Fix handling of streams with no data.
[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 #ifndef SO_MAX_PACING_RATE
32 #define SO_MAX_PACING_RATE 47
33 #endif
34
35 using namespace std;
36
37 extern AccessLogThread *access_log;
38
39 Server::Server()
40 {
41         pthread_mutex_init(&mutex, NULL);
42         pthread_mutex_init(&queued_clients_mutex, NULL);
43
44         epoll_fd = epoll_create(1024);  // Size argument is ignored.
45         if (epoll_fd == -1) {
46                 log_perror("epoll_fd");
47                 exit(1);
48         }
49 }
50
51 Server::~Server()
52 {
53         for (size_t i = 0; i < streams.size(); ++i) {   
54                 delete streams[i];
55         }
56
57         safe_close(epoll_fd);
58 }
59
60 vector<ClientStats> Server::get_client_stats() const
61 {
62         vector<ClientStats> ret;
63
64         MutexLock lock(&mutex);
65         for (map<int, Client>::const_iterator client_it = clients.begin();
66              client_it != clients.end();
67              ++client_it) {
68                 ret.push_back(client_it->second.get_stats());
69         }
70         return ret;
71 }
72
73 void Server::do_work()
74 {
75         while (!should_stop()) {
76                 // Wait until there's activity on at least one of the fds,
77                 // or 20 ms (about one frame at 50 fps) has elapsed.
78                 //
79                 // We could in theory wait forever and rely on wakeup()
80                 // from add_client_deferred() and add_data_deferred(),
81                 // but wakeup is a pretty expensive operation, and the
82                 // two threads might end up fighting over a lock, so it's
83                 // seemingly (much) more efficient to just have a timeout here.
84                 int nfds = epoll_pwait(epoll_fd, events, EPOLL_MAX_EVENTS, EPOLL_TIMEOUT_MS, &sigset_without_usr1_block);
85                 if (nfds == -1 && errno != EINTR) {
86                         log_perror("epoll_wait");
87                         exit(1);
88                 }
89
90                 MutexLock lock(&mutex);  // We release the mutex between iterations.
91         
92                 process_queued_data();
93
94                 for (int i = 0; i < nfds; ++i) {
95                         Client *client = reinterpret_cast<Client *>(events[i].data.u64);
96
97                         if (events[i].events & (EPOLLERR | EPOLLRDHUP | EPOLLHUP)) {
98                                 close_client(client);
99                                 continue;
100                         }
101
102                         process_client(client);
103                 }
104
105                 for (size_t i = 0; i < streams.size(); ++i) {   
106                         vector<Client *> to_process;
107                         swap(streams[i]->to_process, to_process);
108                         for (size_t i = 0; i < to_process.size(); ++i) {
109                                 process_client(to_process[i]);
110                         }
111                 }
112         }
113 }
114
115 CubemapStateProto Server::serialize()
116 {
117         // We don't serialize anything queued, so empty the queues.
118         process_queued_data();
119
120         // Set all clients in a consistent state before serializing
121         // (ie., they have no remaining lost data). Otherwise, increasing
122         // the backlog could take clients into a newly valid area of the backlog,
123         // sending a stream of zeros instead of skipping the data as it should.
124         //
125         // TODO: Do this when clients are added back from serialized state instead;
126         // it would probably be less wasteful.
127         for (map<int, Client>::iterator client_it = clients.begin();
128              client_it != clients.end();
129              ++client_it) {
130                 skip_lost_data(&client_it->second);
131         }
132
133         CubemapStateProto serialized;
134         for (map<int, Client>::const_iterator client_it = clients.begin();
135              client_it != clients.end();
136              ++client_it) {
137                 serialized.add_clients()->MergeFrom(client_it->second.serialize());
138         }
139         for (size_t i = 0; i < streams.size(); ++i) {   
140                 serialized.add_streams()->MergeFrom(streams[i]->serialize());
141         }
142         return serialized;
143 }
144
145 void Server::add_client_deferred(int sock)
146 {
147         MutexLock lock(&queued_clients_mutex);
148         queued_add_clients.push_back(sock);
149 }
150
151 void Server::add_client(int sock)
152 {
153         pair<map<int, Client>::iterator, bool> ret =
154                 clients.insert(make_pair(sock, Client(sock)));
155         assert(ret.second == true);  // Should not already exist.
156         Client *client_ptr = &ret.first->second;
157
158         // Start listening on data from this socket.
159         epoll_event ev;
160         ev.events = EPOLLIN | EPOLLET | EPOLLRDHUP;
161         ev.data.u64 = reinterpret_cast<uint64_t>(client_ptr);
162         if (epoll_ctl(epoll_fd, EPOLL_CTL_ADD, sock, &ev) == -1) {
163                 log_perror("epoll_ctl(EPOLL_CTL_ADD)");
164                 exit(1);
165         }
166
167         process_client(client_ptr);
168 }
169
170 void Server::add_client_from_serialized(const ClientProto &client)
171 {
172         MutexLock lock(&mutex);
173         Stream *stream;
174         int stream_index = lookup_stream_by_url(client.url());
175         if (stream_index == -1) {
176                 assert(client.state() != Client::SENDING_DATA);
177                 stream = NULL;
178         } else {
179                 stream = streams[stream_index];
180         }
181         pair<map<int, Client>::iterator, bool> ret =
182                 clients.insert(make_pair(client.sock(), Client(client, stream)));
183         assert(ret.second == true);  // Should not already exist.
184         Client *client_ptr = &ret.first->second;
185
186         // Start listening on data from this socket.
187         epoll_event ev;
188         if (client.state() == Client::READING_REQUEST) {
189                 ev.events = EPOLLIN | EPOLLET | EPOLLRDHUP;
190         } else {
191                 // If we don't have more data for this client, we'll be putting it into
192                 // the sleeping array again soon.
193                 ev.events = EPOLLOUT | EPOLLET | EPOLLRDHUP;
194         }
195         ev.data.u64 = reinterpret_cast<uint64_t>(client_ptr);
196         if (epoll_ctl(epoll_fd, EPOLL_CTL_ADD, client.sock(), &ev) == -1) {
197                 log_perror("epoll_ctl(EPOLL_CTL_ADD)");
198                 exit(1);
199         }
200
201         if (client_ptr->state == Client::WAITING_FOR_KEYFRAME ||
202             (client_ptr->state == Client::SENDING_DATA &&
203              client_ptr->stream_pos == client_ptr->stream->bytes_received)) {
204                 client_ptr->stream->put_client_to_sleep(client_ptr);
205         } else {
206                 process_client(client_ptr);
207         }
208 }
209
210 int Server::lookup_stream_by_url(const std::string &url) const
211 {
212         map<string, int>::const_iterator url_it = url_map.find(url);
213         if (url_it == url_map.end()) {
214                 return -1;
215         }
216         return url_it->second;
217 }
218
219 int Server::add_stream(const string &url, size_t backlog_size, Stream::Encoding encoding)
220 {
221         MutexLock lock(&mutex);
222         url_map.insert(make_pair(url, streams.size()));
223         streams.push_back(new Stream(url, backlog_size, encoding));
224         return streams.size() - 1;
225 }
226
227 int Server::add_stream_from_serialized(const StreamProto &stream, int data_fd)
228 {
229         MutexLock lock(&mutex);
230         url_map.insert(make_pair(stream.url(), streams.size()));
231         streams.push_back(new Stream(stream, data_fd));
232         return streams.size() - 1;
233 }
234         
235 void Server::set_backlog_size(int stream_index, size_t new_size)
236 {
237         MutexLock lock(&mutex);
238         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
239         streams[stream_index]->set_backlog_size(new_size);
240 }
241         
242 void Server::set_encoding(int stream_index, Stream::Encoding encoding)
243 {
244         MutexLock lock(&mutex);
245         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
246         streams[stream_index]->encoding = encoding;
247 }
248         
249 void Server::set_header(int stream_index, const string &http_header, const string &stream_header)
250 {
251         MutexLock lock(&mutex);
252         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
253         streams[stream_index]->http_header = http_header;
254         streams[stream_index]->stream_header = stream_header;
255 }
256         
257 void Server::set_mark_pool(int stream_index, MarkPool *mark_pool)
258 {
259         MutexLock lock(&mutex);
260         assert(clients.empty());
261         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
262         streams[stream_index]->mark_pool = mark_pool;
263 }
264
265 void Server::set_pacing_rate(int stream_index, uint32_t pacing_rate)
266 {
267         MutexLock lock(&mutex);
268         assert(clients.empty());
269         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
270         streams[stream_index]->pacing_rate = pacing_rate;
271 }
272
273 void Server::add_data_deferred(int stream_index, const char *data, size_t bytes, StreamStartSuitability suitable_for_stream_start)
274 {
275         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
276         streams[stream_index]->add_data_deferred(data, bytes, suitable_for_stream_start);
277 }
278
279 // See the .h file for postconditions after this function.      
280 void Server::process_client(Client *client)
281 {
282         switch (client->state) {
283         case Client::READING_REQUEST: {
284 read_request_again:
285                 // Try to read more of the request.
286                 char buf[1024];
287                 int ret;
288                 do {
289                         ret = read(client->sock, buf, sizeof(buf));
290                 } while (ret == -1 && errno == EINTR);
291
292                 if (ret == -1 && errno == EAGAIN) {
293                         // No more data right now. Nothing to do.
294                         // This is postcondition #2.
295                         return;
296                 }
297                 if (ret == -1) {
298                         log_perror("read");
299                         close_client(client);
300                         return;
301                 }
302                 if (ret == 0) {
303                         // OK, the socket is closed.
304                         close_client(client);
305                         return;
306                 }
307
308                 RequestParseStatus status = wait_for_double_newline(&client->request, buf, ret);
309         
310                 switch (status) {
311                 case RP_OUT_OF_SPACE:
312                         log(WARNING, "[%s] Client sent overlong request!", client->remote_addr.c_str());
313                         close_client(client);
314                         return;
315                 case RP_NOT_FINISHED_YET:
316                         // OK, we don't have the entire header yet. Fine; we'll get it later.
317                         // See if there's more data for us.
318                         goto read_request_again;
319                 case RP_EXTRA_DATA:
320                         log(WARNING, "[%s] Junk data after request!", client->remote_addr.c_str());
321                         close_client(client);
322                         return;
323                 case RP_FINISHED:
324                         break;
325                 }
326
327                 assert(status == RP_FINISHED);
328
329                 int error_code = parse_request(client);
330                 if (error_code == 200) {
331                         construct_header(client);
332                 } else {
333                         construct_error(client, error_code);
334                 }
335
336                 // We've changed states, so fall through.
337                 assert(client->state == Client::SENDING_ERROR ||
338                        client->state == Client::SENDING_HEADER);
339         }
340         case Client::SENDING_ERROR:
341         case Client::SENDING_HEADER: {
342 sending_header_or_error_again:
343                 int ret;
344                 do {
345                         ret = write(client->sock,
346                                     client->header_or_error.data() + client->header_or_error_bytes_sent,
347                                     client->header_or_error.size() - client->header_or_error_bytes_sent);
348                 } while (ret == -1 && errno == EINTR);
349
350                 if (ret == -1 && errno == EAGAIN) {
351                         // We're out of socket space, so now we're at the “low edge” of epoll's
352                         // edge triggering. epoll will tell us when there is more room, so for now,
353                         // just return.
354                         // This is postcondition #4.
355                         return;
356                 }
357
358                 if (ret == -1) {
359                         // Error! Postcondition #1.
360                         log_perror("write");
361                         close_client(client);
362                         return;
363                 }
364                 
365                 client->header_or_error_bytes_sent += ret;
366                 assert(client->header_or_error_bytes_sent <= client->header_or_error.size());
367
368                 if (client->header_or_error_bytes_sent < client->header_or_error.size()) {
369                         // We haven't sent all yet. Fine; go another round.
370                         goto sending_header_or_error_again;
371                 }
372
373                 // We're done sending the header or error! Clear it to release some memory.
374                 client->header_or_error.clear();
375
376                 if (client->state == Client::SENDING_ERROR) {
377                         // We're done sending the error, so now close.  
378                         // This is postcondition #1.
379                         close_client(client);
380                         return;
381                 }
382
383                 // Start sending from the first keyframe we get. In other
384                 // words, we won't send any of the backlog, but we'll start
385                 // sending immediately as we get the next keyframe block.
386                 // This is postcondition #3.
387                 if (client->stream_pos == size_t(-2)) {
388                         client->stream_pos = std::min<size_t>(
389                             client->stream->bytes_received - client->stream->backlog_size,
390                             0);
391                         client->state = Client::SENDING_DATA;
392                 } else {
393                         // client->stream_pos should be -1, but it might not be,
394                         // if we have clients from an older version.
395                         client->stream_pos = client->stream->bytes_received;
396                         client->state = Client::WAITING_FOR_KEYFRAME;
397                 }
398                 client->stream->put_client_to_sleep(client);
399                 return;
400         }
401         case Client::WAITING_FOR_KEYFRAME: {
402                 Stream *stream = client->stream;
403                 if (ssize_t(client->stream_pos) > stream->last_suitable_starting_point) {
404                         // We haven't received a keyframe since this stream started waiting,
405                         // so keep on waiting for one.
406                         // This is postcondition #3.
407                         stream->put_client_to_sleep(client);
408                         return;
409                 }
410                 client->stream_pos = stream->last_suitable_starting_point;
411                 client->state = Client::SENDING_DATA;
412                 // Fall through.
413         }
414         case Client::SENDING_DATA: {
415                 skip_lost_data(client);
416                 Stream *stream = client->stream;
417
418 sending_data_again:
419                 size_t bytes_to_send = stream->bytes_received - client->stream_pos;
420                 assert(bytes_to_send <= stream->backlog_size);
421                 if (bytes_to_send == 0) {
422                         return;
423                 }
424
425                 // See if we need to split across the circular buffer.
426                 bool more_data = false;
427                 if ((client->stream_pos % stream->backlog_size) + bytes_to_send > stream->backlog_size) {
428                         bytes_to_send = stream->backlog_size - (client->stream_pos % stream->backlog_size);
429                         more_data = true;
430                 }
431
432                 ssize_t ret;
433                 do {
434                         off_t offset = client->stream_pos % stream->backlog_size;
435                         ret = sendfile(client->sock, stream->data_fd, &offset, bytes_to_send);
436                 } while (ret == -1 && errno == EINTR);
437
438                 if (ret == -1 && errno == EAGAIN) {
439                         // We're out of socket space, so return; epoll will wake us up
440                         // when there is more room.
441                         // This is postcondition #4.
442                         return;
443                 }
444                 if (ret == -1) {
445                         // Error, close; postcondition #1.
446                         log_perror("sendfile");
447                         close_client(client);
448                         return;
449                 }
450                 client->stream_pos += ret;
451                 client->bytes_sent += ret;
452
453                 if (client->stream_pos == stream->bytes_received) {
454                         // We don't have any more data for this client, so put it to sleep.
455                         // This is postcondition #3.
456                         stream->put_client_to_sleep(client);
457                 } else if (more_data && size_t(ret) == bytes_to_send) {
458                         goto sending_data_again;
459                 }
460                 break;
461         }
462         default:
463                 assert(false);
464         }
465 }
466
467 // See if there's some data we've lost. Ideally, we should drop to a block boundary,
468 // but resync will be the mux's problem.
469 void Server::skip_lost_data(Client *client)
470 {
471         Stream *stream = client->stream;
472         if (stream == NULL) {
473                 return;
474         }
475         size_t bytes_to_send = stream->bytes_received - client->stream_pos;
476         if (bytes_to_send > stream->backlog_size) {
477                 size_t bytes_lost = bytes_to_send - stream->backlog_size;
478                 client->stream_pos = stream->bytes_received - stream->backlog_size;
479                 client->bytes_lost += bytes_lost;
480                 ++client->num_loss_events;
481
482                 double loss_fraction = double(client->bytes_lost) / double(client->bytes_lost + client->bytes_sent);
483                 log(WARNING, "[%s] Client lost %lld bytes (total loss: %.2f%%), maybe too slow connection",
484                         client->remote_addr.c_str(),
485                         (long long int)(bytes_lost),
486                         100.0 * loss_fraction);
487         }
488 }
489
490 int Server::parse_request(Client *client)
491 {
492         vector<string> lines = split_lines(client->request);
493         if (lines.empty()) {
494                 return 400;  // Bad request (empty).
495         }
496
497         vector<string> request_tokens = split_tokens(lines[0]);
498         if (request_tokens.size() < 2) {
499                 return 400;  // Bad request (empty).
500         }
501         if (request_tokens[0] != "GET") {
502                 return 400;  // Should maybe be 405 instead?
503         }
504
505         string url = request_tokens[1];
506         if (url.find("?backlog") == url.size() - 8) {
507                 client->stream_pos = -2;
508                 url = url.substr(0, url.size() - 8);
509         } else {
510                 client->stream_pos = -1;
511         }
512
513         map<string, int>::const_iterator url_map_it = url_map.find(url);
514         if (url_map_it == url_map.end()) {
515                 return 404;  // Not found.
516         }
517
518         Stream *stream = streams[url_map_it->second];
519         if (stream->http_header.empty()) {
520                 return 503;  // Service unavailable.
521         }
522
523         client->url = request_tokens[1];
524         client->stream = stream;
525         if (client->stream->mark_pool != NULL) {
526                 client->fwmark = client->stream->mark_pool->get_mark();
527         } else {
528                 client->fwmark = 0;  // No mark.
529         }
530         if (setsockopt(client->sock, SOL_SOCKET, SO_MARK, &client->fwmark, sizeof(client->fwmark)) == -1) {                          
531                 if (client->fwmark != 0) {
532                         log_perror("setsockopt(SO_MARK)");
533                 }
534         }
535         if (setsockopt(client->sock, SOL_SOCKET, SO_MAX_PACING_RATE, &client->stream->pacing_rate, sizeof(client->stream->pacing_rate)) == -1) {
536                 if (client->stream->pacing_rate != ~0U) {
537                         log_perror("setsockopt(SO_MAX_PACING_RATE)");
538                 }
539         }
540         client->request.clear();
541
542         return 200;  // OK!
543 }
544
545 void Server::construct_header(Client *client)
546 {
547         Stream *stream = client->stream;
548         if (stream->encoding == Stream::STREAM_ENCODING_RAW) {
549                 client->header_or_error = stream->http_header +
550                         "\r\n" +
551                         stream->stream_header;
552         } else if (stream->encoding == Stream::STREAM_ENCODING_METACUBE) {
553                 client->header_or_error = stream->http_header +
554                         "Content-encoding: metacube\r\n" +
555                         "\r\n";
556                 if (!stream->stream_header.empty()) {
557                         metacube2_block_header hdr;
558                         memcpy(hdr.sync, METACUBE2_SYNC, sizeof(hdr.sync));
559                         hdr.size = htonl(stream->stream_header.size());
560                         hdr.flags = htons(METACUBE_FLAGS_HEADER);
561                         hdr.csum = htons(metacube2_compute_crc(&hdr));
562                         client->header_or_error.append(
563                                 string(reinterpret_cast<char *>(&hdr), sizeof(hdr)));
564                 }
565                 client->header_or_error.append(stream->stream_header);
566         } else {
567                 assert(false);
568         }
569
570         // Switch states.
571         client->state = Client::SENDING_HEADER;
572
573         epoll_event ev;
574         ev.events = EPOLLOUT | EPOLLET | EPOLLRDHUP;
575         ev.data.u64 = reinterpret_cast<uint64_t>(client);
576
577         if (epoll_ctl(epoll_fd, EPOLL_CTL_MOD, client->sock, &ev) == -1) {
578                 log_perror("epoll_ctl(EPOLL_CTL_MOD)");
579                 exit(1);
580         }
581 }
582         
583 void Server::construct_error(Client *client, int error_code)
584 {
585         char error[256];
586         snprintf(error, 256, "HTTP/1.0 %d Error\r\nContent-type: text/plain\r\n\r\nSomething went wrong. Sorry.\r\n",
587                 error_code);
588         client->header_or_error = error;
589
590         // Switch states.
591         client->state = Client::SENDING_ERROR;
592
593         epoll_event ev;
594         ev.events = EPOLLOUT | EPOLLET | EPOLLRDHUP;
595         ev.data.u64 = reinterpret_cast<uint64_t>(client);
596
597         if (epoll_ctl(epoll_fd, EPOLL_CTL_MOD, client->sock, &ev) == -1) {
598                 log_perror("epoll_ctl(EPOLL_CTL_MOD)");
599                 exit(1);
600         }
601 }
602
603 template<class T>
604 void delete_from(vector<T> *v, T elem)
605 {
606         typename vector<T>::iterator new_end = remove(v->begin(), v->end(), elem);
607         v->erase(new_end, v->end());
608 }
609         
610 void Server::close_client(Client *client)
611 {
612         if (epoll_ctl(epoll_fd, EPOLL_CTL_DEL, client->sock, NULL) == -1) {
613                 log_perror("epoll_ctl(EPOLL_CTL_DEL)");
614                 exit(1);
615         }
616
617         // This client could be sleeping, so we'll need to fix that. (Argh, O(n).)
618         if (client->stream != NULL) {
619                 delete_from(&client->stream->sleeping_clients, client);
620                 delete_from(&client->stream->to_process, client);
621                 if (client->stream->mark_pool != NULL) {
622                         int fwmark = client->fwmark;
623                         client->stream->mark_pool->release_mark(fwmark);
624                 }
625         }
626
627         // Log to access_log.
628         access_log->write(client->get_stats());
629
630         // Bye-bye!
631         safe_close(client->sock);
632
633         clients.erase(client->sock);
634 }
635         
636 void Server::process_queued_data()
637 {
638         {
639                 MutexLock lock(&queued_clients_mutex);
640
641                 for (size_t i = 0; i < queued_add_clients.size(); ++i) {
642                         add_client(queued_add_clients[i]);
643                 }
644                 queued_add_clients.clear();
645         }
646
647         for (size_t i = 0; i < streams.size(); ++i) {   
648                 streams[i]->process_queued_data();
649         }
650 }