Add support for forced prebuffering.
[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 "metacube2.h"
23 #include "mutexlock.h"
24 #include "parse.h"
25 #include "server.h"
26 #include "state.pb.h"
27 #include "stream.h"
28 #include "util.h"
29
30 #ifndef SO_MAX_PACING_RATE
31 #define SO_MAX_PACING_RATE 47
32 #endif
33
34 using namespace std;
35
36 extern AccessLogThread *access_log;
37
38 Server::Server()
39 {
40         pthread_mutex_init(&mutex, NULL);
41         pthread_mutex_init(&queued_clients_mutex, NULL);
42
43         epoll_fd = epoll_create(1024);  // Size argument is ignored.
44         if (epoll_fd == -1) {
45                 log_perror("epoll_fd");
46                 exit(1);
47         }
48 }
49
50 Server::~Server()
51 {
52         for (size_t i = 0; i < streams.size(); ++i) {   
53                 delete streams[i];
54         }
55
56         safe_close(epoll_fd);
57 }
58
59 vector<ClientStats> Server::get_client_stats() const
60 {
61         vector<ClientStats> ret;
62
63         MutexLock lock(&mutex);
64         for (map<int, Client>::const_iterator client_it = clients.begin();
65              client_it != clients.end();
66              ++client_it) {
67                 ret.push_back(client_it->second.get_stats());
68         }
69         return ret;
70 }
71
72 void Server::do_work()
73 {
74         while (!should_stop()) {
75                 // Wait until there's activity on at least one of the fds,
76                 // or 20 ms (about one frame at 50 fps) has elapsed.
77                 //
78                 // We could in theory wait forever and rely on wakeup()
79                 // from add_client_deferred() and add_data_deferred(),
80                 // but wakeup is a pretty expensive operation, and the
81                 // two threads might end up fighting over a lock, so it's
82                 // seemingly (much) more efficient to just have a timeout here.
83                 int nfds = epoll_pwait(epoll_fd, events, EPOLL_MAX_EVENTS, EPOLL_TIMEOUT_MS, &sigset_without_usr1_block);
84                 if (nfds == -1 && errno != EINTR) {
85                         log_perror("epoll_wait");
86                         exit(1);
87                 }
88
89                 MutexLock lock(&mutex);  // We release the mutex between iterations.
90         
91                 process_queued_data();
92
93                 for (int i = 0; i < nfds; ++i) {
94                         Client *client = reinterpret_cast<Client *>(events[i].data.u64);
95
96                         if (events[i].events & (EPOLLERR | EPOLLRDHUP | EPOLLHUP)) {
97                                 close_client(client);
98                                 continue;
99                         }
100
101                         process_client(client);
102                 }
103
104                 for (size_t i = 0; i < streams.size(); ++i) {   
105                         vector<Client *> to_process;
106                         swap(streams[i]->to_process, to_process);
107                         for (size_t i = 0; i < to_process.size(); ++i) {
108                                 process_client(to_process[i]);
109                         }
110                 }
111         }
112 }
113
114 CubemapStateProto Server::serialize()
115 {
116         // We don't serialize anything queued, so empty the queues.
117         process_queued_data();
118
119         // Set all clients in a consistent state before serializing
120         // (ie., they have no remaining lost data). Otherwise, increasing
121         // the backlog could take clients into a newly valid area of the backlog,
122         // sending a stream of zeros instead of skipping the data as it should.
123         //
124         // TODO: Do this when clients are added back from serialized state instead;
125         // it would probably be less wasteful.
126         for (map<int, Client>::iterator client_it = clients.begin();
127              client_it != clients.end();
128              ++client_it) {
129                 skip_lost_data(&client_it->second);
130         }
131
132         CubemapStateProto serialized;
133         for (map<int, Client>::const_iterator client_it = clients.begin();
134              client_it != clients.end();
135              ++client_it) {
136                 serialized.add_clients()->MergeFrom(client_it->second.serialize());
137         }
138         for (size_t i = 0; i < streams.size(); ++i) {   
139                 serialized.add_streams()->MergeFrom(streams[i]->serialize());
140         }
141         return serialized;
142 }
143
144 void Server::add_client_deferred(int sock)
145 {
146         MutexLock lock(&queued_clients_mutex);
147         queued_add_clients.push_back(sock);
148 }
149
150 void Server::add_client(int sock)
151 {
152         pair<map<int, Client>::iterator, bool> ret =
153                 clients.insert(make_pair(sock, Client(sock)));
154         assert(ret.second == true);  // Should not already exist.
155         Client *client_ptr = &ret.first->second;
156
157         // Start listening on data from this socket.
158         epoll_event ev;
159         ev.events = EPOLLIN | EPOLLET | EPOLLRDHUP;
160         ev.data.u64 = reinterpret_cast<uint64_t>(client_ptr);
161         if (epoll_ctl(epoll_fd, EPOLL_CTL_ADD, sock, &ev) == -1) {
162                 log_perror("epoll_ctl(EPOLL_CTL_ADD)");
163                 exit(1);
164         }
165
166         process_client(client_ptr);
167 }
168
169 void Server::add_client_from_serialized(const ClientProto &client)
170 {
171         MutexLock lock(&mutex);
172         Stream *stream;
173         int stream_index = lookup_stream_by_url(client.url());
174         if (stream_index == -1) {
175                 assert(client.state() != Client::SENDING_DATA);
176                 stream = NULL;
177         } else {
178                 stream = streams[stream_index];
179         }
180         pair<map<int, Client>::iterator, bool> ret =
181                 clients.insert(make_pair(client.sock(), Client(client, stream)));
182         assert(ret.second == true);  // Should not already exist.
183         Client *client_ptr = &ret.first->second;
184
185         // Start listening on data from this socket.
186         epoll_event ev;
187         if (client.state() == Client::READING_REQUEST) {
188                 ev.events = EPOLLIN | EPOLLET | EPOLLRDHUP;
189         } else {
190                 // If we don't have more data for this client, we'll be putting it into
191                 // the sleeping array again soon.
192                 ev.events = EPOLLOUT | EPOLLET | EPOLLRDHUP;
193         }
194         ev.data.u64 = reinterpret_cast<uint64_t>(client_ptr);
195         if (epoll_ctl(epoll_fd, EPOLL_CTL_ADD, client.sock(), &ev) == -1) {
196                 log_perror("epoll_ctl(EPOLL_CTL_ADD)");
197                 exit(1);
198         }
199
200         if (client_ptr->state == Client::WAITING_FOR_KEYFRAME ||
201             client_ptr->state == Client::PREBUFFERING ||
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, size_t prebuffering_bytes, 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, prebuffering_bytes, 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_pacing_rate(int stream_index, uint32_t pacing_rate)
258 {
259         MutexLock lock(&mutex);
260         assert(clients.empty());
261         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
262         streams[stream_index]->pacing_rate = pacing_rate;
263 }
264
265 void Server::add_data_deferred(int stream_index, const char *data, size_t bytes, StreamStartSuitability suitable_for_stream_start)
266 {
267         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
268         streams[stream_index]->add_data_deferred(data, bytes, suitable_for_stream_start);
269 }
270
271 // See the .h file for postconditions after this function.      
272 void Server::process_client(Client *client)
273 {
274         switch (client->state) {
275         case Client::READING_REQUEST: {
276 read_request_again:
277                 // Try to read more of the request.
278                 char buf[1024];
279                 int ret;
280                 do {
281                         ret = read(client->sock, buf, sizeof(buf));
282                 } while (ret == -1 && errno == EINTR);
283
284                 if (ret == -1 && errno == EAGAIN) {
285                         // No more data right now. Nothing to do.
286                         // This is postcondition #2.
287                         return;
288                 }
289                 if (ret == -1) {
290                         log_perror("read");
291                         close_client(client);
292                         return;
293                 }
294                 if (ret == 0) {
295                         // OK, the socket is closed.
296                         close_client(client);
297                         return;
298                 }
299
300                 RequestParseStatus status = wait_for_double_newline(&client->request, buf, ret);
301         
302                 switch (status) {
303                 case RP_OUT_OF_SPACE:
304                         log(WARNING, "[%s] Client sent overlong request!", client->remote_addr.c_str());
305                         close_client(client);
306                         return;
307                 case RP_NOT_FINISHED_YET:
308                         // OK, we don't have the entire header yet. Fine; we'll get it later.
309                         // See if there's more data for us.
310                         goto read_request_again;
311                 case RP_EXTRA_DATA:
312                         log(WARNING, "[%s] Junk data after request!", client->remote_addr.c_str());
313                         close_client(client);
314                         return;
315                 case RP_FINISHED:
316                         break;
317                 }
318
319                 assert(status == RP_FINISHED);
320
321                 int error_code = parse_request(client);
322                 if (error_code == 200) {
323                         construct_header(client);
324                 } else {
325                         construct_error(client, error_code);
326                 }
327
328                 // We've changed states, so fall through.
329                 assert(client->state == Client::SENDING_ERROR ||
330                        client->state == Client::SENDING_HEADER);
331         }
332         case Client::SENDING_ERROR:
333         case Client::SENDING_HEADER: {
334 sending_header_or_error_again:
335                 int ret;
336                 do {
337                         ret = write(client->sock,
338                                     client->header_or_error.data() + client->header_or_error_bytes_sent,
339                                     client->header_or_error.size() - client->header_or_error_bytes_sent);
340                 } while (ret == -1 && errno == EINTR);
341
342                 if (ret == -1 && errno == EAGAIN) {
343                         // We're out of socket space, so now we're at the “low edge” of epoll's
344                         // edge triggering. epoll will tell us when there is more room, so for now,
345                         // just return.
346                         // This is postcondition #4.
347                         return;
348                 }
349
350                 if (ret == -1) {
351                         // Error! Postcondition #1.
352                         log_perror("write");
353                         close_client(client);
354                         return;
355                 }
356                 
357                 client->header_or_error_bytes_sent += ret;
358                 assert(client->header_or_error_bytes_sent <= client->header_or_error.size());
359
360                 if (client->header_or_error_bytes_sent < client->header_or_error.size()) {
361                         // We haven't sent all yet. Fine; go another round.
362                         goto sending_header_or_error_again;
363                 }
364
365                 // We're done sending the header or error! Clear it to release some memory.
366                 client->header_or_error.clear();
367
368                 if (client->state == Client::SENDING_ERROR) {
369                         // We're done sending the error, so now close.  
370                         // This is postcondition #1.
371                         close_client(client);
372                         return;
373                 }
374
375                 // Start sending from the first keyframe we get. In other
376                 // words, we won't send any of the backlog, but we'll start
377                 // sending immediately as we get the next keyframe block.
378                 // This is postcondition #3.
379                 if (client->stream_pos == size_t(-2)) {
380                         client->stream_pos = std::min<size_t>(
381                             client->stream->bytes_received - client->stream->backlog_size,
382                             0);
383                         client->state = Client::SENDING_DATA;
384                 } else {
385                         // client->stream_pos should be -1, but it might not be,
386                         // if we have clients from an older version.
387                         client->stream_pos = client->stream->bytes_received;
388                         client->state = Client::WAITING_FOR_KEYFRAME;
389                 }
390                 client->stream->put_client_to_sleep(client);
391                 return;
392         }
393         case Client::WAITING_FOR_KEYFRAME: {
394                 Stream *stream = client->stream;
395                 if (ssize_t(client->stream_pos) > stream->last_suitable_starting_point) {
396                         // We haven't received a keyframe since this stream started waiting,
397                         // so keep on waiting for one.
398                         // This is postcondition #3.
399                         stream->put_client_to_sleep(client);
400                         return;
401                 }
402                 client->stream_pos = stream->last_suitable_starting_point;
403                 client->state = Client::PREBUFFERING;
404                 // Fall through.
405         }
406         case Client::PREBUFFERING: {
407                 Stream *stream = client->stream;
408                 size_t bytes_to_send = stream->bytes_received - client->stream_pos;
409                 assert(bytes_to_send <= stream->backlog_size);
410                 if (bytes_to_send < stream->prebuffering_bytes) {
411                         // We don't have enough bytes buffered to start this client yet.
412                         stream->put_client_to_sleep(client);
413                         return;
414                 }
415                 client->state = Client::SENDING_DATA;
416                 // Fall through.
417         }
418         case Client::SENDING_DATA: {
419                 skip_lost_data(client);
420                 Stream *stream = client->stream;
421
422 sending_data_again:
423                 size_t bytes_to_send = stream->bytes_received - client->stream_pos;
424                 assert(bytes_to_send <= stream->backlog_size);
425                 if (bytes_to_send == 0) {
426                         return;
427                 }
428
429                 // See if we need to split across the circular buffer.
430                 bool more_data = false;
431                 if ((client->stream_pos % stream->backlog_size) + bytes_to_send > stream->backlog_size) {
432                         bytes_to_send = stream->backlog_size - (client->stream_pos % stream->backlog_size);
433                         more_data = true;
434                 }
435
436                 ssize_t ret;
437                 do {
438                         off_t offset = client->stream_pos % stream->backlog_size;
439                         ret = sendfile(client->sock, stream->data_fd, &offset, bytes_to_send);
440                 } while (ret == -1 && errno == EINTR);
441
442                 if (ret == -1 && errno == EAGAIN) {
443                         // We're out of socket space, so return; epoll will wake us up
444                         // when there is more room.
445                         // This is postcondition #4.
446                         return;
447                 }
448                 if (ret == -1) {
449                         // Error, close; postcondition #1.
450                         log_perror("sendfile");
451                         close_client(client);
452                         return;
453                 }
454                 client->stream_pos += ret;
455                 client->bytes_sent += ret;
456
457                 if (client->stream_pos == stream->bytes_received) {
458                         // We don't have any more data for this client, so put it to sleep.
459                         // This is postcondition #3.
460                         stream->put_client_to_sleep(client);
461                 } else if (more_data && size_t(ret) == bytes_to_send) {
462                         goto sending_data_again;
463                 }
464                 break;
465         }
466         default:
467                 assert(false);
468         }
469 }
470
471 // See if there's some data we've lost. Ideally, we should drop to a block boundary,
472 // but resync will be the mux's problem.
473 void Server::skip_lost_data(Client *client)
474 {
475         Stream *stream = client->stream;
476         if (stream == NULL) {
477                 return;
478         }
479         size_t bytes_to_send = stream->bytes_received - client->stream_pos;
480         if (bytes_to_send > stream->backlog_size) {
481                 size_t bytes_lost = bytes_to_send - stream->backlog_size;
482                 client->stream_pos = stream->bytes_received - stream->backlog_size;
483                 client->bytes_lost += bytes_lost;
484                 ++client->num_loss_events;
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         Stream *stream = streams[url_map_it->second];
517         if (stream->http_header.empty()) {
518                 return 503;  // Service unavailable.
519         }
520
521         client->url = request_tokens[1];
522         client->stream = stream;
523         if (setsockopt(client->sock, SOL_SOCKET, SO_MAX_PACING_RATE, &client->stream->pacing_rate, sizeof(client->stream->pacing_rate)) == -1) {
524                 if (client->stream->pacing_rate != ~0U) {
525                         log_perror("setsockopt(SO_MAX_PACING_RATE)");
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         }
610
611         // Log to access_log.
612         access_log->write(client->get_stats());
613
614         // Bye-bye!
615         safe_close(client->sock);
616
617         clients.erase(client->sock);
618 }
619         
620 void Server::process_queued_data()
621 {
622         {
623                 MutexLock lock(&queued_clients_mutex);
624
625                 for (size_t i = 0; i < queued_add_clients.size(); ++i) {
626                         add_client(queued_add_clients[i]);
627                 }
628                 queued_add_clients.clear();
629         }
630
631         for (size_t i = 0; i < streams.size(); ++i) {   
632                 streams[i]->process_queued_data();
633         }
634 }