Tweak the MutexLock implementation slightly, so as to confuse Coverity less.
[cubemap] / server.cpp
1 #include <assert.h>
2 #include <errno.h>
3 #include <netinet/in.h>
4 #include <pthread.h>
5 #include <stdint.h>
6 #include <stdio.h>
7 #include <stdlib.h>
8 #include <string.h>
9 #include <sys/epoll.h>
10 #include <sys/sendfile.h>
11 #include <sys/socket.h>
12 #include <sys/types.h>
13 #include <unistd.h>
14 #include <algorithm>
15 #include <map>
16 #include <string>
17 #include <utility>
18 #include <vector>
19
20 #include "accesslog.h"
21 #include "log.h"
22 #include "markpool.h"
23 #include "metacube2.h"
24 #include "mutexlock.h"
25 #include "parse.h"
26 #include "server.h"
27 #include "state.pb.h"
28 #include "stream.h"
29 #include "util.h"
30
31 using namespace std;
32
33 extern AccessLogThread *access_log;
34
35 Server::Server()
36 {
37         pthread_mutex_init(&mutex, NULL);
38         pthread_mutex_init(&queued_data_mutex, NULL);
39
40         epoll_fd = epoll_create(1024);  // Size argument is ignored.
41         if (epoll_fd == -1) {
42                 log_perror("epoll_fd");
43                 exit(1);
44         }
45 }
46
47 Server::~Server()
48 {
49         for (size_t i = 0; i < streams.size(); ++i) {   
50                 delete streams[i];
51         }
52
53         safe_close(epoll_fd);
54 }
55
56 vector<ClientStats> Server::get_client_stats() const
57 {
58         vector<ClientStats> ret;
59
60         MutexLock lock(&mutex);
61         for (map<int, Client>::const_iterator client_it = clients.begin();
62              client_it != clients.end();
63              ++client_it) {
64                 ret.push_back(client_it->second.get_stats());
65         }
66         return ret;
67 }
68
69 void Server::do_work()
70 {
71         while (!should_stop()) {
72                 // Wait until there's activity on at least one of the fds,
73                 // or 20 ms (about one frame at 50 fps) has elapsed.
74                 //
75                 // We could in theory wait forever and rely on wakeup()
76                 // from add_client_deferred() and add_data_deferred(),
77                 // but wakeup is a pretty expensive operation, and the
78                 // two threads might end up fighting over a lock, so it's
79                 // seemingly (much) more efficient to just have a timeout here.
80                 int nfds = epoll_pwait(epoll_fd, events, EPOLL_MAX_EVENTS, EPOLL_TIMEOUT_MS, &sigset_without_usr1_block);
81                 if (nfds == -1 && errno != EINTR) {
82                         log_perror("epoll_wait");
83                         exit(1);
84                 }
85
86                 MutexLock lock(&mutex);  // We release the mutex between iterations.
87         
88                 process_queued_data();
89
90                 for (int i = 0; i < nfds; ++i) {
91                         Client *client = reinterpret_cast<Client *>(events[i].data.u64);
92
93                         if (events[i].events & (EPOLLERR | EPOLLRDHUP | EPOLLHUP)) {
94                                 close_client(client);
95                                 continue;
96                         }
97
98                         process_client(client);
99                 }
100
101                 for (size_t i = 0; i < streams.size(); ++i) {   
102                         vector<Client *> to_process;
103                         swap(streams[i]->to_process, to_process);
104                         for (size_t i = 0; i < to_process.size(); ++i) {
105                                 process_client(to_process[i]);
106                         }
107                 }
108         }
109 }
110
111 CubemapStateProto Server::serialize()
112 {
113         // We don't serialize anything queued, so empty the queues.
114         process_queued_data();
115
116         // Set all clients in a consistent state before serializing
117         // (ie., they have no remaining lost data). Otherwise, increasing
118         // the backlog could take clients into a newly valid area of the backlog,
119         // sending a stream of zeros instead of skipping the data as it should.
120         //
121         // TODO: Do this when clients are added back from serialized state instead;
122         // it would probably be less wasteful.
123         for (map<int, Client>::iterator client_it = clients.begin();
124              client_it != clients.end();
125              ++client_it) {
126                 skip_lost_data(&client_it->second);
127         }
128
129         CubemapStateProto serialized;
130         for (map<int, Client>::const_iterator client_it = clients.begin();
131              client_it != clients.end();
132              ++client_it) {
133                 serialized.add_clients()->MergeFrom(client_it->second.serialize());
134         }
135         for (size_t i = 0; i < streams.size(); ++i) {   
136                 serialized.add_streams()->MergeFrom(streams[i]->serialize());
137         }
138         return serialized;
139 }
140
141 void Server::add_client_deferred(int sock)
142 {
143         MutexLock lock(&queued_data_mutex);
144         queued_add_clients.push_back(sock);
145 }
146
147 void Server::add_client(int sock)
148 {
149         pair<map<int, Client>::iterator, bool> ret =
150                 clients.insert(make_pair(sock, Client(sock)));
151         assert(ret.second == true);  // Should not already exist.
152         Client *client_ptr = &ret.first->second;
153
154         // Start listening on data from this socket.
155         epoll_event ev;
156         ev.events = EPOLLIN | EPOLLET | EPOLLRDHUP;
157         ev.data.u64 = reinterpret_cast<uint64_t>(client_ptr);
158         if (epoll_ctl(epoll_fd, EPOLL_CTL_ADD, sock, &ev) == -1) {
159                 log_perror("epoll_ctl(EPOLL_CTL_ADD)");
160                 exit(1);
161         }
162
163         process_client(client_ptr);
164 }
165
166 void Server::add_client_from_serialized(const ClientProto &client)
167 {
168         MutexLock lock(&mutex);
169         Stream *stream;
170         int stream_index = lookup_stream_by_url(client.url());
171         if (stream_index == -1) {
172                 assert(client.state() != Client::SENDING_DATA);
173                 stream = NULL;
174         } else {
175                 stream = streams[stream_index];
176         }
177         pair<map<int, Client>::iterator, bool> ret =
178                 clients.insert(make_pair(client.sock(), Client(client, stream)));
179         assert(ret.second == true);  // Should not already exist.
180         Client *client_ptr = &ret.first->second;
181
182         // Start listening on data from this socket.
183         epoll_event ev;
184         if (client.state() == Client::READING_REQUEST) {
185                 ev.events = EPOLLIN | EPOLLET | EPOLLRDHUP;
186         } else {
187                 // If we don't have more data for this client, we'll be putting it into
188                 // the sleeping array again soon.
189                 ev.events = EPOLLOUT | EPOLLET | EPOLLRDHUP;
190         }
191         ev.data.u64 = reinterpret_cast<uint64_t>(client_ptr);
192         if (epoll_ctl(epoll_fd, EPOLL_CTL_ADD, client.sock(), &ev) == -1) {
193                 log_perror("epoll_ctl(EPOLL_CTL_ADD)");
194                 exit(1);
195         }
196
197         if (client_ptr->state == Client::WAITING_FOR_KEYFRAME ||
198             (client_ptr->state == Client::SENDING_DATA &&
199              client_ptr->stream_pos == client_ptr->stream->bytes_received)) {
200                 client_ptr->stream->put_client_to_sleep(client_ptr);
201         } else {
202                 process_client(client_ptr);
203         }
204 }
205
206 int Server::lookup_stream_by_url(const std::string &url) const
207 {
208         map<string, int>::const_iterator url_it = url_map.find(url);
209         if (url_it == url_map.end()) {
210                 return -1;
211         }
212         return url_it->second;
213 }
214
215 int Server::add_stream(const string &url, size_t backlog_size, Stream::Encoding encoding)
216 {
217         MutexLock lock(&mutex);
218         url_map.insert(make_pair(url, streams.size()));
219         streams.push_back(new Stream(url, backlog_size, encoding));
220         return streams.size() - 1;
221 }
222
223 int Server::add_stream_from_serialized(const StreamProto &stream, int data_fd)
224 {
225         MutexLock lock(&mutex);
226         url_map.insert(make_pair(stream.url(), streams.size()));
227         streams.push_back(new Stream(stream, data_fd));
228         return streams.size() - 1;
229 }
230         
231 void Server::set_backlog_size(int stream_index, size_t new_size)
232 {
233         MutexLock lock(&mutex);
234         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
235         streams[stream_index]->set_backlog_size(new_size);
236 }
237         
238 void Server::set_encoding(int stream_index, Stream::Encoding encoding)
239 {
240         MutexLock lock(&mutex);
241         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
242         streams[stream_index]->encoding = encoding;
243 }
244         
245 void Server::set_header(int stream_index, const string &http_header, const string &stream_header)
246 {
247         MutexLock lock(&mutex);
248         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
249         streams[stream_index]->http_header = http_header;
250         streams[stream_index]->stream_header = stream_header;
251
252         // If there are clients we haven't sent anything to yet, we should give
253         // them the header, so push back into the SENDING_HEADER state.
254         for (map<int, Client>::iterator client_it = clients.begin();
255              client_it != clients.end();
256              ++client_it) {
257                 Client *client = &client_it->second;
258                 if (client->state == Client::SENDING_DATA &&
259                     client->stream_pos == 0) {
260                         construct_header(client);
261                 }
262         }
263 }
264         
265 void Server::set_mark_pool(int stream_index, MarkPool *mark_pool)
266 {
267         MutexLock lock(&mutex);
268         assert(clients.empty());
269         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
270         streams[stream_index]->mark_pool = mark_pool;
271 }
272
273 void Server::add_data_deferred(int stream_index, const char *data, size_t bytes, StreamStartSuitability suitable_for_stream_start)
274 {
275         MutexLock lock(&queued_data_mutex);
276         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
277         streams[stream_index]->add_data_deferred(data, bytes, suitable_for_stream_start);
278 }
279
280 // See the .h file for postconditions after this function.      
281 void Server::process_client(Client *client)
282 {
283         switch (client->state) {
284         case Client::READING_REQUEST: {
285 read_request_again:
286                 // Try to read more of the request.
287                 char buf[1024];
288                 int ret;
289                 do {
290                         ret = read(client->sock, buf, sizeof(buf));
291                 } while (ret == -1 && errno == EINTR);
292
293                 if (ret == -1 && errno == EAGAIN) {
294                         // No more data right now. Nothing to do.
295                         // This is postcondition #2.
296                         return;
297                 }
298                 if (ret == -1) {
299                         log_perror("read");
300                         close_client(client);
301                         return;
302                 }
303                 if (ret == 0) {
304                         // OK, the socket is closed.
305                         close_client(client);
306                         return;
307                 }
308
309                 RequestParseStatus status = wait_for_double_newline(&client->request, buf, ret);
310         
311                 switch (status) {
312                 case RP_OUT_OF_SPACE:
313                         log(WARNING, "[%s] Client sent overlong request!", client->remote_addr.c_str());
314                         close_client(client);
315                         return;
316                 case RP_NOT_FINISHED_YET:
317                         // OK, we don't have the entire header yet. Fine; we'll get it later.
318                         // See if there's more data for us.
319                         goto read_request_again;
320                 case RP_EXTRA_DATA:
321                         log(WARNING, "[%s] Junk data after request!", client->remote_addr.c_str());
322                         close_client(client);
323                         return;
324                 case RP_FINISHED:
325                         break;
326                 }
327
328                 assert(status == RP_FINISHED);
329
330                 int error_code = parse_request(client);
331                 if (error_code == 200) {
332                         construct_header(client);
333                 } else {
334                         construct_error(client, error_code);
335                 }
336
337                 // We've changed states, so fall through.
338                 assert(client->state == Client::SENDING_ERROR ||
339                        client->state == Client::SENDING_HEADER);
340         }
341         case Client::SENDING_ERROR:
342         case Client::SENDING_HEADER: {
343 sending_header_or_error_again:
344                 int ret;
345                 do {
346                         ret = write(client->sock,
347                                     client->header_or_error.data() + client->header_or_error_bytes_sent,
348                                     client->header_or_error.size() - client->header_or_error_bytes_sent);
349                 } while (ret == -1 && errno == EINTR);
350
351                 if (ret == -1 && errno == EAGAIN) {
352                         // We're out of socket space, so now we're at the “low edge” of epoll's
353                         // edge triggering. epoll will tell us when there is more room, so for now,
354                         // just return.
355                         // This is postcondition #4.
356                         return;
357                 }
358
359                 if (ret == -1) {
360                         // Error! Postcondition #1.
361                         log_perror("write");
362                         close_client(client);
363                         return;
364                 }
365                 
366                 client->header_or_error_bytes_sent += ret;
367                 assert(client->header_or_error_bytes_sent <= client->header_or_error.size());
368
369                 if (client->header_or_error_bytes_sent < client->header_or_error.size()) {
370                         // We haven't sent all yet. Fine; go another round.
371                         goto sending_header_or_error_again;
372                 }
373
374                 // We're done sending the header or error! Clear it to release some memory.
375                 client->header_or_error.clear();
376
377                 if (client->state == Client::SENDING_ERROR) {
378                         // We're done sending the error, so now close.  
379                         // This is postcondition #1.
380                         close_client(client);
381                         return;
382                 }
383
384                 // Start sending from the first keyframe we get. In other
385                 // words, we won't send any of the backlog, but we'll start
386                 // sending immediately as we get the next keyframe block.
387                 // This is postcondition #3.
388                 if (client->stream_pos == size_t(-2)) {
389                         client->stream_pos = std::min<size_t>(
390                             client->stream->bytes_received - client->stream->backlog_size,
391                             0);
392                         client->state = Client::SENDING_DATA;
393                 } else {
394                         // client->stream_pos should be -1, but it might not be,
395                         // if we have clients from an older version.
396                         client->stream_pos = client->stream->bytes_received;
397                         client->state = Client::WAITING_FOR_KEYFRAME;
398                 }
399                 client->stream->put_client_to_sleep(client);
400                 return;
401         }
402         case Client::WAITING_FOR_KEYFRAME: {
403                 Stream *stream = client->stream;
404                 if (ssize_t(client->stream_pos) > stream->last_suitable_starting_point) {
405                         // We haven't received a keyframe since this stream started waiting,
406                         // so keep on waiting for one.
407                         // This is postcondition #3.
408                         stream->put_client_to_sleep(client);
409                         return;
410                 }
411                 client->stream_pos = stream->last_suitable_starting_point;
412                 client->state = Client::SENDING_DATA;
413                 // Fall through.
414         }
415         case Client::SENDING_DATA: {
416                 skip_lost_data(client);
417                 Stream *stream = client->stream;
418
419 sending_data_again:
420                 size_t bytes_to_send = stream->bytes_received - client->stream_pos;
421                 assert(bytes_to_send <= stream->backlog_size);
422                 if (bytes_to_send == 0) {
423                         return;
424                 }
425
426                 // See if we need to split across the circular buffer.
427                 bool more_data = false;
428                 if ((client->stream_pos % stream->backlog_size) + bytes_to_send > stream->backlog_size) {
429                         bytes_to_send = stream->backlog_size - (client->stream_pos % stream->backlog_size);
430                         more_data = true;
431                 }
432
433                 ssize_t ret;
434                 do {
435                         loff_t offset = client->stream_pos % stream->backlog_size;
436                         ret = sendfile(client->sock, stream->data_fd, &offset, bytes_to_send);
437                 } while (ret == -1 && errno == EINTR);
438
439                 if (ret == -1 && errno == EAGAIN) {
440                         // We're out of socket space, so return; epoll will wake us up
441                         // when there is more room.
442                         // This is postcondition #4.
443                         return;
444                 }
445                 if (ret == -1) {
446                         // Error, close; postcondition #1.
447                         log_perror("sendfile");
448                         close_client(client);
449                         return;
450                 }
451                 client->stream_pos += ret;
452                 client->bytes_sent += ret;
453
454                 if (client->stream_pos == stream->bytes_received) {
455                         // We don't have any more data for this client, so put it to sleep.
456                         // This is postcondition #3.
457                         stream->put_client_to_sleep(client);
458                 } else if (more_data && size_t(ret) == bytes_to_send) {
459                         goto sending_data_again;
460                 }
461                 break;
462         }
463         default:
464                 assert(false);
465         }
466 }
467
468 // See if there's some data we've lost. Ideally, we should drop to a block boundary,
469 // but resync will be the mux's problem.
470 void Server::skip_lost_data(Client *client)
471 {
472         Stream *stream = client->stream;
473         size_t bytes_to_send = stream->bytes_received - client->stream_pos;
474         if (bytes_to_send > stream->backlog_size) {
475                 size_t bytes_lost = bytes_to_send - stream->backlog_size;
476                 client->stream_pos = stream->bytes_received - stream->backlog_size;
477                 client->bytes_lost += bytes_lost;
478                 ++client->num_loss_events;
479
480                 double loss_fraction = double(client->bytes_lost) / double(client->bytes_lost + client->bytes_sent);
481                 log(WARNING, "[%s] Client lost %lld bytes (total loss: %.2f%%), maybe too slow connection",
482                         client->remote_addr.c_str(),
483                         (long long int)(bytes_lost),
484                         100.0 * loss_fraction);
485         }
486 }
487
488 int Server::parse_request(Client *client)
489 {
490         vector<string> lines = split_lines(client->request);
491         if (lines.empty()) {
492                 return 400;  // Bad request (empty).
493         }
494
495         vector<string> request_tokens = split_tokens(lines[0]);
496         if (request_tokens.size() < 2) {
497                 return 400;  // Bad request (empty).
498         }
499         if (request_tokens[0] != "GET") {
500                 return 400;  // Should maybe be 405 instead?
501         }
502
503         string url = request_tokens[1];
504         if (url.find("?backlog") == url.size() - 8) {
505                 client->stream_pos = -2;
506                 url = url.substr(0, url.size() - 8);
507         } else {
508                 client->stream_pos = -1;
509         }
510
511         map<string, int>::const_iterator url_map_it = url_map.find(url);
512         if (url_map_it == url_map.end()) {
513                 return 404;  // Not found.
514         }
515
516         client->url = request_tokens[1];
517         client->stream = streams[url_map_it->second];
518         if (client->stream->mark_pool != NULL) {
519                 client->fwmark = client->stream->mark_pool->get_mark();
520         } else {
521                 client->fwmark = 0;  // No mark.
522         }
523         if (setsockopt(client->sock, SOL_SOCKET, SO_MARK, &client->fwmark, sizeof(client->fwmark)) == -1) {                          
524                 if (client->fwmark != 0) {
525                         log_perror("setsockopt(SO_MARK)");
526                 }
527         }
528         client->request.clear();
529
530         return 200;  // OK!
531 }
532
533 void Server::construct_header(Client *client)
534 {
535         Stream *stream = client->stream;
536         if (stream->encoding == Stream::STREAM_ENCODING_RAW) {
537                 client->header_or_error = stream->http_header +
538                         "\r\n" +
539                         stream->stream_header;
540         } else if (stream->encoding == Stream::STREAM_ENCODING_METACUBE) {
541                 client->header_or_error = stream->http_header +
542                         "Content-encoding: metacube\r\n" +
543                         "\r\n";
544                 if (!stream->stream_header.empty()) {
545                         metacube2_block_header hdr;
546                         memcpy(hdr.sync, METACUBE2_SYNC, sizeof(hdr.sync));
547                         hdr.size = htonl(stream->stream_header.size());
548                         hdr.flags = htons(METACUBE_FLAGS_HEADER);
549                         hdr.csum = htons(metacube2_compute_crc(&hdr));
550                         client->header_or_error.append(
551                                 string(reinterpret_cast<char *>(&hdr), sizeof(hdr)));
552                 }
553                 client->header_or_error.append(stream->stream_header);
554         } else {
555                 assert(false);
556         }
557
558         // Switch states.
559         client->state = Client::SENDING_HEADER;
560
561         epoll_event ev;
562         ev.events = EPOLLOUT | EPOLLET | EPOLLRDHUP;
563         ev.data.u64 = reinterpret_cast<uint64_t>(client);
564
565         if (epoll_ctl(epoll_fd, EPOLL_CTL_MOD, client->sock, &ev) == -1) {
566                 log_perror("epoll_ctl(EPOLL_CTL_MOD)");
567                 exit(1);
568         }
569 }
570         
571 void Server::construct_error(Client *client, int error_code)
572 {
573         char error[256];
574         snprintf(error, 256, "HTTP/1.0 %d Error\r\nContent-type: text/plain\r\n\r\nSomething went wrong. Sorry.\r\n",
575                 error_code);
576         client->header_or_error = error;
577
578         // Switch states.
579         client->state = Client::SENDING_ERROR;
580
581         epoll_event ev;
582         ev.events = EPOLLOUT | EPOLLET | EPOLLRDHUP;
583         ev.data.u64 = reinterpret_cast<uint64_t>(client);
584
585         if (epoll_ctl(epoll_fd, EPOLL_CTL_MOD, client->sock, &ev) == -1) {
586                 log_perror("epoll_ctl(EPOLL_CTL_MOD)");
587                 exit(1);
588         }
589 }
590
591 template<class T>
592 void delete_from(vector<T> *v, T elem)
593 {
594         typename vector<T>::iterator new_end = remove(v->begin(), v->end(), elem);
595         v->erase(new_end, v->end());
596 }
597         
598 void Server::close_client(Client *client)
599 {
600         if (epoll_ctl(epoll_fd, EPOLL_CTL_DEL, client->sock, NULL) == -1) {
601                 log_perror("epoll_ctl(EPOLL_CTL_DEL)");
602                 exit(1);
603         }
604
605         // This client could be sleeping, so we'll need to fix that. (Argh, O(n).)
606         if (client->stream != NULL) {
607                 delete_from(&client->stream->sleeping_clients, client);
608                 delete_from(&client->stream->to_process, client);
609                 if (client->stream->mark_pool != NULL) {
610                         int fwmark = client->fwmark;
611                         client->stream->mark_pool->release_mark(fwmark);
612                 }
613         }
614
615         // Log to access_log.
616         access_log->write(client->get_stats());
617
618         // Bye-bye!
619         safe_close(client->sock);
620
621         clients.erase(client->sock);
622 }
623         
624 void Server::process_queued_data()
625 {
626         MutexLock lock(&queued_data_mutex);
627
628         for (size_t i = 0; i < queued_add_clients.size(); ++i) {
629                 add_client(queued_add_clients[i]);
630         }
631         queued_add_clients.clear();     
632
633         for (size_t i = 0; i < streams.size(); ++i) {   
634                 streams[i]->process_queued_data();
635         }
636 }