Fix an ineffective assert.
[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 < streams[stream_index]->backlog_size);
439         streams[stream_index]->hls_backlog_margin = hls_backlog_margin;
440 }
441
442 void Server::set_allow_origin(int stream_index, const string &allow_origin)
443 {
444         lock_guard<mutex> lock(mu);
445         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
446         streams[stream_index]->allow_origin = allow_origin;
447 }
448
449 void Server::register_hls_url(int stream_index, const string &hls_url)
450 {
451         lock_guard<mutex> lock(mu);
452         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
453         assert(!hls_url.empty());
454         stream_hls_url_map.insert(make_pair(hls_url, stream_index));
455 }
456         
457 void Server::set_header(int stream_index, const string &http_header, const string &stream_header)
458 {
459         lock_guard<mutex> lock(mu);
460         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
461         Stream *stream = streams[stream_index].get();
462         stream->http_header = http_header;
463
464         if (stream_header != stream->stream_header) {
465                 // We cannot start at any of the older starting points anymore,
466                 // since they'd get the wrong header for the stream (not to mention
467                 // that a changed header probably means the stream restarted,
468                 // which means any client starting on the old one would probably
469                 // stop playing properly at the change point). Next block
470                 // should be a suitable starting point (if not, something is
471                 // pretty strange), so it will fill up again soon enough.
472                 stream->suitable_starting_points.clear();
473
474                 if (!stream->fragments.empty()) {
475                         stream->fragments.clear();
476                         ++stream->discontinuity_counter;
477                         stream->clear_hls_playlist_cache();
478                 }
479         }
480         stream->stream_header = stream_header;
481 }
482         
483 void Server::set_pacing_rate(int stream_index, uint32_t pacing_rate)
484 {
485         lock_guard<mutex> lock(mu);
486         assert(clients.empty());
487         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
488         streams[stream_index]->pacing_rate = pacing_rate;
489 }
490
491 void Server::add_gen204(const std::string &url, const std::string &allow_origin)
492 {
493         lock_guard<mutex> lock(mu);
494         assert(clients.empty());
495         ping_url_map[url] = allow_origin;
496 }
497
498 void Server::create_tls_context_for_acceptor(const Acceptor *acceptor)
499 {
500         assert(acceptor->is_tls());
501
502         bool is_server = true;
503         TLSContext *server_context = tls_create_context(is_server, TLS_V12);
504
505         const string &cert = acceptor->get_certificate_chain();
506         int num_cert = tls_load_certificates(server_context, reinterpret_cast<const unsigned char *>(cert.data()), cert.size());
507         assert(num_cert > 0);  // Should have been checked by config earlier.
508
509         const string &key = acceptor->get_private_key();
510         int num_key = tls_load_private_key(server_context, reinterpret_cast<const unsigned char *>(key.data()), key.size());
511         assert(num_key > 0);  // Should have been checked by config earlier.
512
513         tls_server_contexts.insert(make_pair(acceptor, server_context));
514 }
515
516 void Server::add_data_deferred(int stream_index, const char *data, size_t bytes, uint16_t metacube_flags, const RationalPTS &pts)
517 {
518         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
519         streams[stream_index]->add_data_deferred(data, bytes, metacube_flags, pts);
520 }
521
522 // See the .h file for postconditions after this function.      
523 void Server::process_client(Client *client)
524 {
525         switch (client->state) {
526         case Client::READING_REQUEST: {
527                 if (client->tls_context != nullptr && !client->in_ktls_mode) {
528                         if (send_pending_tls_data(client)) {
529                                 // send_pending_tls_data() hit postconditions #1 or #4.
530                                 return;
531                         }
532                 }
533
534 read_request_again:
535                 // Try to read more of the request.
536                 char buf[1024];
537                 int ret;
538                 if (client->tls_context == nullptr || client->in_ktls_mode) {
539                         ret = read_plain_data(client, buf, sizeof(buf));
540                         if (ret == -1) {
541                                 // read_plain_data() hit postconditions #1 or #2.
542                                 return;
543                         }
544                 } else {
545                         ret = read_tls_data(client, buf, sizeof(buf));
546                         if (ret == -1) {
547                                 // read_tls_data() hit postconditions #1, #2 or #4.
548                                 return;
549                         }
550                 }
551
552                 RequestParseStatus status = wait_for_double_newline(&client->request, buf, ret);
553         
554                 switch (status) {
555                 case RP_OUT_OF_SPACE:
556                         log(WARNING, "[%s] Client sent overlong request!", client->remote_addr.c_str());
557                         close_client(client);
558                         return;
559                 case RP_NOT_FINISHED_YET:
560                         // OK, we don't have the entire header yet. Fine; we'll get it later.
561                         // See if there's more data for us.
562                         goto read_request_again;
563                 case RP_EXTRA_DATA:
564                         log(WARNING, "[%s] Junk data after request!", client->remote_addr.c_str());
565                         close_client(client);
566                         return;
567                 case RP_FINISHED:
568                         break;
569                 }
570
571                 assert(status == RP_FINISHED);
572
573                 int error_code = parse_request(client);
574                 if (error_code == 200) {
575                         if (client->serving_hls_playlist) {
576                                 construct_hls_playlist(client);
577                         } else {
578                                 construct_stream_header(client);
579                         }
580                 } else if (error_code == 204) {
581                         construct_204(client);
582                 } else {
583                         construct_error(client, error_code);
584                 }
585
586                 // We've changed states, so fall through.
587                 assert(client->state == Client::SENDING_SHORT_RESPONSE ||
588                        client->state == Client::SENDING_HEADER);
589         }
590         case Client::SENDING_SHORT_RESPONSE:
591         case Client::SENDING_HEADER: {
592 sending_header_or_short_response_again:
593                 int ret;
594                 do {
595                         ret = write(client->sock,
596                                     client->header_or_short_response->data() + client->header_or_short_response_bytes_sent,
597                                     client->header_or_short_response->size() - client->header_or_short_response_bytes_sent);
598                 } while (ret == -1 && errno == EINTR);
599
600                 if (ret == -1 && errno == EAGAIN) {
601                         // We're out of socket space, so now we're at the “low edge” of epoll's
602                         // edge triggering. epoll will tell us when there is more room, so for now,
603                         // just return.
604                         // This is postcondition #4.
605                         return;
606                 }
607
608                 if (ret == -1) {
609                         // Error! Postcondition #1.
610                         log_perror("write");
611                         close_client(client);
612                         return;
613                 }
614                 
615                 client->header_or_short_response_bytes_sent += ret;
616                 assert(client->header_or_short_response_bytes_sent <= client->header_or_short_response->size());
617
618                 if (client->header_or_short_response_bytes_sent < client->header_or_short_response->size()) {
619                         // We haven't sent all yet. Fine; go another round.
620                         goto sending_header_or_short_response_again;
621                 }
622
623                 // We're done sending the header or error! Clear it to release some memory.
624                 client->header_or_short_response = nullptr;
625                 client->header_or_short_response_holder.clear();
626                 client->header_or_short_response_ref.reset();
627
628                 if (client->state == Client::SENDING_SHORT_RESPONSE) {
629                         if (more_requests(client)) {
630                                 // We're done sending the error, but should keep on reading new requests.
631                                 goto read_request_again;
632                         } else {
633                                 // We're done sending the error, so now close.
634                                 // This is postcondition #1.
635                                 close_client(client);
636                         }
637                         return;
638                 }
639
640                 Stream *stream = client->stream;
641                 hls_zombies.erase(client->get_hls_zombie_key());
642                 if (client->stream_pos == Client::STREAM_POS_AT_START) {
643                         // Start sending from the beginning of the backlog.
644                         client->stream_pos = min<size_t>(
645                             stream->bytes_received - stream->backlog_size,
646                             0);
647                         client->state = Client::SENDING_DATA;
648                         goto sending_data;
649                 } else if (client->stream_pos_end != Client::STREAM_POS_NO_END) {
650                         // We're sending a fragment, and should have all of it,
651                         // so start sending right away.
652                         assert(ssize_t(client->stream_pos) >= 0);
653                         client->state = Client::SENDING_DATA;
654                         goto sending_data;
655                 } else if (stream->prebuffering_bytes == 0) {
656                         // Start sending from the first keyframe we get. In other
657                         // words, we won't send any of the backlog, but we'll start
658                         // sending immediately as we get the next keyframe block.
659                         // Note that this is functionally identical to the next if branch,
660                         // except that we save a binary search.
661                         assert(client->stream_pos == Client::STREAM_POS_AT_END);
662                         assert(client->stream_pos_end == Client::STREAM_POS_NO_END);
663                         client->stream_pos = stream->bytes_received;
664                         client->state = Client::WAITING_FOR_KEYFRAME;
665                 } else {
666                         // We're not going to send anything to the client before we have
667                         // N bytes. However, this wait might be boring; we can just as well
668                         // use it to send older data if we have it. We use lower_bound()
669                         // so that we are conservative and never add extra latency over just
670                         // waiting (assuming CBR or nearly so); otherwise, we could want e.g.
671                         // 100 kB prebuffer but end up sending a 10 MB GOP.
672                         assert(client->stream_pos == Client::STREAM_POS_AT_END);
673                         assert(client->stream_pos_end == Client::STREAM_POS_NO_END);
674                         deque<uint64_t>::const_iterator starting_point_it =
675                                 lower_bound(stream->suitable_starting_points.begin(),
676                                             stream->suitable_starting_points.end(),
677                                             stream->bytes_received - stream->prebuffering_bytes);
678                         if (starting_point_it == stream->suitable_starting_points.end()) {
679                                 // None found. Just put us at the end, and then wait for the
680                                 // first keyframe to appear.
681                                 client->stream_pos = stream->bytes_received;
682                                 client->state = Client::WAITING_FOR_KEYFRAME;
683                         } else {
684                                 client->stream_pos = *starting_point_it;
685                                 client->state = Client::PREBUFFERING;
686                                 goto prebuffering;
687                         }
688                 }
689                 // Fall through.
690         }
691         case Client::WAITING_FOR_KEYFRAME: {
692                 Stream *stream = client->stream;
693                 if (stream->suitable_starting_points.empty() ||
694                     client->stream_pos > stream->suitable_starting_points.back()) {
695                         // We haven't received a keyframe since this stream started waiting,
696                         // so keep on waiting for one.
697                         // This is postcondition #3.
698                         stream->put_client_to_sleep(client);
699                         return;
700                 }
701                 client->stream_pos = stream->suitable_starting_points.back();
702                 client->state = Client::PREBUFFERING;
703                 // Fall through.
704         }
705         case Client::PREBUFFERING: {
706 prebuffering:
707                 Stream *stream = client->stream;
708                 size_t bytes_to_send = stream->bytes_received - client->stream_pos;
709                 assert(bytes_to_send <= stream->backlog_size);
710                 if (bytes_to_send < stream->prebuffering_bytes) {
711                         // We don't have enough bytes buffered to start this client yet.
712                         // This is postcondition #3.
713                         stream->put_client_to_sleep(client);
714                         return;
715                 }
716                 client->state = Client::SENDING_DATA;
717                 // Fall through.
718         }
719         case Client::SENDING_DATA: {
720 sending_data:
721                 skip_lost_data(client);
722                 Stream *stream = client->stream;
723
724 sending_data_again:
725                 size_t bytes_to_send;
726                 if (client->stream_pos_end == Client::STREAM_POS_NO_END) {
727                          bytes_to_send = stream->bytes_received - client->stream_pos;
728                 } else {
729                          bytes_to_send = client->stream_pos_end - client->stream_pos;
730                 }
731                 assert(bytes_to_send <= stream->backlog_size);
732                 if (bytes_to_send == 0) {
733                         if (client->stream_pos == client->stream_pos_end) {  // We have a definite end, and we're at it.
734                                 // Add (or overwrite) a HLS zombie.
735                                 timespec now;
736                                 if (clock_gettime(CLOCK_MONOTONIC_COARSE, &now) == -1) {
737                                         log_perror("clock_gettime(CLOCK_MONOTONIC_COARSE)");
738                                 } else {
739                                         HLSZombie zombie;
740                                         zombie.remote_addr = client->remote_addr;
741                                         zombie.referer = client->referer;
742                                         zombie.user_agent = client->user_agent;
743                                         zombie.url = client->stream->url + "?frag=<idle>";
744                                         zombie.expires = now;
745                                         zombie.expires.tv_sec += client->stream->hls_frag_duration * 3;
746                                         hls_zombies[client->get_hls_zombie_key()] = move(zombie);
747                                 }
748                                 if (more_requests(client)) {
749                                         // We're done sending the fragment, but should keep on reading new requests.
750                                         goto read_request_again;
751                                 } else {
752                                         // We're done sending the fragment, so now close.
753                                         // This is postcondition #1.
754                                         close_client(client);
755                                 }
756                         }
757                         return;
758                 }
759
760                 // See if we need to split across the circular buffer.
761                 bool more_data = false;
762                 if ((client->stream_pos % stream->backlog_size) + bytes_to_send > stream->backlog_size) {
763                         bytes_to_send = stream->backlog_size - (client->stream_pos % stream->backlog_size);
764                         more_data = true;
765                 }
766
767                 ssize_t ret;
768                 do {
769                         off_t offset = client->stream_pos % stream->backlog_size;
770                         ret = sendfile(client->sock, stream->data_fd, &offset, bytes_to_send);
771                 } while (ret == -1 && errno == EINTR);
772
773                 if (ret == -1 && errno == EAGAIN) {
774                         // We're out of socket space, so return; epoll will wake us up
775                         // when there is more room.
776                         // This is postcondition #4.
777                         return;
778                 }
779                 if (ret == -1) {
780                         // Error, close; postcondition #1.
781                         log_perror("sendfile");
782                         close_client(client);
783                         return;
784                 }
785                 client->stream_pos += ret;
786                 client->bytes_sent += ret;
787
788                 assert(client->stream_pos_end == Client::STREAM_POS_NO_END || client->stream_pos <= client->stream_pos_end);
789                 if (client->stream_pos == client->stream_pos_end) {
790                         goto sending_data_again;  // Will see that bytes_to_send == 0 and end.
791                 } else if (client->stream_pos == stream->bytes_received) {
792                         // We don't have any more data for this client, so put it to sleep.
793                         // This is postcondition #3.
794                         stream->put_client_to_sleep(client);
795                 } else if (more_data && size_t(ret) == bytes_to_send) {
796                         goto sending_data_again;
797                 }
798                 // We'll also get here for postcondition #4 (similar to the EAGAIN path above).
799                 break;
800         }
801         default:
802                 assert(false);
803         }
804 }
805
806 namespace {
807
808 void flush_pending_data(int sock)
809 {
810         // Flush pending data, which would otherwise wait for the 200ms TCP_CORK timer
811         // to elapsed; does not cancel out TCP_CORK (since that still takes priority),
812         // but does a one-off flush.
813         int one = 1;
814         if (setsockopt(sock, SOL_TCP, TCP_NODELAY, &one, sizeof(one)) == -1) {
815                 log_perror("setsockopt(TCP_NODELAY)");
816                 // Can still continue.
817         }
818 }
819
820 }  // namespace
821
822 bool Server::send_pending_tls_data(Client *client)
823 {
824         // See if there's data from the TLS library to write.
825         if (client->tls_data_to_send == nullptr) {
826                 client->tls_data_to_send = tls_get_write_buffer(client->tls_context, &client->tls_data_left_to_send);
827                 if (client->tls_data_to_send == nullptr) {
828                         // Really no data to send.
829                         return false;
830                 }
831         }
832
833 send_data_again:
834         int ret;
835         do {
836                 ret = write(client->sock, client->tls_data_to_send, client->tls_data_left_to_send);
837         } while (ret == -1 && errno == EINTR);
838         assert(ret < 0 || size_t(ret) <= client->tls_data_left_to_send);
839
840         if (ret == -1 && errno == EAGAIN) {
841                 // We're out of socket space, so now we're at the “low edge” of epoll's
842                 // edge triggering. epoll will tell us when there is more room, so for now,
843                 // just return.
844                 // This is postcondition #4.
845                 return true;
846         }
847         if (ret == -1) {
848                 // Error! Postcondition #1.
849                 log_perror("write");
850                 close_client(client);
851                 return true;
852         }
853         if (ret > 0 && size_t(ret) == client->tls_data_left_to_send) {
854                 // All data has been sent, so we don't need to go to sleep
855                 // (although we are likely to do so immediately afterwards,
856                 // due to lack of client data).
857                 tls_buffer_clear(client->tls_context);
858                 client->tls_data_to_send = nullptr;
859
860                 // Flush the data we just wrote, since the client probably
861                 // is waiting for it.
862                 flush_pending_data(client->sock);
863                 return false;
864         }
865
866         // More data to send, so try again.
867         client->tls_data_to_send += ret;
868         client->tls_data_left_to_send -= ret;
869         goto send_data_again;
870 }
871
872 int Server::read_plain_data(Client *client, char *buf, size_t max_size)
873 {
874         int ret;
875         do {
876                 ret = read(client->sock, buf, max_size);
877         } while (ret == -1 && errno == EINTR);
878
879         if (ret == -1 && errno == EAGAIN) {
880                 // No more data right now. Nothing to do.
881                 // This is postcondition #2.
882                 return -1;
883         }
884         if (ret == -1) {
885                 log_perror("read");
886                 close_client(client);
887                 return -1;
888         }
889         if (ret == 0) {
890                 // OK, the socket is closed.
891                 close_client(client);
892                 return -1;
893         }
894
895         return ret;
896 }
897
898 int Server::read_tls_data(Client *client, char *buf, size_t max_size)
899 {
900 read_again:
901         assert(!client->in_ktls_mode);
902
903         int ret;
904         do {
905                 ret = read(client->sock, buf, max_size);
906         } while (ret == -1 && errno == EINTR);
907
908         if (ret == -1 && errno == EAGAIN) {
909                 // No more data right now. Nothing to do.
910                 // This is postcondition #2.
911                 return -1;
912         }
913         if (ret == -1) {
914                 log_perror("read");
915                 close_client(client);
916                 return -1;
917         }
918         if (ret == 0) {
919                 // OK, the socket is closed.
920                 close_client(client);
921                 return -1;
922         }
923
924         // Give it to the TLS library.
925         int err = tls_consume_stream(client->tls_context, reinterpret_cast<const unsigned char *>(buf), ret, nullptr);
926         if (err < 0) {
927                 log_tls_error("tls_consume_stream", err);
928                 close_client(client);
929                 return -1;
930         }
931         if (err == 0) {
932                 // Not consumed any data. See if we can read more.
933                 goto read_again;
934         }
935
936         // Read any decrypted data available for us. (We can reuse buf, since it's free now.)
937         ret = tls_read(client->tls_context, reinterpret_cast<unsigned char *>(buf), max_size);
938         if (ret == 0) {
939                 // No decrypted data for us yet, but there might be some more handshaking
940                 // to send. Do that if needed, then look for more data.
941                 if (send_pending_tls_data(client)) {
942                         // send_pending_tls_data() hit postconditions #1 or #4.
943                         return -1;
944                 }
945                 goto read_again;
946         }
947         if (ret < 0) {
948                 log_tls_error("tls_read", ret);
949                 close_client(client);
950                 return -1;
951         }
952
953         if (tls_established(client->tls_context)) {
954                 // We're ready to enter kTLS mode, unless we still have some
955                 // handshake data to send (which then must be sent as non-kTLS).
956                 if (send_pending_tls_data(client)) {
957                         // send_pending_tls_data() hit postconditions #1 or #4.
958                         return -1;
959                 }
960                 int err = tls_make_ktls(client->tls_context, client->sock);  // Don't overwrite ret.
961                 if (err < 0) {
962                         log_tls_error("tls_make_ktls", ret);
963                         close_client(client);
964                         return -1;
965                 }
966                 client->in_ktls_mode = true;
967         }
968
969         assert(ret > 0);
970         return ret;
971 }
972
973 // See if there's some data we've lost. Ideally, we should drop to a block boundary,
974 // but resync will be the mux's problem.
975 void Server::skip_lost_data(Client *client)
976 {
977         Stream *stream = client->stream;
978         if (stream == nullptr) {
979                 return;
980         }
981         size_t bytes_to_send = stream->bytes_received - client->stream_pos;
982         if (bytes_to_send > stream->backlog_size) {
983                 size_t bytes_lost = bytes_to_send - stream->backlog_size;
984                 client->bytes_lost += bytes_lost;
985                 ++client->num_loss_events;
986                 if (!client->close_after_response) {
987                         assert(client->stream_pos_end != Client::STREAM_POS_NO_END);
988
989                         // We've already sent a Content-Length, so we can't just skip data.
990                         // Close the connection immediately and hope the other side
991                         // is able to figure out that there was an error and it needs to skip.
992                         client->close_after_response = true;
993                         client->stream_pos = client->stream_pos_end;
994                 } else {
995                         client->stream_pos = stream->bytes_received - stream->backlog_size;
996                 }
997         }
998 }
999
1000 int Server::parse_request(Client *client)
1001 {
1002         vector<string> lines = split_lines(client->request);
1003         client->request.clear();
1004         if (lines.empty()) {
1005                 return 400;  // Bad request (empty).
1006         }
1007
1008         // Parse the headers, for logging purposes.
1009         HTTPHeaderMultimap headers = extract_headers(lines, client->remote_addr);
1010         const auto referer_it = headers.find("Referer");
1011         if (referer_it != headers.end()) {
1012                 client->referer = referer_it->second;
1013         }
1014         const auto user_agent_it = headers.find("User-Agent");
1015         if (user_agent_it != headers.end()) {
1016                 client->user_agent = user_agent_it->second;
1017         }
1018         const auto x_playback_session_id_it = headers.find("X-Playback-Session-Id");
1019         if (x_playback_session_id_it != headers.end()) {
1020                 client->x_playback_session_id = x_playback_session_id_it->second;
1021         } else {
1022                 client->x_playback_session_id.clear();
1023         }
1024
1025         vector<string> request_tokens = split_tokens(lines[0]);
1026         if (request_tokens.size() < 3) {
1027                 return 400;  // Bad request (empty).
1028         }
1029         if (request_tokens[0] != "GET") {
1030                 return 400;  // Should maybe be 405 instead?
1031         }
1032
1033         string url = request_tokens[1];
1034         client->url = url;
1035         if (url.size() > 8 && url.find("?backlog") == url.size() - 8) {
1036                 client->stream_pos = Client::STREAM_POS_AT_START;
1037                 url = url.substr(0, url.size() - 8);
1038         } else {
1039                 size_t pos = url.find("?frag=");
1040                 if (pos != string::npos) {
1041                         // Parse an endpoint of the type /stream.mp4?frag=1234-5678.
1042                         const char *ptr = url.c_str() + pos + 6;
1043
1044                         // "?frag=header" is special.
1045                         if (strcmp(ptr, "header") == 0) {
1046                                 client->stream_pos = Client::STREAM_POS_HEADER_ONLY;
1047                                 client->stream_pos_end = -1;
1048                         } else {
1049                                 char *endptr;
1050                                 long long frag_start = strtol(ptr, &endptr, 10);
1051                                 if (ptr == endptr || frag_start < 0 || frag_start == LLONG_MAX) {
1052                                         return 400;  // Bad request.
1053                                 }
1054                                 if (*endptr != '-') {
1055                                         return 400;  // Bad request.
1056                                 }
1057                                 ptr = endptr + 1;
1058
1059                                 long long frag_end = strtol(ptr, &endptr, 10);
1060                                 if (ptr == endptr || frag_end < frag_start || frag_end == LLONG_MAX) {
1061                                         return 400;  // Bad request.
1062                                 }
1063
1064                                 if (*endptr != '\0') {
1065                                         return 400;  // Bad request.
1066                                 }
1067
1068                                 client->stream_pos = frag_start;
1069                                 client->stream_pos_end = frag_end;
1070                         }
1071                         url = url.substr(0, pos);
1072                 } else {
1073                         client->stream_pos = -1;
1074                         client->stream_pos_end = -1;
1075                 }
1076         }
1077
1078         // Figure out if we're supposed to close the socket after we've delivered the response.
1079         string protocol = request_tokens[2];
1080         if (protocol.find("HTTP/") != 0) {
1081                 return 400;  // Bad request.
1082         }
1083         client->close_after_response = false;
1084         client->http_11 = true;
1085         if (protocol == "HTTP/1.0") {
1086                 // No persistent connections.
1087                 client->close_after_response = true;
1088                 client->http_11 = false;
1089         } else {
1090                 const auto connection_it = headers.find("Connection");
1091                 if (connection_it != headers.end() && connection_it->second == "close") {
1092                         client->close_after_response = true;
1093                 }
1094         }
1095
1096         const auto stream_url_map_it = stream_url_map.find(url);
1097         if (stream_url_map_it != stream_url_map.end()) {
1098                 // Serve a regular stream..
1099                 client->stream = streams[stream_url_map_it->second].get();
1100                 client->serving_hls_playlist = false;
1101         } else {
1102                 const auto stream_hls_url_map_it = stream_hls_url_map.find(url);
1103                 if (stream_hls_url_map_it != stream_hls_url_map.end()) {
1104                         // Serve HLS playlist.
1105                         client->stream = streams[stream_hls_url_map_it->second].get();
1106                         client->serving_hls_playlist = true;
1107                 } else {
1108                         const auto ping_url_map_it = ping_url_map.find(url);
1109                         if (ping_url_map_it == ping_url_map.end()) {
1110                                 return 404;  // Not found.
1111                         } else {
1112                                 // Serve a ping (204 no error).
1113                                 return 204;
1114                         }
1115                 }
1116         }
1117
1118         Stream *stream = client->stream;
1119         if (stream->http_header.empty()) {
1120                 return 503;  // Service unavailable.
1121         }
1122
1123         if (client->serving_hls_playlist) {
1124                 if (stream->encoding == Stream::STREAM_ENCODING_METACUBE) {
1125                         // This doesn't make any sense, and is hard to implement, too.
1126                         return 404;
1127                 } else {
1128                         return 200;
1129                 }
1130         }
1131
1132         if (client->stream_pos_end == Client::STREAM_POS_NO_END) {
1133                 // This stream won't end, so we don't have a content-length,
1134                 // and can just as well tell the client it's Connection: close
1135                 // (otherwise, we'd have to implement chunking TE for no good reason).
1136                 client->close_after_response = true;
1137         } else {
1138                 if (stream->encoding == Stream::STREAM_ENCODING_METACUBE) {
1139                         // This doesn't make any sense, and is hard to implement, too.
1140                         return 416;  // Range not satisfiable.
1141                 }
1142
1143                 // Check that we have the requested fragment in our backlog.
1144                 size_t buffer_end = stream->bytes_received;
1145                 size_t buffer_start = (buffer_end <= stream->backlog_size) ? 0 : buffer_end - stream->backlog_size;
1146
1147                 if (client->stream_pos_end > buffer_end ||
1148                     client->stream_pos < buffer_start) {
1149                         return 416;  // Range not satisfiable.
1150                 }
1151         }
1152
1153         client->stream = stream;
1154         if (setsockopt(client->sock, SOL_SOCKET, SO_MAX_PACING_RATE, &client->stream->pacing_rate, sizeof(client->stream->pacing_rate)) == -1) {
1155                 if (client->stream->pacing_rate != ~0U) {
1156                         log_perror("setsockopt(SO_MAX_PACING_RATE)");
1157                 }
1158         }
1159         client->request.clear();
1160
1161         return 200;  // OK!
1162 }
1163
1164 void Server::construct_stream_header(Client *client)
1165 {
1166         Stream *stream = client->stream;
1167         string response = stream->http_header;
1168         if (client->stream_pos == Client::STREAM_POS_HEADER_ONLY) {
1169                 char buf[64];
1170                 snprintf(buf, sizeof(buf), "Content-Length: %zu\r\n", stream->stream_header.size());
1171                 response.append(buf);
1172         } else if (client->stream_pos_end != Client::STREAM_POS_NO_END) {
1173                 char buf[64];
1174                 snprintf(buf, sizeof(buf), "Content-Length: %" PRIu64 "\r\n", client->stream_pos_end - client->stream_pos);
1175                 response.append(buf);
1176         }
1177         if (client->http_11) {
1178                 assert(response.find("HTTP/1.0") == 0);
1179                 response[7] = '1';  // Change to HTTP/1.1.
1180                 if (client->close_after_response) {
1181                         response.append("Connection: close\r\n");
1182                 }
1183         } else {
1184                 assert(client->close_after_response);
1185         }
1186         if (!stream->allow_origin.empty()) {
1187                 response.append("Access-Control-Allow-Origin: ");
1188                 response.append(stream->allow_origin);
1189                 response.append("\r\n");
1190         }
1191         if (stream->encoding == Stream::STREAM_ENCODING_RAW) {
1192                 response.append("\r\n");
1193         } else if (stream->encoding == Stream::STREAM_ENCODING_METACUBE) {
1194                 response.append("Content-Encoding: metacube\r\n\r\n");
1195                 if (!stream->stream_header.empty()) {
1196                         metacube2_block_header hdr;
1197                         memcpy(hdr.sync, METACUBE2_SYNC, sizeof(hdr.sync));
1198                         hdr.size = htonl(stream->stream_header.size());
1199                         hdr.flags = htons(METACUBE_FLAGS_HEADER);
1200                         hdr.csum = htons(metacube2_compute_crc(&hdr));
1201                         response.append(string(reinterpret_cast<char *>(&hdr), sizeof(hdr)));
1202                 }
1203         } else {
1204                 assert(false);
1205         }
1206         if (client->stream_pos == Client::STREAM_POS_HEADER_ONLY) {
1207                 client->state = Client::SENDING_SHORT_RESPONSE;
1208                 response.append(stream->stream_header);
1209         } else {
1210                 client->state = Client::SENDING_HEADER;
1211                 if (client->stream_pos_end == Client::STREAM_POS_NO_END) {  // Fragments don't contain stream headers.
1212                         response.append(stream->stream_header);
1213                 }
1214         }
1215
1216         client->header_or_short_response_holder = move(response);
1217         client->header_or_short_response = &client->header_or_short_response_holder;
1218
1219         // Switch states.
1220         change_epoll_events(client, EPOLLOUT | EPOLLET | EPOLLRDHUP);
1221 }
1222         
1223 void Server::construct_error(Client *client, int error_code)
1224 {
1225         char error[256];
1226         if (client->http_11 && client->close_after_response) {
1227                 snprintf(error, sizeof(error),
1228                         "HTTP/1.1 %d Error\r\nContent-Type: text/plain\r\nConnection: close\r\n\r\nSomething went wrong. Sorry.\r\n",
1229                         error_code);
1230         } else {
1231                 snprintf(error, sizeof(error),
1232                         "HTTP/1.%d %d Error\r\nContent-Type: text/plain\r\nContent-Length: 30\r\n\r\nSomething went wrong. Sorry.\r\n",
1233                         client->http_11, error_code);
1234         }
1235         client->header_or_short_response_holder = error;
1236         client->header_or_short_response = &client->header_or_short_response_holder;
1237
1238         // Switch states.
1239         client->state = Client::SENDING_SHORT_RESPONSE;
1240         change_epoll_events(client, EPOLLOUT | EPOLLET | EPOLLRDHUP);
1241 }
1242
1243 void Server::construct_hls_playlist(Client *client)
1244 {
1245         Stream *stream = client->stream;
1246         shared_ptr<const string> *cache;
1247         if (client->http_11) {
1248                 if (client->close_after_response) {
1249                         cache = &stream->hls_playlist_http11_close;
1250                 } else {
1251                         cache = &stream->hls_playlist_http11_persistent;
1252                 }
1253         } else {
1254                 assert(client->close_after_response);
1255                 cache = &stream->hls_playlist_http10;
1256         }
1257
1258         if (*cache == nullptr) {
1259                 *cache = stream->generate_hls_playlist(client->http_11, client->close_after_response);
1260         }
1261         client->header_or_short_response_ref = *cache;
1262         client->header_or_short_response = cache->get();
1263
1264         // Switch states.
1265         client->state = Client::SENDING_SHORT_RESPONSE;
1266         change_epoll_events(client, EPOLLOUT | EPOLLET | EPOLLRDHUP);
1267 }
1268
1269 void Server::construct_204(Client *client)
1270 {
1271         const auto ping_url_map_it = ping_url_map.find(client->url);
1272         assert(ping_url_map_it != ping_url_map.end());
1273
1274         string response;
1275         if (client->http_11) {
1276                 response = "HTTP/1.1 204 No Content\r\n";
1277                 if (client->close_after_response) {
1278                         response.append("Connection: close\r\n");
1279                 }
1280         } else {
1281                 response = "HTTP/1.0 204 No Content\r\n";
1282                 assert(client->close_after_response);
1283         }
1284         if (!ping_url_map_it->second.empty()) {
1285                 response.append("Access-Control-Allow-Origin: ");
1286                 response.append(ping_url_map_it->second);
1287                 response.append("\r\n");
1288         }
1289         response.append("\r\n");
1290
1291         client->header_or_short_response_holder = move(response);
1292         client->header_or_short_response = &client->header_or_short_response_holder;
1293
1294         // Switch states.
1295         client->state = Client::SENDING_SHORT_RESPONSE;
1296         change_epoll_events(client, EPOLLOUT | EPOLLET | EPOLLRDHUP);
1297 }
1298
1299 namespace {
1300
1301 template<class T>
1302 void delete_from(vector<T> *v, T elem)
1303 {
1304         typename vector<T>::iterator new_end = remove(v->begin(), v->end(), elem);
1305         v->erase(new_end, v->end());
1306 }
1307
1308 void send_ktls_close(int sock)
1309 {
1310         uint8_t record_type = 21;  // Alert.
1311         uint8_t body[] = {
1312                 1,   // Warning level (but still fatal!).
1313                 0,   // close_notify.
1314         };
1315
1316         int cmsg_len = sizeof(record_type);
1317         char buf[CMSG_SPACE(cmsg_len)];
1318
1319         msghdr msg = {0};
1320         msg.msg_control = buf;
1321         msg.msg_controllen = sizeof(buf);
1322         cmsghdr *cmsg = CMSG_FIRSTHDR(&msg);
1323         cmsg->cmsg_level = SOL_TLS;
1324         cmsg->cmsg_type = TLS_SET_RECORD_TYPE;
1325         cmsg->cmsg_len = CMSG_LEN(cmsg_len);
1326         *CMSG_DATA(cmsg) = record_type;
1327         msg.msg_controllen = cmsg->cmsg_len;
1328
1329         iovec msg_iov;
1330         msg_iov.iov_base = body;
1331         msg_iov.iov_len = sizeof(body);
1332         msg.msg_iov = &msg_iov;
1333         msg.msg_iovlen = 1;
1334
1335         int err;
1336         do {
1337                 err = sendmsg(sock, &msg, 0);
1338         } while (err == -1 && errno == EINTR);  // Ignore all other errors.
1339 }
1340
1341 }  // namespace
1342         
1343 void Server::close_client(Client *client)
1344 {
1345         if (epoll_ctl(epoll_fd, EPOLL_CTL_DEL, client->sock, nullptr) == -1) {
1346                 log_perror("epoll_ctl(EPOLL_CTL_DEL)");
1347                 exit(1);
1348         }
1349
1350         // This client could be sleeping, so we'll need to fix that. (Argh, O(n).)
1351         if (client->stream != nullptr) {
1352                 delete_from(&client->stream->sleeping_clients, client);
1353                 delete_from(&client->stream->to_process, client);
1354         }
1355
1356         if (client->tls_context) {
1357                 if (client->in_ktls_mode) {
1358                         // Keep GnuTLS happy.
1359                         send_ktls_close(client->sock);
1360                 }
1361                 tls_destroy_context(client->tls_context);
1362         }
1363
1364         // Log to access_log.
1365         access_log->write(client->get_stats());
1366
1367         // Bye-bye!
1368         safe_close(client->sock);
1369
1370         clients.erase(client->sock);
1371 }
1372
1373 void Server::change_epoll_events(Client *client, uint32_t events)
1374 {
1375         epoll_event ev;
1376         ev.events = events;
1377         ev.data.ptr = client;
1378
1379         if (epoll_ctl(epoll_fd, EPOLL_CTL_MOD, client->sock, &ev) == -1) {
1380                 log_perror("epoll_ctl(EPOLL_CTL_MOD)");
1381                 exit(1);
1382         }
1383 }
1384
1385 bool Server::more_requests(Client *client)
1386 {
1387         if (client->close_after_response) {
1388                 return false;
1389         }
1390
1391         // Log to access_log.
1392         access_log->write(client->get_stats());
1393
1394         flush_pending_data(client->sock);
1395
1396         // Switch states and reset the parsers. We don't reset statistics.
1397         client->state = Client::READING_REQUEST;
1398         client->url.clear();
1399         client->stream = NULL;
1400         client->header_or_short_response = nullptr;
1401         client->header_or_short_response_holder.clear();
1402         client->header_or_short_response_ref.reset();
1403         client->header_or_short_response_bytes_sent = 0;
1404         client->bytes_sent = 0;
1405         start_client_timeout_timer(client);
1406
1407         change_epoll_events(client, EPOLLIN | EPOLLET | EPOLLRDHUP);  // No TLS handshake, so no EPOLLOUT needed.
1408
1409         return true;
1410 }
1411
1412 void Server::process_queued_data()
1413 {
1414         {
1415                 lock_guard<mutex> lock(queued_clients_mutex);
1416
1417                 for (const pair<int, Acceptor *> &id_and_acceptor : queued_add_clients) {
1418                         add_client(id_and_acceptor.first, id_and_acceptor.second);
1419                 }
1420                 queued_add_clients.clear();
1421         }
1422
1423         for (unique_ptr<Stream> &stream : streams) {
1424                 stream->process_queued_data();
1425         }
1426 }