Keep information about HLS downloads around for some time afterwards.
[cubemap] / server.cpp
1 #include <assert.h>
2 #include <errno.h>
3 #include <inttypes.h>
4 #include <limits.h>
5 #include <netinet/in.h>
6 #include <netinet/tcp.h>
7 #include <pthread.h>
8 #include <stdint.h>
9 #include <stdio.h>
10 #include <stdlib.h>
11 #include <string.h>
12 #include <sys/epoll.h>
13 #include <sys/sendfile.h>
14 #include <sys/socket.h>
15 #include <sys/types.h>
16 #include <unistd.h>
17 #include <algorithm>
18 #include <map>
19 #include <string>
20 #include <utility>
21 #include <vector>
22
23 #include "ktls.h"
24 #include "tlse.h"
25
26 #include "acceptor.h"
27 #include "accesslog.h"
28 #include "log.h"
29 #include "metacube2.h"
30 #include "parse.h"
31 #include "server.h"
32 #include "state.pb.h"
33 #include "stream.h"
34 #include "util.h"
35
36 #ifndef SO_MAX_PACING_RATE
37 #define SO_MAX_PACING_RATE 47
38 #endif
39
40 using namespace std;
41
42 extern AccessLogThread *access_log;
43
44 namespace {
45
46 inline bool is_equal(timespec a, timespec b)
47 {
48         return a.tv_sec == b.tv_sec &&
49                a.tv_nsec == b.tv_nsec;
50 }
51
52 inline bool is_earlier(timespec a, timespec b)
53 {
54         if (a.tv_sec != b.tv_sec)
55                 return a.tv_sec < b.tv_sec;
56         return a.tv_nsec < b.tv_nsec;
57 }
58
59 }  // namespace
60
61 Server::Server()
62 {
63         epoll_fd = epoll_create(1024);  // Size argument is ignored.
64         if (epoll_fd == -1) {
65                 log_perror("epoll_fd");
66                 exit(1);
67         }
68 }
69
70 Server::~Server()
71 {
72         safe_close(epoll_fd);
73 }
74
75 vector<ClientStats> Server::get_client_stats() const
76 {
77         vector<ClientStats> ret;
78
79         lock_guard<mutex> lock(mu);
80         for (const auto &fd_and_client : clients) {
81                 ret.push_back(fd_and_client.second.get_stats());
82         }
83         return ret;
84 }
85
86 vector<HLSZombie> Server::get_hls_zombies()
87 {
88         vector<HLSZombie> ret;
89
90         timespec now;
91         if (clock_gettime(CLOCK_MONOTONIC_COARSE, &now) == -1) {
92                 log_perror("clock_gettime(CLOCK_MONOTONIC_COARSE)");
93                 return ret;
94         }
95
96         lock_guard<mutex> lock(mu);
97         for (auto it = hls_zombies.begin(); it != hls_zombies.end(); ) {
98                 if (is_earlier(it->second.expires, now)) {
99                         hls_zombies.erase(it++);
100                 } else {
101                         ret.push_back(it->second);
102                         ++it;
103                 }
104         }
105         return ret;
106 }
107
108 void Server::do_work()
109 {
110         while (!should_stop()) {
111                 // Wait until there's activity on at least one of the fds,
112                 // or 20 ms (about one frame at 50 fps) has elapsed.
113                 //
114                 // We could in theory wait forever and rely on wakeup()
115                 // from add_client_deferred() and add_data_deferred(),
116                 // but wakeup is a pretty expensive operation, and the
117                 // two threads might end up fighting over a lock, so it's
118                 // seemingly (much) more efficient to just have a timeout here.
119                 int nfds = epoll_pwait(epoll_fd, events, EPOLL_MAX_EVENTS, EPOLL_TIMEOUT_MS, &sigset_without_usr1_block);
120                 if (nfds == -1 && errno != EINTR) {
121                         log_perror("epoll_wait");
122                         exit(1);
123                 }
124
125                 lock_guard<mutex> lock(mu);  // We release the mutex between iterations.
126         
127                 process_queued_data();
128
129                 // Process each client where we have socket activity.
130                 for (int i = 0; i < nfds; ++i) {
131                         Client *client = reinterpret_cast<Client *>(events[i].data.ptr);
132
133                         if (events[i].events & (EPOLLERR | EPOLLRDHUP | EPOLLHUP)) {
134                                 close_client(client);
135                                 continue;
136                         }
137
138                         process_client(client);
139                 }
140
141                 // Process each client where its stream has new data,
142                 // even if there was no socket activity.
143                 for (unique_ptr<Stream> &stream : streams) {
144                         vector<Client *> to_process;
145                         swap(stream->to_process, to_process);
146                         for (Client *client : to_process) {
147                                 process_client(client);
148                         }
149                 }
150
151                 // Finally, go through each client to see if it's timed out
152                 // in the READING_REQUEST state. (Seemingly there are clients
153                 // that can hold sockets up for days at a time without sending
154                 // anything at all.)
155                 timespec timeout_time;
156                 if (clock_gettime(CLOCK_MONOTONIC_COARSE, &timeout_time) == -1) {
157                         log_perror("clock_gettime(CLOCK_MONOTONIC_COARSE)");
158                         continue;
159                 }
160                 timeout_time.tv_sec -= REQUEST_READ_TIMEOUT_SEC;
161                 while (!clients_ordered_by_connect_time.empty()) {
162                         const pair<timespec, int> &connect_time_and_fd = clients_ordered_by_connect_time.front();
163
164                         // See if we have reached the end of clients to process.
165                         if (is_earlier(timeout_time, connect_time_and_fd.first)) {
166                                 break;
167                         }
168
169                         // If this client doesn't exist anymore, just ignore it
170                         // (it was deleted earlier).
171                         auto client_it = clients.find(connect_time_and_fd.second);
172                         if (client_it == clients.end()) {
173                                 clients_ordered_by_connect_time.pop();
174                                 continue;
175                         }
176                         Client *client = &client_it->second;
177                         if (!is_equal(client->connect_time, connect_time_and_fd.first)) {
178                                 // Another client has taken this fd in the meantime.
179                                 clients_ordered_by_connect_time.pop();
180                                 continue;
181                         }
182
183                         if (client->state != Client::READING_REQUEST) {
184                                 // Only READING_REQUEST can time out.
185                                 clients_ordered_by_connect_time.pop();
186                                 continue;
187                         }
188
189                         // OK, it timed out.
190                         close_client(client);
191                         clients_ordered_by_connect_time.pop();
192                 }
193         }
194 }
195
196 CubemapStateProto Server::serialize(unordered_map<const string *, size_t> *short_response_pool)
197 {
198         // We don't serialize anything queued, so empty the queues.
199         process_queued_data();
200
201         // Set all clients in a consistent state before serializing
202         // (ie., they have no remaining lost data). Otherwise, increasing
203         // the backlog could take clients into a newly valid area of the backlog,
204         // sending a stream of zeros instead of skipping the data as it should.
205         //
206         // TODO: Do this when clients are added back from serialized state instead;
207         // it would probably be less wasteful.
208         for (auto &fd_and_client : clients) {
209                 skip_lost_data(&fd_and_client.second);
210         }
211
212         CubemapStateProto serialized;
213         for (const auto &fd_and_client : clients) {
214                 serialized.add_clients()->MergeFrom(fd_and_client.second.serialize(short_response_pool));
215         }
216         for (unique_ptr<Stream> &stream : streams) {
217                 serialized.add_streams()->MergeFrom(stream->serialize());
218         }
219         for (const auto &key_and_zombie : hls_zombies) {
220                 HLSZombieProto *proto = serialized.add_hls_zombies();
221                 proto->set_key(key_and_zombie.first);
222
223                 const HLSZombie &zombie = key_and_zombie.second;
224                 proto->set_remote_addr(zombie.remote_addr);
225                 proto->set_url(zombie.url);
226                 proto->set_referer(zombie.referer);
227                 proto->set_user_agent(zombie.user_agent);
228                 proto->set_expires_sec(zombie.expires.tv_sec);
229                 proto->set_expires_nsec(zombie.expires.tv_nsec);
230         }
231         return serialized;
232 }
233
234 void Server::add_client_deferred(int sock, Acceptor *acceptor)
235 {
236         lock_guard<mutex> lock(queued_clients_mutex);
237         queued_add_clients.push_back(std::make_pair(sock, acceptor));
238 }
239
240 void Server::add_client(int sock, Acceptor *acceptor)
241 {
242         const bool is_tls = acceptor->is_tls();
243         auto inserted = clients.insert(make_pair(sock, Client(sock)));
244         assert(inserted.second == true);  // Should not already exist.
245         Client *client_ptr = &inserted.first->second;
246
247         start_client_timeout_timer(client_ptr);
248
249         // Start listening on data from this socket.
250         epoll_event ev;
251         if (is_tls) {
252                 // Even in the initial state (READING_REQUEST), TLS needs to
253                 // send data for the handshake, and thus might end up needing
254                 // to know about EPOLLOUT.
255                 ev.events = EPOLLIN | EPOLLOUT | EPOLLET | EPOLLRDHUP;
256         } else {
257                 // EPOLLOUT will be added once we go out of READING_REQUEST.
258                 ev.events = EPOLLIN | EPOLLET | EPOLLRDHUP;
259         }
260         ev.data.ptr = client_ptr;
261         if (epoll_ctl(epoll_fd, EPOLL_CTL_ADD, sock, &ev) == -1) {
262                 log_perror("epoll_ctl(EPOLL_CTL_ADD)");
263                 exit(1);
264         }
265
266         if (is_tls) {
267                 assert(tls_server_contexts.count(acceptor));
268                 client_ptr->tls_context = tls_accept(tls_server_contexts[acceptor]);
269                 if (client_ptr->tls_context == nullptr) {
270                         log(ERROR, "tls_accept() failed");
271                         close_client(client_ptr);
272                         return;
273                 }
274                 tls_make_exportable(client_ptr->tls_context, 1);
275         }
276
277         process_client(client_ptr);
278 }
279
280 void Server::add_client_from_serialized(const ClientProto &client, const vector<shared_ptr<const string>> &short_responses)
281 {
282         lock_guard<mutex> lock(mu);
283         Stream *stream;
284         int stream_index = lookup_stream_by_url(client.url());
285         if (stream_index == -1) {
286                 assert(client.state() != Client::SENDING_DATA);
287                 stream = nullptr;
288         } else {
289                 stream = streams[stream_index].get();
290         }
291         auto inserted = clients.insert(make_pair(client.sock(), Client(client, short_responses, stream)));
292         assert(inserted.second == true);  // Should not already exist.
293         Client *client_ptr = &inserted.first->second;
294
295         // Connection timestamps must be nondecreasing.
296         assert(clients_ordered_by_connect_time.empty() ||
297                !is_earlier(client_ptr->connect_time, clients_ordered_by_connect_time.back().first));
298         clients_ordered_by_connect_time.push(make_pair(client_ptr->connect_time, client.sock()));
299
300         // Start listening on data from this socket.
301         epoll_event ev;
302         if (client.state() == Client::READING_REQUEST) {
303                 // See the corresponding comment in Server::add_client().
304                 if (client.has_tls_context()) {
305                         ev.events = EPOLLIN | EPOLLOUT | EPOLLET | EPOLLRDHUP;
306                 } else {
307                         ev.events = EPOLLIN | EPOLLET | EPOLLRDHUP;
308                 }
309         } else {
310                 // If we don't have more data for this client, we'll be putting it into
311                 // the sleeping array again soon.
312                 ev.events = EPOLLOUT | EPOLLET | EPOLLRDHUP;
313         }
314         ev.data.ptr = client_ptr;
315         if (epoll_ctl(epoll_fd, EPOLL_CTL_ADD, client.sock(), &ev) == -1) {
316                 log_perror("epoll_ctl(EPOLL_CTL_ADD)");
317                 exit(1);
318         }
319
320         if (client_ptr->state == Client::WAITING_FOR_KEYFRAME ||
321             client_ptr->state == Client::PREBUFFERING ||
322             (client_ptr->state == Client::SENDING_DATA &&
323              client_ptr->stream_pos == client_ptr->stream->bytes_received)) {
324                 client_ptr->stream->put_client_to_sleep(client_ptr);
325         } else {
326                 process_client(client_ptr);
327         }
328 }
329
330 void Server::start_client_timeout_timer(Client *client)
331 {
332         // Connection timestamps must be nondecreasing. I can't find any guarantee
333         // that even the monotonic clock can't go backwards by a small amount
334         // (think switching between CPUs with non-synchronized TSCs), so if
335         // this actually should happen, we hack around it by fudging
336         // connect_time.
337         if (clock_gettime(CLOCK_MONOTONIC_COARSE, &client->connect_time) == -1) {
338                 log_perror("clock_gettime(CLOCK_MONOTONIC_COARSE)");
339         } else {
340                 if (!clients_ordered_by_connect_time.empty() &&
341                     is_earlier(client->connect_time, clients_ordered_by_connect_time.back().first)) {
342                         client->connect_time = clients_ordered_by_connect_time.back().first;
343                 }
344                 clients_ordered_by_connect_time.push(make_pair(client->connect_time, client->sock));
345         }
346 }
347
348 int Server::lookup_stream_by_url(const string &url) const
349 {
350         const auto stream_url_it = stream_url_map.find(url);
351         if (stream_url_it == stream_url_map.end()) {
352                 return -1;
353         }
354         return stream_url_it->second;
355 }
356
357 int Server::add_stream(const string &url,
358                        const string &hls_url,
359                        size_t backlog_size,
360                        size_t prebuffering_bytes,
361                        Stream::Encoding encoding,
362                        Stream::Encoding src_encoding,
363                        unsigned hls_frag_duration,
364                        size_t hls_backlog_margin,
365                        const string &allow_origin)
366 {
367         lock_guard<mutex> lock(mu);
368         stream_url_map.insert(make_pair(url, streams.size()));
369         if (!hls_url.empty()) {
370                 stream_hls_url_map.insert(make_pair(hls_url, streams.size()));
371         }
372         streams.emplace_back(new Stream(url, backlog_size, prebuffering_bytes, encoding, src_encoding, hls_frag_duration, hls_backlog_margin, allow_origin));
373         return streams.size() - 1;
374 }
375
376 int Server::add_stream_from_serialized(const StreamProto &stream, int data_fd)
377 {
378         lock_guard<mutex> lock(mu);
379         stream_url_map.insert(make_pair(stream.url(), streams.size()));
380         // stream_hls_url_map will be updated in register_hls_url(), since it is not part
381         // of the serialized state (it will always be picked out from the configuration).
382         streams.emplace_back(new Stream(stream, data_fd));
383         return streams.size() - 1;
384 }
385
386 void Server::add_hls_zombie_from_serialized(const HLSZombieProto &zombie_proto)
387 {
388         lock_guard<mutex> lock(mu);
389         HLSZombie zombie;
390         zombie.remote_addr = zombie_proto.remote_addr();
391         zombie.url = zombie_proto.url();
392         zombie.referer = zombie_proto.referer();
393         zombie.user_agent = zombie_proto.user_agent();
394         zombie.expires.tv_sec = zombie_proto.expires_sec();
395         zombie.expires.tv_nsec = zombie_proto.expires_nsec();
396         hls_zombies[zombie_proto.key()] = move(zombie);
397 }
398
399 void Server::set_backlog_size(int stream_index, size_t new_size)
400 {
401         lock_guard<mutex> lock(mu);
402         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
403         streams[stream_index]->set_backlog_size(new_size);
404 }
405
406 void Server::set_prebuffering_bytes(int stream_index, size_t new_amount)
407 {
408         lock_guard<mutex> lock(mu);
409         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
410         streams[stream_index]->prebuffering_bytes = new_amount;
411 }
412         
413 void Server::set_encoding(int stream_index, Stream::Encoding encoding)
414 {
415         lock_guard<mutex> lock(mu);
416         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
417         streams[stream_index]->encoding = encoding;
418 }
419
420 void Server::set_src_encoding(int stream_index, Stream::Encoding encoding)
421 {
422         lock_guard<mutex> lock(mu);
423         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
424         streams[stream_index]->src_encoding = encoding;
425 }
426
427 void Server::set_hls_frag_duration(int stream_index, unsigned hls_frag_duration)
428 {
429         lock_guard<mutex> lock(mu);
430         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
431         streams[stream_index]->hls_frag_duration = hls_frag_duration;
432 }
433
434 void Server::set_hls_backlog_margin(int stream_index, size_t hls_backlog_margin)
435 {
436         lock_guard<mutex> lock(mu);
437         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
438         assert(hls_backlog_margin >= 0);
439         assert(hls_backlog_margin < streams[stream_index]->backlog_size);
440         streams[stream_index]->hls_backlog_margin = hls_backlog_margin;
441 }
442
443 void Server::set_allow_origin(int stream_index, const string &allow_origin)
444 {
445         lock_guard<mutex> lock(mu);
446         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
447         streams[stream_index]->allow_origin = allow_origin;
448 }
449
450 void Server::register_hls_url(int stream_index, const string &hls_url)
451 {
452         lock_guard<mutex> lock(mu);
453         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
454         assert(!hls_url.empty());
455         stream_hls_url_map.insert(make_pair(hls_url, stream_index));
456 }
457         
458 void Server::set_header(int stream_index, const string &http_header, const string &stream_header)
459 {
460         lock_guard<mutex> lock(mu);
461         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
462         Stream *stream = streams[stream_index].get();
463         stream->http_header = http_header;
464
465         if (stream_header != stream->stream_header) {
466                 // We cannot start at any of the older starting points anymore,
467                 // since they'd get the wrong header for the stream (not to mention
468                 // that a changed header probably means the stream restarted,
469                 // which means any client starting on the old one would probably
470                 // stop playing properly at the change point). Next block
471                 // should be a suitable starting point (if not, something is
472                 // pretty strange), so it will fill up again soon enough.
473                 stream->suitable_starting_points.clear();
474
475                 if (!stream->fragments.empty()) {
476                         stream->fragments.clear();
477                         ++stream->discontinuity_counter;
478                         stream->clear_hls_playlist_cache();
479                 }
480         }
481         stream->stream_header = stream_header;
482 }
483         
484 void Server::set_pacing_rate(int stream_index, uint32_t pacing_rate)
485 {
486         lock_guard<mutex> lock(mu);
487         assert(clients.empty());
488         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
489         streams[stream_index]->pacing_rate = pacing_rate;
490 }
491
492 void Server::add_gen204(const std::string &url, const std::string &allow_origin)
493 {
494         lock_guard<mutex> lock(mu);
495         assert(clients.empty());
496         ping_url_map[url] = allow_origin;
497 }
498
499 void Server::create_tls_context_for_acceptor(const Acceptor *acceptor)
500 {
501         assert(acceptor->is_tls());
502
503         bool is_server = true;
504         TLSContext *server_context = tls_create_context(is_server, TLS_V12);
505
506         const string &cert = acceptor->get_certificate_chain();
507         int num_cert = tls_load_certificates(server_context, reinterpret_cast<const unsigned char *>(cert.data()), cert.size());
508         assert(num_cert > 0);  // Should have been checked by config earlier.
509
510         const string &key = acceptor->get_private_key();
511         int num_key = tls_load_private_key(server_context, reinterpret_cast<const unsigned char *>(key.data()), key.size());
512         assert(num_key > 0);  // Should have been checked by config earlier.
513
514         tls_server_contexts.insert(make_pair(acceptor, server_context));
515 }
516
517 void Server::add_data_deferred(int stream_index, const char *data, size_t bytes, uint16_t metacube_flags, const RationalPTS &pts)
518 {
519         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
520         streams[stream_index]->add_data_deferred(data, bytes, metacube_flags, pts);
521 }
522
523 // See the .h file for postconditions after this function.      
524 void Server::process_client(Client *client)
525 {
526         switch (client->state) {
527         case Client::READING_REQUEST: {
528                 if (client->tls_context != nullptr) {
529                         if (send_pending_tls_data(client)) {
530                                 // send_pending_tls_data() hit postconditions #1 or #4.
531                                 return;
532                         }
533                 }
534
535 read_request_again:
536                 // Try to read more of the request.
537                 char buf[1024];
538                 int ret;
539                 if (client->tls_context == nullptr) {
540                         ret = read_nontls_data(client, buf, sizeof(buf));
541                         if (ret == -1) {
542                                 // read_nontls_data() hit postconditions #1 or #2.
543                                 return;
544                         }
545                 } else {
546                         ret = read_tls_data(client, buf, sizeof(buf));
547                         if (ret == -1) {
548                                 // read_tls_data() hit postconditions #1, #2 or #4.
549                                 return;
550                         }
551                 }
552
553                 RequestParseStatus status = wait_for_double_newline(&client->request, buf, ret);
554         
555                 switch (status) {
556                 case RP_OUT_OF_SPACE:
557                         log(WARNING, "[%s] Client sent overlong request!", client->remote_addr.c_str());
558                         close_client(client);
559                         return;
560                 case RP_NOT_FINISHED_YET:
561                         // OK, we don't have the entire header yet. Fine; we'll get it later.
562                         // See if there's more data for us.
563                         goto read_request_again;
564                 case RP_EXTRA_DATA:
565                         log(WARNING, "[%s] Junk data after request!", client->remote_addr.c_str());
566                         close_client(client);
567                         return;
568                 case RP_FINISHED:
569                         break;
570                 }
571
572                 assert(status == RP_FINISHED);
573
574                 if (client->tls_context && !client->in_ktls_mode && tls_established(client->tls_context)) {
575                         // We're ready to enter kTLS mode, unless we still have some
576                         // handshake data to send (which then must be sent as non-kTLS).
577                         if (send_pending_tls_data(client)) {
578                                 // send_pending_tls_data() hit postconditions #1 or #4.
579                                 return;
580                         }
581                         ret = tls_make_ktls(client->tls_context, client->sock);
582                         if (ret < 0) {
583                                 log_tls_error("tls_make_ktls", ret);
584                                 close_client(client);
585                                 return;
586                         }
587                         client->in_ktls_mode = true;
588                 }
589
590                 int error_code = parse_request(client);
591                 if (error_code == 200) {
592                         if (client->serving_hls_playlist) {
593                                 construct_hls_playlist(client);
594                         } else {
595                                 construct_stream_header(client);
596                         }
597                 } else if (error_code == 204) {
598                         construct_204(client);
599                 } else {
600                         construct_error(client, error_code);
601                 }
602
603                 // We've changed states, so fall through.
604                 assert(client->state == Client::SENDING_SHORT_RESPONSE ||
605                        client->state == Client::SENDING_HEADER);
606         }
607         case Client::SENDING_SHORT_RESPONSE:
608         case Client::SENDING_HEADER: {
609 sending_header_or_short_response_again:
610                 int ret;
611                 do {
612                         ret = write(client->sock,
613                                     client->header_or_short_response->data() + client->header_or_short_response_bytes_sent,
614                                     client->header_or_short_response->size() - client->header_or_short_response_bytes_sent);
615                 } while (ret == -1 && errno == EINTR);
616
617                 if (ret == -1 && errno == EAGAIN) {
618                         // We're out of socket space, so now we're at the “low edge” of epoll's
619                         // edge triggering. epoll will tell us when there is more room, so for now,
620                         // just return.
621                         // This is postcondition #4.
622                         return;
623                 }
624
625                 if (ret == -1) {
626                         // Error! Postcondition #1.
627                         log_perror("write");
628                         close_client(client);
629                         return;
630                 }
631                 
632                 client->header_or_short_response_bytes_sent += ret;
633                 assert(client->header_or_short_response_bytes_sent <= client->header_or_short_response->size());
634
635                 if (client->header_or_short_response_bytes_sent < client->header_or_short_response->size()) {
636                         // We haven't sent all yet. Fine; go another round.
637                         goto sending_header_or_short_response_again;
638                 }
639
640                 // We're done sending the header or error! Clear it to release some memory.
641                 client->header_or_short_response = nullptr;
642                 client->header_or_short_response_holder.clear();
643                 client->header_or_short_response_ref.reset();
644
645                 if (client->state == Client::SENDING_SHORT_RESPONSE) {
646                         if (more_requests(client)) {
647                                 // We're done sending the error, but should keep on reading new requests.
648                                 goto read_request_again;
649                         } else {
650                                 // We're done sending the error, so now close.
651                                 // This is postcondition #1.
652                                 close_client(client);
653                         }
654                         return;
655                 }
656
657                 Stream *stream = client->stream;
658                 hls_zombies.erase(client->get_hls_zombie_key());
659                 if (client->stream_pos == Client::STREAM_POS_AT_START) {
660                         // Start sending from the beginning of the backlog.
661                         client->stream_pos = min<size_t>(
662                             stream->bytes_received - stream->backlog_size,
663                             0);
664                         client->state = Client::SENDING_DATA;
665                         goto sending_data;
666                 } else if (client->stream_pos_end != Client::STREAM_POS_NO_END) {
667                         // We're sending a fragment, and should have all of it,
668                         // so start sending right away.
669                         assert(client->stream_pos >= 0);
670                         client->state = Client::SENDING_DATA;
671                         goto sending_data;
672                 } else if (stream->prebuffering_bytes == 0) {
673                         // Start sending from the first keyframe we get. In other
674                         // words, we won't send any of the backlog, but we'll start
675                         // sending immediately as we get the next keyframe block.
676                         // Note that this is functionally identical to the next if branch,
677                         // except that we save a binary search.
678                         assert(client->stream_pos == Client::STREAM_POS_AT_END);
679                         assert(client->stream_pos_end == Client::STREAM_POS_NO_END);
680                         client->stream_pos = stream->bytes_received;
681                         client->state = Client::WAITING_FOR_KEYFRAME;
682                 } else {
683                         // We're not going to send anything to the client before we have
684                         // N bytes. However, this wait might be boring; we can just as well
685                         // use it to send older data if we have it. We use lower_bound()
686                         // so that we are conservative and never add extra latency over just
687                         // waiting (assuming CBR or nearly so); otherwise, we could want e.g.
688                         // 100 kB prebuffer but end up sending a 10 MB GOP.
689                         assert(client->stream_pos == Client::STREAM_POS_AT_END);
690                         assert(client->stream_pos_end == Client::STREAM_POS_NO_END);
691                         deque<size_t>::const_iterator starting_point_it =
692                                 lower_bound(stream->suitable_starting_points.begin(),
693                                             stream->suitable_starting_points.end(),
694                                             stream->bytes_received - stream->prebuffering_bytes);
695                         if (starting_point_it == stream->suitable_starting_points.end()) {
696                                 // None found. Just put us at the end, and then wait for the
697                                 // first keyframe to appear.
698                                 client->stream_pos = stream->bytes_received;
699                                 client->state = Client::WAITING_FOR_KEYFRAME;
700                         } else {
701                                 client->stream_pos = *starting_point_it;
702                                 client->state = Client::PREBUFFERING;
703                                 goto prebuffering;
704                         }
705                 }
706                 // Fall through.
707         }
708         case Client::WAITING_FOR_KEYFRAME: {
709                 Stream *stream = client->stream;
710                 if (stream->suitable_starting_points.empty() ||
711                     client->stream_pos > stream->suitable_starting_points.back()) {
712                         // We haven't received a keyframe since this stream started waiting,
713                         // so keep on waiting for one.
714                         // This is postcondition #3.
715                         stream->put_client_to_sleep(client);
716                         return;
717                 }
718                 client->stream_pos = stream->suitable_starting_points.back();
719                 client->state = Client::PREBUFFERING;
720                 // Fall through.
721         }
722         case Client::PREBUFFERING: {
723 prebuffering:
724                 Stream *stream = client->stream;
725                 size_t bytes_to_send = stream->bytes_received - client->stream_pos;
726                 assert(bytes_to_send <= stream->backlog_size);
727                 if (bytes_to_send < stream->prebuffering_bytes) {
728                         // We don't have enough bytes buffered to start this client yet.
729                         // This is postcondition #3.
730                         stream->put_client_to_sleep(client);
731                         return;
732                 }
733                 client->state = Client::SENDING_DATA;
734                 // Fall through.
735         }
736         case Client::SENDING_DATA: {
737 sending_data:
738                 skip_lost_data(client);
739                 Stream *stream = client->stream;
740
741 sending_data_again:
742                 size_t bytes_to_send;
743                 if (client->stream_pos_end == Client::STREAM_POS_NO_END) {
744                          bytes_to_send = stream->bytes_received - client->stream_pos;
745                 } else {
746                          bytes_to_send = client->stream_pos_end - client->stream_pos;
747                 }
748                 assert(bytes_to_send <= stream->backlog_size);
749                 if (bytes_to_send == 0) {
750                         if (client->stream_pos == client->stream_pos_end) {  // We have a definite end, and we're at it.
751                                 // Add (or overwrite) a HLS zombie.
752                                 timespec now;
753                                 if (clock_gettime(CLOCK_MONOTONIC_COARSE, &now) == -1) {
754                                         log_perror("clock_gettime(CLOCK_MONOTONIC_COARSE)");
755                                 } else {
756                                         HLSZombie zombie;
757                                         zombie.remote_addr = client->remote_addr;
758                                         zombie.referer = client->referer;
759                                         zombie.user_agent = client->user_agent;
760                                         zombie.url = client->stream->url + "?frag=<idle>";
761                                         zombie.expires = now;
762                                         zombie.expires.tv_sec += client->stream->hls_frag_duration * 3;
763                                         hls_zombies[client->get_hls_zombie_key()] = move(zombie);
764                                 }
765                                 if (more_requests(client)) {
766                                         // We're done sending the fragment, but should keep on reading new requests.
767                                         goto read_request_again;
768                                 } else {
769                                         // We're done sending the fragment, so now close.
770                                         // This is postcondition #1.
771                                         close_client(client);
772                                 }
773                         }
774                         return;
775                 }
776
777                 // See if we need to split across the circular buffer.
778                 bool more_data = false;
779                 if ((client->stream_pos % stream->backlog_size) + bytes_to_send > stream->backlog_size) {
780                         bytes_to_send = stream->backlog_size - (client->stream_pos % stream->backlog_size);
781                         more_data = true;
782                 }
783
784                 ssize_t ret;
785                 do {
786                         off_t offset = client->stream_pos % stream->backlog_size;
787                         ret = sendfile(client->sock, stream->data_fd, &offset, bytes_to_send);
788                 } while (ret == -1 && errno == EINTR);
789
790                 if (ret == -1 && errno == EAGAIN) {
791                         // We're out of socket space, so return; epoll will wake us up
792                         // when there is more room.
793                         // This is postcondition #4.
794                         return;
795                 }
796                 if (ret == -1) {
797                         // Error, close; postcondition #1.
798                         log_perror("sendfile");
799                         close_client(client);
800                         return;
801                 }
802                 client->stream_pos += ret;
803                 client->bytes_sent += ret;
804
805                 assert(client->stream_pos_end == Client::STREAM_POS_NO_END || client->stream_pos <= client->stream_pos_end);
806                 if (client->stream_pos == client->stream_pos_end) {
807                         goto sending_data_again;  // Will see that bytes_to_send == 0 and end.
808                 } else if (client->stream_pos == stream->bytes_received) {
809                         // We don't have any more data for this client, so put it to sleep.
810                         // This is postcondition #3.
811                         stream->put_client_to_sleep(client);
812                 } else if (more_data && size_t(ret) == bytes_to_send) {
813                         goto sending_data_again;
814                 }
815                 // We'll also get here for postcondition #4 (similar to the EAGAIN path above).
816                 break;
817         }
818         default:
819                 assert(false);
820         }
821 }
822
823 namespace {
824
825 void flush_pending_data(int sock)
826 {
827         // Flush pending data, which would otherwise wait for the 200ms TCP_CORK timer
828         // to elapsed; does not cancel out TCP_CORK (since that still takes priority),
829         // but does a one-off flush.
830         int one = 1;
831         if (setsockopt(sock, SOL_TCP, TCP_NODELAY, &one, sizeof(one)) == -1) {
832                 log_perror("setsockopt(TCP_NODELAY)");
833                 // Can still continue.
834         }
835 }
836
837 }  // namespace
838
839 bool Server::send_pending_tls_data(Client *client)
840 {
841         // See if there's data from the TLS library to write.
842         if (client->tls_data_to_send == nullptr) {
843                 client->tls_data_to_send = tls_get_write_buffer(client->tls_context, &client->tls_data_left_to_send);
844                 if (client->tls_data_to_send == nullptr) {
845                         // Really no data to send.
846                         return false;
847                 }
848         }
849
850 send_data_again:
851         int ret;
852         do {
853                 ret = write(client->sock, client->tls_data_to_send, client->tls_data_left_to_send);
854         } while (ret == -1 && errno == EINTR);
855         assert(ret < 0 || size_t(ret) <= client->tls_data_left_to_send);
856
857         if (ret == -1 && errno == EAGAIN) {
858                 // We're out of socket space, so now we're at the “low edge” of epoll's
859                 // edge triggering. epoll will tell us when there is more room, so for now,
860                 // just return.
861                 // This is postcondition #4.
862                 return true;
863         }
864         if (ret == -1) {
865                 // Error! Postcondition #1.
866                 log_perror("write");
867                 close_client(client);
868                 return true;
869         }
870         if (ret > 0 && size_t(ret) == client->tls_data_left_to_send) {
871                 // All data has been sent, so we don't need to go to sleep
872                 // (although we are likely to do so immediately afterwards,
873                 // due to lack of client data).
874                 tls_buffer_clear(client->tls_context);
875                 client->tls_data_to_send = nullptr;
876
877                 // Flush the data we just wrote, since the client probably
878                 // is waiting for it.
879                 flush_pending_data(client->sock);
880                 return false;
881         }
882
883         // More data to send, so try again.
884         client->tls_data_to_send += ret;
885         client->tls_data_left_to_send -= ret;
886         goto send_data_again;
887 }
888
889 int Server::read_nontls_data(Client *client, char *buf, size_t max_size)
890 {
891         int ret;
892         do {
893                 ret = read(client->sock, buf, max_size);
894         } while (ret == -1 && errno == EINTR);
895
896         if (ret == -1 && errno == EAGAIN) {
897                 // No more data right now. Nothing to do.
898                 // This is postcondition #2.
899                 return -1;
900         }
901         if (ret == -1) {
902                 log_perror("read");
903                 close_client(client);
904                 return -1;
905         }
906         if (ret == 0) {
907                 // OK, the socket is closed.
908                 close_client(client);
909                 return -1;
910         }
911
912         return ret;
913 }
914
915 int Server::read_tls_data(Client *client, char *buf, size_t max_size)
916 {
917 read_again:
918         int ret;
919         do {
920                 ret = read(client->sock, buf, max_size);
921         } while (ret == -1 && errno == EINTR);
922
923         if (ret == -1 && errno == EAGAIN) {
924                 // No more data right now. Nothing to do.
925                 // This is postcondition #2.
926                 return -1;
927         }
928         if (ret == -1) {
929                 log_perror("read");
930                 close_client(client);
931                 return -1;
932         }
933         if (ret == 0) {
934                 // OK, the socket is closed.
935                 close_client(client);
936                 return -1;
937         }
938
939         // Give it to the TLS library.
940         int err = tls_consume_stream(client->tls_context, reinterpret_cast<const unsigned char *>(buf), ret, nullptr);
941         if (err < 0) {
942                 log_tls_error("tls_consume_stream", err);
943                 close_client(client);
944                 return -1;
945         }
946         if (err == 0) {
947                 // Not consumed any data. See if we can read more.
948                 goto read_again;
949         }
950
951         // Read any decrypted data available for us. (We can reuse buf, since it's free now.)
952         ret = tls_read(client->tls_context, reinterpret_cast<unsigned char *>(buf), max_size);
953         if (ret == 0) {
954                 // No decrypted data for us yet, but there might be some more handshaking
955                 // to send. Do that if needed, then look for more data.
956                 if (send_pending_tls_data(client)) {
957                         // send_pending_tls_data() hit postconditions #1 or #4.
958                         return -1;
959                 }
960                 goto read_again;
961         }
962         if (ret < 0) {
963                 log_tls_error("tls_read", ret);
964                 close_client(client);
965                 return -1;
966         }
967
968         assert(ret > 0);
969         return ret;
970 }
971
972 // See if there's some data we've lost. Ideally, we should drop to a block boundary,
973 // but resync will be the mux's problem.
974 void Server::skip_lost_data(Client *client)
975 {
976         Stream *stream = client->stream;
977         if (stream == nullptr) {
978                 return;
979         }
980         size_t bytes_to_send = stream->bytes_received - client->stream_pos;
981         if (bytes_to_send > stream->backlog_size) {
982                 size_t bytes_lost = bytes_to_send - stream->backlog_size;
983                 client->bytes_lost += bytes_lost;
984                 ++client->num_loss_events;
985                 if (!client->close_after_response) {
986                         assert(client->stream_pos_end != Client::STREAM_POS_NO_END);
987
988                         // We've already sent a Content-Length, so we can't just skip data.
989                         // Close the connection immediately and hope the other side
990                         // is able to figure out that there was an error and it needs to skip.
991                         client->close_after_response = true;
992                         client->stream_pos = client->stream_pos_end;
993                 } else {
994                         client->stream_pos = stream->bytes_received - stream->backlog_size;
995                 }
996         }
997 }
998
999 int Server::parse_request(Client *client)
1000 {
1001         vector<string> lines = split_lines(client->request);
1002         client->request.clear();
1003         if (lines.empty()) {
1004                 return 400;  // Bad request (empty).
1005         }
1006
1007         // Parse the headers, for logging purposes.
1008         HTTPHeaderMultimap headers = extract_headers(lines, client->remote_addr);
1009         const auto referer_it = headers.find("Referer");
1010         if (referer_it != headers.end()) {
1011                 client->referer = referer_it->second;
1012         }
1013         const auto user_agent_it = headers.find("User-Agent");
1014         if (user_agent_it != headers.end()) {
1015                 client->user_agent = user_agent_it->second;
1016         }
1017         const auto x_playback_session_id_it = headers.find("X-Playback-Session-Id");
1018         if (x_playback_session_id_it != headers.end()) {
1019                 client->x_playback_session_id = x_playback_session_id_it->second;
1020         } else {
1021                 client->x_playback_session_id.clear();
1022         }
1023
1024         vector<string> request_tokens = split_tokens(lines[0]);
1025         if (request_tokens.size() < 3) {
1026                 return 400;  // Bad request (empty).
1027         }
1028         if (request_tokens[0] != "GET") {
1029                 return 400;  // Should maybe be 405 instead?
1030         }
1031
1032         string url = request_tokens[1];
1033         client->url = url;
1034         if (url.size() > 8 && url.find("?backlog") == url.size() - 8) {
1035                 client->stream_pos = Client::STREAM_POS_AT_START;
1036                 url = url.substr(0, url.size() - 8);
1037         } else {
1038                 size_t pos = url.find("?frag=");
1039                 if (pos != string::npos) {
1040                         // Parse an endpoint of the type /stream.mp4?frag=1234-5678.
1041                         const char *ptr = url.c_str() + pos + 6;
1042
1043                         // "?frag=header" is special.
1044                         if (strcmp(ptr, "header") == 0) {
1045                                 client->stream_pos = Client::STREAM_POS_HEADER_ONLY;
1046                                 client->stream_pos_end = -1;
1047                         } else {
1048                                 char *endptr;
1049                                 long long frag_start = strtol(ptr, &endptr, 10);
1050                                 if (ptr == endptr || frag_start < 0 || frag_start == LLONG_MAX) {
1051                                         return 400;  // Bad request.
1052                                 }
1053                                 if (*endptr != '-') {
1054                                         return 400;  // Bad request.
1055                                 }
1056                                 ptr = endptr + 1;
1057
1058                                 long long frag_end = strtol(ptr, &endptr, 10);
1059                                 if (ptr == endptr || frag_end < frag_start || frag_end == LLONG_MAX) {
1060                                         return 400;  // Bad request.
1061                                 }
1062
1063                                 if (*endptr != '\0') {
1064                                         return 400;  // Bad request.
1065                                 }
1066
1067                                 client->stream_pos = frag_start;
1068                                 client->stream_pos_end = frag_end;
1069                         }
1070                         url = url.substr(0, pos);
1071                 } else {
1072                         client->stream_pos = -1;
1073                         client->stream_pos_end = -1;
1074                 }
1075         }
1076
1077         // Figure out if we're supposed to close the socket after we've delivered the response.
1078         string protocol = request_tokens[2];
1079         if (protocol.find("HTTP/") != 0) {
1080                 return 400;  // Bad request.
1081         }
1082         client->close_after_response = false;
1083         client->http_11 = true;
1084         if (protocol == "HTTP/1.0") {
1085                 // No persistent connections.
1086                 client->close_after_response = true;
1087                 client->http_11 = false;
1088         } else {
1089                 const auto connection_it = headers.find("Connection");
1090                 if (connection_it != headers.end() && connection_it->second == "close") {
1091                         client->close_after_response = true;
1092                 }
1093         }
1094
1095         const auto stream_url_map_it = stream_url_map.find(url);
1096         if (stream_url_map_it != stream_url_map.end()) {
1097                 // Serve a regular stream..
1098                 client->stream = streams[stream_url_map_it->second].get();
1099                 client->serving_hls_playlist = false;
1100         } else {
1101                 const auto stream_hls_url_map_it = stream_hls_url_map.find(url);
1102                 if (stream_hls_url_map_it != stream_hls_url_map.end()) {
1103                         // Serve HLS playlist.
1104                         client->stream = streams[stream_hls_url_map_it->second].get();
1105                         client->serving_hls_playlist = true;
1106                 } else {
1107                         const auto ping_url_map_it = ping_url_map.find(url);
1108                         if (ping_url_map_it == ping_url_map.end()) {
1109                                 return 404;  // Not found.
1110                         } else {
1111                                 // Serve a ping (204 no error).
1112                                 return 204;
1113                         }
1114                 }
1115         }
1116
1117         Stream *stream = client->stream;
1118         if (stream->http_header.empty()) {
1119                 return 503;  // Service unavailable.
1120         }
1121
1122         if (client->serving_hls_playlist) {
1123                 if (stream->encoding == Stream::STREAM_ENCODING_METACUBE) {
1124                         // This doesn't make any sense, and is hard to implement, too.
1125                         return 404;
1126                 } else {
1127                         return 200;
1128                 }
1129         }
1130
1131         if (client->stream_pos_end == Client::STREAM_POS_NO_END) {
1132                 // This stream won't end, so we don't have a content-length,
1133                 // and can just as well tell the client it's Connection: close
1134                 // (otherwise, we'd have to implement chunking TE for no good reason).
1135                 client->close_after_response = true;
1136         } else {
1137                 if (stream->encoding == Stream::STREAM_ENCODING_METACUBE) {
1138                         // This doesn't make any sense, and is hard to implement, too.
1139                         return 416;  // Range not satisfiable.
1140                 }
1141
1142                 // Check that we have the requested fragment in our backlog.
1143                 size_t buffer_end = stream->bytes_received;
1144                 size_t buffer_start = (buffer_end <= stream->backlog_size) ? 0 : buffer_end - stream->backlog_size;
1145
1146                 if (client->stream_pos_end > buffer_end ||
1147                     client->stream_pos < buffer_start) {
1148                         return 416;  // Range not satisfiable.
1149                 }
1150         }
1151
1152         client->stream = stream;
1153         if (setsockopt(client->sock, SOL_SOCKET, SO_MAX_PACING_RATE, &client->stream->pacing_rate, sizeof(client->stream->pacing_rate)) == -1) {
1154                 if (client->stream->pacing_rate != ~0U) {
1155                         log_perror("setsockopt(SO_MAX_PACING_RATE)");
1156                 }
1157         }
1158         client->request.clear();
1159
1160         return 200;  // OK!
1161 }
1162
1163 void Server::construct_stream_header(Client *client)
1164 {
1165         Stream *stream = client->stream;
1166         string response = stream->http_header;
1167         if (client->stream_pos == Client::STREAM_POS_HEADER_ONLY) {
1168                 char buf[64];
1169                 snprintf(buf, sizeof(buf), "Content-Length: %zu\r\n", stream->stream_header.size());
1170                 response.append(buf);
1171         } else if (client->stream_pos_end != Client::STREAM_POS_NO_END) {
1172                 char buf[64];
1173                 snprintf(buf, sizeof(buf), "Content-Length: %" PRIu64 "\r\n", client->stream_pos_end - client->stream_pos);
1174                 response.append(buf);
1175         }
1176         if (client->http_11) {
1177                 assert(response.find("HTTP/1.0") == 0);
1178                 response[7] = '1';  // Change to HTTP/1.1.
1179                 if (client->close_after_response) {
1180                         response.append("Connection: close\r\n");
1181                 }
1182         } else {
1183                 assert(client->close_after_response);
1184         }
1185         if (!stream->allow_origin.empty()) {
1186                 response.append("Access-Control-Allow-Origin: ");
1187                 response.append(stream->allow_origin);
1188                 response.append("\r\n");
1189         }
1190         if (stream->encoding == Stream::STREAM_ENCODING_RAW) {
1191                 response.append("\r\n");
1192         } else if (stream->encoding == Stream::STREAM_ENCODING_METACUBE) {
1193                 response.append("Content-Encoding: metacube\r\n\r\n");
1194                 if (!stream->stream_header.empty()) {
1195                         metacube2_block_header hdr;
1196                         memcpy(hdr.sync, METACUBE2_SYNC, sizeof(hdr.sync));
1197                         hdr.size = htonl(stream->stream_header.size());
1198                         hdr.flags = htons(METACUBE_FLAGS_HEADER);
1199                         hdr.csum = htons(metacube2_compute_crc(&hdr));
1200                         response.append(string(reinterpret_cast<char *>(&hdr), sizeof(hdr)));
1201                 }
1202         } else {
1203                 assert(false);
1204         }
1205         if (client->stream_pos == Client::STREAM_POS_HEADER_ONLY) {
1206                 client->state = Client::SENDING_SHORT_RESPONSE;
1207                 response.append(stream->stream_header);
1208         } else {
1209                 client->state = Client::SENDING_HEADER;
1210                 if (client->stream_pos_end == Client::STREAM_POS_NO_END) {  // Fragments don't contain stream headers.
1211                         response.append(stream->stream_header);
1212                 }
1213         }
1214
1215         client->header_or_short_response_holder = move(response);
1216         client->header_or_short_response = &client->header_or_short_response_holder;
1217
1218         // Switch states.
1219         change_epoll_events(client, EPOLLOUT | EPOLLET | EPOLLRDHUP);
1220 }
1221         
1222 void Server::construct_error(Client *client, int error_code)
1223 {
1224         char error[256];
1225         if (client->http_11 && client->close_after_response) {
1226                 snprintf(error, sizeof(error),
1227                         "HTTP/1.1 %d Error\r\nContent-Type: text/plain\r\nConnection: close\r\n\r\nSomething went wrong. Sorry.\r\n",
1228                         error_code);
1229         } else {
1230                 snprintf(error, sizeof(error),
1231                         "HTTP/1.%d %d Error\r\nContent-Type: text/plain\r\nContent-Length: 30\r\n\r\nSomething went wrong. Sorry.\r\n",
1232                         client->http_11, error_code);
1233         }
1234         client->header_or_short_response_holder = error;
1235         client->header_or_short_response = &client->header_or_short_response_holder;
1236
1237         // Switch states.
1238         client->state = Client::SENDING_SHORT_RESPONSE;
1239         change_epoll_events(client, EPOLLOUT | EPOLLET | EPOLLRDHUP);
1240 }
1241
1242 void Server::construct_hls_playlist(Client *client)
1243 {
1244         Stream *stream = client->stream;
1245         shared_ptr<const string> *cache;
1246         if (client->http_11) {
1247                 if (client->close_after_response) {
1248                         cache = &stream->hls_playlist_http11_close;
1249                 } else {
1250                         cache = &stream->hls_playlist_http11_persistent;
1251                 }
1252         } else {
1253                 assert(client->close_after_response);
1254                 cache = &stream->hls_playlist_http10;
1255         }
1256
1257         if (*cache == nullptr) {
1258                 *cache = stream->generate_hls_playlist(client->http_11, client->close_after_response);
1259         }
1260         client->header_or_short_response_ref = *cache;
1261         client->header_or_short_response = cache->get();
1262
1263         // Switch states.
1264         client->state = Client::SENDING_SHORT_RESPONSE;
1265         change_epoll_events(client, EPOLLOUT | EPOLLET | EPOLLRDHUP);
1266 }
1267
1268 void Server::construct_204(Client *client)
1269 {
1270         const auto ping_url_map_it = ping_url_map.find(client->url);
1271         assert(ping_url_map_it != ping_url_map.end());
1272
1273         string response;
1274         if (client->http_11) {
1275                 response = "HTTP/1.1 204 No Content\r\n";
1276                 if (client->close_after_response) {
1277                         response.append("Connection: close\r\n");
1278                 }
1279         } else {
1280                 response = "HTTP/1.0 204 No Content\r\n";
1281                 assert(client->close_after_response);
1282         }
1283         if (!ping_url_map_it->second.empty()) {
1284                 response.append("Access-Control-Allow-Origin: ");
1285                 response.append(ping_url_map_it->second);
1286                 response.append("\r\n");
1287         }
1288         response.append("\r\n");
1289
1290         client->header_or_short_response_holder = move(response);
1291         client->header_or_short_response = &client->header_or_short_response_holder;
1292
1293         // Switch states.
1294         client->state = Client::SENDING_SHORT_RESPONSE;
1295         change_epoll_events(client, EPOLLOUT | EPOLLET | EPOLLRDHUP);
1296 }
1297
1298 namespace {
1299
1300 template<class T>
1301 void delete_from(vector<T> *v, T elem)
1302 {
1303         typename vector<T>::iterator new_end = remove(v->begin(), v->end(), elem);
1304         v->erase(new_end, v->end());
1305 }
1306
1307 void send_ktls_close(int sock)
1308 {
1309         uint8_t record_type = 21;  // Alert.
1310         uint8_t body[] = {
1311                 1,   // Warning level (but still fatal!).
1312                 0,   // close_notify.
1313         };
1314
1315         int cmsg_len = sizeof(record_type);
1316         char buf[CMSG_SPACE(cmsg_len)];
1317
1318         msghdr msg = {0};
1319         msg.msg_control = buf;
1320         msg.msg_controllen = sizeof(buf);
1321         cmsghdr *cmsg = CMSG_FIRSTHDR(&msg);
1322         cmsg->cmsg_level = SOL_TLS;
1323         cmsg->cmsg_type = TLS_SET_RECORD_TYPE;
1324         cmsg->cmsg_len = CMSG_LEN(cmsg_len);
1325         *CMSG_DATA(cmsg) = record_type;
1326         msg.msg_controllen = cmsg->cmsg_len;
1327
1328         iovec msg_iov;
1329         msg_iov.iov_base = body;
1330         msg_iov.iov_len = sizeof(body);
1331         msg.msg_iov = &msg_iov;
1332         msg.msg_iovlen = 1;
1333
1334         int err;
1335         do {
1336                 err = sendmsg(sock, &msg, 0);
1337         } while (err == -1 && errno == EINTR);  // Ignore all other errors.
1338 }
1339
1340 }  // namespace
1341         
1342 void Server::close_client(Client *client)
1343 {
1344         if (epoll_ctl(epoll_fd, EPOLL_CTL_DEL, client->sock, nullptr) == -1) {
1345                 log_perror("epoll_ctl(EPOLL_CTL_DEL)");
1346                 exit(1);
1347         }
1348
1349         // This client could be sleeping, so we'll need to fix that. (Argh, O(n).)
1350         if (client->stream != nullptr) {
1351                 delete_from(&client->stream->sleeping_clients, client);
1352                 delete_from(&client->stream->to_process, client);
1353         }
1354
1355         if (client->tls_context) {
1356                 if (client->in_ktls_mode) {
1357                         // Keep GnuTLS happy.
1358                         send_ktls_close(client->sock);
1359                 }
1360                 tls_destroy_context(client->tls_context);
1361         }
1362
1363         // Log to access_log.
1364         access_log->write(client->get_stats());
1365
1366         // Bye-bye!
1367         safe_close(client->sock);
1368
1369         clients.erase(client->sock);
1370 }
1371
1372 void Server::change_epoll_events(Client *client, uint32_t events)
1373 {
1374         epoll_event ev;
1375         ev.events = events;
1376         ev.data.ptr = client;
1377
1378         if (epoll_ctl(epoll_fd, EPOLL_CTL_MOD, client->sock, &ev) == -1) {
1379                 log_perror("epoll_ctl(EPOLL_CTL_MOD)");
1380                 exit(1);
1381         }
1382 }
1383
1384 bool Server::more_requests(Client *client)
1385 {
1386         if (client->close_after_response) {
1387                 return false;
1388         }
1389
1390         // Log to access_log.
1391         access_log->write(client->get_stats());
1392
1393         flush_pending_data(client->sock);
1394
1395         // Switch states and reset the parsers. We don't reset statistics.
1396         client->state = Client::READING_REQUEST;
1397         client->url.clear();
1398         client->stream = NULL;
1399         client->header_or_short_response = nullptr;
1400         client->header_or_short_response_holder.clear();
1401         client->header_or_short_response_ref.reset();
1402         client->header_or_short_response_bytes_sent = 0;
1403         client->bytes_sent = 0;
1404         start_client_timeout_timer(client);
1405
1406         change_epoll_events(client, EPOLLIN | EPOLLET | EPOLLRDHUP);  // No TLS handshake, so no EPOLLOUT needed.
1407
1408         return true;
1409 }
1410
1411 void Server::process_queued_data()
1412 {
1413         {
1414                 lock_guard<mutex> lock(queued_clients_mutex);
1415
1416                 for (const pair<int, Acceptor *> &id_and_acceptor : queued_add_clients) {
1417                         add_client(id_and_acceptor.first, id_and_acceptor.second);
1418                 }
1419                 queued_add_clients.clear();
1420         }
1421
1422         for (unique_ptr<Stream> &stream : streams) {
1423                 stream->process_queued_data();
1424         }
1425 }