git.sesse.net Git - cubemap/blob - server.cpp

   1 #include <assert.h>
   2 #include <errno.h>
   3 #include <netinet/in.h>
   4 #include <pthread.h>
   5 #include <stdint.h>
   6 #include <stdio.h>
   7 #include <stdlib.h>
   8 #include <string.h>
   9 #include <sys/epoll.h>
  10 #include <sys/sendfile.h>
  11 #include <sys/socket.h>
  12 #include <sys/types.h>
  13 #include <unistd.h>
  14 #include <algorithm>
  15 #include <map>
  16 #include <string>
  17 #include <utility>
  18 #include <vector>
  19
  20 #include "tlse.h"
  21
  22 #include "acceptor.h"
  23 #include "accesslog.h"
  24 #include "log.h"
  25 #include "metacube2.h"
  26 #include "mutexlock.h"
  27 #include "parse.h"
  28 #include "server.h"
  29 #include "state.pb.h"
  30 #include "stream.h"
  31 #include "util.h"
  32
  33 #ifndef SO_MAX_PACING_RATE
  34 #define SO_MAX_PACING_RATE 47
  35 #endif
  36
  37 using namespace std;
  38
  39 extern AccessLogThread *access_log;
  40
  41 namespace {
  42
  43 inline bool is_equal(timespec a, timespec b)
  44 {
  45         return a.tv_sec == b.tv_sec &&
  46                a.tv_nsec == b.tv_nsec;
  47 }
  48
  49 inline bool is_earlier(timespec a, timespec b)
  50 {
  51         if (a.tv_sec != b.tv_sec)
  52                 return a.tv_sec < b.tv_sec;
  53         return a.tv_nsec < b.tv_nsec;
  54 }
  55
  56 }  // namespace
  57
  58 Server::Server()
  59 {
  60         pthread_mutex_init(&mutex, nullptr);
  61         pthread_mutex_init(&queued_clients_mutex, nullptr);
  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         MutexLock lock(&mutex);
  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 void Server::do_work()
  87 {
  88         while (!should_stop()) {
  89                 // Wait until there's activity on at least one of the fds,
  90                 // or 20 ms (about one frame at 50 fps) has elapsed.
  91                 //
  92                 // We could in theory wait forever and rely on wakeup()
  93                 // from add_client_deferred() and add_data_deferred(),
  94                 // but wakeup is a pretty expensive operation, and the
  95                 // two threads might end up fighting over a lock, so it's
  96                 // seemingly (much) more efficient to just have a timeout here.
  97                 int nfds = epoll_pwait(epoll_fd, events, EPOLL_MAX_EVENTS, EPOLL_TIMEOUT_MS, &sigset_without_usr1_block);
  98                 if (nfds == -1 && errno != EINTR) {
  99                         log_perror("epoll_wait");
 100                         exit(1);
 101                 }
 102
 103                 MutexLock lock(&mutex);  // We release the mutex between iterations.
 104
 105                 process_queued_data();
 106
 107                 // Process each client where we have socket activity.
 108                 for (int i = 0; i < nfds; ++i) {
 109                         Client *client = reinterpret_cast<Client *>(events[i].data.ptr);
 110
 111                         if (events[i].events & (EPOLLERR | EPOLLRDHUP | EPOLLHUP)) {
 112                                 close_client(client);
 113                                 continue;
 114                         }
 115
 116                         process_client(client);
 117                 }
 118
 119                 // Process each client where its stream has new data,
 120                 // even if there was no socket activity.
 121                 for (unique_ptr<Stream> &stream : streams) {
 122                         vector<Client *> to_process;
 123                         swap(stream->to_process, to_process);
 124                         for (Client *client : to_process) {
 125                                 process_client(client);
 126                         }
 127                 }
 128
 129                 // Finally, go through each client to see if it's timed out
 130                 // in the READING_REQUEST state. (Seemingly there are clients
 131                 // that can hold sockets up for days at a time without sending
 132                 // anything at all.)
 133                 timespec timeout_time;
 134                 if (clock_gettime(CLOCK_MONOTONIC_COARSE, &timeout_time) == -1) {
 135                         log_perror("clock_gettime(CLOCK_MONOTONIC_COARSE)");
 136                         continue;
 137                 }
 138                 timeout_time.tv_sec -= REQUEST_READ_TIMEOUT_SEC;
 139                 while (!clients_ordered_by_connect_time.empty()) {
 140                         const pair<timespec, int> &connect_time_and_fd = clients_ordered_by_connect_time.front();
 141
 142                         // See if we have reached the end of clients to process.
 143                         if (is_earlier(timeout_time, connect_time_and_fd.first)) {
 144                                 break;
 145                         }
 146
 147                         // If this client doesn't exist anymore, just ignore it
 148                         // (it was deleted earlier).
 149                         auto client_it = clients.find(connect_time_and_fd.second);
 150                         if (client_it == clients.end()) {
 151                                 clients_ordered_by_connect_time.pop();
 152                                 continue;
 153                         }
 154                         Client *client = &client_it->second;
 155                         if (!is_equal(client->connect_time, connect_time_and_fd.first)) {
 156                                 // Another client has taken this fd in the meantime.
 157                                 clients_ordered_by_connect_time.pop();
 158                                 continue;
 159                         }
 160
 161                         if (client->state != Client::READING_REQUEST) {
 162                                 // Only READING_REQUEST can time out.
 163                                 clients_ordered_by_connect_time.pop();
 164                                 continue;
 165                         }
 166
 167                         // OK, it timed out.
 168                         close_client(client);
 169                         clients_ordered_by_connect_time.pop();
 170                 }
 171         }
 172 }
 173
 174 CubemapStateProto Server::serialize()
 175 {
 176         // We don't serialize anything queued, so empty the queues.
 177         process_queued_data();
 178
 179         // Set all clients in a consistent state before serializing
 180         // (ie., they have no remaining lost data). Otherwise, increasing
 181         // the backlog could take clients into a newly valid area of the backlog,
 182         // sending a stream of zeros instead of skipping the data as it should.
 183         //
 184         // TODO: Do this when clients are added back from serialized state instead;
 185         // it would probably be less wasteful.
 186         for (auto &fd_and_client : clients) {
 187                 skip_lost_data(&fd_and_client.second);
 188         }
 189
 190         CubemapStateProto serialized;
 191         for (const auto &fd_and_client : clients) {
 192                 serialized.add_clients()->MergeFrom(fd_and_client.second.serialize());
 193         }
 194         for (unique_ptr<Stream> &stream : streams) {
 195                 serialized.add_streams()->MergeFrom(stream->serialize());
 196         }
 197         return serialized;
 198 }
 199
 200 void Server::add_client_deferred(int sock, Acceptor *acceptor)
 201 {
 202         MutexLock lock(&queued_clients_mutex);
 203         queued_add_clients.push_back(std::make_pair(sock, acceptor));
 204 }
 205
 206 void Server::add_client(int sock, Acceptor *acceptor)
 207 {
 208         const bool is_tls = acceptor->is_tls();
 209         auto inserted = clients.insert(make_pair(sock, Client(sock)));
 210         assert(inserted.second == true);  // Should not already exist.
 211         Client *client_ptr = &inserted.first->second;
 212
 213         // Connection timestamps must be nondecreasing. I can't find any guarantee
 214         // that even the monotonic clock can't go backwards by a small amount
 215         // (think switching between CPUs with non-synchronized TSCs), so if
 216         // this actually should happen, we hack around it by fudging
 217         // connect_time.
 218         if (!clients_ordered_by_connect_time.empty() &&
 219             is_earlier(client_ptr->connect_time, clients_ordered_by_connect_time.back().first)) {
 220                 client_ptr->connect_time = clients_ordered_by_connect_time.back().first;
 221         }
 222         clients_ordered_by_connect_time.push(make_pair(client_ptr->connect_time, sock));
 223
 224         // Start listening on data from this socket.
 225         epoll_event ev;
 226         if (is_tls) {
 227                 // Even in the initial state (READING_REQUEST), TLS needs to
 228                 // send data for the handshake, and thus might end up needing
 229                 // to know about EPOLLOUT.
 230                 ev.events = EPOLLIN | EPOLLOUT | EPOLLET | EPOLLRDHUP;
 231         } else {
 232                 // EPOLLOUT will be added once we go out of READING_REQUEST.
 233                 ev.events = EPOLLIN | EPOLLET | EPOLLRDHUP;
 234         }
 235         ev.data.ptr = client_ptr;
 236         if (epoll_ctl(epoll_fd, EPOLL_CTL_ADD, sock, &ev) == -1) {
 237                 log_perror("epoll_ctl(EPOLL_CTL_ADD)");
 238                 exit(1);
 239         }
 240
 241         if (is_tls) {
 242                 assert(tls_server_contexts.count(acceptor));
 243                 client_ptr->tls_context = tls_accept(tls_server_contexts[acceptor]);
 244                 if (client_ptr->tls_context == nullptr) {
 245                         log(ERROR, "tls_accept() failed");
 246                         close_client(client_ptr);
 247                         return;
 248                 }
 249                 tls_make_exportable(client_ptr->tls_context, 1);
 250         }
 251
 252         process_client(client_ptr);
 253 }
 254
 255 void Server::add_client_from_serialized(const ClientProto &client)
 256 {
 257         MutexLock lock(&mutex);
 258         Stream *stream;
 259         int stream_index = lookup_stream_by_url(client.url());
 260         if (stream_index == -1) {
 261                 assert(client.state() != Client::SENDING_DATA);
 262                 stream = nullptr;
 263         } else {
 264                 stream = streams[stream_index].get();
 265         }
 266         auto inserted = clients.insert(make_pair(client.sock(), Client(client, stream)));
 267         assert(inserted.second == true);  // Should not already exist.
 268         Client *client_ptr = &inserted.first->second;
 269
 270         // Connection timestamps must be nondecreasing.
 271         assert(clients_ordered_by_connect_time.empty() ||
 272                !is_earlier(client_ptr->connect_time, clients_ordered_by_connect_time.back().first));
 273         clients_ordered_by_connect_time.push(make_pair(client_ptr->connect_time, client.sock()));
 274
 275         // Start listening on data from this socket.
 276         epoll_event ev;
 277         if (client.state() == Client::READING_REQUEST) {
 278                 // See the corresponding comment in Server::add_client().
 279                 if (client.has_tls_context()) {
 280                         ev.events = EPOLLIN | EPOLLOUT | EPOLLET | EPOLLRDHUP;
 281                 } else {
 282                         ev.events = EPOLLIN | EPOLLET | EPOLLRDHUP;
 283                 }
 284         } else {
 285                 // If we don't have more data for this client, we'll be putting it into
 286                 // the sleeping array again soon.
 287                 ev.events = EPOLLOUT | EPOLLET | EPOLLRDHUP;
 288         }
 289         ev.data.ptr = client_ptr;
 290         if (epoll_ctl(epoll_fd, EPOLL_CTL_ADD, client.sock(), &ev) == -1) {
 291                 log_perror("epoll_ctl(EPOLL_CTL_ADD)");
 292                 exit(1);
 293         }
 294
 295         if (client_ptr->state == Client::WAITING_FOR_KEYFRAME ||
 296             client_ptr->state == Client::PREBUFFERING ||
 297             (client_ptr->state == Client::SENDING_DATA &&
 298              client_ptr->stream_pos == client_ptr->stream->bytes_received)) {
 299                 client_ptr->stream->put_client_to_sleep(client_ptr);
 300         } else {
 301                 process_client(client_ptr);
 302         }
 303 }
 304
 305 int Server::lookup_stream_by_url(const string &url) const
 306 {
 307         map<string, int>::const_iterator stream_url_it = stream_url_map.find(url);
 308         if (stream_url_it == stream_url_map.end()) {
 309                 return -1;
 310         }
 311         return stream_url_it->second;
 312 }
 313
 314 int Server::add_stream(const string &url, size_t backlog_size, size_t prebuffering_bytes, Stream::Encoding encoding, Stream::Encoding src_encoding)
 315 {
 316         MutexLock lock(&mutex);
 317         stream_url_map.insert(make_pair(url, streams.size()));
 318         streams.emplace_back(new Stream(url, backlog_size, prebuffering_bytes, encoding, src_encoding));
 319         return streams.size() - 1;
 320 }
 321
 322 int Server::add_stream_from_serialized(const StreamProto &stream, int data_fd)
 323 {
 324         MutexLock lock(&mutex);
 325         stream_url_map.insert(make_pair(stream.url(), streams.size()));
 326         streams.emplace_back(new Stream(stream, data_fd));
 327         return streams.size() - 1;
 328 }
 329
 330 void Server::set_backlog_size(int stream_index, size_t new_size)
 331 {
 332         MutexLock lock(&mutex);
 333         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
 334         streams[stream_index]->set_backlog_size(new_size);
 335 }
 336
 337 void Server::set_prebuffering_bytes(int stream_index, size_t new_amount)
 338 {
 339         MutexLock lock(&mutex);
 340         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
 341         streams[stream_index]->prebuffering_bytes = new_amount;
 342 }
 343
 344 void Server::set_encoding(int stream_index, Stream::Encoding encoding)
 345 {
 346         MutexLock lock(&mutex);
 347         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
 348         streams[stream_index]->encoding = encoding;
 349 }
 350
 351 void Server::set_src_encoding(int stream_index, Stream::Encoding encoding)
 352 {
 353         MutexLock lock(&mutex);
 354         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
 355         streams[stream_index]->src_encoding = encoding;
 356 }
 357
 358 void Server::set_header(int stream_index, const string &http_header, const string &stream_header)
 359 {
 360         MutexLock lock(&mutex);
 361         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
 362         streams[stream_index]->http_header = http_header;
 363
 364         if (stream_header != streams[stream_index]->stream_header) {
 365                 // We cannot start at any of the older starting points anymore,
 366                 // since they'd get the wrong header for the stream (not to mention
 367                 // that a changed header probably means the stream restarted,
 368                 // which means any client starting on the old one would probably
 369                 // stop playing properly at the change point). Next block
 370                 // should be a suitable starting point (if not, something is
 371                 // pretty strange), so it will fill up again soon enough.
 372                 streams[stream_index]->suitable_starting_points.clear();
 373         }
 374         streams[stream_index]->stream_header = stream_header;
 375 }
 376
 377 void Server::set_pacing_rate(int stream_index, uint32_t pacing_rate)
 378 {
 379         MutexLock lock(&mutex);
 380         assert(clients.empty());
 381         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
 382         streams[stream_index]->pacing_rate = pacing_rate;
 383 }
 384
 385 void Server::add_gen204(const std::string &url, const std::string &allow_origin)
 386 {
 387         MutexLock lock(&mutex);
 388         assert(clients.empty());
 389         ping_url_map[url] = allow_origin;
 390 }
 391
 392 void Server::create_tls_context_for_acceptor(const Acceptor *acceptor)
 393 {
 394         assert(acceptor->is_tls());
 395
 396         bool is_server = true;
 397         TLSContext *server_context = tls_create_context(is_server, TLS_V12);
 398
 399         const string &cert = acceptor->get_certificate_chain();
 400         int num_cert = tls_load_certificates(server_context, reinterpret_cast<const unsigned char *>(cert.data()), cert.size());
 401         assert(num_cert > 0);  // Should have been checked by config earlier.
 402
 403         const string &key = acceptor->get_private_key();
 404         int num_key = tls_load_private_key(server_context, reinterpret_cast<const unsigned char *>(key.data()), key.size());
 405         assert(num_key > 0);  // Should have been checked by config earlier.
 406
 407         tls_server_contexts.insert(make_pair(acceptor, server_context));
 408 }
 409
 410 void Server::add_data_deferred(int stream_index, const char *data, size_t bytes, uint16_t metacube_flags)
 411 {
 412         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
 413         streams[stream_index]->add_data_deferred(data, bytes, metacube_flags);
 414 }
 415
 416 // See the .h file for postconditions after this function.
 417 void Server::process_client(Client *client)
 418 {
 419         switch (client->state) {
 420         case Client::READING_REQUEST: {
 421                 if (client->tls_context != nullptr) {
 422                         if (send_pending_tls_data(client)) {
 423                                 // send_pending_tls_data() hit postconditions #1 or #4.
 424                                 return;
 425                         }
 426                 }
 427
 428 read_request_again:
 429                 // Try to read more of the request.
 430                 char buf[1024];
 431                 int ret;
 432                 if (client->tls_context == nullptr) {
 433                         ret = read_nontls_data(client, buf, sizeof(buf));
 434                         if (ret == -1) {
 435                                 // read_nontls_data() hit postconditions #1 or #2.
 436                                 return;
 437                         }
 438                 } else {
 439                         ret = read_tls_data(client, buf, sizeof(buf));
 440                         if (ret == -1) {
 441                                 // read_tls_data() hit postconditions #1, #2 or #4.
 442                                 return;
 443                         }
 444                 }
 445
 446                 RequestParseStatus status = wait_for_double_newline(&client->request, buf, ret);
 447
 448                 switch (status) {
 449                 case RP_OUT_OF_SPACE:
 450                         log(WARNING, "[%s] Client sent overlong request!", client->remote_addr.c_str());
 451                         close_client(client);
 452                         return;
 453                 case RP_NOT_FINISHED_YET:
 454                         // OK, we don't have the entire header yet. Fine; we'll get it later.
 455                         // See if there's more data for us.
 456                         goto read_request_again;
 457                 case RP_EXTRA_DATA:
 458                         log(WARNING, "[%s] Junk data after request!", client->remote_addr.c_str());
 459                         close_client(client);
 460                         return;
 461                 case RP_FINISHED:
 462                         break;
 463                 }
 464
 465                 assert(status == RP_FINISHED);
 466
 467                 if (client->tls_context && !client->in_ktls_mode && tls_established(client->tls_context)) {
 468                         // We're ready to enter kTLS mode, unless we still have some
 469                         // handshake data to send (which then must be sent as non-kTLS).
 470                         if (send_pending_tls_data(client)) {
 471                                 // send_pending_tls_data() hit postconditions #1 or #4.
 472                                 return;
 473                         }
 474                         ret = tls_make_ktls(client->tls_context, client->sock);
 475                         if (ret < 0) {
 476                                 log_tls_error("tls_make_ktls", ret);
 477                                 close_client(client);
 478                                 return;
 479                         }
 480                         client->in_ktls_mode = true;
 481                 }
 482
 483                 int error_code = parse_request(client);
 484                 if (error_code == 200) {
 485                         construct_header(client);
 486                 } else if (error_code == 204) {
 487                         construct_204(client);
 488                 } else {
 489                         construct_error(client, error_code);
 490                 }
 491
 492                 // We've changed states, so fall through.
 493                 assert(client->state == Client::SENDING_SHORT_RESPONSE ||
 494                        client->state == Client::SENDING_HEADER);
 495         }
 496         case Client::SENDING_SHORT_RESPONSE:
 497         case Client::SENDING_HEADER: {
 498 sending_header_or_short_response_again:
 499                 int ret;
 500                 do {
 501                         ret = write(client->sock,
 502                                     client->header_or_short_response.data() + client->header_or_short_response_bytes_sent,
 503                                     client->header_or_short_response.size() - client->header_or_short_response_bytes_sent);
 504                 } while (ret == -1 && errno == EINTR);
 505
 506                 if (ret == -1 && errno == EAGAIN) {
 507                         // We're out of socket space, so now we're at the “low edge” of epoll's
 508                         // edge triggering. epoll will tell us when there is more room, so for now,
 509                         // just return.
 510                         // This is postcondition #4.
 511                         return;
 512                 }
 513
 514                 if (ret == -1) {
 515                         // Error! Postcondition #1.
 516                         log_perror("write");
 517                         close_client(client);
 518                         return;
 519                 }
 520
 521                 client->header_or_short_response_bytes_sent += ret;
 522                 assert(client->header_or_short_response_bytes_sent <= client->header_or_short_response.size());
 523
 524                 if (client->header_or_short_response_bytes_sent < client->header_or_short_response.size()) {
 525                         // We haven't sent all yet. Fine; go another round.
 526                         goto sending_header_or_short_response_again;
 527                 }
 528
 529                 // We're done sending the header or error! Clear it to release some memory.
 530                 client->header_or_short_response.clear();
 531
 532                 if (client->state == Client::SENDING_SHORT_RESPONSE) {
 533                         // We're done sending the error, so now close.
 534                         // This is postcondition #1.
 535                         close_client(client);
 536                         return;
 537                 }
 538
 539                 Stream *stream = client->stream;
 540                 if (client->stream_pos == size_t(-2)) {
 541                         // Start sending from the beginning of the backlog.
 542                         client->stream_pos = min<size_t>(
 543                             stream->bytes_received - stream->backlog_size,
 544                             0);
 545                         client->state = Client::SENDING_DATA;
 546                         goto sending_data;
 547                 } else if (stream->prebuffering_bytes == 0) {
 548                         // Start sending from the first keyframe we get. In other
 549                         // words, we won't send any of the backlog, but we'll start
 550                         // sending immediately as we get the next keyframe block.
 551                         // Note that this is functionally identical to the next if branch,
 552                         // except that we save a binary search.
 553                         client->stream_pos = stream->bytes_received;
 554                         client->state = Client::WAITING_FOR_KEYFRAME;
 555                 } else {
 556                         // We're not going to send anything to the client before we have
 557                         // N bytes. However, this wait might be boring; we can just as well
 558                         // use it to send older data if we have it. We use lower_bound()
 559                         // so that we are conservative and never add extra latency over just
 560                         // waiting (assuming CBR or nearly so); otherwise, we could want e.g.
 561                         // 100 kB prebuffer but end up sending a 10 MB GOP.
 562                         deque<size_t>::const_iterator starting_point_it =
 563                                 lower_bound(stream->suitable_starting_points.begin(),
 564                                             stream->suitable_starting_points.end(),
 565                                             stream->bytes_received - stream->prebuffering_bytes);
 566                         if (starting_point_it == stream->suitable_starting_points.end()) {
 567                                 // None found. Just put us at the end, and then wait for the
 568                                 // first keyframe to appear.
 569                                 client->stream_pos = stream->bytes_received;
 570                                 client->state = Client::WAITING_FOR_KEYFRAME;
 571                         } else {
 572                                 client->stream_pos = *starting_point_it;
 573                                 client->state = Client::PREBUFFERING;
 574                                 goto prebuffering;
 575                         }
 576                 }
 577                 // Fall through.
 578         }
 579         case Client::WAITING_FOR_KEYFRAME: {
 580                 Stream *stream = client->stream;
 581                 if (stream->suitable_starting_points.empty() ||
 582                     client->stream_pos > stream->suitable_starting_points.back()) {
 583                         // We haven't received a keyframe since this stream started waiting,
 584                         // so keep on waiting for one.
 585                         // This is postcondition #3.
 586                         stream->put_client_to_sleep(client);
 587                         return;
 588                 }
 589                 client->stream_pos = stream->suitable_starting_points.back();
 590                 client->state = Client::PREBUFFERING;
 591                 // Fall through.
 592         }
 593         case Client::PREBUFFERING: {
 594 prebuffering:
 595                 Stream *stream = client->stream;
 596                 size_t bytes_to_send = stream->bytes_received - client->stream_pos;
 597                 assert(bytes_to_send <= stream->backlog_size);
 598                 if (bytes_to_send < stream->prebuffering_bytes) {
 599                         // We don't have enough bytes buffered to start this client yet.
 600                         // This is postcondition #3.
 601                         stream->put_client_to_sleep(client);
 602                         return;
 603                 }
 604                 client->state = Client::SENDING_DATA;
 605                 // Fall through.
 606         }
 607         case Client::SENDING_DATA: {
 608 sending_data:
 609                 skip_lost_data(client);
 610                 Stream *stream = client->stream;
 611
 612 sending_data_again:
 613                 size_t bytes_to_send = stream->bytes_received - client->stream_pos;
 614                 assert(bytes_to_send <= stream->backlog_size);
 615                 if (bytes_to_send == 0) {
 616                         return;
 617                 }
 618
 619                 // See if we need to split across the circular buffer.
 620                 bool more_data = false;
 621                 if ((client->stream_pos % stream->backlog_size) + bytes_to_send > stream->backlog_size) {
 622                         bytes_to_send = stream->backlog_size - (client->stream_pos % stream->backlog_size);
 623                         more_data = true;
 624                 }
 625
 626                 ssize_t ret;
 627                 do {
 628                         off_t offset = client->stream_pos % stream->backlog_size;
 629                         ret = sendfile(client->sock, stream->data_fd, &offset, bytes_to_send);
 630                 } while (ret == -1 && errno == EINTR);
 631
 632                 if (ret == -1 && errno == EAGAIN) {
 633                         // We're out of socket space, so return; epoll will wake us up
 634                         // when there is more room.
 635                         // This is postcondition #4.
 636                         return;
 637                 }
 638                 if (ret == -1) {
 639                         // Error, close; postcondition #1.
 640                         log_perror("sendfile");
 641                         close_client(client);
 642                         return;
 643                 }
 644                 client->stream_pos += ret;
 645                 client->bytes_sent += ret;
 646
 647                 if (client->stream_pos == stream->bytes_received) {
 648                         // We don't have any more data for this client, so put it to sleep.
 649                         // This is postcondition #3.
 650                         stream->put_client_to_sleep(client);
 651                 } else if (more_data && size_t(ret) == bytes_to_send) {
 652                         goto sending_data_again;
 653                 }
 654                 break;
 655         }
 656         default:
 657                 assert(false);
 658         }
 659 }
 660
 661 bool Server::send_pending_tls_data(Client *client)
 662 {
 663         // See if there's data from the TLS library to write.
 664         if (client->tls_data_to_send == nullptr) {
 665                 client->tls_data_to_send = tls_get_write_buffer(client->tls_context, &client->tls_data_left_to_send);
 666                 if (client->tls_data_to_send == nullptr) {
 667                         // Really no data to send.
 668                         return false;
 669                 }
 670         }
 671
 672 send_data_again:
 673         int ret;
 674         do {
 675                 ret = write(client->sock, client->tls_data_to_send, client->tls_data_left_to_send);
 676         } while (ret == -1 && errno == EINTR);
 677         assert(ret < 0 || size_t(ret) <= client->tls_data_left_to_send);
 678
 679         if (ret == -1 && errno == EAGAIN) {
 680                 // We're out of socket space, so now we're at the “low edge” of epoll's
 681                 // edge triggering. epoll will tell us when there is more room, so for now,
 682                 // just return.
 683                 // This is postcondition #4.
 684                 return true;
 685         }
 686         if (ret == -1) {
 687                 // Error! Postcondition #1.
 688                 log_perror("write");
 689                 close_client(client);
 690                 return true;
 691         }
 692         if (ret > 0 && size_t(ret) == client->tls_data_left_to_send) {
 693                 // All data has been sent, so we don't need to go to sleep.
 694                 tls_buffer_clear(client->tls_context);
 695                 client->tls_data_to_send = nullptr;
 696                 return false;
 697         }
 698
 699         // More data to send, so try again.
 700         client->tls_data_to_send += ret;
 701         client->tls_data_left_to_send -= ret;
 702         goto send_data_again;
 703 }
 704
 705 int Server::read_nontls_data(Client *client, char *buf, size_t max_size)
 706 {
 707         int ret;
 708         do {
 709                 ret = read(client->sock, buf, max_size);
 710         } while (ret == -1 && errno == EINTR);
 711
 712         if (ret == -1 && errno == EAGAIN) {
 713                 // No more data right now. Nothing to do.
 714                 // This is postcondition #2.
 715                 return -1;
 716         }
 717         if (ret == -1) {
 718                 log_perror("read");
 719                 close_client(client);
 720                 return -1;
 721         }
 722         if (ret == 0) {
 723                 // OK, the socket is closed.
 724                 close_client(client);
 725                 return -1;
 726         }
 727
 728         return ret;
 729 }
 730
 731 int Server::read_tls_data(Client *client, char *buf, size_t max_size)
 732 {
 733 read_again:
 734         int ret;
 735         do {
 736                 ret = read(client->sock, buf, max_size);
 737         } while (ret == -1 && errno == EINTR);
 738
 739         if (ret == -1 && errno == EAGAIN) {
 740                 // No more data right now. Nothing to do.
 741                 // This is postcondition #2.
 742                 return -1;
 743         }
 744         if (ret == -1) {
 745                 log_perror("read");
 746                 close_client(client);
 747                 return -1;
 748         }
 749         if (ret == 0) {
 750                 // OK, the socket is closed.
 751                 close_client(client);
 752                 return -1;
 753         }
 754
 755         // Give it to the TLS library.
 756         int err = tls_consume_stream(client->tls_context, reinterpret_cast<const unsigned char *>(buf), ret, nullptr);
 757         if (err < 0) {
 758                 log_tls_error("tls_consume_stream", err);
 759                 close_client(client);
 760                 return -1;
 761         }
 762         if (err == 0) {
 763                 // Not consumed any data. See if we can read more.
 764                 goto read_again;
 765         }
 766
 767         // Read any decrypted data available for us. (We can reuse buf, since it's free now.)
 768         ret = tls_read(client->tls_context, reinterpret_cast<unsigned char *>(buf), max_size);
 769         if (ret == 0) {
 770                 // No decrypted data for us yet, but there might be some more handshaking
 771                 // to send. Do that if needed, then look for more data.
 772                 if (send_pending_tls_data(client)) {
 773                         // send_pending_tls_data() hit postconditions #1 or #4.
 774                         return -1;
 775                 }
 776                 goto read_again;
 777         }
 778         if (ret < 0) {
 779                 log_tls_error("tls_read", ret);
 780                 close_client(client);
 781                 return -1;
 782         }
 783
 784         assert(ret > 0);
 785         return ret;
 786 }
 787
 788 // See if there's some data we've lost. Ideally, we should drop to a block boundary,
 789 // but resync will be the mux's problem.
 790 void Server::skip_lost_data(Client *client)
 791 {
 792         Stream *stream = client->stream;
 793         if (stream == nullptr) {
 794                 return;
 795         }
 796         size_t bytes_to_send = stream->bytes_received - client->stream_pos;
 797         if (bytes_to_send > stream->backlog_size) {
 798                 size_t bytes_lost = bytes_to_send - stream->backlog_size;
 799                 client->stream_pos = stream->bytes_received - stream->backlog_size;
 800                 client->bytes_lost += bytes_lost;
 801                 ++client->num_loss_events;
 802         }
 803 }
 804
 805 int Server::parse_request(Client *client)
 806 {
 807         vector<string> lines = split_lines(client->request);
 808         if (lines.empty()) {
 809                 return 400;  // Bad request (empty).
 810         }
 811
 812         // Parse the headers, for logging purposes.
 813         // TODO: Case-insensitivity.
 814         multimap<string, string> headers = extract_headers(lines, client->remote_addr);
 815         multimap<string, string>::const_iterator referer_it = headers.find("Referer");
 816         if (referer_it != headers.end()) {
 817                 client->referer = referer_it->second;
 818         }
 819         multimap<string, string>::const_iterator user_agent_it = headers.find("User-Agent");
 820         if (user_agent_it != headers.end()) {
 821                 client->user_agent = user_agent_it->second;
 822         }
 823
 824         vector<string> request_tokens = split_tokens(lines[0]);
 825         if (request_tokens.size() < 2) {
 826                 return 400;  // Bad request (empty).
 827         }
 828         if (request_tokens[0] != "GET") {
 829                 return 400;  // Should maybe be 405 instead?
 830         }
 831
 832         string url = request_tokens[1];
 833         client->url = url;
 834         if (url.size() > 8 && url.find("?backlog") == url.size() - 8) {
 835                 client->stream_pos = -2;
 836                 url = url.substr(0, url.size() - 8);
 837         } else {
 838                 client->stream_pos = -1;
 839         }
 840
 841         map<string, int>::const_iterator stream_url_map_it = stream_url_map.find(url);
 842         if (stream_url_map_it == stream_url_map.end()) {
 843                 map<string, string>::const_iterator ping_url_map_it = ping_url_map.find(url);
 844                 if (ping_url_map_it == ping_url_map.end()) {
 845                         return 404;  // Not found.
 846                 } else {
 847                         return 204;  // No error.
 848                 }
 849         }
 850
 851         Stream *stream = streams[stream_url_map_it->second].get();
 852         if (stream->http_header.empty()) {
 853                 return 503;  // Service unavailable.
 854         }
 855
 856         client->stream = stream;
 857         if (setsockopt(client->sock, SOL_SOCKET, SO_MAX_PACING_RATE, &client->stream->pacing_rate, sizeof(client->stream->pacing_rate)) == -1) {
 858                 if (client->stream->pacing_rate != ~0U) {
 859                         log_perror("setsockopt(SO_MAX_PACING_RATE)");
 860                 }
 861         }
 862         client->request.clear();
 863
 864         return 200;  // OK!
 865 }
 866
 867 void Server::construct_header(Client *client)
 868 {
 869         Stream *stream = client->stream;
 870         if (stream->encoding == Stream::STREAM_ENCODING_RAW) {
 871                 client->header_or_short_response = stream->http_header +
 872                         "\r\n" +
 873                         stream->stream_header;
 874         } else if (stream->encoding == Stream::STREAM_ENCODING_METACUBE) {
 875                 client->header_or_short_response = stream->http_header +
 876                         "Content-encoding: metacube\r\n" +
 877                         "\r\n";
 878                 if (!stream->stream_header.empty()) {
 879                         metacube2_block_header hdr;
 880                         memcpy(hdr.sync, METACUBE2_SYNC, sizeof(hdr.sync));
 881                         hdr.size = htonl(stream->stream_header.size());
 882                         hdr.flags = htons(METACUBE_FLAGS_HEADER);
 883                         hdr.csum = htons(metacube2_compute_crc(&hdr));
 884                         client->header_or_short_response.append(
 885                                 string(reinterpret_cast<char *>(&hdr), sizeof(hdr)));
 886                 }
 887                 client->header_or_short_response.append(stream->stream_header);
 888         } else {
 889                 assert(false);
 890         }
 891
 892         // Switch states.
 893         client->state = Client::SENDING_HEADER;
 894         change_epoll_events(client, EPOLLOUT | EPOLLET | EPOLLRDHUP);
 895 }
 896
 897 void Server::construct_error(Client *client, int error_code)
 898 {
 899         char error[256];
 900         snprintf(error, 256, "HTTP/1.0 %d Error\r\nContent-type: text/plain\r\n\r\nSomething went wrong. Sorry.\r\n",
 901                 error_code);
 902         client->header_or_short_response = error;
 903
 904         // Switch states.
 905         client->state = Client::SENDING_SHORT_RESPONSE;
 906         change_epoll_events(client, EPOLLOUT | EPOLLET | EPOLLRDHUP);
 907 }
 908
 909 void Server::construct_204(Client *client)
 910 {
 911         map<string, string>::const_iterator ping_url_map_it = ping_url_map.find(client->url);
 912         assert(ping_url_map_it != ping_url_map.end());
 913
 914         if (ping_url_map_it->second.empty()) {
 915                 client->header_or_short_response =
 916                         "HTTP/1.0 204 No Content\r\n"
 917                         "\r\n";
 918         } else {
 919                 char response[256];
 920                 snprintf(response, 256,
 921                          "HTTP/1.0 204 No Content\r\n"
 922                          "Access-Control-Allow-Origin: %s\r\n"
 923                          "\r\n",
 924                          ping_url_map_it->second.c_str());
 925                 client->header_or_short_response = response;
 926         }
 927
 928         // Switch states.
 929         client->state = Client::SENDING_SHORT_RESPONSE;
 930         change_epoll_events(client, EPOLLOUT | EPOLLET | EPOLLRDHUP);
 931 }
 932
 933 template<class T>
 934 void delete_from(vector<T> *v, T elem)
 935 {
 936         typename vector<T>::iterator new_end = remove(v->begin(), v->end(), elem);
 937         v->erase(new_end, v->end());
 938 }
 939
 940 void Server::close_client(Client *client)
 941 {
 942         if (epoll_ctl(epoll_fd, EPOLL_CTL_DEL, client->sock, nullptr) == -1) {
 943                 log_perror("epoll_ctl(EPOLL_CTL_DEL)");
 944                 exit(1);
 945         }
 946
 947         // This client could be sleeping, so we'll need to fix that. (Argh, O(n).)
 948         if (client->stream != nullptr) {
 949                 delete_from(&client->stream->sleeping_clients, client);
 950                 delete_from(&client->stream->to_process, client);
 951         }
 952
 953         if (client->tls_context) {
 954                 tls_destroy_context(client->tls_context);
 955         }
 956
 957         // Log to access_log.
 958         access_log->write(client->get_stats());
 959
 960         // Bye-bye!
 961         safe_close(client->sock);
 962
 963         clients.erase(client->sock);
 964 }
 965
 966 void Server::change_epoll_events(Client *client, uint32_t events)
 967 {
 968         epoll_event ev;
 969         ev.events = events;
 970         ev.data.ptr = client;
 971
 972         if (epoll_ctl(epoll_fd, EPOLL_CTL_MOD, client->sock, &ev) == -1) {
 973                 log_perror("epoll_ctl(EPOLL_CTL_MOD)");
 974                 exit(1);
 975         }
 976 }
 977
 978 void Server::process_queued_data()
 979 {
 980         {
 981                 MutexLock lock(&queued_clients_mutex);
 982
 983                 for (const pair<int, Acceptor *> &id_and_acceptor : queued_add_clients) {
 984                         add_client(id_and_acceptor.first, id_and_acceptor.second);
 985                 }
 986                 queued_add_clients.clear();
 987         }
 988
 989         for (unique_ptr<Stream> &stream : streams) {
 990                 stream->process_queued_data();
 991         }
 992 }