0aa91f38becb8ab9b0d9a6cd75a6a11278e997d0
[cubemap] / main.cpp
1 #include <assert.h>
2 #include <errno.h>
3 #include <getopt.h>
4 #include <limits.h>
5 #include <signal.h>
6 #include <stddef.h>
7 #include <stdio.h>
8 #include <stdlib.h>
9 #include <string.h>
10 #include <sys/time.h>
11 #include <sys/wait.h>
12 #include <unistd.h>
13 #include <map>
14 #include <set>
15 #include <string>
16 #include <utility>
17 #include <vector>
18
19 #include "acceptor.h"
20 #include "accesslog.h"
21 #include "config.h"
22 #include "input.h"
23 #include "log.h"
24 #include "markpool.h"
25 #include "serverpool.h"
26 #include "state.pb.h"
27 #include "stats.h"
28 #include "stream.h"
29 #include "util.h"
30 #include "version.h"
31
32 using namespace std;
33
34 AccessLogThread *access_log = NULL;
35 ServerPool *servers = NULL;
36 vector<MarkPool *> mark_pools;
37 volatile bool hupped = false;
38 volatile bool stopped = false;
39
40 struct InputWithRefcount {
41         Input *input;
42         int refcount;
43 };
44
45 void hup(int signum)
46 {
47         hupped = true;
48         if (signum == SIGINT) {
49                 stopped = true;
50         }
51 }
52
53 void do_nothing(int signum)
54 {
55 }
56
57 CubemapStateProto collect_state(const timeval &serialize_start,
58                                 const vector<Acceptor *> acceptors,
59                                 const multimap<string, InputWithRefcount> inputs,
60                                 ServerPool *servers)
61 {
62         CubemapStateProto state = servers->serialize();  // Fills streams() and clients().
63         state.set_serialize_start_sec(serialize_start.tv_sec);
64         state.set_serialize_start_usec(serialize_start.tv_usec);
65         
66         for (size_t i = 0; i < acceptors.size(); ++i) {
67                 state.add_acceptors()->MergeFrom(acceptors[i]->serialize());
68         }
69
70         for (multimap<string, InputWithRefcount>::const_iterator input_it = inputs.begin();
71              input_it != inputs.end();
72              ++input_it) {
73                 state.add_inputs()->MergeFrom(input_it->second.input->serialize());
74         }
75
76         return state;
77 }
78
79 // Find all port statements in the configuration file, and create acceptors for htem.
80 vector<Acceptor *> create_acceptors(
81         const Config &config,
82         map<int, Acceptor *> *deserialized_acceptors)
83 {
84         vector<Acceptor *> acceptors;
85         for (unsigned i = 0; i < config.acceptors.size(); ++i) {
86                 const AcceptorConfig &acceptor_config = config.acceptors[i];
87                 Acceptor *acceptor = NULL;
88                 map<int, Acceptor *>::iterator deserialized_acceptor_it =
89                         deserialized_acceptors->find(acceptor_config.port);
90                 if (deserialized_acceptor_it != deserialized_acceptors->end()) {
91                         acceptor = deserialized_acceptor_it->second;
92                         deserialized_acceptors->erase(deserialized_acceptor_it);
93                 } else {
94                         int server_sock = create_server_socket(acceptor_config.port, TCP_SOCKET);
95                         acceptor = new Acceptor(server_sock, acceptor_config.port);
96                 }
97                 acceptor->run();
98                 acceptors.push_back(acceptor);
99         }
100
101         // Close all acceptors that are no longer in the configuration file.
102         for (map<int, Acceptor *>::iterator acceptor_it = deserialized_acceptors->begin();
103              acceptor_it != deserialized_acceptors->end();
104              ++acceptor_it) {
105                 acceptor_it->second->close_socket();
106                 delete acceptor_it->second;
107         }
108
109         return acceptors;
110 }
111
112 // Find all streams in the configuration file, and create inputs for them.
113 void create_config_inputs(const Config &config, multimap<string, InputWithRefcount> *inputs)
114 {
115         for (unsigned i = 0; i < config.streams.size(); ++i) {
116                 const StreamConfig &stream_config = config.streams[i];
117                 if (stream_config.src.empty()) {
118                         continue;
119                 }
120
121                 string src = stream_config.src;
122                 if (inputs->count(src) != 0) {
123                         continue;
124                 }
125
126                 InputWithRefcount iwr;
127                 iwr.input = create_input(src);
128                 if (iwr.input == NULL) {
129                         log(ERROR, "did not understand URL '%s', clients will not get any data.",
130                                 src.c_str());
131                         continue;
132                 }
133                 iwr.refcount = 0;
134                 inputs->insert(make_pair(src, iwr));
135         }
136 }
137
138 void create_streams(const Config &config,
139                     const set<string> &deserialized_stream_ids,
140                     multimap<string, InputWithRefcount> *inputs)
141 {
142         for (unsigned i = 0; i < config.mark_pools.size(); ++i) {
143                 const MarkPoolConfig &mp_config = config.mark_pools[i];
144                 mark_pools.push_back(new MarkPool(mp_config.from, mp_config.to));
145         }
146
147         set<string> expecting_stream_ids = deserialized_stream_ids;
148         for (unsigned i = 0; i < config.streams.size(); ++i) {
149                 const StreamConfig &stream_config = config.streams[i];
150                 if (deserialized_stream_ids.count(stream_config.stream_id) == 0) {
151                         servers->add_stream(stream_config.stream_id,
152                                             stream_config.backlog_size,
153                                             Stream::Encoding(stream_config.encoding));
154                 } else {
155                         servers->set_backlog_size(stream_config.stream_id, stream_config.backlog_size);
156                 }
157                 servers->set_encoding(stream_config.stream_id,
158                                       Stream::Encoding(stream_config.encoding));
159                 expecting_stream_ids.erase(stream_config.stream_id);
160
161                 if (stream_config.mark_pool != -1) {
162                         servers->set_mark_pool(stream_config.stream_id,
163                                                mark_pools[stream_config.mark_pool]);
164                 }
165
166                 string src = stream_config.src;
167                 if (!src.empty()) {
168                         multimap<string, InputWithRefcount>::iterator input_it = inputs->find(src);
169                         assert(input_it != inputs->end());
170                         input_it->second.input->add_destination(stream_config.stream_id);
171                         ++input_it->second.refcount;
172                 }
173         }
174
175         // Warn about any servers we've lost.
176         // TODO: Make an option (delete=yes?) to actually shut down streams.
177         for (set<string>::const_iterator stream_it = expecting_stream_ids.begin();
178              stream_it != expecting_stream_ids.end();
179              ++stream_it) {
180                 string stream_id = *stream_it;
181                 log(WARNING, "stream '%s' disappeared from the configuration file. "
182                              "It will not be deleted, but clients will not get any new inputs.",
183                              stream_id.c_str());
184         }
185 }
186         
187 void open_logs(const vector<LogConfig> &log_destinations)
188 {
189         for (size_t i = 0; i < log_destinations.size(); ++i) {
190                 if (log_destinations[i].type == LogConfig::LOG_TYPE_FILE) {
191                         add_log_destination_file(log_destinations[i].filename);
192                 } else if (log_destinations[i].type == LogConfig::LOG_TYPE_CONSOLE) {
193                         add_log_destination_console();
194                 } else if (log_destinations[i].type == LogConfig::LOG_TYPE_SYSLOG) {
195                         add_log_destination_syslog();
196                 } else {
197                         assert(false);
198                 }
199         }
200         start_logging();
201 }
202         
203 bool dry_run_config(const std::string &argv0, const std::string &config_filename)
204 {
205         char *argv0_copy = strdup(argv0.c_str());
206         char *config_filename_copy = strdup(config_filename.c_str());
207
208         pid_t pid = fork();
209         switch (pid) {
210         case -1:
211                 log_perror("fork()");
212                 free(argv0_copy);
213                 free(config_filename_copy);
214                 return false;
215         case 0:
216                 // Child.
217                 execlp(argv0_copy, argv0_copy, "--test-config", config_filename_copy, NULL);
218                 log_perror(argv0_copy);
219                 _exit(1);
220         default:
221                 // Parent.
222                 break;
223         }
224                 
225         free(argv0_copy);
226         free(config_filename_copy);
227
228         int status;
229         pid_t err;
230         do {
231                 err = waitpid(pid, &status, 0);
232         } while (err == -1 && errno == EINTR);
233
234         if (err == -1) {
235                 log_perror("waitpid()");
236                 return false;
237         }       
238
239         return (WIFEXITED(status) && WEXITSTATUS(status) == 0);
240 }
241
242 int main(int argc, char **argv)
243 {
244         signal(SIGHUP, hup);
245         signal(SIGINT, hup);
246         signal(SIGUSR1, do_nothing);  // Used in internal signalling.
247         signal(SIGPIPE, SIG_IGN);
248         
249         // Parse options.
250         int state_fd = -1;
251         bool test_config = false;
252         for ( ;; ) {
253                 static const option long_options[] = {
254                         { "state", required_argument, 0, 's' },
255                         { "test-config", no_argument, 0, 't' },
256                         { 0, 0, 0, 0 }
257                 };
258                 int option_index = 0;
259                 int c = getopt_long(argc, argv, "s:t", long_options, &option_index);
260      
261                 if (c == -1) {
262                         break;
263                 }
264                 switch (c) {
265                 case 's':
266                         state_fd = atoi(optarg);
267                         break;
268                 case 't':
269                         test_config = true;
270                         break;
271                 default:
272                         fprintf(stderr, "Unknown option '%s'\n", argv[option_index]);
273                         exit(1);
274                 }
275         }
276
277         string config_filename = "cubemap.config";
278         if (optind < argc) {
279                 config_filename = argv[optind++];
280         }
281
282         // Canonicalize argv[0] and config_filename.
283         char argv0_canon[PATH_MAX];
284         char config_filename_canon[PATH_MAX];
285
286         if (realpath(argv[0], argv0_canon) == NULL) {
287                 log_perror(argv[0]);
288                 exit(1);
289         }
290         if (realpath(config_filename.c_str(), config_filename_canon) == NULL) {
291                 log_perror(config_filename.c_str());
292                 exit(1);
293         }
294
295         // Now parse the configuration file.
296         Config config;
297         if (!parse_config(config_filename_canon, &config)) {
298                 exit(1);
299         }
300         if (test_config) {
301                 exit(0);
302         }
303         
304         // Ideally we'd like to daemonize only when we've started up all threads etc.,
305         // but daemon() forks, which is not good in multithreaded software, so we'll
306         // have to do it here.
307         if (config.daemonize) {
308                 if (daemon(0, 0) == -1) {
309                         log_perror("daemon");
310                         exit(1);
311                 }
312         }
313
314 start:
315         // Open logs as soon as possible.
316         open_logs(config.log_destinations);
317
318         log(INFO, "Cubemap " SERVER_VERSION " starting.");
319         if (config.access_log_file.empty()) {
320                 // Create a dummy logger.
321                 access_log = new AccessLogThread();
322         } else {
323                 access_log = new AccessLogThread(config.access_log_file);
324         }
325         access_log->run();
326
327         servers = new ServerPool(config.num_servers);
328
329         CubemapStateProto loaded_state;
330         struct timeval serialize_start;
331         set<string> deserialized_stream_ids;
332         map<int, Acceptor *> deserialized_acceptors;
333         multimap<string, InputWithRefcount> inputs;  // multimap due to older versions without deduplication.
334         if (state_fd != -1) {
335                 log(INFO, "Deserializing state from previous process...");
336                 string serialized;
337                 if (!read_tempfile(state_fd, &serialized)) {
338                         exit(1);
339                 }
340                 if (!loaded_state.ParseFromString(serialized)) {
341                         log(ERROR, "Failed deserialization of state.");
342                         exit(1);
343                 }
344
345                 serialize_start.tv_sec = loaded_state.serialize_start_sec();
346                 serialize_start.tv_usec = loaded_state.serialize_start_usec();
347
348                 // Deserialize the streams.
349                 for (int i = 0; i < loaded_state.streams_size(); ++i) {
350                         const StreamProto &stream = loaded_state.streams(i);
351
352                         vector<int> data_fds;
353                         for (int j = 0; j < stream.data_fds_size(); ++j) {
354                                 data_fds.push_back(stream.data_fds(j));
355                         }
356
357                         // Older versions stored the data once in the protobuf instead of
358                         // sending around file descriptors.
359                         if (data_fds.empty() && stream.has_data()) {
360                                 data_fds.push_back(make_tempfile(stream.data()));
361                         }
362
363                         servers->add_stream_from_serialized(stream, data_fds);
364                         deserialized_stream_ids.insert(stream.stream_id());
365                 }
366
367                 // Deserialize the inputs. Note that we don't actually add them to any stream yet.
368                 for (int i = 0; i < loaded_state.inputs_size(); ++i) {
369                         InputWithRefcount iwr;
370                         iwr.input = create_input(loaded_state.inputs(i));
371                         iwr.refcount = 0;
372                         inputs.insert(make_pair(loaded_state.inputs(i).url(), iwr));
373                 } 
374
375                 // Deserialize the acceptors.
376                 for (int i = 0; i < loaded_state.acceptors_size(); ++i) {
377                         deserialized_acceptors.insert(make_pair(
378                                 loaded_state.acceptors(i).port(),
379                                 new Acceptor(loaded_state.acceptors(i))));
380                 }
381
382                 log(INFO, "Deserialization done.");
383         }
384
385         // Add any new inputs coming from the config.
386         create_config_inputs(config, &inputs);
387         
388         // Find all streams in the configuration file, create them, and connect to the inputs.
389         create_streams(config, deserialized_stream_ids, &inputs);
390         vector<Acceptor *> acceptors = create_acceptors(config, &deserialized_acceptors);
391         
392         // Put back the existing clients. It doesn't matter which server we
393         // allocate them to, so just do round-robin. However, we need to add
394         // them after the mark pools have been set up.
395         for (int i = 0; i < loaded_state.clients_size(); ++i) {
396                 servers->add_client_from_serialized(loaded_state.clients(i));
397         }
398         
399         servers->run();
400
401         // Now delete all inputs that are longer in use, and start the others.
402         for (multimap<string, InputWithRefcount>::iterator input_it = inputs.begin();
403              input_it != inputs.end(); ) {
404                 if (input_it->second.refcount == 0) {
405                         log(WARNING, "Input '%s' no longer in use, closing.",
406                             input_it->first.c_str());
407                         input_it->second.input->close_socket();
408                         delete input_it->second.input;
409                         inputs.erase(input_it++);
410                 } else {
411                         input_it->second.input->run();
412                         ++input_it;
413                 }
414         }
415
416         // Start writing statistics.
417         StatsThread *stats_thread = NULL;
418         if (!config.stats_file.empty()) {
419                 stats_thread = new StatsThread(config.stats_file, config.stats_interval);
420                 stats_thread->run();
421         }
422
423         struct timeval server_start;
424         gettimeofday(&server_start, NULL);
425         if (state_fd != -1) {
426                 // Measure time from we started deserializing (below) to now, when basically everything
427                 // is up and running. This is, in other words, a conservative estimate of how long our
428                 // “glitch” period was, not counting of course reconnects if the configuration changed.
429                 double glitch_time = server_start.tv_sec - serialize_start.tv_sec +
430                         1e-6 * (server_start.tv_usec - serialize_start.tv_usec);
431                 log(INFO, "Re-exec happened in approx. %.0f ms.", glitch_time * 1000.0);
432         }
433
434         while (!hupped) {
435                 usleep(100000);
436         }
437
438         // OK, we've been HUPed. Time to shut down everything, serialize, and re-exec.
439         gettimeofday(&serialize_start, NULL);
440
441         if (stats_thread != NULL) {
442                 stats_thread->stop();
443                 delete stats_thread;
444         }
445         for (size_t i = 0; i < acceptors.size(); ++i) {
446                 acceptors[i]->stop();
447         }
448         for (multimap<string, InputWithRefcount>::iterator input_it = inputs.begin();
449              input_it != inputs.end();
450              ++input_it) {
451                 input_it->second.input->stop();
452         }
453         servers->stop();
454
455         CubemapStateProto state;
456         if (stopped) {
457                 log(INFO, "Shutting down.");
458         } else {
459                 log(INFO, "Serializing state and re-execing...");
460                 state = collect_state(
461                         serialize_start, acceptors, inputs, servers);
462                 string serialized;
463                 state.SerializeToString(&serialized);
464                 state_fd = make_tempfile(serialized);
465                 if (state_fd == -1) {
466                         exit(1);
467                 }
468         }
469         delete servers;
470
471         for (unsigned i = 0; i < mark_pools.size(); ++i) {
472                 delete mark_pools[i];
473         }
474         mark_pools.clear();
475
476         access_log->stop();
477         delete access_log;
478         shut_down_logging();
479
480         if (stopped) {
481                 exit(0);
482         }
483
484         // OK, so the signal was SIGHUP. Check that the new config is okay, then exec the new binary.
485         if (!dry_run_config(argv0_canon, config_filename_canon)) {
486                 open_logs(config.log_destinations);
487                 log(ERROR, "%s --test-config failed. Restarting old version instead of new.", argv[0]);
488                 hupped = false;
489                 shut_down_logging();
490                 goto start;
491         }
492          
493         char buf[16];
494         sprintf(buf, "%d", state_fd);
495
496         for ( ;; ) {
497                 execlp(argv0_canon, argv0_canon, config_filename_canon, "--state", buf, NULL);
498                 open_logs(config.log_destinations);
499                 log_perror("execlp");
500                 log(ERROR, "re-exec of %s failed. Waiting 0.2 seconds and trying again...", argv0_canon);
501                 shut_down_logging();
502                 usleep(200000);
503         }
504 }