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