projects / lttng-tools.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
Fix: session conditions not evaluated at subscription/registration
[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,
88 /* An error occured while processing an event on the connection. */
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
fd20dac9 861 if (!revents) {
7591bab1
MD		 862 /*
863 * No activity for this FD (poll
864 * implementation).
865 */
fd20dac9
MD		 866 continue;
867 }
868
b8aa1682
JD		 869 /* Thread quit pipe has been closed. Killing thread. */
870 ret = check_thread_quit_pipe(pollfd, revents);
871 if (ret) {
095a4ae5
MD		 872 err = 0;
873 goto exit;
b8aa1682
JD		 874 }
875
03e43155 876 if (revents & LPOLLIN) {
4b7f17b2 877 /*
7591bab1
MD		 878 * A new connection is requested, therefore a
879 * sessiond/consumerd connection is allocated in
880 * this thread, enqueued to a global queue and
881 * dequeued (and freed) in the worker thread.
4b7f17b2 882 */
58eb9381
DG		 883 int val = 1;
884 struct relay_connection *new_conn;
4b7f17b2 885 struct lttcomm_sock *newsock;
7591bab1 886 enum connection_type type;
b8aa1682
JD		 887
888 if (pollfd == data_sock->fd) {
7591bab1 889 type = RELAY_DATA;
b8aa1682 890 newsock = data_sock->ops->accept(data_sock);
58eb9381
DG		 891 DBG("Relay data connection accepted, socket %d",
892 newsock->fd);
4b7f17b2
MD		 893 } else {
894 assert(pollfd == control_sock->fd);
7591bab1 895 type = RELAY_CONTROL;
b8aa1682 896 newsock = control_sock->ops->accept(control_sock);
58eb9381
DG		 897 DBG("Relay control connection accepted, socket %d",
898 newsock->fd);
b8aa1682 899 }
58eb9381
DG		 900 if (!newsock) {
901 PERROR("accepting sock");
58eb9381
DG		 902 goto error;
903 }
904
905 ret = setsockopt(newsock->fd, SOL_SOCKET, SO_REUSEADDR, &val,
906 sizeof(val));
b8aa1682
JD		 907 if (ret < 0) {
908 PERROR("setsockopt inet");
4b7f17b2 909 lttcomm_destroy_sock(newsock);
b8aa1682
JD		 910 goto error;
911 }
f056029c
JR		 912
913 ret = socket_apply_keep_alive_config(newsock->fd);
914 if (ret < 0) {
915 ERR("Failed to apply TCP keep-alive configuration on socket (%i)",
916 newsock->fd);
917 lttcomm_destroy_sock(newsock);
918 goto error;
919 }
920
7591bab1
MD		 921 new_conn = connection_create(newsock, type);
922 if (!new_conn) {
923 lttcomm_destroy_sock(newsock);
924 goto error;
925 }
58eb9381
DG		 926
927 /* Enqueue request for the dispatcher thread. */
8bdee6e2
SM		 928 cds_wfcq_enqueue(&relay_conn_queue.head, &relay_conn_queue.tail,
929 &new_conn->qnode);
b8aa1682
JD		 930
931 /*
7591bab1
MD		 932 * Wake the dispatch queue futex.
933 * Implicit memory barrier with the
934 * exchange in cds_wfcq_enqueue.
b8aa1682 935 */
58eb9381 936 futex_nto1_wake(&relay_conn_queue.futex);
03e43155
MD		 937 } else if (revents & (LPOLLERR | LPOLLHUP | LPOLLRDHUP)) {
938 ERR("socket poll error");
939 goto error;
940 } else {
941 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
942 goto error;
b8aa1682
JD		 943 }
944 }
945 }
946
095a4ae5 947exit:
b8aa1682
JD		 948error:
949error_poll_add:
9b5e0863 950error_testpoint:
b8aa1682
JD		 951 lttng_poll_clean(&events);
952error_create_poll:
095a4ae5
MD		 953 if (data_sock->fd >= 0) {
954 ret = data_sock->ops->close(data_sock);
b8aa1682
JD		 955 if (ret) {
956 PERROR("close");
957 }
b8aa1682 958 }
095a4ae5
MD		 959 lttcomm_destroy_sock(data_sock);
960error_sock_relay:
961 if (control_sock->fd >= 0) {
962 ret = control_sock->ops->close(control_sock);
b8aa1682
JD		 963 if (ret) {
964 PERROR("close");
965 }
b8aa1682 966 }
095a4ae5
MD		 967 lttcomm_destroy_sock(control_sock);
968error_sock_control:
969 if (err) {
f385ae0a
MD		 970 health_error();
971 ERR("Health error occurred in %s", __func__);
095a4ae5 972 }
55706a7d 973 health_unregister(health_relayd);
b8aa1682 974 DBG("Relay listener thread cleanup complete");
b4aacfdc 975 lttng_relay_stop_threads();
b8aa1682
JD		 976 return NULL;
977}
978
979/*
980 * This thread manages the dispatching of the requests to worker threads
981 */
7591bab1 982static void *relay_thread_dispatcher(void *data)
b8aa1682 983{
6cd525e8
MD		 984 int err = -1;
985 ssize_t ret;
8bdee6e2 986 struct cds_wfcq_node *node;
58eb9381 987 struct relay_connection *new_conn = NULL;
b8aa1682
JD		 988
989 DBG("[thread] Relay dispatcher started");
990
55706a7d
MD		 991 health_register(health_relayd, HEALTH_RELAYD_TYPE_DISPATCHER);
992
9b5e0863
MD		 993 if (testpoint(relayd_thread_dispatcher)) {
994 goto error_testpoint;
995 }
996
f385ae0a
MD		 997 health_code_update();
998
0ed3b1a8 999 for (;;) {
f385ae0a
MD		 1000 health_code_update();
1001
b8aa1682 1002 /* Atomically prepare the queue futex */
58eb9381 1003 futex_nto1_prepare(&relay_conn_queue.futex);
b8aa1682 1004
0ed3b1a8
MD		 1005 if (CMM_LOAD_SHARED(dispatch_thread_exit)) {
1006 break;
1007 }
1008
b8aa1682 1009 do {
f385ae0a
MD		 1010 health_code_update();
1011
b8aa1682 1012 /* Dequeue commands */
8bdee6e2
SM		 1013 node = cds_wfcq_dequeue_blocking(&relay_conn_queue.head,
1014 &relay_conn_queue.tail);
b8aa1682
JD		 1015 if (node == NULL) {
1016 DBG("Woken up but nothing in the relay command queue");
1017 /* Continue thread execution */
1018 break;
1019 }
58eb9381 1020 new_conn = caa_container_of(node, struct relay_connection, qnode);
b8aa1682 1021
58eb9381 1022 DBG("Dispatching request waiting on sock %d", new_conn->sock->fd);
b8aa1682
JD		 1023
1024 /*
7591bab1
MD		 1025 * Inform worker thread of the new request. This
1026 * call is blocking so we can be assured that
1027 * the data will be read at some point in time
1028 * or wait to the end of the world :)
b8aa1682 1029 */
58eb9381
DG		 1030 ret = lttng_write(relay_conn_pipe[1], &new_conn, sizeof(new_conn));
1031 if (ret < 0) {
1032 PERROR("write connection pipe");
7591bab1 1033 connection_put(new_conn);
b8aa1682
JD		 1034 goto error;
1035 }
1036 } while (node != NULL);
1037
1038 /* Futex wait on queue. Blocking call on futex() */
f385ae0a 1039 health_poll_entry();
58eb9381 1040 futex_nto1_wait(&relay_conn_queue.futex);
f385ae0a 1041 health_poll_exit();
b8aa1682
JD		 1042 }
1043
f385ae0a
MD		 1044 /* Normal exit, no error */
1045 err = 0;
1046
b8aa1682 1047error:
9b5e0863 1048error_testpoint:
f385ae0a
MD		 1049 if (err) {
1050 health_error();
1051 ERR("Health error occurred in %s", __func__);
1052 }
55706a7d 1053 health_unregister(health_relayd);
b8aa1682 1054 DBG("Dispatch thread dying");
b4aacfdc 1055 lttng_relay_stop_threads();
b8aa1682
JD		 1056 return NULL;
1057}
1058
b8aa1682 1059/*
7591bab1 1060 * Set index data from the control port to a given index object.
b8aa1682 1061 */
7591bab1 1062static int set_index_control_data(struct relay_index *index,
234cd636
JD		 1063 struct lttcomm_relayd_index *data,
1064 struct relay_connection *conn)
1c20f0e2 1065{
7591bab1 1066 struct ctf_packet_index index_data;
1c20f0e2
JD		 1067
1068 /*
5312a3ed 1069 * The index on disk is encoded in big endian.
1c20f0e2 1070 */
5312a3ed
JG		 1071 index_data.packet_size = htobe64(data->packet_size);
1072 index_data.content_size = htobe64(data->content_size);
1073 index_data.timestamp_begin = htobe64(data->timestamp_begin);
1074 index_data.timestamp_end = htobe64(data->timestamp_end);
1075 index_data.events_discarded = htobe64(data->events_discarded);
1076 index_data.stream_id = htobe64(data->stream_id);
234cd636
JD		 1077
1078 if (conn->minor >= 8) {
5312a3ed
JG		 1079 index->index_data.stream_instance_id = htobe64(data->stream_instance_id);
1080 index->index_data.packet_seq_num = htobe64(data->packet_seq_num);
234cd636
JD		 1081 }
1082
7591bab1 1083 return relay_index_set_data(index, &index_data);
1c20f0e2
JD		 1084}
1085
c5b6f4f0
DG		 1086/*
1087 * Handle the RELAYD_CREATE_SESSION command.
1088 *
1089 * On success, send back the session id or else return a negative value.
1090 */
5312a3ed
JG		 1091static int relay_create_session(const struct lttcomm_relayd_hdr *recv_hdr,
1092 struct relay_connection *conn,
1093 const struct lttng_buffer_view *payload)
c5b6f4f0 1094{
5312a3ed
JG		 1095 int ret = 0;
1096 ssize_t send_ret;
c5b6f4f0
DG		 1097 struct relay_session *session;
1098 struct lttcomm_relayd_status_session reply;
36d2e35d 1099 char session_name[LTTNG_NAME_MAX];
9cf3eb20 1100 char hostname[LTTNG_HOST_NAME_MAX];
7591bab1
MD		 1101 uint32_t live_timer = 0;
1102 bool snapshot = false;
c5b6f4f0 1103
36d2e35d 1104 memset(session_name, 0, LTTNG_NAME_MAX);
9cf3eb20 1105 memset(hostname, 0, LTTNG_HOST_NAME_MAX);
c5b6f4f0
DG		 1106
1107 memset(&reply, 0, sizeof(reply));
1108
f86f6389
JR		 1109 if (conn->minor < 4) {
1110 /* From 2.1 to 2.3 */
1111 ret = 0;
1112 } else if (conn->minor >= 4 && conn->minor < 11) {
1113 /* From 2.4 to 2.10 */
5312a3ed 1114 ret = cmd_create_session_2_4(payload, session_name,
7591bab1 1115 hostname, &live_timer, &snapshot);
f86f6389
JR		 1116 } else {
1117 /* From 2.11 to ... */
1118 ret = cmd_create_session_2_11(payload, session_name,
1119 hostname, &live_timer, &snapshot);
7591bab1 1120 }
f86f6389 1121
7591bab1
MD		 1122 if (ret < 0) {
1123 goto send_reply;
d3e2ba59
JD		 1124 }
1125
7591bab1
MD		 1126 session = session_create(session_name, hostname, live_timer,
1127 snapshot, conn->major, conn->minor);
1128 if (!session) {
1129 ret = -1;
1130 goto send_reply;
1131 }
1132 assert(!conn->session);
1133 conn->session = session;
c5b6f4f0
DG		 1134 DBG("Created session %" PRIu64, session->id);
1135
7591bab1
MD		 1136 reply.session_id = htobe64(session->id);
1137
1138send_reply:
c5b6f4f0
DG		 1139 if (ret < 0) {
1140 reply.ret_code = htobe32(LTTNG_ERR_FATAL);
1141 } else {
1142 reply.ret_code = htobe32(LTTNG_OK);
1143 }
1144
58eb9381 1145 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply, sizeof(reply), 0);
5312a3ed
JG		 1146 if (send_ret < (ssize_t) sizeof(reply)) {
1147 ERR("Failed to send \"create session\" command reply (ret = %zd)",
1148 send_ret);
1149 ret = -1;
c5b6f4f0
DG		 1150 }
1151
1152 return ret;
1153}
1154
a4baae1b
JD		 1155/*
1156 * When we have received all the streams and the metadata for a channel,
1157 * we make them visible to the viewer threads.
1158 */
7591bab1 1159static void publish_connection_local_streams(struct relay_connection *conn)
a4baae1b 1160{
7591bab1
MD		 1161 struct relay_stream *stream;
1162 struct relay_session *session = conn->session;
a4baae1b 1163
7591bab1
MD		 1164 /*
1165 * We publish all streams belonging to a session atomically wrt
1166 * session lock.
1167 */
1168 pthread_mutex_lock(&session->lock);
1169 rcu_read_lock();
1170 cds_list_for_each_entry_rcu(stream, &session->recv_list,
1171 recv_node) {
1172 stream_publish(stream);
a4baae1b 1173 }
7591bab1 1174 rcu_read_unlock();
a4baae1b 1175
7591bab1
MD		 1176 /*
1177 * Inform the viewer that there are new streams in the session.
1178 */
1179 if (session->viewer_attached) {
1180 uatomic_set(&session->new_streams, 1);
1181 }
1182 pthread_mutex_unlock(&session->lock);
a4baae1b
JD		 1183}
1184
b8aa1682
JD		 1185/*
1186 * relay_add_stream: allocate a new stream for a session
1187 */
5312a3ed
JG		 1188static int relay_add_stream(const struct lttcomm_relayd_hdr *recv_hdr,
1189 struct relay_connection *conn,
1190 const struct lttng_buffer_view *payload)
b8aa1682 1191{
7591bab1
MD		 1192 int ret;
1193 ssize_t send_ret;
58eb9381 1194 struct relay_session *session = conn->session;
b8aa1682
JD		 1195 struct relay_stream *stream = NULL;
1196 struct lttcomm_relayd_status_stream reply;
4030a636 1197 struct ctf_trace *trace = NULL;
7591bab1
MD		 1198 uint64_t stream_handle = -1ULL;
1199 char *path_name = NULL, *channel_name = NULL;
1200 uint64_t tracefile_size = 0, tracefile_count = 0;
81164b6b 1201 struct relay_stream_chunk_id stream_chunk_id = { 0 };
b8aa1682 1202
5312a3ed 1203 if (!session || !conn->version_check_done) {
b8aa1682
JD		 1204 ERR("Trying to add a stream before version check");
1205 ret = -1;
1206 goto end_no_session;
1207 }
1208
2f21a469
JR		 1209 if (session->minor == 1) {
1210 /* For 2.1 */
5312a3ed 1211 ret = cmd_recv_stream_2_1(payload, &path_name,
7591bab1 1212 &channel_name);
2f21a469
JR		 1213 } else if (session->minor > 1 && session->minor < 11) {
1214 /* From 2.2 to 2.10 */
5312a3ed 1215 ret = cmd_recv_stream_2_2(payload, &path_name,
7591bab1 1216 &channel_name, &tracefile_size, &tracefile_count);
2f21a469
JR		 1217 } else {
1218 /* From 2.11 to ... */
1219 ret = cmd_recv_stream_2_11(payload, &path_name,
0b50e4b3
JG		 1220 &channel_name, &tracefile_size, &tracefile_count,
1221 &stream_chunk_id.value);
1222 stream_chunk_id.is_set = true;
0f907de1 1223 }
2f21a469 1224
0f907de1 1225 if (ret < 0) {
7591bab1 1226 goto send_reply;
b8aa1682
JD		 1227 }
1228
7591bab1 1229 trace = ctf_trace_get_by_path_or_create(session, path_name);
2a174661 1230 if (!trace) {
7591bab1 1231 goto send_reply;
2a174661 1232 }
7591bab1 1233 /* This stream here has one reference on the trace. */
2a174661 1234
7591bab1
MD		 1235 pthread_mutex_lock(&last_relay_stream_id_lock);
1236 stream_handle = ++last_relay_stream_id;
1237 pthread_mutex_unlock(&last_relay_stream_id_lock);
d3e2ba59 1238
7591bab1
MD		 1239 /* We pass ownership of path_name and channel_name. */
1240 stream = stream_create(trace, stream_handle, path_name,
81164b6b
JG		 1241 channel_name, tracefile_size, tracefile_count,
1242 &stream_chunk_id);
7591bab1
MD		 1243 path_name = NULL;
1244 channel_name = NULL;
a4baae1b 1245
2a174661 1246 /*
7591bab1
MD		 1247 * Streams are the owners of their trace. Reference to trace is
1248 * kept within stream_create().
2a174661 1249 */
7591bab1 1250 ctf_trace_put(trace);
d3e2ba59 1251
7591bab1 1252send_reply:
53efb85a 1253 memset(&reply, 0, sizeof(reply));
7591bab1
MD		 1254 reply.handle = htobe64(stream_handle);
1255 if (!stream) {
f73fabfd 1256 reply.ret_code = htobe32(LTTNG_ERR_UNK);
b8aa1682 1257 } else {
f73fabfd 1258 reply.ret_code = htobe32(LTTNG_OK);
b8aa1682 1259 }
5af40280 1260
58eb9381 1261 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply,
b8aa1682 1262 sizeof(struct lttcomm_relayd_status_stream), 0);
5312a3ed
JG		 1263 if (send_ret < (ssize_t) sizeof(reply)) {
1264 ERR("Failed to send \"add stream\" command reply (ret = %zd)",
1265 send_ret);
1266 ret = -1;
b8aa1682
JD		 1267 }
1268
1269end_no_session:
7591bab1
MD		 1270 free(path_name);
1271 free(channel_name);
0f907de1 1272 return ret;
b8aa1682
JD		 1273}
1274
173af62f
DG		 1275/*
1276 * relay_close_stream: close a specific stream
1277 */
5312a3ed
JG		 1278static int relay_close_stream(const struct lttcomm_relayd_hdr *recv_hdr,
1279 struct relay_connection *conn,
1280 const struct lttng_buffer_view *payload)
173af62f 1281{
5312a3ed
JG		 1282 int ret;
1283 ssize_t send_ret;
58eb9381 1284 struct relay_session *session = conn->session;
173af62f
DG		 1285 struct lttcomm_relayd_close_stream stream_info;
1286 struct lttcomm_relayd_generic_reply reply;
1287 struct relay_stream *stream;
173af62f
DG		 1288
1289 DBG("Close stream received");
1290
5312a3ed 1291 if (!session || !conn->version_check_done) {
173af62f
DG		 1292 ERR("Trying to close a stream before version check");
1293 ret = -1;
1294 goto end_no_session;
1295 }
1296
5312a3ed
JG		 1297 if (payload->size < sizeof(stream_info)) {
1298 ERR("Unexpected payload size in \"relay_close_stream\": expected >= %zu bytes, got %zu bytes",
1299 sizeof(stream_info), payload->size);
173af62f
DG		 1300 ret = -1;
1301 goto end_no_session;
1302 }
5312a3ed
JG		 1303 memcpy(&stream_info, payload->data, sizeof(stream_info));
1304 stream_info.stream_id = be64toh(stream_info.stream_id);
1305 stream_info.last_net_seq_num = be64toh(stream_info.last_net_seq_num);
173af62f 1306
5312a3ed 1307 stream = stream_get_by_id(stream_info.stream_id);
173af62f
DG		 1308 if (!stream) {
1309 ret = -1;
7591bab1 1310 goto end;
173af62f 1311 }
77f7bd85
MD		 1312
1313 /*
1314 * Set last_net_seq_num before the close flag. Required by data
1315 * pending check.
1316 */
7591bab1 1317 pthread_mutex_lock(&stream->lock);
5312a3ed 1318 stream->last_net_seq_num = stream_info.last_net_seq_num;
77f7bd85
MD		 1319 pthread_mutex_unlock(&stream->lock);
1320
bda7c7b9
JG		 1321 /*
1322 * This is one of the conditions which may trigger a stream close
1323 * with the others being:
1324 * 1) A close command is received for a stream
1325 * 2) The control connection owning the stream is closed
1326 * 3) We have received all of the stream's data _after_ a close
1327 * request.
1328 */
1329 try_stream_close(stream);
7591bab1
MD		 1330 if (stream->is_metadata) {
1331 struct relay_viewer_stream *vstream;
173af62f 1332
7591bab1
MD		 1333 vstream = viewer_stream_get_by_id(stream->stream_handle);
1334 if (vstream) {
1335 if (vstream->metadata_sent == stream->metadata_received) {
1336 /*
1337 * Since all the metadata has been sent to the
1338 * viewer and that we have a request to close
1339 * its stream, we can safely teardown the
1340 * corresponding metadata viewer stream.
1341 */
1342 viewer_stream_put(vstream);
1343 }
1344 /* Put local reference. */
1345 viewer_stream_put(vstream);
1346 }
1347 }
7591bab1 1348 stream_put(stream);
5312a3ed 1349 ret = 0;
173af62f 1350
7591bab1 1351end:
53efb85a 1352 memset(&reply, 0, sizeof(reply));
173af62f 1353 if (ret < 0) {
f73fabfd 1354 reply.ret_code = htobe32(LTTNG_ERR_UNK);
173af62f 1355 } else {
f73fabfd 1356 reply.ret_code = htobe32(LTTNG_OK);
173af62f 1357 }
58eb9381 1358 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply,
173af62f 1359 sizeof(struct lttcomm_relayd_generic_reply), 0);
5312a3ed
JG		 1360 if (send_ret < (ssize_t) sizeof(reply)) {
1361 ERR("Failed to send \"close stream\" command reply (ret = %zd)",
1362 send_ret);
1363 ret = -1;
173af62f
DG		 1364 }
1365
1366end_no_session:
1367 return ret;
1368}
1369
93ec662e
JD		 1370/*
1371 * relay_reset_metadata: reset a metadata stream
1372 */
1373static
5312a3ed
JG		 1374int relay_reset_metadata(const struct lttcomm_relayd_hdr *recv_hdr,
1375 struct relay_connection *conn,
1376 const struct lttng_buffer_view *payload)
93ec662e 1377{
5312a3ed
JG		 1378 int ret;
1379 ssize_t send_ret;
93ec662e
JD		 1380 struct relay_session *session = conn->session;
1381 struct lttcomm_relayd_reset_metadata stream_info;
1382 struct lttcomm_relayd_generic_reply reply;
1383 struct relay_stream *stream;
1384
1385 DBG("Reset metadata received");
1386
5312a3ed 1387 if (!session || !conn->version_check_done) {
93ec662e
JD		 1388 ERR("Trying to reset a metadata stream before version check");
1389 ret = -1;
1390 goto end_no_session;
1391 }
1392
5312a3ed
JG		 1393 if (payload->size < sizeof(stream_info)) {
1394 ERR("Unexpected payload size in \"relay_reset_metadata\": expected >= %zu bytes, got %zu bytes",
1395 sizeof(stream_info), payload->size);
93ec662e
JD		 1396 ret = -1;
1397 goto end_no_session;
1398 }
5312a3ed
JG		 1399 memcpy(&stream_info, payload->data, sizeof(stream_info));
1400 stream_info.stream_id = be64toh(stream_info.stream_id);
1401 stream_info.version = be64toh(stream_info.version);
1402
1403 DBG("Update metadata to version %" PRIu64, stream_info.version);
93ec662e
JD		 1404
1405 /* Unsupported for live sessions for now. */
1406 if (session->live_timer != 0) {
1407 ret = -1;
1408 goto end;
1409 }
1410
5312a3ed 1411 stream = stream_get_by_id(stream_info.stream_id);
93ec662e
JD		 1412 if (!stream) {
1413 ret = -1;
1414 goto end;
1415 }
1416 pthread_mutex_lock(&stream->lock);
1417 if (!stream->is_metadata) {
1418 ret = -1;
1419 goto end_unlock;
1420 }
1421
1422 ret = utils_rotate_stream_file(stream->path_name, stream->channel_name,
1423 0, 0, -1, -1, stream->stream_fd->fd, NULL,
1424 &stream->stream_fd->fd);
1425 if (ret < 0) {
1426 ERR("Failed to rotate metadata file %s of channel %s",
1427 stream->path_name, stream->channel_name);
1428 goto end_unlock;
1429 }
1430
1431end_unlock:
1432 pthread_mutex_unlock(&stream->lock);
1433 stream_put(stream);
1434
1435end:
1436 memset(&reply, 0, sizeof(reply));
1437 if (ret < 0) {
1438 reply.ret_code = htobe32(LTTNG_ERR_UNK);
1439 } else {
1440 reply.ret_code = htobe32(LTTNG_OK);
1441 }
1442 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply,
1443 sizeof(struct lttcomm_relayd_generic_reply), 0);
5312a3ed
JG		 1444 if (send_ret < (ssize_t) sizeof(reply)) {
1445 ERR("Failed to send \"reset metadata\" command reply (ret = %zd)",
1446 send_ret);
1447 ret = -1;
93ec662e
JD		 1448 }
1449
1450end_no_session:
1451 return ret;
1452}
1453
b8aa1682
JD		 1454/*
1455 * relay_unknown_command: send -1 if received unknown command
1456 */
7591bab1 1457static void relay_unknown_command(struct relay_connection *conn)
b8aa1682
JD		 1458{
1459 struct lttcomm_relayd_generic_reply reply;
5312a3ed 1460 ssize_t send_ret;
b8aa1682 1461
53efb85a 1462 memset(&reply, 0, sizeof(reply));
f73fabfd 1463 reply.ret_code = htobe32(LTTNG_ERR_UNK);
5312a3ed
JG		 1464 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply, sizeof(reply), 0);
1465 if (send_ret < sizeof(reply)) {
1466 ERR("Failed to send \"unknown command\" command reply (ret = %zd)", send_ret);
b8aa1682
JD		 1467 }
1468}
1469
1470/*
1471 * relay_start: send an acknowledgment to the client to tell if we are
1472 * ready to receive data. We are ready if a session is established.
1473 */
5312a3ed
JG		 1474static int relay_start(const struct lttcomm_relayd_hdr *recv_hdr,
1475 struct relay_connection *conn,
1476 const struct lttng_buffer_view *payload)
b8aa1682 1477{
5312a3ed
JG		 1478 int ret = 0;
1479 ssize_t send_ret;
b8aa1682 1480 struct lttcomm_relayd_generic_reply reply;
58eb9381 1481 struct relay_session *session = conn->session;
b8aa1682
JD		 1482
1483 if (!session) {
1484 DBG("Trying to start the streaming without a session established");
f73fabfd 1485 ret = htobe32(LTTNG_ERR_UNK);
b8aa1682
JD		 1486 }
1487
53efb85a 1488 memset(&reply, 0, sizeof(reply));
5312a3ed
JG		 1489 reply.ret_code = htobe32(LTTNG_OK);
1490 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply,
1491 sizeof(reply), 0);
1492 if (send_ret < (ssize_t) sizeof(reply)) {
1493 ERR("Failed to send \"relay_start\" command reply (ret = %zd)",
1494 send_ret);
1495 ret = -1;
b8aa1682
JD		 1496 }
1497
1498 return ret;
1499}
1500
1d4dfdef
DG		 1501/*
1502 * Append padding to the file pointed by the file descriptor fd.
1503 */
1504static int write_padding_to_file(int fd, uint32_t size)
1505{
6cd525e8 1506 ssize_t ret = 0;
1d4dfdef
DG		 1507 char *zeros;
1508
1509 if (size == 0) {
1510 goto end;
1511 }
1512
1513 zeros = zmalloc(size);
1514 if (zeros == NULL) {
1515 PERROR("zmalloc zeros for padding");
1516 ret = -1;
1517 goto end;
1518 }
1519
6cd525e8
MD		 1520 ret = lttng_write(fd, zeros, size);
1521 if (ret < size) {
1d4dfdef
DG		 1522 PERROR("write padding to file");
1523 }
1524
e986c7a1
DG		 1525 free(zeros);
1526
1d4dfdef
DG		 1527end:
1528 return ret;
1529}
1530
d3ecc550
JD		 1531/*
1532 * Close the current index file if it is open, and create a new one.
1533 *
1534 * Return 0 on success, -1 on error.
1535 */
1536static
1537int create_rotate_index_file(struct relay_stream *stream)
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;
1549 stream->index_file = lttng_index_file_create(stream->path_name,
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
1567int do_rotate_stream(struct relay_stream *stream)
1568{
1569 int ret;
1570
1571 /* Perform the stream rotation. */
1572 ret = utils_rotate_stream_file(stream->path_name,
1573 stream->channel_name, stream->tracefile_size,
1574 stream->tracefile_count, -1,
1575 -1, stream->stream_fd->fd,
1576 NULL, &stream->stream_fd->fd);
1577 if (ret < 0) {
1578 ERR("Rotating stream output file");
1579 goto end;
1580 }
1581 stream->tracefile_size_current = 0;
1582
1583 /* Rotate also the index if the stream is not a metadata stream. */
1584 if (!stream->is_metadata) {
1585 ret = create_rotate_index_file(stream);
1586 if (ret < 0) {
1587 ERR("Failed to rotate index file");
1588 goto end;
1589 }
1590 }
1591
1592 stream->rotate_at_seq_num = -1ULL;
1593 stream->pos_after_last_complete_data_index = 0;
1594
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
1606 * before the rotation command on the control connection arrives. We don't need
1607 * to update the index because its order is guaranteed with the rotation
1608 * command message.
1609 */
1610static
1611int rotate_truncate_stream(struct relay_stream *stream)
1612{
1613 int ret, new_fd;
a5df8828 1614 off_t lseek_ret;
d3ecc550
JD		 1615 uint64_t diff, pos = 0;
1616 char buf[FILE_COPY_BUFFER_SIZE];
1617
1618 assert(!stream->is_metadata);
1619
1620 assert(stream->tracefile_size_current >
1621 stream->pos_after_last_complete_data_index);
1622 diff = stream->tracefile_size_current -
1623 stream->pos_after_last_complete_data_index;
1624
1625 /* Create the new tracefile. */
1626 new_fd = utils_create_stream_file(stream->path_name,
1627 stream->channel_name,
1628 stream->tracefile_size, stream->tracefile_count,
1629 /* uid */ -1, /* gid */ -1, /* suffix */ NULL);
1630 if (new_fd < 0) {
1631 ERR("Failed to create new stream file at path %s for channel %s",
1632 stream->path_name, stream->channel_name);
1633 ret = -1;
1634 goto end;
1635 }
1636
1637 /*
1638 * Rewind the current tracefile to the position at which the rotation
1639 * should have occured.
1640 */
a5df8828 1641 lseek_ret = lseek(stream->stream_fd->fd,
d3ecc550 1642 stream->pos_after_last_complete_data_index, SEEK_SET);
a5df8828 1643 if (lseek_ret < 0) {
d3ecc550 1644 PERROR("seek truncate stream");
a5df8828 1645 ret = -1;
d3ecc550
JD		 1646 goto end;
1647 }
1648
1649 /* Move data from the old file to the new file. */
1650 while (pos < diff) {
1651 uint64_t count, bytes_left;
1652 ssize_t io_ret;
1653
1654 bytes_left = diff - pos;
1655 count = bytes_left > sizeof(buf) ? sizeof(buf) : bytes_left;
1656 assert(count <= SIZE_MAX);
1657
1658 io_ret = lttng_read(stream->stream_fd->fd, buf, count);
1659 if (io_ret < (ssize_t) count) {
1660 char error_string[256];
1661
1662 snprintf(error_string, sizeof(error_string),
1663 "Failed to read %" PRIu64 " bytes from fd %i in rotate_truncate_stream(), returned %zi",
1664 count, stream->stream_fd->fd, io_ret);
1665 if (io_ret == -1) {
1666 PERROR("%s", error_string);
1667 } else {
1668 ERR("%s", error_string);
1669 }
1670 ret = -1;
1671 goto end;
1672 }
1673
1674 io_ret = lttng_write(new_fd, buf, count);
1675 if (io_ret < (ssize_t) count) {
1676 char error_string[256];
1677
1678 snprintf(error_string, sizeof(error_string),
1679 "Failed to write %" PRIu64 " bytes from fd %i in rotate_truncate_stream(), returned %zi",
1680 count, new_fd, io_ret);
1681 if (io_ret == -1) {
1682 PERROR("%s", error_string);
1683 } else {
1684 ERR("%s", error_string);
1685 }
1686 ret = -1;
1687 goto end;
1688 }
1689
1690 pos += count;
1691 }
1692
1693 /* Truncate the file to get rid of the excess data. */
1694 ret = ftruncate(stream->stream_fd->fd,
1695 stream->pos_after_last_complete_data_index);
1696 if (ret) {
1697 PERROR("ftruncate");
1698 goto end;
1699 }
1700
1701 ret = close(stream->stream_fd->fd);
1702 if (ret < 0) {
1703 PERROR("Closing tracefile");
1704 goto end;
1705 }
1706
1707 ret = create_rotate_index_file(stream);
1708 if (ret < 0) {
1709 ERR("Rotate stream index file");
1710 goto end;
1711 }
1712
1713 /*
1714 * Update the offset and FD of all the eventual indexes created by the
1715 * data connection before the rotation command arrived.
1716 */
1717 ret = relay_index_switch_all_files(stream);
1718 if (ret < 0) {
1719 ERR("Failed to rotate index file");
1720 goto end;
1721 }
1722
1723 stream->stream_fd->fd = new_fd;
1724 stream->tracefile_size_current = diff;
1725 stream->pos_after_last_complete_data_index = 0;
1726 stream->rotate_at_seq_num = -1ULL;
1727
1728 ret = 0;
1729
1730end:
1731 return ret;
1732}
1733
1734/*
1735 * Check if a stream should perform a rotation (for session rotation).
1736 * Must be called with the stream lock held.
1737 *
1738 * Return 0 on success, a negative value on error.
1739 */
1740static
1741int try_rotate_stream(struct relay_stream *stream)
1742{
1743 int ret = 0;
1744
1745 /* No rotation expected. */
1746 if (stream->rotate_at_seq_num == -1ULL) {
1747 goto end;
1748 }
1749
3765c8cc
JG		 1750 if (stream->prev_seq < stream->rotate_at_seq_num ||
1751 stream->prev_seq == -1ULL) {
d3ecc550
JD		 1752 DBG("Stream %" PRIu64 " no yet ready for rotation",
1753 stream->stream_handle);
1754 goto end;
1755 } else if (stream->prev_seq > stream->rotate_at_seq_num) {
1756 DBG("Rotation after too much data has been written in tracefile "
1757 "for stream %" PRIu64 ", need to truncate before "
1758 "rotating", stream->stream_handle);
1759 ret = rotate_truncate_stream(stream);
1760 if (ret) {
1761 ERR("Failed to truncate stream");
1762 goto end;
1763 }
1764 } else {
1765 /* stream->prev_seq == stream->rotate_at_seq_num */
1766 DBG("Stream %" PRIu64 " ready for rotation",
1767 stream->stream_handle);
1768 ret = do_rotate_stream(stream);
1769 }
1770
1771end:
1772 return ret;
1773}
1774
b8aa1682 1775/*
7591bab1 1776 * relay_recv_metadata: receive the metadata for the session.
b8aa1682 1777 */
5312a3ed
JG		 1778static int relay_recv_metadata(const struct lttcomm_relayd_hdr *recv_hdr,
1779 struct relay_connection *conn,
1780 const struct lttng_buffer_view *payload)
b8aa1682 1781{
32d1569c 1782 int ret = 0;
6cd525e8 1783 ssize_t size_ret;
58eb9381 1784 struct relay_session *session = conn->session;
5312a3ed 1785 struct lttcomm_relayd_metadata_payload metadata_payload_header;
b8aa1682 1786 struct relay_stream *metadata_stream;
5312a3ed 1787 uint64_t metadata_payload_size;
b8aa1682
JD		 1788
1789 if (!session) {
1790 ERR("Metadata sent before version check");
1791 ret = -1;
1792 goto end;
1793 }
1794
5312a3ed 1795 if (recv_hdr->data_size < sizeof(struct lttcomm_relayd_metadata_payload)) {
f6416125
MD		 1796 ERR("Incorrect data size");
1797 ret = -1;
1798 goto end;
1799 }
5312a3ed
JG		 1800 metadata_payload_size = recv_hdr->data_size -
1801 sizeof(struct lttcomm_relayd_metadata_payload);
f6416125 1802
5312a3ed
JG		 1803 memcpy(&metadata_payload_header, payload->data,
1804 sizeof(metadata_payload_header));
1805 metadata_payload_header.stream_id = be64toh(
1806 metadata_payload_header.stream_id);
1807 metadata_payload_header.padding_size = be32toh(
1808 metadata_payload_header.padding_size);
9d1bbf21 1809
5312a3ed 1810 metadata_stream = stream_get_by_id(metadata_payload_header.stream_id);
b8aa1682
JD		 1811 if (!metadata_stream) {
1812 ret = -1;
7591bab1 1813 goto end;
b8aa1682
JD		 1814 }
1815
7591bab1
MD		 1816 pthread_mutex_lock(&metadata_stream->lock);
1817
5312a3ed
JG		 1818 size_ret = lttng_write(metadata_stream->stream_fd->fd,
1819 payload->data + sizeof(metadata_payload_header),
1820 metadata_payload_size);
1821 if (size_ret < metadata_payload_size) {
b8aa1682
JD		 1822 ERR("Relay error writing metadata on file");
1823 ret = -1;
7591bab1 1824 goto end_put;
b8aa1682 1825 }
1d4dfdef 1826
32d1569c 1827 size_ret = write_padding_to_file(metadata_stream->stream_fd->fd,
5312a3ed 1828 metadata_payload_header.padding_size);
715e6fb1 1829 if (size_ret < (int64_t) metadata_payload_header.padding_size) {
5312a3ed 1830 ret = -1;
7591bab1 1831 goto end_put;
1d4dfdef 1832 }
2a174661 1833
7591bab1 1834 metadata_stream->metadata_received +=
5312a3ed 1835 metadata_payload_size + metadata_payload_header.padding_size;
7591bab1
MD		 1836 DBG2("Relay metadata written. Updated metadata_received %" PRIu64,
1837 metadata_stream->metadata_received);
1d4dfdef 1838
d3ecc550
JD		 1839 ret = try_rotate_stream(metadata_stream);
1840 if (ret < 0) {
1841 goto end_put;
1842 }
1843
7591bab1
MD		 1844end_put:
1845 pthread_mutex_unlock(&metadata_stream->lock);
1846 stream_put(metadata_stream);
b8aa1682
JD		 1847end:
1848 return ret;
1849}
1850
1851/*
1852 * relay_send_version: send relayd version number
1853 */
5312a3ed
JG		 1854static int relay_send_version(const struct lttcomm_relayd_hdr *recv_hdr,
1855 struct relay_connection *conn,
1856 const struct lttng_buffer_view *payload)
b8aa1682 1857{
7f51dcba 1858 int ret;
5312a3ed 1859 ssize_t send_ret;
092b6259 1860 struct lttcomm_relayd_version reply, msg;
87cb6359 1861 bool compatible = true;
b8aa1682 1862
5312a3ed 1863 conn->version_check_done = true;
b8aa1682 1864
092b6259 1865 /* Get version from the other side. */
5312a3ed
JG		 1866 if (payload->size < sizeof(msg)) {
1867 ERR("Unexpected payload size in \"relay_send_version\": expected >= %zu bytes, got %zu bytes",
1868 sizeof(msg), payload->size);
092b6259 1869 ret = -1;
092b6259
DG		 1870 goto end;
1871 }
1872
5312a3ed
JG		 1873 memcpy(&msg, payload->data, sizeof(msg));
1874 msg.major = be32toh(msg.major);
1875 msg.minor = be32toh(msg.minor);
1876
53efb85a 1877 memset(&reply, 0, sizeof(reply));
d83a952c
MD		 1878 reply.major = RELAYD_VERSION_COMM_MAJOR;
1879 reply.minor = RELAYD_VERSION_COMM_MINOR;
d4519fa3
JD		 1880
1881 /* Major versions must be the same */
5312a3ed 1882 if (reply.major != msg.major) {
6151a90f 1883 DBG("Incompatible major versions (%u vs %u), deleting session",
5312a3ed 1884 reply.major, msg.major);
87cb6359 1885 compatible = false;
d4519fa3
JD		 1886 }
1887
58eb9381 1888 conn->major = reply.major;
0f907de1 1889 /* We adapt to the lowest compatible version */
5312a3ed 1890 if (reply.minor <= msg.minor) {
58eb9381 1891 conn->minor = reply.minor;
0f907de1 1892 } else {
5312a3ed 1893 conn->minor = msg.minor;
0f907de1
JD		 1894 }
1895
6151a90f
JD		 1896 reply.major = htobe32(reply.major);
1897 reply.minor = htobe32(reply.minor);
5312a3ed
JG		 1898 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply,
1899 sizeof(reply), 0);
1900 if (send_ret < (ssize_t) sizeof(reply)) {
1901 ERR("Failed to send \"send version\" command reply (ret = %zd)",
1902 send_ret);
1903 ret = -1;
1904 goto end;
1905 } else {
1906 ret = 0;
6151a90f
JD		 1907 }
1908
87cb6359
JD		 1909 if (!compatible) {
1910 ret = -1;
1911 goto end;
1912 }
1913
58eb9381
DG		 1914 DBG("Version check done using protocol %u.%u", conn->major,
1915 conn->minor);
b8aa1682
JD		 1916
1917end:
1918 return ret;
1919}
1920
c8f59ee5 1921/*
6d805429 1922 * Check for data pending for a given stream id from the session daemon.
c8f59ee5 1923 */
5312a3ed
JG		 1924static int relay_data_pending(const struct lttcomm_relayd_hdr *recv_hdr,
1925 struct relay_connection *conn,
1926 const struct lttng_buffer_view *payload)
c8f59ee5 1927{
58eb9381 1928 struct relay_session *session = conn->session;
6d805429 1929 struct lttcomm_relayd_data_pending msg;
c8f59ee5
DG		 1930 struct lttcomm_relayd_generic_reply reply;
1931 struct relay_stream *stream;
5312a3ed 1932 ssize_t send_ret;
c8f59ee5 1933 int ret;
c8f59ee5 1934
6d805429 1935 DBG("Data pending command received");
c8f59ee5 1936
5312a3ed 1937 if (!session || !conn->version_check_done) {
c8f59ee5
DG		 1938 ERR("Trying to check for data before version check");
1939 ret = -1;
1940 goto end_no_session;
1941 }
1942
5312a3ed
JG		 1943 if (payload->size < sizeof(msg)) {
1944 ERR("Unexpected payload size in \"relay_data_pending\": expected >= %zu bytes, got %zu bytes",
1945 sizeof(msg), payload->size);
c8f59ee5
DG		 1946 ret = -1;
1947 goto end_no_session;
1948 }
5312a3ed
JG		 1949 memcpy(&msg, payload->data, sizeof(msg));
1950 msg.stream_id = be64toh(msg.stream_id);
1951 msg.last_net_seq_num = be64toh(msg.last_net_seq_num);
c8f59ee5 1952
5312a3ed 1953 stream = stream_get_by_id(msg.stream_id);
de91f48a 1954 if (stream == NULL) {
c8f59ee5 1955 ret = -1;
7591bab1 1956 goto end;
c8f59ee5
DG		 1957 }
1958
7591bab1
MD		 1959 pthread_mutex_lock(&stream->lock);
1960
6d805429 1961 DBG("Data pending for stream id %" PRIu64 " prev_seq %" PRIu64
5312a3ed
JG		 1962 " and last_seq %" PRIu64, msg.stream_id,
1963 stream->prev_seq, msg.last_net_seq_num);
c8f59ee5 1964
33832e64 1965 /* Avoid wrapping issue */
5312a3ed 1966 if (((int64_t) (stream->prev_seq - msg.last_net_seq_num)) >= 0) {
6d805429 1967 /* Data has in fact been written and is NOT pending */
c8f59ee5 1968 ret = 0;
6d805429
DG		 1969 } else {
1970 /* Data still being streamed thus pending */
1971 ret = 1;
c8f59ee5
DG		 1972 }
1973
7591bab1
MD		 1974 stream->data_pending_check_done = true;
1975 pthread_mutex_unlock(&stream->lock);
f7079f67 1976
7591bab1
MD		 1977 stream_put(stream);
1978end:
c8f59ee5 1979
53efb85a 1980 memset(&reply, 0, sizeof(reply));
c8f59ee5 1981 reply.ret_code = htobe32(ret);
5312a3ed
JG		 1982 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply, sizeof(reply), 0);
1983 if (send_ret < (ssize_t) sizeof(reply)) {
1984 ERR("Failed to send \"data pending\" command reply (ret = %zd)",
1985 send_ret);
1986 ret = -1;
c8f59ee5
DG		 1987 }
1988
1989end_no_session:
1990 return ret;
1991}
1992
1993/*
1994 * Wait for the control socket to reach a quiescent state.
1995 *
7591bab1
MD		 1996 * Note that for now, when receiving this command from the session
1997 * daemon, this means that every subsequent commands or data received on
1998 * the control socket has been handled. So, this is why we simply return
1999 * OK here.
c8f59ee5 2000 */
5312a3ed
JG		 2001static int relay_quiescent_control(const struct lttcomm_relayd_hdr *recv_hdr,
2002 struct relay_connection *conn,
2003 const struct lttng_buffer_view *payload)
c8f59ee5
DG		 2004{
2005 int ret;
5312a3ed 2006 ssize_t send_ret;
ad7051c0 2007 struct relay_stream *stream;
ad7051c0 2008 struct lttcomm_relayd_quiescent_control msg;
c8f59ee5
DG		 2009 struct lttcomm_relayd_generic_reply reply;
2010
2011 DBG("Checking quiescent state on control socket");
2012
5312a3ed 2013 if (!conn->session || !conn->version_check_done) {
ad7051c0
DG		 2014 ERR("Trying to check for data before version check");
2015 ret = -1;
2016 goto end_no_session;
2017 }
2018
5312a3ed
JG		 2019 if (payload->size < sizeof(msg)) {
2020 ERR("Unexpected payload size in \"relay_quiescent_control\": expected >= %zu bytes, got %zu bytes",
2021 sizeof(msg), payload->size);
ad7051c0
DG		 2022 ret = -1;
2023 goto end_no_session;
2024 }
5312a3ed
JG		 2025 memcpy(&msg, payload->data, sizeof(msg));
2026 msg.stream_id = be64toh(msg.stream_id);
ad7051c0 2027
5312a3ed 2028 stream = stream_get_by_id(msg.stream_id);
7591bab1
MD		 2029 if (!stream) {
2030 goto reply;
2031 }
2032 pthread_mutex_lock(&stream->lock);
2033 stream->data_pending_check_done = true;
2034 pthread_mutex_unlock(&stream->lock);
5312a3ed
JG		 2035
2036 DBG("Relay quiescent control pending flag set to %" PRIu64, msg.stream_id);
7591bab1
MD		 2037 stream_put(stream);
2038reply:
53efb85a 2039 memset(&reply, 0, sizeof(reply));
c8f59ee5 2040 reply.ret_code = htobe32(LTTNG_OK);
5312a3ed
JG		 2041 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply, sizeof(reply), 0);
2042 if (send_ret < (ssize_t) sizeof(reply)) {
2043 ERR("Failed to send \"quiescent control\" command reply (ret = %zd)",
2044 send_ret);
2045 ret = -1;
2046 } else {
2047 ret = 0;
c8f59ee5
DG		 2048 }
2049
ad7051c0 2050end_no_session:
c8f59ee5
DG		 2051 return ret;
2052}
2053
f7079f67 2054/*
7591bab1
MD		 2055 * Initialize a data pending command. This means that a consumer is about
2056 * to ask for data pending for each stream it holds. Simply iterate over
2057 * all streams of a session and set the data_pending_check_done flag.
f7079f67
DG		 2058 *
2059 * This command returns to the client a LTTNG_OK code.
2060 */
5312a3ed
JG		 2061static int relay_begin_data_pending(const struct lttcomm_relayd_hdr *recv_hdr,
2062 struct relay_connection *conn,
2063 const struct lttng_buffer_view *payload)
f7079f67
DG		 2064{
2065 int ret;
5312a3ed 2066 ssize_t send_ret;
f7079f67
DG		 2067 struct lttng_ht_iter iter;
2068 struct lttcomm_relayd_begin_data_pending msg;
2069 struct lttcomm_relayd_generic_reply reply;
2070 struct relay_stream *stream;
f7079f67
DG		 2071
2072 assert(recv_hdr);
58eb9381 2073 assert(conn);
f7079f67
DG		 2074
2075 DBG("Init streams for data pending");
2076
5312a3ed 2077 if (!conn->session || !conn->version_check_done) {
f7079f67
DG		 2078 ERR("Trying to check for data before version check");
2079 ret = -1;
2080 goto end_no_session;
2081 }
2082
5312a3ed
JG		 2083 if (payload->size < sizeof(msg)) {
2084 ERR("Unexpected payload size in \"relay_begin_data_pending\": expected >= %zu bytes, got %zu bytes",
2085 sizeof(msg), payload->size);
f7079f67
DG		 2086 ret = -1;
2087 goto end_no_session;
2088 }
5312a3ed
JG		 2089 memcpy(&msg, payload->data, sizeof(msg));
2090 msg.session_id = be64toh(msg.session_id);
f7079f67
DG		 2091
2092 /*
7591bab1
MD		 2093 * Iterate over all streams to set the begin data pending flag.
2094 * For now, the streams are indexed by stream handle so we have
2095 * to iterate over all streams to find the one associated with
2096 * the right session_id.
f7079f67
DG		 2097 */
2098 rcu_read_lock();
d3e2ba59 2099 cds_lfht_for_each_entry(relay_streams_ht->ht, &iter.iter, stream,
2a174661 2100 node.node) {
7591bab1
MD		 2101 if (!stream_get(stream)) {
2102 continue;
2103 }
5312a3ed 2104 if (stream->trace->session->id == msg.session_id) {
7591bab1
MD		 2105 pthread_mutex_lock(&stream->lock);
2106 stream->data_pending_check_done = false;
2107 pthread_mutex_unlock(&stream->lock);
f7079f67
DG		 2108 DBG("Set begin data pending flag to stream %" PRIu64,
2109 stream->stream_handle);
2110 }
7591bab1 2111 stream_put(stream);
f7079f67
DG		 2112 }
2113 rcu_read_unlock();
2114
53efb85a 2115 memset(&reply, 0, sizeof(reply));
f7079f67
DG		 2116 /* All good, send back reply. */
2117 reply.ret_code = htobe32(LTTNG_OK);
2118
5312a3ed
JG		 2119 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply, sizeof(reply), 0);
2120 if (send_ret < (ssize_t) sizeof(reply)) {
2121 ERR("Failed to send \"begin data pending\" command reply (ret = %zd)",
2122 send_ret);
2123 ret = -1;
2124 } else {
2125 ret = 0;
f7079f67
DG		 2126 }
2127
2128end_no_session:
2129 return ret;
2130}
2131
2132/*
7591bab1
MD		 2133 * End data pending command. This will check, for a given session id, if
2134 * each stream associated with it has its data_pending_check_done flag
2135 * set. If not, this means that the client lost track of the stream but
2136 * the data is still being streamed on our side. In this case, we inform
2137 * the client that data is in flight.
f7079f67
DG		 2138 *
2139 * Return to the client if there is data in flight or not with a ret_code.
2140 */
5312a3ed
JG		 2141static int relay_end_data_pending(const struct lttcomm_relayd_hdr *recv_hdr,
2142 struct relay_connection *conn,
2143 const struct lttng_buffer_view *payload)
f7079f67
DG		 2144{
2145 int ret;
5312a3ed 2146 ssize_t send_ret;
f7079f67
DG		 2147 struct lttng_ht_iter iter;
2148 struct lttcomm_relayd_end_data_pending msg;
2149 struct lttcomm_relayd_generic_reply reply;
2150 struct relay_stream *stream;
f7079f67
DG		 2151 uint32_t is_data_inflight = 0;
2152
f7079f67
DG		 2153 DBG("End data pending command");
2154
5312a3ed 2155 if (!conn->session || !conn->version_check_done) {
f7079f67
DG		 2156 ERR("Trying to check for data before version check");
2157 ret = -1;
2158 goto end_no_session;
2159 }
2160
5312a3ed
JG		 2161 if (payload->size < sizeof(msg)) {
2162 ERR("Unexpected payload size in \"relay_end_data_pending\": expected >= %zu bytes, got %zu bytes",
2163 sizeof(msg), payload->size);
f7079f67
DG		 2164 ret = -1;
2165 goto end_no_session;
2166 }
5312a3ed
JG		 2167 memcpy(&msg, payload->data, sizeof(msg));
2168 msg.session_id = be64toh(msg.session_id);
f7079f67 2169
7591bab1
MD		 2170 /*
2171 * Iterate over all streams to see if the begin data pending
2172 * flag is set.
2173 */
f7079f67 2174 rcu_read_lock();
d3e2ba59 2175 cds_lfht_for_each_entry(relay_streams_ht->ht, &iter.iter, stream,
2a174661 2176 node.node) {
7591bab1
MD		 2177 if (!stream_get(stream)) {
2178 continue;
2179 }
5312a3ed 2180 if (stream->trace->session->id != msg.session_id) {
7591bab1
MD		 2181 stream_put(stream);
2182 continue;
2183 }
2184 pthread_mutex_lock(&stream->lock);
2185 if (!stream->data_pending_check_done) {
2186 if (!stream->closed || !(((int64_t) (stream->prev_seq - stream->last_net_seq_num)) >= 0)) {
2187 is_data_inflight = 1;
2188 DBG("Data is still in flight for stream %" PRIu64,
2189 stream->stream_handle);
2190 pthread_mutex_unlock(&stream->lock);
2191 stream_put(stream);
2192 break;
2193 }
f7079f67 2194 }
7591bab1
MD		 2195 pthread_mutex_unlock(&stream->lock);
2196 stream_put(stream);
f7079f67
DG		 2197 }
2198 rcu_read_unlock();
2199
53efb85a 2200 memset(&reply, 0, sizeof(reply));
f7079f67
DG		 2201 /* All good, send back reply. */
2202 reply.ret_code = htobe32(is_data_inflight);
2203
5312a3ed
JG		 2204 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply, sizeof(reply), 0);
2205 if (send_ret < (ssize_t) sizeof(reply)) {
2206 ERR("Failed to send \"end data pending\" command reply (ret = %zd)",
2207 send_ret);
2208 ret = -1;
2209 } else {
2210 ret = 0;
f7079f67
DG		 2211 }
2212
2213end_no_session:
2214 return ret;
2215}
2216
1c20f0e2
JD		 2217/*
2218 * Receive an index for a specific stream.
2219 *
2220 * Return 0 on success else a negative value.
2221 */
5312a3ed
JG		 2222static int relay_recv_index(const struct lttcomm_relayd_hdr *recv_hdr,
2223 struct relay_connection *conn,
2224 const struct lttng_buffer_view *payload)
1c20f0e2 2225{
5312a3ed
JG		 2226 int ret;
2227 ssize_t send_ret;
58eb9381 2228 struct relay_session *session = conn->session;
1c20f0e2 2229 struct lttcomm_relayd_index index_info;
7591bab1 2230 struct relay_index *index;
1c20f0e2
JD		 2231 struct lttcomm_relayd_generic_reply reply;
2232 struct relay_stream *stream;
f8f3885c 2233 size_t msg_len;
1c20f0e2 2234
58eb9381 2235 assert(conn);
1c20f0e2
JD		 2236
2237 DBG("Relay receiving index");
2238
5312a3ed 2239 if (!session || !conn->version_check_done) {
1c20f0e2
JD		 2240 ERR("Trying to close a stream before version check");
2241 ret = -1;
2242 goto end_no_session;
2243 }
2244
f8f3885c
MD		 2245 msg_len = lttcomm_relayd_index_len(
2246 lttng_to_index_major(conn->major, conn->minor),
2247 lttng_to_index_minor(conn->major, conn->minor));
5312a3ed
JG		 2248 if (payload->size < msg_len) {
2249 ERR("Unexpected payload size in \"relay_recv_index\": expected >= %zu bytes, got %zu bytes",
2250 msg_len, payload->size);
1c20f0e2
JD		 2251 ret = -1;
2252 goto end_no_session;
2253 }
5312a3ed
JG		 2254 memcpy(&index_info, payload->data, msg_len);
2255 index_info.relay_stream_id = be64toh(index_info.relay_stream_id);
2256 index_info.net_seq_num = be64toh(index_info.net_seq_num);
2257 index_info.packet_size = be64toh(index_info.packet_size);
2258 index_info.content_size = be64toh(index_info.content_size);
2259 index_info.timestamp_begin = be64toh(index_info.timestamp_begin);
2260 index_info.timestamp_end = be64toh(index_info.timestamp_end);
2261 index_info.events_discarded = be64toh(index_info.events_discarded);
2262 index_info.stream_id = be64toh(index_info.stream_id);
81df238b
JR		 2263
2264 if (conn->minor >= 8) {
2265 index_info.stream_instance_id =
2266 be64toh(index_info.stream_instance_id);
2267 index_info.packet_seq_num = be64toh(index_info.packet_seq_num);
2268 }
5312a3ed
JG		 2269
2270 stream = stream_get_by_id(index_info.relay_stream_id);
1c20f0e2 2271 if (!stream) {
7591bab1 2272 ERR("stream_get_by_id not found");
1c20f0e2 2273 ret = -1;
7591bab1 2274 goto end;
1c20f0e2 2275 }
7591bab1 2276 pthread_mutex_lock(&stream->lock);
1c20f0e2 2277
d3e2ba59
JD		 2278 /* Live beacon handling */
2279 if (index_info.packet_size == 0) {
7591bab1
MD		 2280 DBG("Received live beacon for stream %" PRIu64,
2281 stream->stream_handle);
d3e2ba59
JD		 2282
2283 /*
7591bab1
MD		 2284 * Only flag a stream inactive when it has already
2285 * received data and no indexes are in flight.
d3e2ba59 2286 */
a44ca2ca 2287 if (stream->index_received_seqcount > 0
7591bab1 2288 && stream->indexes_in_flight == 0) {
5312a3ed 2289 stream->beacon_ts_end = index_info.timestamp_end;
d3e2ba59
JD		 2290 }
2291 ret = 0;
7591bab1 2292 goto end_stream_put;
d3e2ba59
JD		 2293 } else {
2294 stream->beacon_ts_end = -1ULL;
2295 }
2296
528f2ffa 2297 if (stream->ctf_stream_id == -1ULL) {
5312a3ed 2298 stream->ctf_stream_id = index_info.stream_id;
528f2ffa 2299 }
5312a3ed 2300 index = relay_index_get_by_id_or_create(stream, index_info.net_seq_num);
7591bab1
MD		 2301 if (!index) {
2302 ret = -1;
2303 ERR("relay_index_get_by_id_or_create index NULL");
2304 goto end_stream_put;
1c20f0e2 2305 }
234cd636 2306 if (set_index_control_data(index, &index_info, conn)) {
7591bab1
MD		 2307 ERR("set_index_control_data error");
2308 relay_index_put(index);
2309 ret = -1;
2310 goto end_stream_put;
2311 }
2312 ret = relay_index_try_flush(index);
2313 if (ret == 0) {
a44ca2ca
MD		 2314 tracefile_array_commit_seq(stream->tfa);
2315 stream->index_received_seqcount++;
d3ecc550 2316 stream->pos_after_last_complete_data_index += index->total_size;
7591bab1
MD		 2317 } else if (ret > 0) {
2318 /* no flush. */
2319 ret = 0;
2320 } else {
245f8bec
JR		 2321 /*
2322 * ret < 0
2323 *
2324 * relay_index_try_flush is responsible for the self-reference
2325 * put of the index object on error.
2326 */
7591bab1 2327 ERR("relay_index_try_flush error %d", ret);
7591bab1 2328 ret = -1;
1c20f0e2
JD		 2329 }
2330
7591bab1
MD		 2331end_stream_put:
2332 pthread_mutex_unlock(&stream->lock);
2333 stream_put(stream);
2334
2335end:
1c20f0e2 2336
53efb85a 2337 memset(&reply, 0, sizeof(reply));
1c20f0e2
JD		 2338 if (ret < 0) {
2339 reply.ret_code = htobe32(LTTNG_ERR_UNK);
2340 } else {
2341 reply.ret_code = htobe32(LTTNG_OK);
2342 }
58eb9381 2343 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply, sizeof(reply), 0);
5312a3ed
JG		 2344 if (send_ret < (ssize_t) sizeof(reply)) {
2345 ERR("Failed to send \"recv index\" command reply (ret = %zd)", send_ret);
2346 ret = -1;
1c20f0e2
JD		 2347 }
2348
2349end_no_session:
2350 return ret;
2351}
2352
a4baae1b
JD		 2353/*
2354 * Receive the streams_sent message.
2355 *
2356 * Return 0 on success else a negative value.
2357 */
5312a3ed
JG		 2358static int relay_streams_sent(const struct lttcomm_relayd_hdr *recv_hdr,
2359 struct relay_connection *conn,
2360 const struct lttng_buffer_view *payload)
a4baae1b 2361{
5312a3ed
JG		 2362 int ret;
2363 ssize_t send_ret;
a4baae1b
JD		 2364 struct lttcomm_relayd_generic_reply reply;
2365
58eb9381 2366 assert(conn);
a4baae1b
JD		 2367
2368 DBG("Relay receiving streams_sent");
2369
5312a3ed 2370 if (!conn->session || !conn->version_check_done) {
a4baae1b
JD		 2371 ERR("Trying to close a stream before version check");
2372 ret = -1;
2373 goto end_no_session;
2374 }
2375
2376 /*
7591bab1
MD		 2377 * Publish every pending stream in the connection recv list which are
2378 * now ready to be used by the viewer.
4a9daf17 2379 */
7591bab1 2380 publish_connection_local_streams(conn);
4a9daf17 2381
53efb85a 2382 memset(&reply, 0, sizeof(reply));
a4baae1b 2383 reply.ret_code = htobe32(LTTNG_OK);
58eb9381 2384 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply, sizeof(reply), 0);
5312a3ed
JG		 2385 if (send_ret < (ssize_t) sizeof(reply)) {
2386 ERR("Failed to send \"streams sent\" command reply (ret = %zd)",
2387 send_ret);
2388 ret = -1;
a4baae1b
JD		 2389 } else {
2390 /* Success. */
2391 ret = 0;
2392 }
2393
2394end_no_session:
2395 return ret;
2396}
2397
d3ecc550 2398/*
5312a3ed 2399 * relay_rotate_session_stream: rotate a stream to a new tracefile for the session
d3ecc550
JD		 2400 * rotation feature (not the tracefile rotation feature).
2401 */
5312a3ed
JG		 2402static int relay_rotate_session_stream(const struct lttcomm_relayd_hdr *recv_hdr,
2403 struct relay_connection *conn,
2404 const struct lttng_buffer_view *payload)
d3ecc550 2405{
5312a3ed
JG		 2406 int ret;
2407 ssize_t send_ret;
d3ecc550
JD		 2408 struct relay_session *session = conn->session;
2409 struct lttcomm_relayd_rotate_stream stream_info;
2410 struct lttcomm_relayd_generic_reply reply;
2411 struct relay_stream *stream;
5312a3ed
JG		 2412 size_t header_len;
2413 size_t path_len;
2414 struct lttng_buffer_view new_path_view;
d3ecc550
JD		 2415
2416 DBG("Rotate stream received");
2417
2418 if (!session || !conn->version_check_done) {
2419 ERR("Trying to rotate a stream before version check");
2420 ret = -1;
2421 goto end_no_reply;
2422 }
2423
2424 if (session->major == 2 && session->minor < 11) {
2425 ERR("Unsupported feature before 2.11");
2426 ret = -1;
2427 goto end_no_reply;
2428 }
2429
5312a3ed 2430 header_len = sizeof(struct lttcomm_relayd_rotate_stream);
d3ecc550 2431
5312a3ed
JG		 2432 if (payload->size < header_len) {
2433 ERR("Unexpected payload size in \"relay_rotate_session_stream\": expected >= %zu bytes, got %zu bytes",
2434 header_len, payload->size);
d3ecc550
JD		 2435 ret = -1;
2436 goto end_no_reply;
2437 }
2438
5312a3ed
JG		 2439 memcpy(&stream_info, payload->data, header_len);
2440
2441 /* Convert to host */
2442 stream_info.pathname_length = be32toh(stream_info.pathname_length);
2443 stream_info.stream_id = be64toh(stream_info.stream_id);
2444 stream_info.new_chunk_id = be64toh(stream_info.new_chunk_id);
2445 stream_info.rotate_at_seq_num = be64toh(stream_info.rotate_at_seq_num);
d3ecc550 2446
5312a3ed
JG		 2447 path_len = stream_info.pathname_length;
2448 if (payload->size < header_len + path_len) {
2449 ERR("Unexpected payload size in \"relay_rotate_session_stream\" including path: expected >= %zu bytes, got %zu bytes",
2450 header_len + path_len, payload->size);
2451 ret = -1;
2452 goto end_no_reply;
2453 }
2454
d3ecc550 2455 /* Ensure it fits in local filename length. */
5312a3ed 2456 if (path_len >= LTTNG_PATH_MAX) {
d3ecc550
JD		 2457 ret = -ENAMETOOLONG;
2458 ERR("Length of relay_rotate_session_stream command's path name (%zu bytes) exceeds the maximal allowed length of %i bytes",
5312a3ed 2459 path_len, LTTNG_PATH_MAX);
d3ecc550
JD		 2460 goto end;
2461 }
2462
5312a3ed
JG		 2463 new_path_view = lttng_buffer_view_from_view(payload, header_len,
2464 stream_info.pathname_length);
d3ecc550 2465
5312a3ed
JG		 2466 stream = stream_get_by_id(stream_info.stream_id);
2467 if (!stream) {
d3ecc550 2468 ret = -1;
5312a3ed 2469 goto end;
d3ecc550
JD		 2470 }
2471
2472 pthread_mutex_lock(&stream->lock);
2473
2474 /*
2475 * Update the trace path (just the folder, the stream name does not
2476 * change).
2477 */
2478 free(stream->path_name);
5312a3ed 2479 stream->path_name = create_output_path(new_path_view.data);
d3ecc550
JD		 2480 if (!stream->path_name) {
2481 ERR("Failed to create a new output path");
5312a3ed 2482 ret = -1;
d3ecc550
JD		 2483 goto end_stream_unlock;
2484 }
2485 ret = utils_mkdir_recursive(stream->path_name, S_IRWXU | S_IRWXG,
2486 -1, -1);
2487 if (ret < 0) {
2488 ERR("relay creating output directory");
5312a3ed 2489 ret = -1;
d3ecc550
JD		 2490 goto end_stream_unlock;
2491 }
5312a3ed 2492
81164b6b
JG		 2493 assert(stream->current_chunk_id.is_set);
2494 stream->current_chunk_id.value = stream_info.new_chunk_id;
d3ecc550
JD		 2495
2496 if (stream->is_metadata) {
2497 /*
2498 * The metadata stream is sent only over the control connection
2499 * so we know we have all the data to perform the stream
2500 * rotation.
2501 */
2502 ret = do_rotate_stream(stream);
2503 } else {
5312a3ed 2504 stream->rotate_at_seq_num = stream_info.rotate_at_seq_num;
d3ecc550
JD		 2505 ret = try_rotate_stream(stream);
2506 }
2507 if (ret < 0) {
2508 goto end_stream_unlock;
2509 }
2510
2511end_stream_unlock:
2512 pthread_mutex_unlock(&stream->lock);
2513 stream_put(stream);
2514end:
2515 memset(&reply, 0, sizeof(reply));
2516 if (ret < 0) {
2517 reply.ret_code = htobe32(LTTNG_ERR_UNK);
2518 } else {
2519 reply.ret_code = htobe32(LTTNG_OK);
2520 }
2521 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply,
2522 sizeof(struct lttcomm_relayd_generic_reply), 0);
5312a3ed
JG		 2523 if (send_ret < (ssize_t) sizeof(reply)) {
2524 ERR("Failed to send \"rotate session stream\" command reply (ret = %zd)",
2525 send_ret);
2526 ret = -1;
d3ecc550
JD		 2527 }
2528
2529end_no_reply:
d3ecc550
JD		 2530 return ret;
2531}
2532
a1ae2ea5
JD		 2533/*
2534 * relay_mkdir: Create a folder on the disk.
2535 */
5312a3ed
JG		 2536static int relay_mkdir(const struct lttcomm_relayd_hdr *recv_hdr,
2537 struct relay_connection *conn,
2538 const struct lttng_buffer_view *payload)
a1ae2ea5
JD		 2539{
2540 int ret;
a1ae2ea5
JD		 2541 struct relay_session *session = conn->session;
2542 struct lttcomm_relayd_mkdir path_info_header;
a1ae2ea5
JD		 2543 struct lttcomm_relayd_generic_reply reply;
2544 char *path = NULL;
5312a3ed
JG		 2545 size_t header_len;
2546 ssize_t send_ret;
2547 struct lttng_buffer_view path_view;
a1ae2ea5
JD		 2548
2549 if (!session || !conn->version_check_done) {
00fb02ac 2550 ERR("Trying to create a directory before version check");
a1ae2ea5
JD		 2551 ret = -1;
2552 goto end_no_session;
2553 }
2554
2555 if (session->major == 2 && session->minor < 11) {
2556 /*
2557 * This client is not supposed to use this command since
2558 * it predates its introduction.
2559 */
2560 ERR("relay_mkdir command is unsupported before LTTng 2.11");
2561 ret = -1;
2562 goto end_no_session;
2563 }
2564
5312a3ed
JG		 2565 header_len = sizeof(path_info_header);
2566 if (payload->size < header_len) {
2567 ERR("Unexpected payload size in \"relay_mkdir\": expected >= %zu bytes, got %zu bytes",
2568 header_len, payload->size);
a1ae2ea5
JD		 2569 ret = -1;
2570 goto end_no_session;
2571 }
2572
5312a3ed
JG		 2573 memcpy(&path_info_header, payload->data, header_len);
2574
a1ae2ea5
JD		 2575 path_info_header.length = be32toh(path_info_header.length);
2576
5312a3ed
JG		 2577 if (payload->size < header_len + path_info_header.length) {
2578 ERR("Unexpected payload size in \"relay_mkdir\" including path: expected >= %zu bytes, got %zu bytes",
2579 header_len + path_info_header.length, payload->size);
2580 ret = -1;
2581 goto end_no_session;
2582 }
2583
a1ae2ea5
JD		 2584 /* Ensure that it fits in local path length. */
2585 if (path_info_header.length >= LTTNG_PATH_MAX) {
2586 ret = -ENAMETOOLONG;
2587 ERR("Path name argument of mkdir command (%" PRIu32 " bytes) exceeds the maximal length allowed (%d bytes)",
2588 path_info_header.length, LTTNG_PATH_MAX);
2589 goto end;
2590 }
2591
5312a3ed
JG		 2592 path_view = lttng_buffer_view_from_view(payload, header_len,
2593 path_info_header.length);
a1ae2ea5 2594
5312a3ed 2595 path = create_output_path(path_view.data);
a1ae2ea5
JD		 2596 if (!path) {
2597 ERR("Failed to create output path");
2598 ret = -1;
2599 goto end;
2600 }
2601
2602 ret = utils_mkdir_recursive(path, S_IRWXU | S_IRWXG, -1, -1);
2603 if (ret < 0) {
2604 ERR("relay creating output directory");
2605 goto end;
2606 }
2607
2608 ret = 0;
2609
2610end:
2611 memset(&reply, 0, sizeof(reply));
2612 if (ret < 0) {
2613 reply.ret_code = htobe32(LTTNG_ERR_UNK);
2614 } else {
2615 reply.ret_code = htobe32(LTTNG_OK);
2616 }
5312a3ed
JG		 2617 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply, sizeof(reply), 0);
2618 if (send_ret < (ssize_t) sizeof(reply)) {
2619 ERR("Failed to send \"mkdir\" command reply (ret = %zd)", send_ret);
a1ae2ea5
JD		 2620 ret = -1;
2621 }
2622
2623end_no_session:
2624 free(path);
a1ae2ea5
JD		 2625 return ret;
2626}
2627
00fb02ac
JD		 2628static int validate_rotate_rename_path_length(const char *path_type,
2629 uint32_t path_length)
2630{
2631 int ret = 0;
2632
2633 if (path_length > LTTNG_PATH_MAX) {
2634 ret = -ENAMETOOLONG;
2635 ERR("rotate rename \"%s\" path name length (%" PRIu32 " bytes) exceeds the allowed size of %i bytes",
2636 path_type, path_length, LTTNG_PATH_MAX);
2637 } else if (path_length == 0) {
2638 ret = -EINVAL;
2639 ERR("rotate rename \"%s\" path name has an illegal length of 0", path_type);
2640 }
2641 return ret;
2642}
2643
2644/*
2645 * relay_rotate_rename: rename the trace folder after a rotation is
2646 * completed. We are not closing any fd here, just moving the folder, so it
2647 * works even if data is still in-flight.
2648 */
5312a3ed
JG		 2649static int relay_rotate_rename(const struct lttcomm_relayd_hdr *recv_hdr,
2650 struct relay_connection *conn,
2651 const struct lttng_buffer_view *payload)
00fb02ac
JD		 2652{
2653 int ret;
5312a3ed 2654 ssize_t send_ret;
00fb02ac
JD		 2655 struct relay_session *session = conn->session;
2656 struct lttcomm_relayd_generic_reply reply;
2657 struct lttcomm_relayd_rotate_rename header;
5312a3ed 2658 size_t header_len;
00fb02ac 2659 size_t received_paths_size;
00fb02ac 2660 char *complete_old_path = NULL, *complete_new_path = NULL;
5312a3ed
JG		 2661 struct lttng_buffer_view old_path_view;
2662 struct lttng_buffer_view new_path_view;
00fb02ac
JD		 2663
2664 if (!session || !conn->version_check_done) {
2665 ERR("Trying to rename a trace folder before version check");
2666 ret = -1;
2667 goto end_no_reply;
2668 }
2669
2670 if (session->major == 2 && session->minor < 11) {
2671 ERR("relay_rotate_rename command is unsupported before LTTng 2.11");
2672 ret = -1;
2673 goto end_no_reply;
2674 }
2675
5312a3ed
JG		 2676 header_len = sizeof(header);
2677 if (payload->size < header_len) {
2678 ERR("Unexpected payload size in \"relay_rotate_rename\": expected >= %zu bytes, got %zu bytes",
2679 header_len, payload->size);
00fb02ac
JD		 2680 ret = -1;
2681 goto end_no_reply;
2682 }
2683
5312a3ed
JG		 2684 memcpy(&header, payload->data, header_len);
2685
00fb02ac
JD		 2686 header.old_path_length = be32toh(header.old_path_length);
2687 header.new_path_length = be32toh(header.new_path_length);
2688 received_paths_size = header.old_path_length + header.new_path_length;
2689
5312a3ed
JG		 2690 if (payload->size < header_len + received_paths_size) {
2691 ERR("Unexpected payload size in \"relay_rotate_rename\" including paths: expected >= %zu bytes, got %zu bytes",
2692 header_len, payload->size);
2693 ret = -1;
2694 goto end_no_reply;
2695 }
2696
00fb02ac
JD		 2697 /* Ensure the paths don't exceed their allowed size. */
2698 ret = validate_rotate_rename_path_length("old", header.old_path_length);
2699 if (ret) {
2700 goto end;
2701 }
2702 ret = validate_rotate_rename_path_length("new", header.new_path_length);
2703 if (ret) {
2704 goto end;
2705 }
2706
5312a3ed
JG		 2707 old_path_view = lttng_buffer_view_from_view(payload, header_len,
2708 header.old_path_length);
2709 new_path_view = lttng_buffer_view_from_view(payload,
2710 header_len + header.old_path_length,
2711 header.new_path_length);
00fb02ac
JD		 2712
2713 /* Validate that both paths received are NULL terminated. */
5312a3ed 2714 if (old_path_view.data[old_path_view.size - 1] != '\0') {
00fb02ac
JD		 2715 ERR("relay_rotate_rename command's \"old\" path is invalid (not NULL terminated)");
2716 ret = -1;
2717 goto end;
2718 }
5312a3ed 2719 if (new_path_view.data[new_path_view.size - 1] != '\0') {
00fb02ac
JD		 2720 ERR("relay_rotate_rename command's \"new\" path is invalid (not NULL terminated)");
2721 ret = -1;
2722 goto end;
2723 }
2724
5312a3ed 2725 complete_old_path = create_output_path(old_path_view.data);
00fb02ac
JD		 2726 if (!complete_old_path) {
2727 ERR("Failed to build old output path in rotate_rename command");
2728 ret = -1;
2729 goto end;
2730 }
2731
5312a3ed 2732 complete_new_path = create_output_path(new_path_view.data);
00fb02ac
JD		 2733 if (!complete_new_path) {
2734 ERR("Failed to build new output path in rotate_rename command");
2735 ret = -1;
2736 goto end;
2737 }
2738
2739 ret = utils_mkdir_recursive(complete_new_path, S_IRWXU | S_IRWXG,
2740 -1, -1);
2741 if (ret < 0) {
2742 ERR("Failed to mkdir() rotate_rename's \"new\" output directory at \"%s\"",
2743 complete_new_path);
2744 goto end;
2745 }
2746
2747 /*
2748 * If a domain has not yet created its channel, the domain-specific
2749 * folder might not exist, but this is not an error.
2750 */
2751 ret = rename(complete_old_path, complete_new_path);
2752 if (ret < 0 && errno != ENOENT) {
2753 PERROR("Renaming chunk in rotate_rename command from \"%s\" to \"%s\"",
2754 complete_old_path, complete_new_path);
2755 goto end;
2756 }
2757 ret = 0;
2758
2759end:
2760 memset(&reply, 0, sizeof(reply));
2761 if (ret < 0) {
2762 reply.ret_code = htobe32(LTTNG_ERR_UNK);
2763 } else {
2764 reply.ret_code = htobe32(LTTNG_OK);
2765 }
5312a3ed
JG		 2766 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply,
2767 sizeof(reply), 0);
2768 if (send_ret < sizeof(reply)) {
2769 ERR("Failed to send \"rotate rename\" command reply (ret = %zd)",
2770 send_ret);
00fb02ac
JD		 2771 ret = -1;
2772 }
2773
2774end_no_reply:
00fb02ac
JD		 2775 free(complete_old_path);
2776 free(complete_new_path);
2777 return ret;
2778}
2779
e2acc8d2
JD		 2780/*
2781 * Check if all the streams in the session have completed the last rotation.
2782 * The chunk_id value is used to distinguish the cases where a stream was
2783 * closed on the consumerd before the rotation started but it still active on
2784 * the relayd, and the case where a stream appeared on the consumerd/relayd
2785 * after the last rotation started (in that case, it is already writing in the
2786 * new chunk folder).
2787 */
2788static
5312a3ed
JG		 2789int relay_rotate_pending(const struct lttcomm_relayd_hdr *recv_hdr,
2790 struct relay_connection *conn,
2791 const struct lttng_buffer_view *payload)
e2acc8d2
JD		 2792{
2793 struct relay_session *session = conn->session;
2794 struct lttcomm_relayd_rotate_pending msg;
d88744a4 2795 struct lttcomm_relayd_rotate_pending_reply reply;
e2acc8d2
JD		 2796 struct lttng_ht_iter iter;
2797 struct relay_stream *stream;
4f5fb4c3 2798 int ret = 0;
5312a3ed 2799 ssize_t send_ret;
e2acc8d2
JD		 2800 uint64_t chunk_id;
2801 bool rotate_pending = false;
2802
2803 DBG("Rotate pending command received");
2804
2805 if (!session || !conn->version_check_done) {
2806 ERR("Trying to check for data before version check");
2807 ret = -1;
2808 goto end_no_reply;
2809 }
2810
2811 if (session->major == 2 && session->minor < 11) {
2812 ERR("Unsupported feature before 2.11");
2813 ret = -1;
2814 goto end_no_reply;
2815 }
2816
5312a3ed
JG		 2817 if (payload->size < sizeof(msg)) {
2818 ERR("Unexpected payload size in \"relay_rotate_pending\": expected >= %zu bytes, got %zu bytes",
2819 sizeof(msg), payload->size);
e2acc8d2
JD		 2820 ret = -1;
2821 goto end_no_reply;
2822 }
2823
5312a3ed
JG		 2824 memcpy(&msg, payload->data, sizeof(msg));
2825
e2acc8d2 2826 chunk_id = be64toh(msg.chunk_id);
5312a3ed 2827
ce6877c9
JG		 2828 DBG("Evaluating rotate pending for session \"%s\" and chunk id %" PRIu64,
2829 session->session_name, chunk_id);
e2acc8d2
JD		 2830
2831 /*
2832 * Iterate over all the streams in the session and check if they are
2833 * still waiting for data to perform their rotation.
2834 */
2835 rcu_read_lock();
2836 cds_lfht_for_each_entry(relay_streams_ht->ht, &iter.iter, stream,
2837 node.node) {
2838 if (!stream_get(stream)) {
2839 continue;
2840 }
2841 if (stream->trace->session != session) {
2842 stream_put(stream);
2843 continue;
2844 }
2845 pthread_mutex_lock(&stream->lock);
2846 if (stream->rotate_at_seq_num != -1ULL) {
2847 /* We have not yet performed the rotation. */
2848 rotate_pending = true;
2849 DBG("Stream %" PRIu64 " is still rotating",
2850 stream->stream_handle);
81164b6b 2851 } else if (stream->current_chunk_id.value < chunk_id) {
e2acc8d2
JD		 2852 /*
2853 * Stream closed on the consumer but still active on the
2854 * relay.
2855 */
2856 rotate_pending = true;
2857 DBG("Stream %" PRIu64 " did not exist on the consumer "
2858 "when the last rotation started, but is"
2859 "still waiting for data before getting"
2860 "closed",
2861 stream->stream_handle);
2862 }
2863 pthread_mutex_unlock(&stream->lock);
2864 stream_put(stream);
2865 if (rotate_pending) {
2866 goto send_reply;
2867 }
2868 }
2869
2870send_reply:
2871 rcu_read_unlock();
2872 memset(&reply, 0, sizeof(reply));
d88744a4
JD		 2873 reply.generic.ret_code = htobe32((uint32_t) LTTNG_OK);
2874 reply.is_pending = (uint8_t) !!rotate_pending;
5312a3ed 2875 send_ret = conn->sock->ops->sendmsg(conn->sock, &reply,
e2acc8d2 2876 sizeof(reply), 0);
5312a3ed
JG		 2877 if (send_ret < (ssize_t) sizeof(reply)) {
2878 ERR("Failed to send \"rotate pending\" command reply (ret = %zd)",
2879 send_ret);
e2acc8d2
JD		 2880 ret = -1;
2881 }
2882
2883end_no_reply:
2884 return ret;
2885}
2886
5312a3ed
JG		 2887#define DBG_CMD(cmd_name, conn) \
2888 DBG3("Processing \"%s\" command for socket %i", cmd_name, conn->sock->fd);
2889
2890static int relay_process_control_command(struct relay_connection *conn,
2891 const struct lttcomm_relayd_hdr *header,
2892 const struct lttng_buffer_view *payload)
b8aa1682
JD		 2893{
2894 int ret = 0;
2895
5312a3ed 2896 switch (header->cmd) {
b8aa1682 2897 case RELAYD_CREATE_SESSION:
5312a3ed
JG		 2898 DBG_CMD("RELAYD_CREATE_SESSION", conn);
2899 ret = relay_create_session(header, conn, payload);
b8aa1682 2900 break;
b8aa1682 2901 case RELAYD_ADD_STREAM:
5312a3ed
JG		 2902 DBG_CMD("RELAYD_ADD_STREAM", conn);
2903 ret = relay_add_stream(header, conn, payload);
b8aa1682
JD		 2904 break;
2905 case RELAYD_START_DATA:
5312a3ed
JG		 2906 DBG_CMD("RELAYD_START_DATA", conn);
2907 ret = relay_start(header, conn, payload);
b8aa1682
JD		 2908 break;
2909 case RELAYD_SEND_METADATA:
5312a3ed
JG		 2910 DBG_CMD("RELAYD_SEND_METADATA", conn);
2911 ret = relay_recv_metadata(header, conn, payload);
b8aa1682
JD		 2912 break;
2913 case RELAYD_VERSION:
5312a3ed
JG		 2914 DBG_CMD("RELAYD_VERSION", conn);
2915 ret = relay_send_version(header, conn, payload);
b8aa1682 2916 break;
173af62f 2917 case RELAYD_CLOSE_STREAM:
5312a3ed
JG		 2918 DBG_CMD("RELAYD_CLOSE_STREAM", conn);
2919 ret = relay_close_stream(header, conn, payload);
173af62f 2920 break;
6d805429 2921 case RELAYD_DATA_PENDING:
5312a3ed
JG		 2922 DBG_CMD("RELAYD_DATA_PENDING", conn);
2923 ret = relay_data_pending(header, conn, payload);
c8f59ee5
DG		 2924 break;
2925 case RELAYD_QUIESCENT_CONTROL:
5312a3ed
JG		 2926 DBG_CMD("RELAYD_QUIESCENT_CONTROL", conn);
2927 ret = relay_quiescent_control(header, conn, payload);
c8f59ee5 2928 break;
f7079f67 2929 case RELAYD_BEGIN_DATA_PENDING:
5312a3ed
JG		 2930 DBG_CMD("RELAYD_BEGIN_DATA_PENDING", conn);
2931 ret = relay_begin_data_pending(header, conn, payload);
f7079f67
DG		 2932 break;
2933 case RELAYD_END_DATA_PENDING:
5312a3ed
JG		 2934 DBG_CMD("RELAYD_END_DATA_PENDING", conn);
2935 ret = relay_end_data_pending(header, conn, payload);
f7079f67 2936 break;
1c20f0e2 2937 case RELAYD_SEND_INDEX:
5312a3ed
JG		 2938 DBG_CMD("RELAYD_SEND_INDEX", conn);
2939 ret = relay_recv_index(header, conn, payload);
1c20f0e2 2940 break;
a4baae1b 2941 case RELAYD_STREAMS_SENT:
5312a3ed
JG		 2942 DBG_CMD("RELAYD_STREAMS_SENT", conn);
2943 ret = relay_streams_sent(header, conn, payload);
a4baae1b 2944 break;
93ec662e 2945 case RELAYD_RESET_METADATA:
5312a3ed
JG		 2946 DBG_CMD("RELAYD_RESET_METADATA", conn);
2947 ret = relay_reset_metadata(header, conn, payload);
93ec662e 2948 break;
d3ecc550 2949 case RELAYD_ROTATE_STREAM:
5312a3ed
JG		 2950 DBG_CMD("RELAYD_ROTATE_STREAM", conn);
2951 ret = relay_rotate_session_stream(header, conn, payload);
d3ecc550 2952 break;
00fb02ac 2953 case RELAYD_ROTATE_RENAME:
5312a3ed
JG		 2954 DBG_CMD("RELAYD_ROTATE_RENAME", conn);
2955 ret = relay_rotate_rename(header, conn, payload);
00fb02ac 2956 break;
e2acc8d2 2957 case RELAYD_ROTATE_PENDING:
5312a3ed
JG		 2958 DBG_CMD("RELAYD_ROTATE_PENDING", conn);
2959 ret = relay_rotate_pending(header, conn, payload);
e2acc8d2 2960 break;
a1ae2ea5 2961 case RELAYD_MKDIR:
5312a3ed
JG		 2962 DBG_CMD("RELAYD_MKDIR", conn);
2963 ret = relay_mkdir(header, conn, payload);
a1ae2ea5 2964 break;
b8aa1682
JD		 2965 case RELAYD_UPDATE_SYNC_INFO:
2966 default:
5312a3ed 2967 ERR("Received unknown command (%u)", header->cmd);
58eb9381 2968 relay_unknown_command(conn);
b8aa1682
JD		 2969 ret = -1;
2970 goto end;
2971 }
2972
2973end:
2974 return ret;
2975}
2976
5569b118
JG		 2977static enum relay_connection_status relay_process_control_receive_payload(
2978 struct relay_connection *conn)
5312a3ed
JG		 2979{
2980 int ret = 0;
5569b118 2981 enum relay_connection_status status = RELAY_CONNECTION_STATUS_OK;
5312a3ed
JG		 2982 struct lttng_dynamic_buffer *reception_buffer =
2983 &conn->protocol.ctrl.reception_buffer;
2984 struct ctrl_connection_state_receive_payload *state =
2985 &conn->protocol.ctrl.state.receive_payload;
2986 struct lttng_buffer_view payload_view;
2987
2988 if (state->left_to_receive == 0) {
2989 /* Short-circuit for payload-less commands. */
2990 goto reception_complete;
2991 }
2992
2993 ret = conn->sock->ops->recvmsg(conn->sock,
2994 reception_buffer->data + state->received,
2995 state->left_to_receive, MSG_DONTWAIT);
2996 if (ret < 0) {
5569b118
JG		 2997 if (errno != EAGAIN && errno != EWOULDBLOCK) {
2998 PERROR("Unable to receive command payload on sock %d",
2999 conn->sock->fd);
3000 status = RELAY_CONNECTION_STATUS_ERROR;
3001 }
5312a3ed
JG		 3002 goto end;
3003 } else if (ret == 0) {
3004 DBG("Socket %d performed an orderly shutdown (received EOF)", conn->sock->fd);
5569b118 3005 status = RELAY_CONNECTION_STATUS_CLOSED;
5312a3ed
JG		 3006 goto end;
3007 }
3008
3009 assert(ret > 0);
3010 assert(ret <= state->left_to_receive);
3011
3012 state->left_to_receive -= ret;
3013 state->received += ret;
3014
3015 if (state->left_to_receive > 0) {
3016 /*
3017 * Can't transition to the protocol's next state, wait to
3018 * receive the rest of the header.
3019 */
3020 DBG3("Partial reception of control connection protocol payload (received %" PRIu64 " bytes, %" PRIu64 " bytes left to receive, fd = %i)",
3021 state->received, state->left_to_receive,
3022 conn->sock->fd);
5312a3ed
JG		 3023 goto end;
3024 }
3025
3026reception_complete:
3027 DBG("Done receiving control command payload: fd = %i, payload size = %" PRIu64 " bytes",
3028 conn->sock->fd, state->received);
3029 /*
3030 * The payload required to process the command has been received.
3031 * A view to the reception buffer is forwarded to the various
3032 * commands and the state of the control is reset on success.
3033 *
3034 * Commands are responsible for sending their reply to the peer.
3035 */
3036 payload_view = lttng_buffer_view_from_dynamic_buffer(reception_buffer,
3037 0, -1);
3038 ret = relay_process_control_command(conn,
3039 &state->header, &payload_view);
3040 if (ret < 0) {
5569b118 3041 status = RELAY_CONNECTION_STATUS_ERROR;
5312a3ed
JG		 3042 goto end;
3043 }
3044
3045 ret = connection_reset_protocol_state(conn);
5569b118
JG		 3046 if (ret) {
3047 status = RELAY_CONNECTION_STATUS_ERROR;
3048 }
5312a3ed 3049end:
5569b118 3050 return status;
5312a3ed
JG		 3051}
3052
5569b118
JG		 3053static enum relay_connection_status relay_process_control_receive_header(
3054 struct relay_connection *conn)
5312a3ed
JG		 3055{
3056 int ret = 0;
5569b118 3057 enum relay_connection_status status = RELAY_CONNECTION_STATUS_OK;
5312a3ed
JG		 3058 struct lttcomm_relayd_hdr header;
3059 struct lttng_dynamic_buffer *reception_buffer =
3060 &conn->protocol.ctrl.reception_buffer;
3061 struct ctrl_connection_state_receive_header *state =
3062 &conn->protocol.ctrl.state.receive_header;
3063
3064 assert(state->left_to_receive != 0);
3065
3066 ret = conn->sock->ops->recvmsg(conn->sock,
3067 reception_buffer->data + state->received,
3068 state->left_to_receive, MSG_DONTWAIT);
3069 if (ret < 0) {
5569b118
JG		 3070 if (errno != EAGAIN && errno != EWOULDBLOCK) {
3071 PERROR("Unable to receive control command header on sock %d",
3072 conn->sock->fd);
3073 status = RELAY_CONNECTION_STATUS_ERROR;
3074 }
5312a3ed
JG		 3075 goto end;
3076 } else if (ret == 0) {
3077 DBG("Socket %d performed an orderly shutdown (received EOF)", conn->sock->fd);
5569b118 3078 status = RELAY_CONNECTION_STATUS_CLOSED;
5312a3ed
JG		 3079 goto end;
3080 }
3081
3082 assert(ret > 0);
3083 assert(ret <= state->left_to_receive);
3084
3085 state->left_to_receive -= ret;
3086 state->received += ret;
3087
3088 if (state->left_to_receive > 0) {
3089 /*
3090 * Can't transition to the protocol's next state, wait to
3091 * receive the rest of the header.
3092 */
3093 DBG3("Partial reception of control connection protocol header (received %" PRIu64 " bytes, %" PRIu64 " bytes left to receive, fd = %i)",
3094 state->received, state->left_to_receive,
3095 conn->sock->fd);
5312a3ed
JG		 3096 goto end;
3097 }
3098
3099 /* Transition to next state: receiving the command's payload. */
3100 conn->protocol.ctrl.state_id =
3101 CTRL_CONNECTION_STATE_RECEIVE_PAYLOAD;
3102 memcpy(&header, reception_buffer->data, sizeof(header));
3103 header.circuit_id = be64toh(header.circuit_id);
3104 header.data_size = be64toh(header.data_size);
3105 header.cmd = be32toh(header.cmd);
3106 header.cmd_version = be32toh(header.cmd_version);
3107 memcpy(&conn->protocol.ctrl.state.receive_payload.header,
3108 &header, sizeof(header));
3109
3110 DBG("Done receiving control command header: fd = %i, cmd = %" PRIu32 ", cmd_version = %" PRIu32 ", payload size = %" PRIu64 " bytes",
3111 conn->sock->fd, header.cmd, header.cmd_version,
3112 header.data_size);
3113
715e6fb1 3114 if (header.data_size > DEFAULT_NETWORK_RELAYD_CTRL_MAX_PAYLOAD_SIZE) {
5312a3ed
JG		 3115 ERR("Command header indicates a payload (%" PRIu64 " bytes) that exceeds the maximal payload size allowed on a control connection.",
3116 header.data_size);
5569b118 3117 status = RELAY_CONNECTION_STATUS_ERROR;
5312a3ed
JG		 3118 goto end;
3119 }
3120
3121 conn->protocol.ctrl.state.receive_payload.left_to_receive =
3122 header.data_size;
3123 conn->protocol.ctrl.state.receive_payload.received = 0;
3124 ret = lttng_dynamic_buffer_set_size(reception_buffer,
3125 header.data_size);
3126 if (ret) {
5569b118 3127 status = RELAY_CONNECTION_STATUS_ERROR;
5312a3ed
JG		 3128 goto end;
3129 }
3130
3131 if (header.data_size == 0) {
3132 /*
3133 * Manually invoke the next state as the poll loop
3134 * will not wake-up to allow us to proceed further.
3135 */
5569b118 3136 status = relay_process_control_receive_payload(conn);
5312a3ed
JG		 3137 }
3138end:
5569b118 3139 return status;
5312a3ed
JG		 3140}
3141
3142/*
3143 * Process the commands received on the control socket
3144 */
5569b118
JG		 3145static enum relay_connection_status relay_process_control(
3146 struct relay_connection *conn)
5312a3ed 3147{
5569b118 3148 enum relay_connection_status status;
5312a3ed
JG		 3149
3150 switch (conn->protocol.ctrl.state_id) {
3151 case CTRL_CONNECTION_STATE_RECEIVE_HEADER:
5569b118 3152 status = relay_process_control_receive_header(conn);
5312a3ed
JG		 3153 break;
3154 case CTRL_CONNECTION_STATE_RECEIVE_PAYLOAD:
5569b118 3155 status = relay_process_control_receive_payload(conn);
5312a3ed
JG		 3156 break;
3157 default:
3158 ERR("Unknown control connection protocol state encountered.");
3159 abort();
3160 }
3161
5569b118 3162 return status;
5312a3ed
JG		 3163}
3164
7d2f7452
DG		 3165/*
3166 * Handle index for a data stream.
3167 *
7591bab1 3168 * Called with the stream lock held.
7d2f7452
DG		 3169 *
3170 * Return 0 on success else a negative value.
3171 */
3172static int handle_index_data(struct relay_stream *stream, uint64_t net_seq_num,
5312a3ed 3173 bool rotate_index, bool *flushed, uint64_t total_size)
7d2f7452 3174{
7591bab1
MD		 3175 int ret = 0;
3176 uint64_t data_offset;
3177 struct relay_index *index;
7d2f7452 3178
7d2f7452
DG		 3179 /* Get data offset because we are about to update the index. */
3180 data_offset = htobe64(stream->tracefile_size_current);
3181
65d66b19
MD		 3182 DBG("handle_index_data: stream %" PRIu64 " net_seq_num %" PRIu64 " data offset %" PRIu64,
3183 stream->stream_handle, net_seq_num, stream->tracefile_size_current);
7591bab1 3184
7d2f7452 3185 /*
7591bab1
MD		 3186 * Lookup for an existing index for that stream id/sequence
3187 * number. If it exists, the control thread has already received the
3188 * data for it, thus we need to write it to disk.
7d2f7452 3189 */
7591bab1 3190 index = relay_index_get_by_id_or_create(stream, net_seq_num);
7d2f7452 3191 if (!index) {
7591bab1
MD		 3192 ret = -1;
3193 goto end;
7d2f7452
DG		 3194 }
3195
f8f3885c 3196 if (rotate_index || !stream->index_file) {
d3ecc550
JD		 3197 ret = create_rotate_index_file(stream);
3198 if (ret < 0) {
3199 ERR("Failed to rotate index");
7591bab1
MD		 3200 /* Put self-ref for this index due to error. */
3201 relay_index_put(index);
f8f3885c 3202 index = NULL;
7591bab1 3203 goto end;
7d2f7452 3204 }
7d2f7452
DG		 3205 }
3206
f8f3885c 3207 if (relay_index_set_file(index, stream->index_file, data_offset)) {
7591bab1
MD		 3208 ret = -1;
3209 /* Put self-ref for this index due to error. */
3210 relay_index_put(index);
f8f3885c 3211 index = NULL;
7591bab1 3212 goto end;
7d2f7452
DG		 3213 }
3214
7591bab1
MD		 3215 ret = relay_index_try_flush(index);
3216 if (ret == 0) {
a44ca2ca
MD		 3217 tracefile_array_commit_seq(stream->tfa);
3218 stream->index_received_seqcount++;
d3ecc550 3219 *flushed = true;
7591bab1 3220 } else if (ret > 0) {
d3ecc550 3221 index->total_size = total_size;
7591bab1
MD		 3222 /* No flush. */
3223 ret = 0;
3224 } else {
245f8bec
JR		 3225 /*
3226 * ret < 0
3227 *
3228 * relay_index_try_flush is responsible for the self-reference
3229 * put of the index object on error.
3230 */
3231 ERR("relay_index_try_flush error %d", ret);
7591bab1
MD		 3232 ret = -1;
3233 }
3234end:
7d2f7452
DG		 3235 return ret;
3236}
3237
5569b118
JG		 3238static enum relay_connection_status relay_process_data_receive_header(
3239 struct relay_connection *conn)
b8aa1682 3240{
5312a3ed 3241 int ret;
5569b118 3242 enum relay_connection_status status = RELAY_CONNECTION_STATUS_OK;
5312a3ed
JG		 3243 struct data_connection_state_receive_header *state =
3244 &conn->protocol.data.state.receive_header;
3245 struct lttcomm_relayd_data_hdr header;
b8aa1682 3246 struct relay_stream *stream;
5312a3ed
JG		 3247
3248 assert(state->left_to_receive != 0);
3249
3250 ret = conn->sock->ops->recvmsg(conn->sock,
3251 state->header_reception_buffer + state->received,
3252 state->left_to_receive, MSG_DONTWAIT);
3253 if (ret < 0) {
5569b118
JG		 3254 if (errno != EAGAIN && errno != EWOULDBLOCK) {
3255 PERROR("Unable to receive data header on sock %d", conn->sock->fd);
3256 status = RELAY_CONNECTION_STATUS_ERROR;
3257 }
5312a3ed
JG		 3258 goto end;
3259 } else if (ret == 0) {
3260 /* Orderly shutdown. Not necessary to print an error. */
3261 DBG("Socket %d performed an orderly shutdown (received EOF)", conn->sock->fd);
5569b118 3262 status = RELAY_CONNECTION_STATUS_CLOSED;
b8aa1682
JD		 3263 goto end;
3264 }
3265
5312a3ed
JG		 3266 assert(ret > 0);
3267 assert(ret <= state->left_to_receive);
3268
3269 state->left_to_receive -= ret;
3270 state->received += ret;
3271
3272 if (state->left_to_receive > 0) {
3273 /*
3274 * Can't transition to the protocol's next state, wait to
3275 * receive the rest of the header.
3276 */
3277 DBG3("Partial reception of data connection header (received %" PRIu64 " bytes, %" PRIu64 " bytes left to receive, fd = %i)",
3278 state->received, state->left_to_receive,
3279 conn->sock->fd);
7591bab1 3280 goto end;
b8aa1682 3281 }
b8aa1682 3282
5312a3ed
JG		 3283 /* Transition to next state: receiving the payload. */
3284 conn->protocol.data.state_id = DATA_CONNECTION_STATE_RECEIVE_PAYLOAD;
173af62f 3285
5312a3ed
JG		 3286 memcpy(&header, state->header_reception_buffer, sizeof(header));
3287 header.circuit_id = be64toh(header.circuit_id);
3288 header.stream_id = be64toh(header.stream_id);
3289 header.data_size = be32toh(header.data_size);
3290 header.net_seq_num = be64toh(header.net_seq_num);
3291 header.padding_size = be32toh(header.padding_size);
3292 memcpy(&conn->protocol.data.state.receive_payload.header, &header, sizeof(header));
3293
3294 conn->protocol.data.state.receive_payload.left_to_receive =
3295 header.data_size;
3296 conn->protocol.data.state.receive_payload.received = 0;
3297 conn->protocol.data.state.receive_payload.rotate_index = false;
3298
3299 DBG("Received data connection header on fd %i: circuit_id = %" PRIu64 ", stream_id = %" PRIu64 ", data_size = %" PRIu32 ", net_seq_num = %" PRIu64 ", padding_size = %" PRIu32,
3300 conn->sock->fd, header.circuit_id,
3301 header.stream_id, header.data_size,
3302 header.net_seq_num, header.padding_size);
3303
3304 stream = stream_get_by_id(header.stream_id);
3305 if (!stream) {
3306 DBG("relay_process_data_receive_payload: Cannot find stream %" PRIu64,
3307 header.stream_id);
5569b118
JG		 3308 /* Protocol error. */
3309 status = RELAY_CONNECTION_STATUS_ERROR;
5312a3ed
JG		 3310 goto end;
3311 }
b8aa1682 3312
7591bab1
MD		 3313 pthread_mutex_lock(&stream->lock);
3314
1c20f0e2 3315 /* Check if a rotation is needed. */
0f907de1 3316 if (stream->tracefile_size > 0 &&
5312a3ed 3317 (stream->tracefile_size_current + header.data_size) >
0f907de1 3318 stream->tracefile_size) {
a44ca2ca
MD		 3319 uint64_t old_id, new_id;
3320
3321 old_id = tracefile_array_get_file_index_head(stream->tfa);
3322 tracefile_array_file_rotate(stream->tfa);
3323
3324 /* new_id is updated by utils_rotate_stream_file. */
3325 new_id = old_id;
6b6b9a5a 3326
7591bab1
MD		 3327 ret = utils_rotate_stream_file(stream->path_name,
3328 stream->channel_name, stream->tracefile_size,
3329 stream->tracefile_count, -1,
3330 -1, stream->stream_fd->fd,
a44ca2ca 3331 &new_id, &stream->stream_fd->fd);
0f907de1 3332 if (ret < 0) {
5312a3ed 3333 ERR("Failed to rotate stream output file");
5569b118 3334 status = RELAY_CONNECTION_STATUS_ERROR;
7591bab1
MD		 3335 goto end_stream_unlock;
3336 }
5312a3ed 3337
7591bab1
MD		 3338 /*
3339 * Reset current size because we just performed a stream
3340 * rotation.
3341 */
a6976990 3342 stream->tracefile_size_current = 0;
5312a3ed 3343 conn->protocol.data.state.receive_payload.rotate_index = true;
1c20f0e2
JD		 3344 }
3345
5312a3ed
JG		 3346end_stream_unlock:
3347 pthread_mutex_unlock(&stream->lock);
3348 stream_put(stream);
3349end:
5569b118 3350 return status;
5312a3ed
JG		 3351}
3352
5569b118
JG		 3353static enum relay_connection_status relay_process_data_receive_payload(
3354 struct relay_connection *conn)
5312a3ed
JG		 3355{
3356 int ret;
5569b118 3357 enum relay_connection_status status = RELAY_CONNECTION_STATUS_OK;
5312a3ed
JG		 3358 struct relay_stream *stream;
3359 struct data_connection_state_receive_payload *state =
3360 &conn->protocol.data.state.receive_payload;
3361 const size_t chunk_size = RECV_DATA_BUFFER_SIZE;
3362 char data_buffer[chunk_size];
3363 bool partial_recv = false;
3364 bool new_stream = false, close_requested = false, index_flushed = false;
3365 uint64_t left_to_receive = state->left_to_receive;
3366 struct relay_session *session;
3367
fd0f1e3e
JR		 3368 DBG3("Receiving data for stream id %" PRIu64 " seqnum %" PRIu64 ", %" PRIu64" bytes received, %" PRIu64 " bytes left to receive",
3369 state->header.stream_id, state->header.net_seq_num,
3370 state->received, left_to_receive);
3371
5312a3ed
JG		 3372 stream = stream_get_by_id(state->header.stream_id);
3373 if (!stream) {
5569b118 3374 /* Protocol error. */
fd0f1e3e 3375 ERR("relay_process_data_receive_payload: cannot find stream %" PRIu64,
5312a3ed 3376 state->header.stream_id);
5569b118 3377 status = RELAY_CONNECTION_STATUS_ERROR;
5312a3ed 3378 goto end;
1c20f0e2
JD		 3379 }
3380
5312a3ed
JG		 3381 pthread_mutex_lock(&stream->lock);
3382 session = stream->trace->session;
fd0f1e3e
JR		 3383 if (!conn->session) {
3384 ret = connection_set_session(conn, session);
3385 if (ret) {
3386 status = RELAY_CONNECTION_STATUS_ERROR;
3387 goto end_stream_unlock;
3388 }
3389 }
5312a3ed
JG		 3390
3391 /*
3392 * The size of the "chunk" received on any iteration is bounded by:
3393 * - the data left to receive,
3394 * - the data immediately available on the socket,
3395 * - the on-stack data buffer
3396 */
3397 while (left_to_receive > 0 && !partial_recv) {
3398 ssize_t write_ret;
3399 size_t recv_size = min(left_to_receive, chunk_size);
3400
3401 ret = conn->sock->ops->recvmsg(conn->sock, data_buffer,
3402 recv_size, MSG_DONTWAIT);
3403 if (ret < 0) {
5569b118
JG		 3404 if (errno != EAGAIN && errno != EWOULDBLOCK) {
3405 PERROR("Socket %d error", conn->sock->fd);
3406 status = RELAY_CONNECTION_STATUS_ERROR;
3407 }
0848dba7 3408 goto end_stream_unlock;
5312a3ed
JG		 3409 } else if (ret == 0) {
3410 /* No more data ready to be consumed on socket. */
3411 DBG3("No more data ready for consumption on data socket of stream id %" PRIu64,
3412 state->header.stream_id);
5569b118 3413 status = RELAY_CONNECTION_STATUS_CLOSED;
5312a3ed
JG		 3414 break;
3415 } else if (ret < (int) recv_size) {
3416 /*
3417 * All the data available on the socket has been
3418 * consumed.
3419 */
3420 partial_recv = true;
0848dba7
MD		 3421 }
3422
5312a3ed
JG		 3423 recv_size = ret;
3424
0848dba7 3425 /* Write data to stream output fd. */
5312a3ed 3426 write_ret = lttng_write(stream->stream_fd->fd, data_buffer,
0848dba7 3427 recv_size);
5312a3ed 3428 if (write_ret < (ssize_t) recv_size) {
0848dba7 3429 ERR("Relay error writing data to file");
5569b118 3430 status = RELAY_CONNECTION_STATUS_ERROR;
0848dba7
MD		 3431 goto end_stream_unlock;
3432 }
3433
5312a3ed
JG		 3434 left_to_receive -= recv_size;
3435 state->received += recv_size;
3436 state->left_to_receive = left_to_receive;
3437
0848dba7 3438 DBG2("Relay wrote %zd bytes to tracefile for stream id %" PRIu64,
5312a3ed
JG		 3439 write_ret, stream->stream_handle);
3440 }
3441
3442 if (state->left_to_receive > 0) {
3443 /*
3444 * Did not receive all the data expected, wait for more data to
3445 * become available on the socket.
3446 */
3447 DBG3("Partial receive on data connection of stream id %" PRIu64 ", %" PRIu64 " bytes received, %" PRIu64 " bytes left to receive",
3448 state->header.stream_id, state->received,
3449 state->left_to_receive);
5312a3ed 3450 goto end_stream_unlock;
0848dba7 3451 }
5ab7344e 3452
7591bab1 3453 ret = write_padding_to_file(stream->stream_fd->fd,
5312a3ed 3454 state->header.padding_size);
46f4eaa5 3455 if ((int64_t) ret < (int64_t) state->header.padding_size) {
65d66b19 3456 ERR("write_padding_to_file: fail stream %" PRIu64 " net_seq_num %" PRIu64 " ret %d",
5312a3ed
JG		 3457 stream->stream_handle,
3458 state->header.net_seq_num, ret);
5569b118 3459 status = RELAY_CONNECTION_STATUS_ERROR;
7591bab1 3460 goto end_stream_unlock;
1d4dfdef 3461 }
5312a3ed
JG		 3462
3463
3464 if (session->minor >= 4 && !session->snapshot) {
3465 ret = handle_index_data(stream, state->header.net_seq_num,
3466 state->rotate_index, &index_flushed, state->header.data_size + state->header.padding_size);
3467 if (ret < 0) {
3468 ERR("handle_index_data: fail stream %" PRIu64 " net_seq_num %" PRIu64 " ret %d",
3469 stream->stream_handle,
3470 state->header.net_seq_num, ret);
5569b118 3471 status = RELAY_CONNECTION_STATUS_ERROR;
5312a3ed
JG		 3472 goto end_stream_unlock;
3473 }
3474 }
3475
3476 stream->tracefile_size_current += state->header.data_size +
3477 state->header.padding_size;
3478
c0bae11d
MD		 3479 if (stream->prev_seq == -1ULL) {
3480 new_stream = true;
3481 }
d3ecc550
JD		 3482 if (index_flushed) {
3483 stream->pos_after_last_complete_data_index =
3484 stream->tracefile_size_current;
3485 }
c0bae11d 3486
5312a3ed
JG		 3487 stream->prev_seq = state->header.net_seq_num;
3488
3489 /*
3490 * Resetting the protocol state (to RECEIVE_HEADER) will trash the
3491 * contents of *state which are aliased (union) to the same location as
3492 * the new state. Don't use it beyond this point.
3493 */
3494 connection_reset_protocol_state(conn);
3495 state = NULL;
173af62f 3496
d3ecc550
JD		 3497 ret = try_rotate_stream(stream);
3498 if (ret < 0) {
5569b118 3499 status = RELAY_CONNECTION_STATUS_ERROR;
d3ecc550
JD		 3500 goto end_stream_unlock;
3501 }
3502
7591bab1 3503end_stream_unlock:
bda7c7b9 3504 close_requested = stream->close_requested;
7591bab1 3505 pthread_mutex_unlock(&stream->lock);
5312a3ed 3506 if (close_requested && left_to_receive == 0) {
bda7c7b9
JG		 3507 try_stream_close(stream);
3508 }
3509
c0bae11d
MD		 3510 if (new_stream) {
3511 pthread_mutex_lock(&session->lock);
3512 uatomic_set(&session->new_streams, 1);
3513 pthread_mutex_unlock(&session->lock);
3514 }
5312a3ed 3515
7591bab1 3516 stream_put(stream);
b8aa1682 3517end:
5569b118 3518 return status;
b8aa1682
JD		 3519}
3520
5312a3ed
JG		 3521/*
3522 * relay_process_data: Process the data received on the data socket
3523 */
5569b118
JG		 3524static enum relay_connection_status relay_process_data(
3525 struct relay_connection *conn)
5312a3ed 3526{
5569b118 3527 enum relay_connection_status status;
5312a3ed
JG		 3528
3529 switch (conn->protocol.data.state_id) {
3530 case DATA_CONNECTION_STATE_RECEIVE_HEADER:
5569b118 3531 status = relay_process_data_receive_header(conn);
5312a3ed
JG		 3532 break;
3533 case DATA_CONNECTION_STATE_RECEIVE_PAYLOAD:
5569b118 3534 status = relay_process_data_receive_payload(conn);
5312a3ed
JG		 3535 break;
3536 default:
3537 ERR("Unexpected data connection communication state.");
3538 abort();
3539 }
3540
5569b118 3541 return status;
5312a3ed
JG		 3542}
3543
7591bab1 3544static void cleanup_connection_pollfd(struct lttng_poll_event *events, int pollfd)
b8aa1682
JD		 3545{
3546 int ret;
3547
58eb9381 3548 (void) lttng_poll_del(events, pollfd);
b8aa1682
JD		 3549
3550 ret = close(pollfd);
3551 if (ret < 0) {
3552 ERR("Closing pollfd %d", pollfd);
3553 }
3554}
3555
7591bab1
MD		 3556static void relay_thread_close_connection(struct lttng_poll_event *events,
3557 int pollfd, struct relay_connection *conn)
9d1bbf21 3558{
7591bab1 3559 const char *type_str;
2a174661 3560
7591bab1
MD		 3561 switch (conn->type) {
3562 case RELAY_DATA:
3563 type_str = "Data";
3564 break;
3565 case RELAY_CONTROL:
3566 type_str = "Control";
3567 break;
3568 case RELAY_VIEWER_COMMAND:
3569 type_str = "Viewer Command";
3570 break;
3571 case RELAY_VIEWER_NOTIFICATION:
3572 type_str = "Viewer Notification";
3573 break;
3574 default:
3575 type_str = "Unknown";
9d1bbf21 3576 }
7591bab1
MD		 3577 cleanup_connection_pollfd(events, pollfd);
3578 connection_put(conn);
3579 DBG("%s connection closed with %d", type_str, pollfd);
b8aa1682
JD		 3580}
3581
3582/*
3583 * This thread does the actual work
3584 */
7591bab1 3585static void *relay_thread_worker(void *data)
b8aa1682 3586{
beaad64c
DG		 3587 int ret, err = -1, last_seen_data_fd = -1;
3588 uint32_t nb_fd;
b8aa1682
JD		 3589 struct lttng_poll_event events;
3590 struct lttng_ht *relay_connections_ht;
b8aa1682 3591 struct lttng_ht_iter iter;
90e7d72f 3592 struct relay_connection *destroy_conn = NULL;
b8aa1682
JD		 3593
3594 DBG("[thread] Relay worker started");
3595
9d1bbf21
MD		 3596 rcu_register_thread();
3597
55706a7d
MD		 3598 health_register(health_relayd, HEALTH_RELAYD_TYPE_WORKER);
3599
9b5e0863
MD		 3600 if (testpoint(relayd_thread_worker)) {
3601 goto error_testpoint;
3602 }
3603
f385ae0a
MD		 3604 health_code_update();
3605
b8aa1682
JD		 3606 /* table of connections indexed on socket */
3607 relay_connections_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG);
095a4ae5
MD		 3608 if (!relay_connections_ht) {
3609 goto relay_connections_ht_error;
3610 }
b8aa1682 3611
b8aa1682
JD		 3612 ret = create_thread_poll_set(&events, 2);
3613 if (ret < 0) {
3614 goto error_poll_create;
3615 }
3616
58eb9381 3617 ret = lttng_poll_add(&events, relay_conn_pipe[0], LPOLLIN | LPOLLRDHUP);
b8aa1682
JD		 3618 if (ret < 0) {
3619 goto error;
3620 }
3621
beaad64c 3622restart:
b8aa1682 3623 while (1) {
beaad64c
DG		 3624 int idx = -1, i, seen_control = 0, last_notdel_data_fd = -1;
3625
f385ae0a
MD		 3626 health_code_update();
3627
b8aa1682 3628 /* Infinite blocking call, waiting for transmission */
87c1611d 3629 DBG3("Relayd worker thread polling...");
f385ae0a 3630 health_poll_entry();
b8aa1682 3631 ret = lttng_poll_wait(&events, -1);
f385ae0a 3632 health_poll_exit();
b8aa1682
JD		 3633 if (ret < 0) {
3634 /*
3635 * Restart interrupted system call.
3636 */
3637 if (errno == EINTR) {
3638 goto restart;
3639 }
3640 goto error;
3641 }
3642
0d9c5d77
DG		 3643 nb_fd = ret;
3644
beaad64c 3645 /*
7591bab1
MD		 3646 * Process control. The control connection is
3647 * prioritized so we don't starve it with high
3648 * throughput tracing data on the data connection.
beaad64c 3649 */
b8aa1682
JD		 3650 for (i = 0; i < nb_fd; i++) {
3651 /* Fetch once the poll data */
beaad64c
DG		 3652 uint32_t revents = LTTNG_POLL_GETEV(&events, i);
3653 int pollfd = LTTNG_POLL_GETFD(&events, i);
b8aa1682 3654
f385ae0a
MD		 3655 health_code_update();
3656
fd20dac9 3657 if (!revents) {
7591bab1
MD		 3658 /*
3659 * No activity for this FD (poll
3660 * implementation).
3661 */
fd20dac9
MD		 3662 continue;
3663 }
3664
b8aa1682
JD		 3665 /* Thread quit pipe has been closed. Killing thread. */
3666 ret = check_thread_quit_pipe(pollfd, revents);
3667 if (ret) {
095a4ae5
MD		 3668 err = 0;
3669 goto exit;
b8aa1682
JD		 3670 }
3671
58eb9381
DG		 3672 /* Inspect the relay conn pipe for new connection */
3673 if (pollfd == relay_conn_pipe[0]) {
03e43155 3674 if (revents & LPOLLIN) {
90e7d72f
JG		 3675 struct relay_connection *conn;
3676
58eb9381 3677 ret = lttng_read(relay_conn_pipe[0], &conn, sizeof(conn));
b8aa1682
JD		 3678 if (ret < 0) {
3679 goto error;
3680 }
58eb9381
DG		 3681 lttng_poll_add(&events, conn->sock->fd,
3682 LPOLLIN | LPOLLRDHUP);
7591bab1 3683 connection_ht_add(relay_connections_ht, conn);
58eb9381 3684 DBG("Connection socket %d added", conn->sock->fd);
03e43155
MD		 3685 } else if (revents & (LPOLLERR | LPOLLHUP | LPOLLRDHUP)) {
3686 ERR("Relay connection pipe error");
3687 goto error;
3688 } else {
3689 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
3690 goto error;
b8aa1682 3691 }
58eb9381 3692 } else {
90e7d72f
JG		 3693 struct relay_connection *ctrl_conn;
3694
7591bab1 3695 ctrl_conn = connection_get_by_sock(relay_connections_ht, pollfd);
58eb9381 3696 /* If not found, there is a synchronization issue. */
90e7d72f 3697 assert(ctrl_conn);
58eb9381 3698
03e43155
MD		 3699 if (ctrl_conn->type == RELAY_DATA) {
3700 if (revents & LPOLLIN) {
beaad64c
DG		 3701 /*
3702 * Flag the last seen data fd not deleted. It will be
3703 * used as the last seen fd if any fd gets deleted in
3704 * this first loop.
3705 */
3706 last_notdel_data_fd = pollfd;
3707 }
03e43155
MD		 3708 goto put_ctrl_connection;
3709 }
3710 assert(ctrl_conn->type == RELAY_CONTROL);
3711
3712 if (revents & LPOLLIN) {
5569b118
JG		 3713 enum relay_connection_status status;
3714
3715 status = relay_process_control(ctrl_conn);
3716 if (status != RELAY_CONNECTION_STATUS_OK) {
fd0f1e3e
JR		 3717 /*
3718 * On socket error flag the session as aborted to force
3719 * the cleanup of its stream otherwise it can leak
3720 * during the lifetime of the relayd.
3721 *
3722 * This prevents situations in which streams can be
3723 * left opened because an index was received, the
3724 * control connection is closed, and the data
3725 * connection is closed (uncleanly) before the packet's
3726 * data provided.
3727 *
3728 * Since the control connection encountered an error,
3729 * it is okay to be conservative and close the
3730 * session right now as we can't rely on the protocol
3731 * being respected anymore.
3732 */
3733 if (status == RELAY_CONNECTION_STATUS_ERROR) {
3734 session_abort(ctrl_conn->session);
3735 }
3736
5569b118 3737 /* Clear the connection on error or close. */
5312a3ed
JG		 3738 relay_thread_close_connection(&events,
3739 pollfd,
03e43155 3740 ctrl_conn);
03e43155 3741 }
5312a3ed 3742 seen_control = 1;
03e43155
MD		 3743 } else if (revents & (LPOLLERR | LPOLLHUP | LPOLLRDHUP)) {
3744 relay_thread_close_connection(&events,
3745 pollfd, ctrl_conn);
3746 if (last_seen_data_fd == pollfd) {
3747 last_seen_data_fd = last_notdel_data_fd;
3748 }
58eb9381 3749 } else {
03e43155
MD		 3750 ERR("Unexpected poll events %u for control sock %d",
3751 revents, pollfd);
3752 connection_put(ctrl_conn);
3753 goto error;
beaad64c 3754 }
03e43155 3755 put_ctrl_connection:
7591bab1 3756 connection_put(ctrl_conn);
beaad64c
DG		 3757 }
3758 }
3759
3760 /*
3761 * The last loop handled a control request, go back to poll to make
3762 * sure we prioritise the control socket.
3763 */
3764 if (seen_control) {
3765 continue;
3766 }
3767
3768 if (last_seen_data_fd >= 0) {
3769 for (i = 0; i < nb_fd; i++) {
3770 int pollfd = LTTNG_POLL_GETFD(&events, i);
f385ae0a
MD		 3771
3772 health_code_update();
3773
beaad64c
DG		 3774 if (last_seen_data_fd == pollfd) {
3775 idx = i;
3776 break;
3777 }
3778 }
3779 }
3780
3781 /* Process data connection. */
3782 for (i = idx + 1; i < nb_fd; i++) {
3783 /* Fetch the poll data. */
3784 uint32_t revents = LTTNG_POLL_GETEV(&events, i);
3785 int pollfd = LTTNG_POLL_GETFD(&events, i);
90e7d72f 3786 struct relay_connection *data_conn;
beaad64c 3787
f385ae0a
MD		 3788 health_code_update();
3789
fd20dac9
MD		 3790 if (!revents) {
3791 /* No activity for this FD (poll implementation). */
3792 continue;
3793 }
3794
beaad64c 3795 /* Skip the command pipe. It's handled in the first loop. */
58eb9381 3796 if (pollfd == relay_conn_pipe[0]) {
beaad64c
DG		 3797 continue;
3798 }
3799
7591bab1 3800 data_conn = connection_get_by_sock(relay_connections_ht, pollfd);
90e7d72f 3801 if (!data_conn) {
fd20dac9 3802 /* Skip it. Might be removed before. */
fd20dac9
MD		 3803 continue;
3804 }
03e43155
MD		 3805 if (data_conn->type == RELAY_CONTROL) {
3806 goto put_data_connection;
3807 }
3808 assert(data_conn->type == RELAY_DATA);
fd20dac9
MD		 3809
3810 if (revents & LPOLLIN) {
5569b118
JG		 3811 enum relay_connection_status status;
3812
3813 status = relay_process_data(data_conn);
3814 /* Connection closed or error. */
3815 if (status != RELAY_CONNECTION_STATUS_OK) {
fd0f1e3e
JR		 3816 /*
3817 * On socket error flag the session as aborted to force
3818 * the cleanup of its stream otherwise it can leak
3819 * during the lifetime of the relayd.
3820 *
3821 * This prevents situations in which streams can be
3822 * left opened because an index was received, the
3823 * control connection is closed, and the data
3824 * connection is closed (uncleanly) before the packet's
3825 * data provided.
3826 *
3827 * Since the data connection encountered an error,
3828 * it is okay to be conservative and close the
3829 * session right now as we can't rely on the protocol
3830 * being respected anymore.
3831 */
3832 if (status == RELAY_CONNECTION_STATUS_ERROR) {
3833 session_abort(data_conn->session);
3834 }
7591bab1 3835 relay_thread_close_connection(&events, pollfd,
03e43155 3836 data_conn);
fd20dac9
MD		 3837 /*
3838 * Every goto restart call sets the last seen fd where
3839 * here we don't really care since we gracefully
3840 * continue the loop after the connection is deleted.
3841 */
3842 } else {
3843 /* Keep last seen port. */
3844 last_seen_data_fd = pollfd;
7591bab1 3845 connection_put(data_conn);
fd20dac9 3846 goto restart;
b8aa1682 3847 }
03e43155
MD		 3848 } else if (revents & (LPOLLERR | LPOLLHUP | LPOLLRDHUP)) {
3849 relay_thread_close_connection(&events, pollfd,
3850 data_conn);
3851 } else {
3852 ERR("Unknown poll events %u for data sock %d",
3853 revents, pollfd);
b8aa1682 3854 }
03e43155 3855 put_data_connection:
7591bab1 3856 connection_put(data_conn);
b8aa1682 3857 }
beaad64c 3858 last_seen_data_fd = -1;
b8aa1682
JD		 3859 }
3860
f385ae0a
MD		 3861 /* Normal exit, no error */
3862 ret = 0;
3863
095a4ae5 3864exit:
b8aa1682 3865error:
71efa8ef 3866 /* Cleanup remaining connection object. */
9d1bbf21 3867 rcu_read_lock();
90e7d72f
JG		 3868 cds_lfht_for_each_entry(relay_connections_ht->ht, &iter.iter,
3869 destroy_conn,
58eb9381 3870 sock_n.node) {
f385ae0a 3871 health_code_update();
98ba050e 3872
fd0f1e3e 3873 session_abort(destroy_conn->session);
98ba050e 3874
7591bab1
MD		 3875 /*
3876 * No need to grab another ref, because we own
3877 * destroy_conn.
3878 */
3879 relay_thread_close_connection(&events, destroy_conn->sock->fd,
3880 destroy_conn);
b8aa1682 3881 }
94d49140 3882 rcu_read_unlock();
7591bab1
MD		 3883
3884 lttng_poll_clean(&events);
7d2f7452 3885error_poll_create:
b8aa1682 3886 lttng_ht_destroy(relay_connections_ht);
095a4ae5 3887relay_connections_ht_error:
58eb9381
DG		 3888 /* Close relay conn pipes */
3889 utils_close_pipe(relay_conn_pipe);
095a4ae5
MD		 3890 if (err) {
3891 DBG("Thread exited with error");
3892 }
b8aa1682 3893 DBG("Worker thread cleanup complete");
9b5e0863 3894error_testpoint:
f385ae0a
MD		 3895 if (err) {
3896 health_error();
3897 ERR("Health error occurred in %s", __func__);
3898 }
3899 health_unregister(health_relayd);
9d1bbf21 3900 rcu_unregister_thread();
b4aacfdc 3901 lttng_relay_stop_threads();
b8aa1682
JD		 3902 return NULL;
3903}
3904
3905/*
3906 * Create the relay command pipe to wake thread_manage_apps.
3907 * Closed in cleanup().
3908 */
58eb9381 3909static int create_relay_conn_pipe(void)
b8aa1682 3910{
a02de639 3911 int ret;
b8aa1682 3912
58eb9381 3913 ret = utils_create_pipe_cloexec(relay_conn_pipe);
b8aa1682 3914
b8aa1682
JD		 3915 return ret;
3916}
3917
3918/*
3919 * main
3920 */
3921int main(int argc, char **argv)
3922{
178a0557 3923 int ret = 0, retval = 0;
b8aa1682
JD		 3924 void *status;
3925
b8aa1682
JD		 3926 /* Parse arguments */
3927 progname = argv[0];
178a0557
MD		 3928 if (set_options(argc, argv)) {
3929 retval = -1;
3930 goto exit_options;
b8aa1682
JD		 3931 }
3932
178a0557
MD		 3933 if (set_signal_handler()) {
3934 retval = -1;
3935 goto exit_options;
b8aa1682
JD		 3936 }
3937
4d513a50
DG		 3938 /* Try to create directory if -o, --output is specified. */
3939 if (opt_output_path) {
994fa64f
DG		 3940 if (*opt_output_path != '/') {
3941 ERR("Please specify an absolute path for -o, --output PATH");
178a0557
MD		 3942 retval = -1;
3943 goto exit_options;
994fa64f
DG		 3944 }
3945
d77dded2
JG		 3946 ret = utils_mkdir_recursive(opt_output_path, S_IRWXU | S_IRWXG,
3947 -1, -1);
4d513a50
DG		 3948 if (ret < 0) {
3949 ERR("Unable to create %s", opt_output_path);
178a0557
MD		 3950 retval = -1;
3951 goto exit_options;
4d513a50
DG		 3952 }
3953 }
3954
b8aa1682 3955 /* Daemonize */
b5218ffb 3956 if (opt_daemon || opt_background) {
3fd27398
MD		 3957 int i;
3958
3959 ret = lttng_daemonize(&child_ppid, &recv_child_signal,
3960 !opt_background);
b8aa1682 3961 if (ret < 0) {
178a0557
MD		 3962 retval = -1;
3963 goto exit_options;
b8aa1682 3964 }
3fd27398
MD		 3965
3966 /*
3967 * We are in the child. Make sure all other file
3968 * descriptors are closed, in case we are called with
3969 * more opened file descriptors than the standard ones.
3970 */
3971 for (i = 3; i < sysconf(_SC_OPEN_MAX); i++) {
3972 (void) close(i);
3973 }
3974 }
3975
178a0557
MD		 3976 /* Initialize thread health monitoring */
3977 health_relayd = health_app_create(NR_HEALTH_RELAYD_TYPES);
3978 if (!health_relayd) {
3979 PERROR("health_app_create error");
3980 retval = -1;
3981 goto exit_health_app_create;
3982 }
3983
3fd27398 3984 /* Create thread quit pipe */
178a0557
MD		 3985 if (init_thread_quit_pipe()) {
3986 retval = -1;
3987 goto exit_init_data;
b8aa1682
JD		 3988 }
3989
b8aa1682 3990 /* Setup the thread apps communication pipe. */
178a0557
MD		 3991 if (create_relay_conn_pipe()) {
3992 retval = -1;
3993 goto exit_init_data;
b8aa1682
JD		 3994 }
3995
3996 /* Init relay command queue. */
8bdee6e2 3997 cds_wfcq_init(&relay_conn_queue.head, &relay_conn_queue.tail);
b8aa1682 3998
554831e7
MD		 3999 /* Initialize communication library */
4000 lttcomm_init();
87e45c13 4001 lttcomm_inet_init();
554831e7 4002
d3e2ba59 4003 /* tables of sessions indexed by session ID */
7591bab1
MD		 4004 sessions_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
4005 if (!sessions_ht) {
178a0557
MD		 4006 retval = -1;
4007 goto exit_init_data;
d3e2ba59
JD		 4008 }
4009
4010 /* tables of streams indexed by stream ID */
2a174661 4011 relay_streams_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
d3e2ba59 4012 if (!relay_streams_ht) {
178a0557
MD		 4013 retval = -1;
4014 goto exit_init_data;
d3e2ba59
JD		 4015 }
4016
4017 /* tables of streams indexed by stream ID */
92c6ca54
DG		 4018 viewer_streams_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
4019 if (!viewer_streams_ht) {
178a0557
MD		 4020 retval = -1;
4021 goto exit_init_data;
55706a7d
MD		 4022 }
4023
65931c8b 4024 ret = utils_create_pipe(health_quit_pipe);
178a0557
MD		 4025 if (ret) {
4026 retval = -1;
4027 goto exit_health_quit_pipe;
65931c8b
MD		 4028 }
4029
4030 /* Create thread to manage the client socket */
1a1a34b4 4031 ret = pthread_create(&health_thread, default_pthread_attr(),
65931c8b 4032 thread_manage_health, (void *) NULL);
178a0557
MD		 4033 if (ret) {
4034 errno = ret;
65931c8b 4035 PERROR("pthread_create health");
178a0557
MD		 4036 retval = -1;
4037 goto exit_health_thread;
65931c8b
MD		 4038 }
4039
b8aa1682 4040 /* Setup the dispatcher thread */
1a1a34b4 4041 ret = pthread_create(&dispatcher_thread, default_pthread_attr(),
b8aa1682 4042 relay_thread_dispatcher, (void *) NULL);
178a0557
MD		 4043 if (ret) {
4044 errno = ret;
b8aa1682 4045 PERROR("pthread_create dispatcher");
178a0557
MD		 4046 retval = -1;
4047 goto exit_dispatcher_thread;
b8aa1682
JD		 4048 }
4049
4050 /* Setup the worker thread */
1a1a34b4 4051 ret = pthread_create(&worker_thread, default_pthread_attr(),
7591bab1 4052 relay_thread_worker, NULL);
178a0557
MD		 4053 if (ret) {
4054 errno = ret;
b8aa1682 4055 PERROR("pthread_create worker");
178a0557
MD		 4056 retval = -1;
4057 goto exit_worker_thread;
b8aa1682
JD		 4058 }
4059
4060 /* Setup the listener thread */
1a1a34b4 4061 ret = pthread_create(&listener_thread, default_pthread_attr(),
b8aa1682 4062 relay_thread_listener, (void *) NULL);
178a0557
MD		 4063 if (ret) {
4064 errno = ret;
b8aa1682 4065 PERROR("pthread_create listener");
178a0557
MD		 4066 retval = -1;
4067 goto exit_listener_thread;
b8aa1682
JD		 4068 }
4069
7591bab1 4070 ret = relayd_live_create(live_uri);
178a0557 4071 if (ret) {
d3e2ba59 4072 ERR("Starting live viewer threads");
178a0557 4073 retval = -1;
50138f51 4074 goto exit_live;
d3e2ba59
JD		 4075 }
4076
178a0557
MD		 4077 /*
4078 * This is where we start awaiting program completion (e.g. through
4079 * signal that asks threads to teardown).
4080 */
4081
4082 ret = relayd_live_join();
4083 if (ret) {
4084 retval = -1;
4085 }
50138f51 4086exit_live:
178a0557 4087
b8aa1682 4088 ret = pthread_join(listener_thread, &status);
178a0557
MD		 4089 if (ret) {
4090 errno = ret;
4091 PERROR("pthread_join listener_thread");
4092 retval = -1;
b8aa1682
JD		 4093 }
4094
178a0557 4095exit_listener_thread:
b8aa1682 4096 ret = pthread_join(worker_thread, &status);
178a0557
MD		 4097 if (ret) {
4098 errno = ret;
4099 PERROR("pthread_join worker_thread");
4100 retval = -1;
b8aa1682
JD		 4101 }
4102
178a0557 4103exit_worker_thread:
b8aa1682 4104 ret = pthread_join(dispatcher_thread, &status);
178a0557
MD		 4105 if (ret) {
4106 errno = ret;
4107 PERROR("pthread_join dispatcher_thread");
4108 retval = -1;
b8aa1682 4109 }
178a0557 4110exit_dispatcher_thread:
42415026 4111
65931c8b 4112 ret = pthread_join(health_thread, &status);
178a0557
MD		 4113 if (ret) {
4114 errno = ret;
4115 PERROR("pthread_join health_thread");
4116 retval = -1;
65931c8b 4117 }
178a0557 4118exit_health_thread:
65931c8b 4119
65931c8b 4120 utils_close_pipe(health_quit_pipe);
178a0557 4121exit_health_quit_pipe:
65931c8b 4122
178a0557 4123exit_init_data:
55706a7d 4124 health_app_destroy(health_relayd);
55706a7d 4125exit_health_app_create:
178a0557 4126exit_options:
4d62fbf8
MD		 4127 /*
4128 * Wait for all pending call_rcu work to complete before tearing
4129 * down data structures. call_rcu worker may be trying to
4130 * perform lookups in those structures.
4131 */
4132 rcu_barrier();
7591bab1
MD		 4133 relayd_cleanup();
4134
4135 /* Ensure all prior call_rcu are done. */
4136 rcu_barrier();
d3e2ba59 4137
178a0557 4138 if (!retval) {
b8aa1682 4139 exit(EXIT_SUCCESS);
178a0557
MD		 4140 } else {
4141 exit(EXIT_FAILURE);
b8aa1682 4142 }
b8aa1682 4143}
LTTng 2.0 tools and control repository
This page took 0.279971 seconds and 4 git commands to generate.