907de1 3464 stream->tracefile_size) {
a44ca2ca
MD		 3465 uint64_t old_id, new_id;
3466
3467 old_id = tracefile_array_get_file_index_head(stream->tfa);
3468 tracefile_array_file_rotate(stream->tfa);
3469
3470 /* new_id is updated by utils_rotate_stream_file. */
3471 new_id = old_id;
6b6b9a5a 3472
7591bab1
MD		 3473 ret = utils_rotate_stream_file(stream->path_name,
3474 stream->channel_name, stream->tracefile_size,
3475 stream->tracefile_count, -1,
3476 -1, stream->stream_fd->fd,
a44ca2ca 3477 &new_id, &stream->stream_fd->fd);
0f907de1 3478 if (ret < 0) {
5312a3ed 3479 ERR("Failed to rotate stream output file");
5569b118 3480 status = RELAY_CONNECTION_STATUS_ERROR;
7591bab1
MD		 3481 goto end_stream_unlock;
3482 }
5312a3ed 3483
7591bab1
MD		 3484 /*
3485 * Reset current size because we just performed a stream
3486 * rotation.
3487 */
a6976990 3488 stream->tracefile_size_current = 0;
5312a3ed 3489 conn->protocol.data.state.receive_payload.rotate_index = true;
1c20f0e2
JD		 3490 }
3491
5312a3ed
JG		 3492end_stream_unlock:
3493 pthread_mutex_unlock(&stream->lock);
3494 stream_put(stream);
3495end:
5569b118 3496 return status;
5312a3ed
JG		 3497}
3498
5569b118
JG		 3499static enum relay_connection_status relay_process_data_receive_payload(
3500 struct relay_connection *conn)
5312a3ed
JG		 3501{
3502 int ret;
5569b118 3503 enum relay_connection_status status = RELAY_CONNECTION_STATUS_OK;
5312a3ed
JG		 3504 struct relay_stream *stream;
3505 struct data_connection_state_receive_payload *state =
3506 &conn->protocol.data.state.receive_payload;
3507 const size_t chunk_size = RECV_DATA_BUFFER_SIZE;
3508 char data_buffer[chunk_size];
3509 bool partial_recv = false;
3510 bool new_stream = false, close_requested = false, index_flushed = false;
3511 uint64_t left_to_receive = state->left_to_receive;
3512 struct relay_session *session;
3513
fd0f1e3e
JR		 3514 DBG3("Receiving data for stream id %" PRIu64 " seqnum %" PRIu64 ", %" PRIu64" bytes received, %" PRIu64 " bytes left to receive",
3515 state->header.stream_id, state->header.net_seq_num,
3516 state->received, left_to_receive);
3517
5312a3ed
JG		 3518 stream = stream_get_by_id(state->header.stream_id);
3519 if (!stream) {
5569b118 3520 /* Protocol error. */
fd0f1e3e 3521 ERR("relay_process_data_receive_payload: cannot find stream %" PRIu64,
5312a3ed 3522 state->header.stream_id);
5569b118 3523 status = RELAY_CONNECTION_STATUS_ERROR;
5312a3ed 3524 goto end;
1c20f0e2
JD		 3525 }
3526
5312a3ed
JG		 3527 pthread_mutex_lock(&stream->lock);
3528 session = stream->trace->session;
fd0f1e3e
JR		 3529 if (!conn->session) {
3530 ret = connection_set_session(conn, session);
3531 if (ret) {
3532 status = RELAY_CONNECTION_STATUS_ERROR;
3533 goto end_stream_unlock;
3534 }
3535 }
5312a3ed
JG		 3536
3537 /*
3538 * The size of the "chunk" received on any iteration is bounded by:
3539 * - the data left to receive,
3540 * - the data immediately available on the socket,
3541 * - the on-stack data buffer
3542 */
3543 while (left_to_receive > 0 && !partial_recv) {
3544 ssize_t write_ret;
3545 size_t recv_size = min(left_to_receive, chunk_size);
3546
3547 ret = conn->sock->ops->recvmsg(conn->sock, data_buffer,
3548 recv_size, MSG_DONTWAIT);
3549 if (ret < 0) {
5569b118
JG		 3550 if (errno != EAGAIN && errno != EWOULDBLOCK) {
3551 PERROR("Socket %d error", conn->sock->fd);
3552 status = RELAY_CONNECTION_STATUS_ERROR;
3553 }
0848dba7 3554 goto end_stream_unlock;
5312a3ed
JG		 3555 } else if (ret == 0) {
3556 /* No more data ready to be consumed on socket. */
3557 DBG3("No more data ready for consumption on data socket of stream id %" PRIu64,
3558 state->header.stream_id);
5569b118 3559 status = RELAY_CONNECTION_STATUS_CLOSED;
5312a3ed
JG		 3560 break;
3561 } else if (ret < (int) recv_size) {
3562 /*
3563 * All the data available on the socket has been
3564 * consumed.
3565 */
3566 partial_recv = true;
0848dba7
MD		 3567 }
3568
5312a3ed
JG		 3569 recv_size = ret;
3570
0848dba7 3571 /* Write data to stream output fd. */
5312a3ed 3572 write_ret = lttng_write(stream->stream_fd->fd, data_buffer,
0848dba7 3573 recv_size);
5312a3ed 3574 if (write_ret < (ssize_t) recv_size) {
0848dba7 3575 ERR("Relay error writing data to file");
5569b118 3576 status = RELAY_CONNECTION_STATUS_ERROR;
0848dba7
MD		 3577 goto end_stream_unlock;
3578 }
3579
5312a3ed
JG		 3580 left_to_receive -= recv_size;
3581 state->received += recv_size;
3582 state->left_to_receive = left_to_receive;
3583
0848dba7 3584 DBG2("Relay wrote %zd bytes to tracefile for stream id %" PRIu64,
5312a3ed
JG		 3585 write_ret, stream->stream_handle);
3586 }
3587
3588 if (state->left_to_receive > 0) {
3589 /*
3590 * Did not receive all the data expected, wait for more data to
3591 * become available on the socket.
3592 */
3593 DBG3("Partial receive on data connection of stream id %" PRIu64 ", %" PRIu64 " bytes received, %" PRIu64 " bytes left to receive",
3594 state->header.stream_id, state->received,
3595 state->left_to_receive);
5312a3ed 3596 goto end_stream_unlock;
0848dba7 3597 }
5ab7344e 3598
7591bab1 3599 ret = write_padding_to_file(stream->stream_fd->fd,
5312a3ed 3600 state->header.padding_size);
46f4eaa5 3601 if ((int64_t) ret < (int64_t) state->header.padding_size) {
65d66b19 3602 ERR("write_padding_to_file: fail stream %" PRIu64 " net_seq_num %" PRIu64 " ret %d",
5312a3ed
JG		 3603 stream->stream_handle,
3604 state->header.net_seq_num, ret);
5569b118 3605 status = RELAY_CONNECTION_STATUS_ERROR;
7591bab1 3606 goto end_stream_unlock;
1d4dfdef 3607 }
5312a3ed
JG		 3608
3609
298a25ca 3610 if (session_streams_have_index(session)) {
5312a3ed
JG		 3611 ret = handle_index_data(stream, state->header.net_seq_num,
3612 state->rotate_index, &index_flushed, state->header.data_size + state->header.padding_size);
3613 if (ret < 0) {
3614 ERR("handle_index_data: fail stream %" PRIu64 " net_seq_num %" PRIu64 " ret %d",
3615 stream->stream_handle,
3616 state->header.net_seq_num, ret);
5569b118 3617 status = RELAY_CONNECTION_STATUS_ERROR;
5312a3ed
JG		 3618 goto end_stream_unlock;
3619 }
3620 }
3621
3622 stream->tracefile_size_current += state->header.data_size +
3623 state->header.padding_size;
3624
a8f9f353 3625 if (stream->prev_data_seq == -1ULL) {
c0bae11d
MD		 3626 new_stream = true;
3627 }
d3ecc550
JD		 3628 if (index_flushed) {
3629 stream->pos_after_last_complete_data_index =
3630 stream->tracefile_size_current;
7a45c7e6 3631 stream->prev_index_seq = state->header.net_seq_num;
c6db3843
JG		 3632 ret = try_rotate_stream_index(stream);
3633 if (ret < 0) {
3634 goto end_stream_unlock;
3635 }
d3ecc550 3636 }
c0bae11d 3637
a8f9f353 3638 stream->prev_data_seq = state->header.net_seq_num;
5312a3ed
JG		 3639
3640 /*
3641 * Resetting the protocol state (to RECEIVE_HEADER) will trash the
3642 * contents of *state which are aliased (union) to the same location as
3643 * the new state. Don't use it beyond this point.
3644 */
3645 connection_reset_protocol_state(conn);
3646 state = NULL;
173af62f 3647
c6db3843 3648 ret = try_rotate_stream_data(stream);
d3ecc550 3649 if (ret < 0) {
5569b118 3650 status = RELAY_CONNECTION_STATUS_ERROR;
d3ecc550
JD		 3651 goto end_stream_unlock;
3652 }
3653
7591bab1 3654end_stream_unlock:
bda7c7b9 3655 close_requested = stream->close_requested;
7591bab1 3656 pthread_mutex_unlock(&stream->lock);
5312a3ed 3657 if (close_requested && left_to_receive == 0) {
bda7c7b9
JG		 3658 try_stream_close(stream);
3659 }
3660
c0bae11d
MD		 3661 if (new_stream) {
3662 pthread_mutex_lock(&session->lock);
3663 uatomic_set(&session->new_streams, 1);
3664 pthread_mutex_unlock(&session->lock);
3665 }
5312a3ed 3666
7591bab1 3667 stream_put(stream);
b8aa1682 3668end:
5569b118 3669 return status;
b8aa1682
JD		 3670}
3671
5312a3ed
JG		 3672/*
3673 * relay_process_data: Process the data received on the data socket
3674 */
5569b118
JG		 3675static enum relay_connection_status relay_process_data(
3676 struct relay_connection *conn)
5312a3ed 3677{
5569b118 3678 enum relay_connection_status status;
5312a3ed
JG		 3679
3680 switch (conn->protocol.data.state_id) {
3681 case DATA_CONNECTION_STATE_RECEIVE_HEADER:
5569b118 3682 status = relay_process_data_receive_header(conn);
5312a3ed
JG		 3683 break;
3684 case DATA_CONNECTION_STATE_RECEIVE_PAYLOAD:
5569b118 3685 status = relay_process_data_receive_payload(conn);
5312a3ed
JG		 3686 break;
3687 default:
3688 ERR("Unexpected data connection communication state.");
3689 abort();
3690 }
3691
5569b118 3692 return status;
5312a3ed
JG		 3693}
3694
7591bab1 3695static void cleanup_connection_pollfd(struct lttng_poll_event *events, int pollfd)
b8aa1682
JD		 3696{
3697 int ret;
3698
58eb9381 3699 (void) lttng_poll_del(events, pollfd);
b8aa1682
JD		 3700
3701 ret = close(pollfd);
3702 if (ret < 0) {
3703 ERR("Closing pollfd %d", pollfd);
3704 }
3705}
3706
7591bab1
MD		 3707static void relay_thread_close_connection(struct lttng_poll_event *events,
3708 int pollfd, struct relay_connection *conn)
9d1bbf21 3709{
7591bab1 3710 const char *type_str;
2a174661 3711
7591bab1
MD		 3712 switch (conn->type) {
3713 case RELAY_DATA:
3714 type_str = "Data";
3715 break;
3716 case RELAY_CONTROL:
3717 type_str = "Control";
3718 break;
3719 case RELAY_VIEWER_COMMAND:
3720 type_str = "Viewer Command";
3721 break;
3722 case RELAY_VIEWER_NOTIFICATION:
3723 type_str = "Viewer Notification";
3724 break;
3725 default:
3726 type_str = "Unknown";
9d1bbf21 3727 }
7591bab1
MD		 3728 cleanup_connection_pollfd(events, pollfd);
3729 connection_put(conn);
3730 DBG("%s connection closed with %d", type_str, pollfd);
b8aa1682
JD		 3731}
3732
3733/*
3734 * This thread does the actual work
3735 */
7591bab1 3736static void *relay_thread_worker(void *data)
b8aa1682 3737{
beaad64c
DG		 3738 int ret, err = -1, last_seen_data_fd = -1;
3739 uint32_t nb_fd;
b8aa1682
JD		 3740 struct lttng_poll_event events;
3741 struct lttng_ht *relay_connections_ht;
b8aa1682 3742 struct lttng_ht_iter iter;
90e7d72f 3743 struct relay_connection *destroy_conn = NULL;
b8aa1682
JD		 3744
3745 DBG("[thread] Relay worker started");
3746
9d1bbf21
MD		 3747 rcu_register_thread();
3748
55706a7d
MD		 3749 health_register(health_relayd, HEALTH_RELAYD_TYPE_WORKER);
3750
9b5e0863
MD		 3751 if (testpoint(relayd_thread_worker)) {
3752 goto error_testpoint;
3753 }
3754
f385ae0a
MD		 3755 health_code_update();
3756
b8aa1682
JD		 3757 /* table of connections indexed on socket */
3758 relay_connections_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG);
095a4ae5
MD		 3759 if (!relay_connections_ht) {
3760 goto relay_connections_ht_error;
3761 }
b8aa1682 3762
b8aa1682
JD		 3763 ret = create_thread_poll_set(&events, 2);
3764 if (ret < 0) {
3765 goto error_poll_create;
3766 }
3767
58eb9381 3768 ret = lttng_poll_add(&events, relay_conn_pipe[0], LPOLLIN | LPOLLRDHUP);
b8aa1682
JD		 3769 if (ret < 0) {
3770 goto error;
3771 }
3772
beaad64c 3773restart:
b8aa1682 3774 while (1) {
beaad64c
DG		 3775 int idx = -1, i, seen_control = 0, last_notdel_data_fd = -1;
3776
f385ae0a
MD		 3777 health_code_update();
3778
b8aa1682 3779 /* Infinite blocking call, waiting for transmission */
87c1611d 3780 DBG3("Relayd worker thread polling...");
f385ae0a 3781 health_poll_entry();
b8aa1682 3782 ret = lttng_poll_wait(&events, -1);
f385ae0a 3783 health_poll_exit();
b8aa1682
JD		 3784 if (ret < 0) {
3785 /*
3786 * Restart interrupted system call.
3787 */
3788 if (errno == EINTR) {
3789 goto restart;
3790 }
3791 goto error;
3792 }
3793
0d9c5d77
DG		 3794 nb_fd = ret;
3795
beaad64c 3796 /*
7591bab1
MD		 3797 * Process control. The control connection is
3798 * prioritized so we don't starve it with high
3799 * throughput tracing data on the data connection.
beaad64c 3800 */
b8aa1682
JD		 3801 for (i = 0; i < nb_fd; i++) {
3802 /* Fetch once the poll data */
beaad64c
DG		 3803 uint32_t revents = LTTNG_POLL_GETEV(&events, i);
3804 int pollfd = LTTNG_POLL_GETFD(&events, i);
b8aa1682 3805
f385ae0a
MD		 3806 health_code_update();
3807
b8aa1682
JD		 3808 /* Thread quit pipe has been closed. Killing thread. */
3809 ret = check_thread_quit_pipe(pollfd, revents);
3810 if (ret) {
095a4ae5
MD		 3811 err = 0;
3812 goto exit;
b8aa1682
JD		 3813 }
3814
58eb9381
DG		 3815 /* Inspect the relay conn pipe for new connection */
3816 if (pollfd == relay_conn_pipe[0]) {
03e43155 3817 if (revents & LPOLLIN) {
90e7d72f
JG		 3818 struct relay_connection *conn;
3819
58eb9381 3820 ret = lttng_read(relay_conn_pipe[0], &conn, sizeof(conn));
b8aa1682
JD		 3821 if (ret < 0) {
3822 goto error;
3823 }
58eb9381
DG		 3824 lttng_poll_add(&events, conn->sock->fd,
3825 LPOLLIN | LPOLLRDHUP);
7591bab1 3826 connection_ht_add(relay_connections_ht, conn);
58eb9381 3827 DBG("Connection socket %d added", conn->sock->fd);
03e43155
MD		 3828 } else if (revents & (LPOLLERR | LPOLLHUP | LPOLLRDHUP)) {
3829 ERR("Relay connection pipe error");
3830 goto error;
3831 } else {
3832 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
3833 goto error;
b8aa1682 3834 }
58eb9381 3835 } else {
90e7d72f
JG		 3836 struct relay_connection *ctrl_conn;
3837
7591bab1 3838 ctrl_conn = connection_get_by_sock(relay_connections_ht, pollfd);
58eb9381 3839 /* If not found, there is a synchronization issue. */
90e7d72f 3840 assert(ctrl_conn);
58eb9381 3841
03e43155
MD		 3842 if (ctrl_conn->type == RELAY_DATA) {
3843 if (revents & LPOLLIN) {
beaad64c
DG		 3844 /*
3845 * Flag the last seen data fd not deleted. It will be
3846 * used as the last seen fd if any fd gets deleted in
3847 * this first loop.
3848 */
3849 last_notdel_data_fd = pollfd;
3850 }
03e43155
MD		 3851 goto put_ctrl_connection;
3852 }
3853 assert(ctrl_conn->type == RELAY_CONTROL);
3854
3855 if (revents & LPOLLIN) {
5569b118
JG		 3856 enum relay_connection_status status;
3857
3858 status = relay_process_control(ctrl_conn);
3859 if (status != RELAY_CONNECTION_STATUS_OK) {
fd0f1e3e
JR		 3860 /*
3861 * On socket error flag the session as aborted to force
3862 * the cleanup of its stream otherwise it can leak
3863 * during the lifetime of the relayd.
3864 *
3865 * This prevents situations in which streams can be
3866 * left opened because an index was received, the
3867 * control connection is closed, and the data
3868 * connection is closed (uncleanly) before the packet's
3869 * data provided.
3870 *
3871 * Since the control connection encountered an error,
3872 * it is okay to be conservative and close the
3873 * session right now as we can't rely on the protocol
3874 * being respected anymore.
3875 */
3876 if (status == RELAY_CONNECTION_STATUS_ERROR) {
3877 session_abort(ctrl_conn->session);
3878 }
3879
5569b118 3880 /* Clear the connection on error or close. */
5312a3ed
JG		 3881 relay_thread_close_connection(&events,
3882 pollfd,
03e43155 3883 ctrl_conn);
03e43155 3884 }
5312a3ed 3885 seen_control = 1;
03e43155
MD		 3886 } else if (revents & (LPOLLERR | LPOLLHUP | LPOLLRDHUP)) {
3887 relay_thread_close_connection(&events,
3888 pollfd, ctrl_conn);
3889 if (last_seen_data_fd == pollfd) {
3890 last_seen_data_fd = last_notdel_data_fd;
3891 }
58eb9381 3892 } else {
03e43155
MD		 3893 ERR("Unexpected poll events %u for control sock %d",
3894 revents, pollfd);
3895 connection_put(ctrl_conn);
3896 goto error;
beaad64c 3897 }
03e43155 3898 put_ctrl_connection:
7591bab1 3899 connection_put(ctrl_conn);
beaad64c
DG		 3900 }
3901 }
3902
3903 /*
3904 * The last loop handled a control request, go back to poll to make
3905 * sure we prioritise the control socket.
3906 */
3907 if (seen_control) {
3908 continue;
3909 }
3910
3911 if (last_seen_data_fd >= 0) {
3912 for (i = 0; i < nb_fd; i++) {
3913 int pollfd = LTTNG_POLL_GETFD(&events, i);
f385ae0a
MD		 3914
3915 health_code_update();
3916
beaad64c
DG		 3917 if (last_seen_data_fd == pollfd) {
3918 idx = i;
3919 break;
3920 }
3921 }
3922 }
3923
3924 /* Process data connection. */
3925 for (i = idx + 1; i < nb_fd; i++) {
3926 /* Fetch the poll data. */
3927 uint32_t revents = LTTNG_POLL_GETEV(&events, i);
3928 int pollfd = LTTNG_POLL_GETFD(&events, i);
90e7d72f 3929 struct relay_connection *data_conn;
beaad64c 3930
f385ae0a
MD		 3931 health_code_update();
3932
fd20dac9
MD		 3933 if (!revents) {
3934 /* No activity for this FD (poll implementation). */
3935 continue;
3936 }
3937
beaad64c 3938 /* Skip the command pipe. It's handled in the first loop. */
58eb9381 3939 if (pollfd == relay_conn_pipe[0]) {
beaad64c
DG		 3940 continue;
3941 }
3942
7591bab1 3943 data_conn = connection_get_by_sock(relay_connections_ht, pollfd);
90e7d72f 3944 if (!data_conn) {
fd20dac9 3945 /* Skip it. Might be removed before. */
fd20dac9
MD		 3946 continue;
3947 }
03e43155
MD		 3948 if (data_conn->type == RELAY_CONTROL) {
3949 goto put_data_connection;
3950 }
3951 assert(data_conn->type == RELAY_DATA);
fd20dac9
MD		 3952
3953 if (revents & LPOLLIN) {
5569b118
JG		 3954 enum relay_connection_status status;
3955
3956 status = relay_process_data(data_conn);
3957 /* Connection closed or error. */
3958 if (status != RELAY_CONNECTION_STATUS_OK) {
fd0f1e3e
JR		 3959 /*
3960 * On socket error flag the session as aborted to force
3961 * the cleanup of its stream otherwise it can leak
3962 * during the lifetime of the relayd.
3963 *
3964 * This prevents situations in which streams can be
3965 * left opened because an index was received, the
3966 * control connection is closed, and the data
3967 * connection is closed (uncleanly) before the packet's
3968 * data provided.
3969 *
3970 * Since the data connection encountered an error,
3971 * it is okay to be conservative and close the
3972 * session right now as we can't rely on the protocol
3973 * being respected anymore.
3974 */
3975 if (status == RELAY_CONNECTION_STATUS_ERROR) {
3976 session_abort(data_conn->session);
3977 }
7591bab1 3978 relay_thread_close_connection(&events, pollfd,
03e43155 3979 data_conn);
fd20dac9
MD		 3980 /*
3981 * Every goto restart call sets the last seen fd where
3982 * here we don't really care since we gracefully
3983 * continue the loop after the connection is deleted.
3984 */
3985 } else {
3986 /* Keep last seen port. */
3987 last_seen_data_fd = pollfd;
7591bab1 3988 connection_put(data_conn);
fd20dac9 3989 goto restart;
b8aa1682 3990 }
03e43155
MD		 3991 } else if (revents & (LPOLLERR | LPOLLHUP | LPOLLRDHUP)) {
3992 relay_thread_close_connection(&events, pollfd,
3993 data_conn);
3994 } else {
3995 ERR("Unknown poll events %u for data sock %d",
3996 revents, pollfd);
b8aa1682 3997 }
03e43155 3998 put_data_connection:
7591bab1 3999 connection_put(data_conn);
b8aa1682 4000 }
beaad64c 4001 last_seen_data_fd = -1;
b8aa1682
JD		 4002 }
4003
f385ae0a
MD		 4004 /* Normal exit, no error */
4005 ret = 0;
4006
095a4ae5 4007exit:
b8aa1682 4008error:
71efa8ef 4009 /* Cleanup remaining connection object. */
9d1bbf21 4010 rcu_read_lock();
90e7d72f
JG		 4011 cds_lfht_for_each_entry(relay_connections_ht->ht, &iter.iter,
4012 destroy_conn,
58eb9381 4013 sock_n.node) {
f385ae0a 4014 health_code_update();
98ba050e 4015
fd0f1e3e 4016 session_abort(destroy_conn->session);
98ba050e 4017
7591bab1
MD		 4018 /*
4019 * No need to grab another ref, because we own
4020 * destroy_conn.
4021 */
4022 relay_thread_close_connection(&events, destroy_conn->sock->fd,
4023 destroy_conn);
b8aa1682 4024 }
94d49140 4025 rcu_read_unlock();
7591bab1
MD		 4026
4027 lttng_poll_clean(&events);
7d2f7452 4028error_poll_create:
b8aa1682 4029 lttng_ht_destroy(relay_connections_ht);
095a4ae5 4030relay_connections_ht_error:
58eb9381
DG		 4031 /* Close relay conn pipes */
4032 utils_close_pipe(relay_conn_pipe);
095a4ae5
MD		 4033 if (err) {
4034 DBG("Thread exited with error");
4035 }
b8aa1682 4036 DBG("Worker thread cleanup complete");
9b5e0863 4037error_testpoint:
f385ae0a
MD		 4038 if (err) {
4039 health_error();
4040 ERR("Health error occurred in %s", __func__);
4041 }
4042 health_unregister(health_relayd);
9d1bbf21 4043 rcu_unregister_thread();
b4aacfdc 4044 lttng_relay_stop_threads();
b8aa1682
JD		 4045 return NULL;
4046}
4047
4048/*
4049 * Create the relay command pipe to wake thread_manage_apps.
4050 * Closed in cleanup().
4051 */
58eb9381 4052static int create_relay_conn_pipe(void)
b8aa1682 4053{
a02de639 4054 int ret;
b8aa1682 4055
58eb9381 4056 ret = utils_create_pipe_cloexec(relay_conn_pipe);
b8aa1682 4057
b8aa1682
JD		 4058 return ret;
4059}
4060
4061/*
4062 * main
4063 */
4064int main(int argc, char **argv)
4065{
178a0557 4066 int ret = 0, retval = 0;
b8aa1682
JD		 4067 void *status;
4068
b8aa1682
JD		 4069 /* Parse arguments */
4070 progname = argv[0];
178a0557
MD		 4071 if (set_options(argc, argv)) {
4072 retval = -1;
4073 goto exit_options;
b8aa1682
JD		 4074 }
4075
178a0557
MD		 4076 if (set_signal_handler()) {
4077 retval = -1;
4078 goto exit_options;
b8aa1682
JD		 4079 }
4080
4d513a50
DG		 4081 /* Try to create directory if -o, --output is specified. */
4082 if (opt_output_path) {
994fa64f
DG		 4083 if (*opt_output_path != '/') {
4084 ERR("Please specify an absolute path for -o, --output PATH");
178a0557
MD		 4085 retval = -1;
4086 goto exit_options;
994fa64f
DG		 4087 }
4088
d77dded2
JG		 4089 ret = utils_mkdir_recursive(opt_output_path, S_IRWXU | S_IRWXG,
4090 -1, -1);
4d513a50
DG		 4091 if (ret < 0) {
4092 ERR("Unable to create %s", opt_output_path);
178a0557
MD		 4093 retval = -1;
4094 goto exit_options;
4d513a50
DG		 4095 }
4096 }
4097
b8aa1682 4098 /* Daemonize */
b5218ffb 4099 if (opt_daemon || opt_background) {
3fd27398
MD		 4100 int i;
4101
4102 ret = lttng_daemonize(&child_ppid, &recv_child_signal,
4103 !opt_background);
b8aa1682 4104 if (ret < 0) {
178a0557
MD		 4105 retval = -1;
4106 goto exit_options;
b8aa1682 4107 }
3fd27398
MD		 4108
4109 /*
4110 * We are in the child. Make sure all other file
4111 * descriptors are closed, in case we are called with
4112 * more opened file descriptors than the standard ones.
4113 */
4114 for (i = 3; i < sysconf(_SC_OPEN_MAX); i++) {
4115 (void) close(i);
4116 }
4117 }
4118
178a0557
MD		 4119 /* Initialize thread health monitoring */
4120 health_relayd = health_app_create(NR_HEALTH_RELAYD_TYPES);
4121 if (!health_relayd) {
4122 PERROR("health_app_create error");
4123 retval = -1;
4124 goto exit_health_app_create;
4125 }
4126
3fd27398 4127 /* Create thread quit pipe */
178a0557
MD		 4128 if (init_thread_quit_pipe()) {
4129 retval = -1;
4130 goto exit_init_data;
b8aa1682
JD		 4131 }
4132
b8aa1682 4133 /* Setup the thread apps communication pipe. */
178a0557
MD		 4134 if (create_relay_conn_pipe()) {
4135 retval = -1;
4136 goto exit_init_data;
b8aa1682
JD		 4137 }
4138
4139 /* Init relay command queue. */
8bdee6e2 4140 cds_wfcq_init(&relay_conn_queue.head, &relay_conn_queue.tail);
b8aa1682 4141
554831e7
MD		 4142 /* Initialize communication library */
4143 lttcomm_init();
87e45c13 4144 lttcomm_inet_init();
554831e7 4145
d3e2ba59 4146 /* tables of sessions indexed by session ID */
7591bab1
MD		 4147 sessions_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
4148 if (!sessions_ht) {
178a0557
MD		 4149 retval = -1;
4150 goto exit_init_data;
d3e2ba59
JD		 4151 }
4152
4153 /* tables of streams indexed by stream ID */
2a174661 4154 relay_streams_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
d3e2ba59 4155 if (!relay_streams_ht) {
178a0557
MD		 4156 retval = -1;
4157 goto exit_init_data;
d3e2ba59
JD		 4158 }
4159
4160 /* tables of streams indexed by stream ID */
92c6ca54
DG		 4161 viewer_streams_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
4162 if (!viewer_streams_ht) {
178a0557
MD		 4163 retval = -1;
4164 goto exit_init_data;
55706a7d
MD		 4165 }
4166
65931c8b 4167 ret = utils_create_pipe(health_quit_pipe);
178a0557
MD		 4168 if (ret) {
4169 retval = -1;
4170 goto exit_health_quit_pipe;
65931c8b
MD		 4171 }
4172
4173 /* Create thread to manage the client socket */
1a1a34b4 4174 ret = pthread_create(&health_thread, default_pthread_attr(),
65931c8b 4175 thread_manage_health, (void *) NULL);
178a0557
MD		 4176 if (ret) {
4177 errno = ret;
65931c8b 4178 PERROR("pthread_create health");
178a0557
MD		 4179 retval = -1;
4180 goto exit_health_thread;
65931c8b
MD		 4181 }
4182
b8aa1682 4183 /* Setup the dispatcher thread */
1a1a34b4 4184 ret = pthread_create(&dispatcher_thread, default_pthread_attr(),
b8aa1682 4185 relay_thread_dispatcher, (void *) NULL);
178a0557
MD		 4186 if (ret) {
4187 errno = ret;
b8aa1682 4188 PERROR("pthread_create dispatcher");
178a0557
MD		 4189 retval = -1;
4190 goto exit_dispatcher_thread;
b8aa1682
JD		 4191 }
4192
4193 /* Setup the worker thread */
1a1a34b4 4194 ret = pthread_create(&worker_thread, default_pthread_attr(),
7591bab1 4195 relay_thread_worker, NULL);
178a0557
MD		 4196 if (ret) {
4197 errno = ret;
b8aa1682 4198 PERROR("pthread_create worker");
178a0557
MD		 4199 retval = -1;
4200 goto exit_worker_thread;
b8aa1682
JD		 4201 }
4202
4203 /* Setup the listener thread */
1a1a34b4 4204 ret = pthread_create(&listener_thread, default_pthread_attr(),
b8aa1682 4205 relay_thread_listener, (void *) NULL);
178a0557
MD		 4206 if (ret) {
4207 errno = ret;
b8aa1682 4208 PERROR("pthread_create listener");
178a0557
MD		 4209 retval = -1;
4210 goto exit_listener_thread;
b8aa1682
JD		 4211 }
4212
7591bab1 4213 ret = relayd_live_create(live_uri);
178a0557 4214 if (ret) {
d3e2ba59 4215 ERR("Starting live viewer threads");
178a0557 4216 retval = -1;
50138f51 4217 goto exit_live;
d3e2ba59
JD		 4218 }
4219
178a0557
MD		 4220 /*
4221 * This is where we start awaiting program completion (e.g. through
4222 * signal that asks threads to teardown).
4223 */
4224
4225 ret = relayd_live_join();
4226 if (ret) {
4227 retval = -1;
4228 }
50138f51 4229exit_live:
178a0557 4230
b8aa1682 4231 ret = pthread_join(listener_thread, &status);
178a0557
MD		 4232 if (ret) {
4233 errno = ret;
4234 PERROR("pthread_join listener_thread");
4235 retval = -1;
b8aa1682
JD		 4236 }
4237
178a0557 4238exit_listener_thread:
b8aa1682 4239 ret = pthread_join(worker_thread, &status);
178a0557
MD		 4240 if (ret) {
4241 errno = ret;
4242 PERROR("pthread_join worker_thread");
4243 retval = -1;
b8aa1682
JD		 4244 }
4245
178a0557 4246exit_worker_thread:
b8aa1682 4247 ret = pthread_join(dispatcher_thread, &status);
178a0557
MD		 4248 if (ret) {
4249 errno = ret;
4250 PERROR("pthread_join dispatcher_thread");
4251 retval = -1;
b8aa1682 4252 }
178a0557 4253exit_dispatcher_thread:
42415026 4254
65931c8b 4255 ret = pthread_join(health_thread, &status);
178a0557
MD		 4256 if (ret) {
4257 errno = ret;
4258 PERROR("pthread_join health_thread");
4259 retval = -1;
65931c8b 4260 }
178a0557 4261exit_health_thread:
65931c8b 4262
65931c8b 4263 utils_close_pipe(health_quit_pipe);
178a0557 4264exit_health_quit_pipe:
65931c8b 4265
178a0557 4266exit_init_data:
55706a7d 4267 health_app_destroy(health_relayd);
55706a7d 4268exit_health_app_create:
178a0557 4269exit_options:
4d62fbf8
MD		 4270 /*
4271 * Wait for all pending call_rcu work to complete before tearing
4272 * down data structures. call_rcu worker may be trying to
4273 * perform lookups in those structures.
4274 */
4275 rcu_barrier();
7591bab1
MD		 4276 relayd_cleanup();
4277
4278 /* Ensure all prior call_rcu are done. */
4279 rcu_barrier();
d3e2ba59 4280
178a0557 4281 if (!retval) {
b8aa1682 4282 exit(EXIT_SUCCESS);
178a0557
MD		 4283 } else {
4284 exit(EXIT_FAILURE);
b8aa1682 4285 }
b8aa1682 4286}
LTTng 2.0 tools and control repository
This page took 0.292388 seconds and 4 git commands to generate.