Add support for deduplicating headers/short responses.
[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 "tlse.h"
21
22 #include "acceptor.h"
23 #include "accesslog.h"
24 #include "log.h"
25 #include "metacube2.h"
26 #include "parse.h"
27 #include "server.h"
28 #include "state.pb.h"
29 #include "stream.h"
30 #include "util.h"
31
32 #ifndef SO_MAX_PACING_RATE
33 #define SO_MAX_PACING_RATE 47
34 #endif
35
36 using namespace std;
37
38 extern AccessLogThread *access_log;
39
40 namespace {
41
42 inline bool is_equal(timespec a, timespec b)
43 {
44         return a.tv_sec == b.tv_sec &&
45                a.tv_nsec == b.tv_nsec;
46 }
47
48 inline bool is_earlier(timespec a, timespec b)
49 {
50         if (a.tv_sec != b.tv_sec)
51                 return a.tv_sec < b.tv_sec;
52         return a.tv_nsec < b.tv_nsec;
53 }
54
55 }  // namespace
56
57 Server::Server()
58 {
59         epoll_fd = epoll_create(1024);  // Size argument is ignored.
60         if (epoll_fd == -1) {
61                 log_perror("epoll_fd");
62                 exit(1);
63         }
64 }
65
66 Server::~Server()
67 {
68         safe_close(epoll_fd);
69 }
70
71 vector<ClientStats> Server::get_client_stats() const
72 {
73         vector<ClientStats> ret;
74
75         lock_guard<mutex> lock(mu);
76         for (const auto &fd_and_client : clients) {
77                 ret.push_back(fd_and_client.second.get_stats());
78         }
79         return ret;
80 }
81
82 void Server::do_work()
83 {
84         while (!should_stop()) {
85                 // Wait until there's activity on at least one of the fds,
86                 // or 20 ms (about one frame at 50 fps) has elapsed.
87                 //
88                 // We could in theory wait forever and rely on wakeup()
89                 // from add_client_deferred() and add_data_deferred(),
90                 // but wakeup is a pretty expensive operation, and the
91                 // two threads might end up fighting over a lock, so it's
92                 // seemingly (much) more efficient to just have a timeout here.
93                 int nfds = epoll_pwait(epoll_fd, events, EPOLL_MAX_EVENTS, EPOLL_TIMEOUT_MS, &sigset_without_usr1_block);
94                 if (nfds == -1 && errno != EINTR) {
95                         log_perror("epoll_wait");
96                         exit(1);
97                 }
98
99                 lock_guard<mutex> lock(mu);  // We release the mutex between iterations.
100         
101                 process_queued_data();
102
103                 // Process each client where we have socket activity.
104                 for (int i = 0; i < nfds; ++i) {
105                         Client *client = reinterpret_cast<Client *>(events[i].data.ptr);
106
107                         if (events[i].events & (EPOLLERR | EPOLLRDHUP | EPOLLHUP)) {
108                                 close_client(client);
109                                 continue;
110                         }
111
112                         process_client(client);
113                 }
114
115                 // Process each client where its stream has new data,
116                 // even if there was no socket activity.
117                 for (unique_ptr<Stream> &stream : streams) {
118                         vector<Client *> to_process;
119                         swap(stream->to_process, to_process);
120                         for (Client *client : to_process) {
121                                 process_client(client);
122                         }
123                 }
124
125                 // Finally, go through each client to see if it's timed out
126                 // in the READING_REQUEST state. (Seemingly there are clients
127                 // that can hold sockets up for days at a time without sending
128                 // anything at all.)
129                 timespec timeout_time;
130                 if (clock_gettime(CLOCK_MONOTONIC_COARSE, &timeout_time) == -1) {
131                         log_perror("clock_gettime(CLOCK_MONOTONIC_COARSE)");
132                         continue;
133                 }
134                 timeout_time.tv_sec -= REQUEST_READ_TIMEOUT_SEC;
135                 while (!clients_ordered_by_connect_time.empty()) {
136                         const pair<timespec, int> &connect_time_and_fd = clients_ordered_by_connect_time.front();
137
138                         // See if we have reached the end of clients to process.
139                         if (is_earlier(timeout_time, connect_time_and_fd.first)) {
140                                 break;
141                         }
142
143                         // If this client doesn't exist anymore, just ignore it
144                         // (it was deleted earlier).
145                         auto client_it = clients.find(connect_time_and_fd.second);
146                         if (client_it == clients.end()) {
147                                 clients_ordered_by_connect_time.pop();
148                                 continue;
149                         }
150                         Client *client = &client_it->second;
151                         if (!is_equal(client->connect_time, connect_time_and_fd.first)) {
152                                 // Another client has taken this fd in the meantime.
153                                 clients_ordered_by_connect_time.pop();
154                                 continue;
155                         }
156
157                         if (client->state != Client::READING_REQUEST) {
158                                 // Only READING_REQUEST can time out.
159                                 clients_ordered_by_connect_time.pop();
160                                 continue;
161                         }
162
163                         // OK, it timed out.
164                         close_client(client);
165                         clients_ordered_by_connect_time.pop();
166                 }
167         }
168 }
169
170 CubemapStateProto Server::serialize(unordered_map<const string *, size_t> *short_response_pool)
171 {
172         // We don't serialize anything queued, so empty the queues.
173         process_queued_data();
174
175         // Set all clients in a consistent state before serializing
176         // (ie., they have no remaining lost data). Otherwise, increasing
177         // the backlog could take clients into a newly valid area of the backlog,
178         // sending a stream of zeros instead of skipping the data as it should.
179         //
180         // TODO: Do this when clients are added back from serialized state instead;
181         // it would probably be less wasteful.
182         for (auto &fd_and_client : clients) {
183                 skip_lost_data(&fd_and_client.second);
184         }
185
186         CubemapStateProto serialized;
187         for (const auto &fd_and_client : clients) {
188                 serialized.add_clients()->MergeFrom(fd_and_client.second.serialize(short_response_pool));
189         }
190         for (unique_ptr<Stream> &stream : streams) {
191                 serialized.add_streams()->MergeFrom(stream->serialize());
192         }
193         return serialized;
194 }
195
196 void Server::add_client_deferred(int sock, Acceptor *acceptor)
197 {
198         lock_guard<mutex> lock(queued_clients_mutex);
199         queued_add_clients.push_back(std::make_pair(sock, acceptor));
200 }
201
202 void Server::add_client(int sock, Acceptor *acceptor)
203 {
204         const bool is_tls = acceptor->is_tls();
205         auto inserted = clients.insert(make_pair(sock, Client(sock)));
206         assert(inserted.second == true);  // Should not already exist.
207         Client *client_ptr = &inserted.first->second;
208
209         // Connection timestamps must be nondecreasing. I can't find any guarantee
210         // that even the monotonic clock can't go backwards by a small amount
211         // (think switching between CPUs with non-synchronized TSCs), so if
212         // this actually should happen, we hack around it by fudging
213         // connect_time.
214         if (!clients_ordered_by_connect_time.empty() &&
215             is_earlier(client_ptr->connect_time, clients_ordered_by_connect_time.back().first)) {
216                 client_ptr->connect_time = clients_ordered_by_connect_time.back().first;
217         }
218         clients_ordered_by_connect_time.push(make_pair(client_ptr->connect_time, sock));
219
220         // Start listening on data from this socket.
221         epoll_event ev;
222         if (is_tls) {
223                 // Even in the initial state (READING_REQUEST), TLS needs to
224                 // send data for the handshake, and thus might end up needing
225                 // to know about EPOLLOUT.
226                 ev.events = EPOLLIN | EPOLLOUT | EPOLLET | EPOLLRDHUP;
227         } else {
228                 // EPOLLOUT will be added once we go out of READING_REQUEST.
229                 ev.events = EPOLLIN | EPOLLET | EPOLLRDHUP;
230         }
231         ev.data.ptr = client_ptr;
232         if (epoll_ctl(epoll_fd, EPOLL_CTL_ADD, sock, &ev) == -1) {
233                 log_perror("epoll_ctl(EPOLL_CTL_ADD)");
234                 exit(1);
235         }
236
237         if (is_tls) {
238                 assert(tls_server_contexts.count(acceptor));
239                 client_ptr->tls_context = tls_accept(tls_server_contexts[acceptor]);
240                 if (client_ptr->tls_context == nullptr) {
241                         log(ERROR, "tls_accept() failed");
242                         close_client(client_ptr);
243                         return;
244                 }
245                 tls_make_exportable(client_ptr->tls_context, 1);
246         }
247
248         process_client(client_ptr);
249 }
250
251 void Server::add_client_from_serialized(const ClientProto &client, const vector<shared_ptr<const string>> &short_responses)
252 {
253         lock_guard<mutex> lock(mu);
254         Stream *stream;
255         int stream_index = lookup_stream_by_url(client.url());
256         if (stream_index == -1) {
257                 assert(client.state() != Client::SENDING_DATA);
258                 stream = nullptr;
259         } else {
260                 stream = streams[stream_index].get();
261         }
262         auto inserted = clients.insert(make_pair(client.sock(), Client(client, short_responses, stream)));
263         assert(inserted.second == true);  // Should not already exist.
264         Client *client_ptr = &inserted.first->second;
265
266         // Connection timestamps must be nondecreasing.
267         assert(clients_ordered_by_connect_time.empty() ||
268                !is_earlier(client_ptr->connect_time, clients_ordered_by_connect_time.back().first));
269         clients_ordered_by_connect_time.push(make_pair(client_ptr->connect_time, client.sock()));
270
271         // Start listening on data from this socket.
272         epoll_event ev;
273         if (client.state() == Client::READING_REQUEST) {
274                 // See the corresponding comment in Server::add_client().
275                 if (client.has_tls_context()) {
276                         ev.events = EPOLLIN | EPOLLOUT | EPOLLET | EPOLLRDHUP;
277                 } else {
278                         ev.events = EPOLLIN | EPOLLET | EPOLLRDHUP;
279                 }
280         } else {
281                 // If we don't have more data for this client, we'll be putting it into
282                 // the sleeping array again soon.
283                 ev.events = EPOLLOUT | EPOLLET | EPOLLRDHUP;
284         }
285         ev.data.ptr = client_ptr;
286         if (epoll_ctl(epoll_fd, EPOLL_CTL_ADD, client.sock(), &ev) == -1) {
287                 log_perror("epoll_ctl(EPOLL_CTL_ADD)");
288                 exit(1);
289         }
290
291         if (client_ptr->state == Client::WAITING_FOR_KEYFRAME ||
292             client_ptr->state == Client::PREBUFFERING ||
293             (client_ptr->state == Client::SENDING_DATA &&
294              client_ptr->stream_pos == client_ptr->stream->bytes_received)) {
295                 client_ptr->stream->put_client_to_sleep(client_ptr);
296         } else {
297                 process_client(client_ptr);
298         }
299 }
300
301 int Server::lookup_stream_by_url(const string &url) const
302 {
303         map<string, int>::const_iterator stream_url_it = stream_url_map.find(url);
304         if (stream_url_it == stream_url_map.end()) {
305                 return -1;
306         }
307         return stream_url_it->second;
308 }
309
310 int Server::add_stream(const string &url, size_t backlog_size, size_t prebuffering_bytes, Stream::Encoding encoding, Stream::Encoding src_encoding)
311 {
312         lock_guard<mutex> lock(mu);
313         stream_url_map.insert(make_pair(url, streams.size()));
314         streams.emplace_back(new Stream(url, backlog_size, prebuffering_bytes, encoding, src_encoding));
315         return streams.size() - 1;
316 }
317
318 int Server::add_stream_from_serialized(const StreamProto &stream, int data_fd)
319 {
320         lock_guard<mutex> lock(mu);
321         stream_url_map.insert(make_pair(stream.url(), streams.size()));
322         streams.emplace_back(new Stream(stream, data_fd));
323         return streams.size() - 1;
324 }
325         
326 void Server::set_backlog_size(int stream_index, size_t new_size)
327 {
328         lock_guard<mutex> lock(mu);
329         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
330         streams[stream_index]->set_backlog_size(new_size);
331 }
332
333 void Server::set_prebuffering_bytes(int stream_index, size_t new_amount)
334 {
335         lock_guard<mutex> lock(mu);
336         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
337         streams[stream_index]->prebuffering_bytes = new_amount;
338 }
339         
340 void Server::set_encoding(int stream_index, Stream::Encoding encoding)
341 {
342         lock_guard<mutex> lock(mu);
343         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
344         streams[stream_index]->encoding = encoding;
345 }
346
347 void Server::set_src_encoding(int stream_index, Stream::Encoding encoding)
348 {
349         lock_guard<mutex> lock(mu);
350         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
351         streams[stream_index]->src_encoding = encoding;
352 }
353         
354 void Server::set_header(int stream_index, const string &http_header, const string &stream_header)
355 {
356         lock_guard<mutex> lock(mu);
357         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
358         streams[stream_index]->http_header = http_header;
359
360         if (stream_header != streams[stream_index]->stream_header) {
361                 // We cannot start at any of the older starting points anymore,
362                 // since they'd get the wrong header for the stream (not to mention
363                 // that a changed header probably means the stream restarted,
364                 // which means any client starting on the old one would probably
365                 // stop playing properly at the change point). Next block
366                 // should be a suitable starting point (if not, something is
367                 // pretty strange), so it will fill up again soon enough.
368                 streams[stream_index]->suitable_starting_points.clear();
369         }
370         streams[stream_index]->stream_header = stream_header;
371 }
372         
373 void Server::set_pacing_rate(int stream_index, uint32_t pacing_rate)
374 {
375         lock_guard<mutex> lock(mu);
376         assert(clients.empty());
377         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
378         streams[stream_index]->pacing_rate = pacing_rate;
379 }
380
381 void Server::add_gen204(const std::string &url, const std::string &allow_origin)
382 {
383         lock_guard<mutex> lock(mu);
384         assert(clients.empty());
385         ping_url_map[url] = allow_origin;
386 }
387
388 void Server::create_tls_context_for_acceptor(const Acceptor *acceptor)
389 {
390         assert(acceptor->is_tls());
391
392         bool is_server = true;
393         TLSContext *server_context = tls_create_context(is_server, TLS_V12);
394
395         const string &cert = acceptor->get_certificate_chain();
396         int num_cert = tls_load_certificates(server_context, reinterpret_cast<const unsigned char *>(cert.data()), cert.size());
397         assert(num_cert > 0);  // Should have been checked by config earlier.
398
399         const string &key = acceptor->get_private_key();
400         int num_key = tls_load_private_key(server_context, reinterpret_cast<const unsigned char *>(key.data()), key.size());
401         assert(num_key > 0);  // Should have been checked by config earlier.
402
403         tls_server_contexts.insert(make_pair(acceptor, server_context));
404 }
405
406 void Server::add_data_deferred(int stream_index, const char *data, size_t bytes, uint16_t metacube_flags)
407 {
408         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
409         streams[stream_index]->add_data_deferred(data, bytes, metacube_flags);
410 }
411
412 // See the .h file for postconditions after this function.      
413 void Server::process_client(Client *client)
414 {
415         switch (client->state) {
416         case Client::READING_REQUEST: {
417                 if (client->tls_context != nullptr) {
418                         if (send_pending_tls_data(client)) {
419                                 // send_pending_tls_data() hit postconditions #1 or #4.
420                                 return;
421                         }
422                 }
423
424 read_request_again:
425                 // Try to read more of the request.
426                 char buf[1024];
427                 int ret;
428                 if (client->tls_context == nullptr) {
429                         ret = read_nontls_data(client, buf, sizeof(buf));
430                         if (ret == -1) {
431                                 // read_nontls_data() hit postconditions #1 or #2.
432                                 return;
433                         }
434                 } else {
435                         ret = read_tls_data(client, buf, sizeof(buf));
436                         if (ret == -1) {
437                                 // read_tls_data() hit postconditions #1, #2 or #4.
438                                 return;
439                         }
440                 }
441
442                 RequestParseStatus status = wait_for_double_newline(&client->request, buf, ret);
443         
444                 switch (status) {
445                 case RP_OUT_OF_SPACE:
446                         log(WARNING, "[%s] Client sent overlong request!", client->remote_addr.c_str());
447                         close_client(client);
448                         return;
449                 case RP_NOT_FINISHED_YET:
450                         // OK, we don't have the entire header yet. Fine; we'll get it later.
451                         // See if there's more data for us.
452                         goto read_request_again;
453                 case RP_EXTRA_DATA:
454                         log(WARNING, "[%s] Junk data after request!", client->remote_addr.c_str());
455                         close_client(client);
456                         return;
457                 case RP_FINISHED:
458                         break;
459                 }
460
461                 assert(status == RP_FINISHED);
462
463                 if (client->tls_context && !client->in_ktls_mode && tls_established(client->tls_context)) {
464                         // We're ready to enter kTLS mode, unless we still have some
465                         // handshake data to send (which then must be sent as non-kTLS).
466                         if (send_pending_tls_data(client)) {
467                                 // send_pending_tls_data() hit postconditions #1 or #4.
468                                 return;
469                         }
470                         ret = tls_make_ktls(client->tls_context, client->sock);
471                         if (ret < 0) {
472                                 log_tls_error("tls_make_ktls", ret);
473                                 close_client(client);
474                                 return;
475                         }
476                         client->in_ktls_mode = true;
477                 }
478
479                 int error_code = parse_request(client);
480                 if (error_code == 200) {
481                         construct_header(client);
482                 } else if (error_code == 204) {
483                         construct_204(client);
484                 } else {
485                         construct_error(client, error_code);
486                 }
487
488                 // We've changed states, so fall through.
489                 assert(client->state == Client::SENDING_SHORT_RESPONSE ||
490                        client->state == Client::SENDING_HEADER);
491         }
492         case Client::SENDING_SHORT_RESPONSE:
493         case Client::SENDING_HEADER: {
494 sending_header_or_short_response_again:
495                 int ret;
496                 do {
497                         ret = write(client->sock,
498                                     client->header_or_short_response->data() + client->header_or_short_response_bytes_sent,
499                                     client->header_or_short_response->size() - client->header_or_short_response_bytes_sent);
500                 } while (ret == -1 && errno == EINTR);
501
502                 if (ret == -1 && errno == EAGAIN) {
503                         // We're out of socket space, so now we're at the “low edge” of epoll's
504                         // edge triggering. epoll will tell us when there is more room, so for now,
505                         // just return.
506                         // This is postcondition #4.
507                         return;
508                 }
509
510                 if (ret == -1) {
511                         // Error! Postcondition #1.
512                         log_perror("write");
513                         close_client(client);
514                         return;
515                 }
516                 
517                 client->header_or_short_response_bytes_sent += ret;
518                 assert(client->header_or_short_response_bytes_sent <= client->header_or_short_response->size());
519
520                 if (client->header_or_short_response_bytes_sent < client->header_or_short_response->size()) {
521                         // We haven't sent all yet. Fine; go another round.
522                         goto sending_header_or_short_response_again;
523                 }
524
525                 // We're done sending the header or error! Clear it to release some memory.
526                 client->header_or_short_response = nullptr;
527                 client->header_or_short_response_holder.clear();
528                 client->header_or_short_response_ref.reset();
529
530                 if (client->state == Client::SENDING_SHORT_RESPONSE) {
531                         if (more_requests(client)) {
532                                 // We're done sending the error, but should keep on reading new requests.
533                                 goto read_request_again;
534                         } else {
535                                 // We're done sending the error, so now close.
536                                 // This is postcondition #1.
537                                 close_client(client);
538                         }
539                         return;
540                 }
541
542                 Stream *stream = client->stream;
543                 if (client->stream_pos == Client::STREAM_POS_AT_START) {
544                         // Start sending from the beginning of the backlog.
545                         client->stream_pos = min<size_t>(
546                             stream->bytes_received - stream->backlog_size,
547                             0);
548                         client->state = Client::SENDING_DATA;
549                         goto sending_data;
550                 } else if (client->stream_pos_end != Client::STREAM_POS_NO_END) {
551                         // We're sending a fragment, and should have all of it,
552                         // so start sending right away.
553                         assert(client->stream_pos >= 0);
554                         client->state = Client::SENDING_DATA;
555                         goto sending_data;
556                 } else if (stream->prebuffering_bytes == 0) {
557                         // Start sending from the first keyframe we get. In other
558                         // words, we won't send any of the backlog, but we'll start
559                         // sending immediately as we get the next keyframe block.
560                         // Note that this is functionally identical to the next if branch,
561                         // except that we save a binary search.
562                         assert(client->stream_pos == Client::STREAM_POS_AT_END);
563                         assert(client->stream_pos_end == Client::STREAM_POS_NO_END);
564                         client->stream_pos = stream->bytes_received;
565                         client->state = Client::WAITING_FOR_KEYFRAME;
566                 } else {
567                         // We're not going to send anything to the client before we have
568                         // N bytes. However, this wait might be boring; we can just as well
569                         // use it to send older data if we have it. We use lower_bound()
570                         // so that we are conservative and never add extra latency over just
571                         // waiting (assuming CBR or nearly so); otherwise, we could want e.g.
572                         // 100 kB prebuffer but end up sending a 10 MB GOP.
573                         assert(client->stream_pos == Client::STREAM_POS_AT_END);
574                         assert(client->stream_pos_end == Client::STREAM_POS_NO_END);
575                         deque<size_t>::const_iterator starting_point_it =
576                                 lower_bound(stream->suitable_starting_points.begin(),
577                                             stream->suitable_starting_points.end(),
578                                             stream->bytes_received - stream->prebuffering_bytes);
579                         if (starting_point_it == stream->suitable_starting_points.end()) {
580                                 // None found. Just put us at the end, and then wait for the
581                                 // first keyframe to appear.
582                                 client->stream_pos = stream->bytes_received;
583                                 client->state = Client::WAITING_FOR_KEYFRAME;
584                         } else {
585                                 client->stream_pos = *starting_point_it;
586                                 client->state = Client::PREBUFFERING;
587                                 goto prebuffering;
588                         }
589                 }
590                 // Fall through.
591         }
592         case Client::WAITING_FOR_KEYFRAME: {
593                 Stream *stream = client->stream;
594                 if (stream->suitable_starting_points.empty() ||
595                     client->stream_pos > stream->suitable_starting_points.back()) {
596                         // We haven't received a keyframe since this stream started waiting,
597                         // so keep on waiting for one.
598                         // This is postcondition #3.
599                         stream->put_client_to_sleep(client);
600                         return;
601                 }
602                 client->stream_pos = stream->suitable_starting_points.back();
603                 client->state = Client::PREBUFFERING;
604                 // Fall through.
605         }
606         case Client::PREBUFFERING: {
607 prebuffering:
608                 Stream *stream = client->stream;
609                 size_t bytes_to_send = stream->bytes_received - client->stream_pos;
610                 assert(bytes_to_send <= stream->backlog_size);
611                 if (bytes_to_send < stream->prebuffering_bytes) {
612                         // We don't have enough bytes buffered to start this client yet.
613                         // This is postcondition #3.
614                         stream->put_client_to_sleep(client);
615                         return;
616                 }
617                 client->state = Client::SENDING_DATA;
618                 // Fall through.
619         }
620         case Client::SENDING_DATA: {
621 sending_data:
622                 skip_lost_data(client);
623                 Stream *stream = client->stream;
624
625 sending_data_again:
626                 size_t bytes_to_send;
627                 if (client->stream_pos_end == Client::STREAM_POS_NO_END) {
628                          bytes_to_send = stream->bytes_received - client->stream_pos;
629                 } else {
630                          bytes_to_send = client->stream_pos_end - client->stream_pos;
631                 }
632                 assert(bytes_to_send <= stream->backlog_size);
633                 if (bytes_to_send == 0) {
634                         if (client->stream_pos == client->stream_pos_end) {  // We have a definite end, and we're at it.
635                                 if (more_requests(client)) {
636                                         // We're done sending the fragment, but should keep on reading new requests.
637                                         goto read_request_again;
638                                 } else {
639                                         // We're done sending the fragment, so now close.
640                                         // This is postcondition #1.
641                                         close_client(client);
642                                 }
643                         }
644                         return;
645                 }
646
647                 // See if we need to split across the circular buffer.
648                 bool more_data = false;
649                 if ((client->stream_pos % stream->backlog_size) + bytes_to_send > stream->backlog_size) {
650                         bytes_to_send = stream->backlog_size - (client->stream_pos % stream->backlog_size);
651                         more_data = true;
652                 }
653
654                 ssize_t ret;
655                 do {
656                         off_t offset = client->stream_pos % stream->backlog_size;
657                         ret = sendfile(client->sock, stream->data_fd, &offset, bytes_to_send);
658                 } while (ret == -1 && errno == EINTR);
659
660                 if (ret == -1 && errno == EAGAIN) {
661                         // We're out of socket space, so return; epoll will wake us up
662                         // when there is more room.
663                         // This is postcondition #4.
664                         return;
665                 }
666                 if (ret == -1) {
667                         // Error, close; postcondition #1.
668                         log_perror("sendfile");
669                         close_client(client);
670                         return;
671                 }
672                 client->stream_pos += ret;
673                 client->bytes_sent += ret;
674
675                 assert(client->stream_pos_end == Client::STREAM_POS_NO_END || client->stream_pos <= client->stream_pos_end);
676                 if (client->stream_pos == client->stream_pos_end) {
677                         goto sending_data_again;  // Will see that bytes_to_send == 0 and end.
678                 } else if (client->stream_pos == stream->bytes_received) {
679                         // We don't have any more data for this client, so put it to sleep.
680                         // This is postcondition #3.
681                         stream->put_client_to_sleep(client);
682                 } else if (more_data && size_t(ret) == bytes_to_send) {
683                         goto sending_data_again;
684                 }
685                 // We'll also get here for postcondition #4 (similar to the EAGAIN path above).
686                 break;
687         }
688         default:
689                 assert(false);
690         }
691 }
692
693 bool Server::send_pending_tls_data(Client *client)
694 {
695         // See if there's data from the TLS library to write.
696         if (client->tls_data_to_send == nullptr) {
697                 client->tls_data_to_send = tls_get_write_buffer(client->tls_context, &client->tls_data_left_to_send);
698                 if (client->tls_data_to_send == nullptr) {
699                         // Really no data to send.
700                         return false;
701                 }
702         }
703
704 send_data_again:
705         int ret;
706         do {
707                 ret = write(client->sock, client->tls_data_to_send, client->tls_data_left_to_send);
708         } while (ret == -1 && errno == EINTR);
709         assert(ret < 0 || size_t(ret) <= client->tls_data_left_to_send);
710
711         if (ret == -1 && errno == EAGAIN) {
712                 // We're out of socket space, so now we're at the “low edge” of epoll's
713                 // edge triggering. epoll will tell us when there is more room, so for now,
714                 // just return.
715                 // This is postcondition #4.
716                 return true;
717         }
718         if (ret == -1) {
719                 // Error! Postcondition #1.
720                 log_perror("write");
721                 close_client(client);
722                 return true;
723         }
724         if (ret > 0 && size_t(ret) == client->tls_data_left_to_send) {
725                 // All data has been sent, so we don't need to go to sleep.
726                 tls_buffer_clear(client->tls_context);
727                 client->tls_data_to_send = nullptr;
728                 return false;
729         }
730
731         // More data to send, so try again.
732         client->tls_data_to_send += ret;
733         client->tls_data_left_to_send -= ret;
734         goto send_data_again;
735 }
736
737 int Server::read_nontls_data(Client *client, char *buf, size_t max_size)
738 {
739         int ret;
740         do {
741                 ret = read(client->sock, buf, max_size);
742         } while (ret == -1 && errno == EINTR);
743
744         if (ret == -1 && errno == EAGAIN) {
745                 // No more data right now. Nothing to do.
746                 // This is postcondition #2.
747                 return -1;
748         }
749         if (ret == -1) {
750                 log_perror("read");
751                 close_client(client);
752                 return -1;
753         }
754         if (ret == 0) {
755                 // OK, the socket is closed.
756                 close_client(client);
757                 return -1;
758         }
759
760         return ret;
761 }
762
763 int Server::read_tls_data(Client *client, char *buf, size_t max_size)
764 {
765 read_again:
766         int ret;
767         do {
768                 ret = read(client->sock, buf, max_size);
769         } while (ret == -1 && errno == EINTR);
770
771         if (ret == -1 && errno == EAGAIN) {
772                 // No more data right now. Nothing to do.
773                 // This is postcondition #2.
774                 return -1;
775         }
776         if (ret == -1) {
777                 log_perror("read");
778                 close_client(client);
779                 return -1;
780         }
781         if (ret == 0) {
782                 // OK, the socket is closed.
783                 close_client(client);
784                 return -1;
785         }
786
787         // Give it to the TLS library.
788         int err = tls_consume_stream(client->tls_context, reinterpret_cast<const unsigned char *>(buf), ret, nullptr);
789         if (err < 0) {
790                 log_tls_error("tls_consume_stream", err);
791                 close_client(client);
792                 return -1;
793         }
794         if (err == 0) {
795                 // Not consumed any data. See if we can read more.
796                 goto read_again;
797         }
798
799         // Read any decrypted data available for us. (We can reuse buf, since it's free now.)
800         ret = tls_read(client->tls_context, reinterpret_cast<unsigned char *>(buf), max_size);
801         if (ret == 0) {
802                 // No decrypted data for us yet, but there might be some more handshaking
803                 // to send. Do that if needed, then look for more data.
804                 if (send_pending_tls_data(client)) {
805                         // send_pending_tls_data() hit postconditions #1 or #4.
806                         return -1;
807                 }
808                 goto read_again;
809         }
810         if (ret < 0) {
811                 log_tls_error("tls_read", ret);
812                 close_client(client);
813                 return -1;
814         }
815
816         assert(ret > 0);
817         return ret;
818 }
819
820 // See if there's some data we've lost. Ideally, we should drop to a block boundary,
821 // but resync will be the mux's problem.
822 void Server::skip_lost_data(Client *client)
823 {
824         Stream *stream = client->stream;
825         if (stream == nullptr) {
826                 return;
827         }
828         size_t bytes_to_send = stream->bytes_received - client->stream_pos;
829         if (bytes_to_send > stream->backlog_size) {
830                 size_t bytes_lost = bytes_to_send - stream->backlog_size;
831                 client->bytes_lost += bytes_lost;
832                 ++client->num_loss_events;
833                 if (!client->close_after_response) {
834                         assert(client->stream_pos_end != Client::STREAM_POS_NO_END);
835
836                         // We've already sent a Content-length, so we can't just skip data.
837                         // Close the connection immediately and hope the other side
838                         // is able to figure out that there was an error and it needs to skip.
839                         client->close_after_response = true;
840                         client->stream_pos = client->stream_pos_end;
841                 } else {
842                         client->stream_pos = stream->bytes_received - stream->backlog_size;
843                 }
844         }
845 }
846
847 int Server::parse_request(Client *client)
848 {
849         vector<string> lines = split_lines(client->request);
850         client->request.clear();
851         if (lines.empty()) {
852                 return 400;  // Bad request (empty).
853         }
854
855         // Parse the headers, for logging purposes.
856         // TODO: Case-insensitivity.
857         multimap<string, string> headers = extract_headers(lines, client->remote_addr);
858         multimap<string, string>::const_iterator referer_it = headers.find("Referer");
859         if (referer_it != headers.end()) {
860                 client->referer = referer_it->second;
861         }
862         multimap<string, string>::const_iterator user_agent_it = headers.find("User-Agent");
863         if (user_agent_it != headers.end()) {
864                 client->user_agent = user_agent_it->second;
865         }
866
867         vector<string> request_tokens = split_tokens(lines[0]);
868         if (request_tokens.size() < 3) {
869                 return 400;  // Bad request (empty).
870         }
871         if (request_tokens[0] != "GET") {
872                 return 400;  // Should maybe be 405 instead?
873         }
874
875         string url = request_tokens[1];
876         client->url = url;
877         if (url.size() > 8 && url.find("?backlog") == url.size() - 8) {
878                 client->stream_pos = Client::STREAM_POS_AT_START;
879                 url = url.substr(0, url.size() - 8);
880         } else {
881                 size_t pos = url.find("?frag=");
882                 if (pos != string::npos) {
883                         // Parse an endpoint of the type /stream.mp4?frag=1234-5678.
884                         const char *ptr = url.c_str() + pos + 6;
885
886                         // "?frag=header" is special.
887                         if (strcmp(ptr, "header") == 0) {
888                                 client->stream_pos = Client::STREAM_POS_HEADER_ONLY;
889                                 client->stream_pos_end = -1;
890                         } else {
891                                 char *endptr;
892                                 long long frag_start = strtol(ptr, &endptr, 10);
893                                 if (ptr == endptr || frag_start < 0 || frag_start == LLONG_MAX) {
894                                         return 400;  // Bad request.
895                                 }
896                                 if (*endptr != '-') {
897                                         return 400;  // Bad request.
898                                 }
899                                 ptr = endptr + 1;
900
901                                 long long frag_end = strtol(ptr, &endptr, 10);
902                                 if (ptr == endptr || frag_end < frag_start || frag_end == LLONG_MAX) {
903                                         return 400;  // Bad request.
904                                 }
905
906                                 if (*endptr != '\0') {
907                                         return 400;  // Bad request.
908                                 }
909
910                                 client->stream_pos = frag_start;
911                                 client->stream_pos_end = frag_end;
912                         }
913                         url = url.substr(0, pos);
914                 } else {
915                         client->stream_pos = -1;
916                         client->stream_pos_end = -1;
917                 }
918         }
919
920         // Figure out if we're supposed to close the socket after we've delivered the response.
921         string protocol = request_tokens[2];
922         if (protocol.find("HTTP/") != 0) {
923                 return 400;  // Bad request.
924         }
925         client->close_after_response = false;
926         client->http_11 = true;
927         if (protocol == "HTTP/1.0") {
928                 // No persistent connections.
929                 client->close_after_response = true;
930                 client->http_11 = false;
931         } else {
932                 multimap<string, string>::const_iterator connection_it = headers.find("Connection");
933                 if (connection_it != headers.end() && connection_it->second == "close") {
934                         client->close_after_response = true;
935                 }
936         }
937
938         map<string, int>::const_iterator stream_url_map_it = stream_url_map.find(url);
939         if (stream_url_map_it == stream_url_map.end()) {
940                 map<string, string>::const_iterator ping_url_map_it = ping_url_map.find(url);
941                 if (ping_url_map_it == ping_url_map.end()) {
942                         return 404;  // Not found.
943                 } else {
944                         return 204;  // No error.
945                 }
946         }
947
948         Stream *stream = streams[stream_url_map_it->second].get();
949         if (stream->http_header.empty()) {
950                 return 503;  // Service unavailable.
951         }
952
953         if (client->stream_pos_end == Client::STREAM_POS_NO_END) {
954                 // This stream won't end, so we don't have a content-length,
955                 // and can just as well tell the client it's Connection: close
956                 // (otherwise, we'd have to implement chunking TE for no good reason).
957                 client->close_after_response = true;
958         } else {
959                 if (stream->encoding == Stream::STREAM_ENCODING_METACUBE) {
960                         // This doesn't make any sense, and is hard to implement, too.
961                         return 416;  // Range not satisfiable.
962                 }
963
964                 // Check that we have the requested fragment in our backlog.
965                 size_t buffer_end = stream->bytes_received;
966                 size_t buffer_start = (buffer_end <= stream->backlog_size) ? 0 : buffer_end - stream->backlog_size;
967
968                 if (client->stream_pos_end > buffer_end ||
969                     client->stream_pos < buffer_start) {
970                         return 416;  // Range not satisfiable.
971                 }
972         }
973
974         client->stream = stream;
975         if (setsockopt(client->sock, SOL_SOCKET, SO_MAX_PACING_RATE, &client->stream->pacing_rate, sizeof(client->stream->pacing_rate)) == -1) {
976                 if (client->stream->pacing_rate != ~0U) {
977                         log_perror("setsockopt(SO_MAX_PACING_RATE)");
978                 }
979         }
980         client->request.clear();
981
982         return 200;  // OK!
983 }
984
985 void Server::construct_header(Client *client)
986 {
987         Stream *stream = client->stream;
988         string response = stream->http_header;
989         if (client->stream_pos == Client::STREAM_POS_HEADER_ONLY) {
990                 char buf[64];
991                 snprintf(buf, sizeof(buf), "Content-length: %zu\r\n", stream->stream_header.size());
992                 response.append(buf);
993         } else if (client->stream_pos_end != Client::STREAM_POS_NO_END) {
994                 char buf[64];
995                 snprintf(buf, sizeof(buf), "Content-length: %zu\r\n", client->stream_pos_end - client->stream_pos);
996                 response.append(buf);
997         }
998         if (client->http_11) {
999                 assert(response.find("HTTP/1.0") == 0);
1000                 response[7] = '1';  // Change to HTTP/1.1.
1001                 if (client->close_after_response) {
1002                         response.append("Connection: close\r\n");
1003                 }
1004         } else {
1005                 assert(client->close_after_response);
1006         }
1007         if (stream->encoding == Stream::STREAM_ENCODING_RAW) {
1008                 response.append("\r\n");
1009         } else if (stream->encoding == Stream::STREAM_ENCODING_METACUBE) {
1010                 response.append("Content-encoding: metacube\r\n\r\n");
1011                 if (!stream->stream_header.empty()) {
1012                         metacube2_block_header hdr;
1013                         memcpy(hdr.sync, METACUBE2_SYNC, sizeof(hdr.sync));
1014                         hdr.size = htonl(stream->stream_header.size());
1015                         hdr.flags = htons(METACUBE_FLAGS_HEADER);
1016                         hdr.csum = htons(metacube2_compute_crc(&hdr));
1017                         response.append(string(reinterpret_cast<char *>(&hdr), sizeof(hdr)));
1018                 }
1019         } else {
1020                 assert(false);
1021         }
1022         if (client->stream_pos == Client::STREAM_POS_HEADER_ONLY) {
1023                 client->state = Client::SENDING_SHORT_RESPONSE;
1024                 response.append(stream->stream_header);
1025         } else {
1026                 client->state = Client::SENDING_HEADER;
1027                 if (client->stream_pos_end == Client::STREAM_POS_NO_END) {  // Fragments don't contain stream headers.
1028                         response.append(stream->stream_header);
1029                 }
1030         }
1031
1032         client->header_or_short_response_holder = move(response);
1033         client->header_or_short_response = &client->header_or_short_response_holder;
1034
1035         // Switch states.
1036         change_epoll_events(client, EPOLLOUT | EPOLLET | EPOLLRDHUP);
1037 }
1038         
1039 void Server::construct_error(Client *client, int error_code)
1040 {
1041         char error[256];
1042         if (client->http_11 && client->close_after_response) {
1043                 snprintf(error, sizeof(error),
1044                         "HTTP/1.1 %d Error\r\nContent-type: text/plain\r\nConnection: close\r\n\r\nSomething went wrong. Sorry.\r\n",
1045                         error_code);
1046         } else {
1047                 snprintf(error, sizeof(error),
1048                         "HTTP/1.%d %d Error\r\nContent-type: text/plain\r\nContent-length: 30\r\n\r\nSomething went wrong. Sorry.\r\n",
1049                         client->http_11, error_code);
1050         }
1051         client->header_or_short_response_holder = error;
1052         client->header_or_short_response = &client->header_or_short_response_holder;
1053
1054         // Switch states.
1055         client->state = Client::SENDING_SHORT_RESPONSE;
1056         change_epoll_events(client, EPOLLOUT | EPOLLET | EPOLLRDHUP);
1057 }
1058
1059 void Server::construct_204(Client *client)
1060 {
1061         map<string, string>::const_iterator ping_url_map_it = ping_url_map.find(client->url);
1062         assert(ping_url_map_it != ping_url_map.end());
1063
1064         string response;
1065         if (client->http_11) {
1066                 response = "HTTP/1.1 204 No Content\r\n";
1067                 if (client->close_after_response) {
1068                         response.append("Connection: close\r\n");
1069                 }
1070         } else {
1071                 response = "HTTP/1.0 204 No Content\r\n";
1072                 assert(client->close_after_response);
1073         }
1074         if (!ping_url_map_it->second.empty()) {
1075                 response.append("Access-Control-Allow-Origin: ");
1076                 response.append(ping_url_map_it->second);
1077                 response.append("\r\n");
1078         }
1079         response.append("\r\n");
1080
1081         client->header_or_short_response_holder = move(response);
1082         client->header_or_short_response = &client->header_or_short_response_holder;
1083
1084         // Switch states.
1085         client->state = Client::SENDING_SHORT_RESPONSE;
1086         change_epoll_events(client, EPOLLOUT | EPOLLET | EPOLLRDHUP);
1087 }
1088
1089 template<class T>
1090 void delete_from(vector<T> *v, T elem)
1091 {
1092         typename vector<T>::iterator new_end = remove(v->begin(), v->end(), elem);
1093         v->erase(new_end, v->end());
1094 }
1095         
1096 void Server::close_client(Client *client)
1097 {
1098         if (epoll_ctl(epoll_fd, EPOLL_CTL_DEL, client->sock, nullptr) == -1) {
1099                 log_perror("epoll_ctl(EPOLL_CTL_DEL)");
1100                 exit(1);
1101         }
1102
1103         // This client could be sleeping, so we'll need to fix that. (Argh, O(n).)
1104         if (client->stream != nullptr) {
1105                 delete_from(&client->stream->sleeping_clients, client);
1106                 delete_from(&client->stream->to_process, client);
1107         }
1108
1109         if (client->tls_context) {
1110                 tls_destroy_context(client->tls_context);
1111         }
1112
1113         // Log to access_log.
1114         access_log->write(client->get_stats());
1115
1116         // Bye-bye!
1117         safe_close(client->sock);
1118
1119         clients.erase(client->sock);
1120 }
1121
1122 void Server::change_epoll_events(Client *client, uint32_t events)
1123 {
1124         epoll_event ev;
1125         ev.events = events;
1126         ev.data.ptr = client;
1127
1128         if (epoll_ctl(epoll_fd, EPOLL_CTL_MOD, client->sock, &ev) == -1) {
1129                 log_perror("epoll_ctl(EPOLL_CTL_MOD)");
1130                 exit(1);
1131         }
1132 }
1133
1134 bool Server::more_requests(Client *client)
1135 {
1136         if (client->close_after_response) {
1137                 return false;
1138         }
1139
1140         // Switch states and reset the parsers. We don't reset statistics.
1141         client->state = Client::READING_REQUEST;
1142         client->url.clear();
1143         client->stream = NULL;
1144         client->header_or_short_response = nullptr;
1145         client->header_or_short_response_holder.clear();
1146         client->header_or_short_response_ref.reset();
1147         client->header_or_short_response_bytes_sent = 0;
1148
1149         change_epoll_events(client, EPOLLIN | EPOLLET | EPOLLRDHUP);  // No TLS handshake, so no EPOLLOUT needed.
1150
1151         return true;
1152 }
1153
1154 void Server::process_queued_data()
1155 {
1156         {
1157                 lock_guard<mutex> lock(queued_clients_mutex);
1158
1159                 for (const pair<int, Acceptor *> &id_and_acceptor : queued_add_clients) {
1160                         add_client(id_and_acceptor.first, id_and_acceptor.second);
1161                 }
1162                 queued_add_clients.clear();
1163         }
1164
1165         for (unique_ptr<Stream> &stream : streams) {
1166                 stream->process_queued_data();
1167         }
1168 }