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