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 "parse.h"
  27 #include "server.h"
  28 #include "state.pb.h"
  29 #include "stream.h"
  30 #include "util.h"
  31
  32 #ifndef SO_MAX_PACING_RATE
  33 #define SO_MAX_PACING_RATE 47
  34 #endif
  35
  36 using namespace std;
  37
  38 extern AccessLogThread *access_log;
  39
  40 namespace {
  41
  42 inline bool is_equal(timespec a, timespec b)
  43 {
  44         return a.tv_sec == b.tv_sec &&
  45                a.tv_nsec == b.tv_nsec;
  46 }
  47
  48 inline bool is_earlier(timespec a, timespec b)
  49 {
  50         if (a.tv_sec != b.tv_sec)
  51                 return a.tv_sec < b.tv_sec;
  52         return a.tv_nsec < b.tv_nsec;
  53 }
  54
  55 }  // namespace
  56
  57 Server::Server()
  58 {
  59         epoll_fd = epoll_create(1024);  // Size argument is ignored.
  60         if (epoll_fd == -1) {
  61                 log_perror("epoll_fd");
  62                 exit(1);
  63         }
  64 }
  65
  66 Server::~Server()
  67 {
  68         safe_close(epoll_fd);
  69 }
  70
  71 vector<ClientStats> Server::get_client_stats() const
  72 {
  73         vector<ClientStats> ret;
  74
  75         lock_guard<mutex> lock(mu);
  76         for (const auto &fd_and_client : clients) {
  77                 ret.push_back(fd_and_client.second.get_stats());
  78         }
  79         return ret;
  80 }
  81
  82 void Server::do_work()
  83 {
  84         while (!should_stop()) {
  85                 // Wait until there's activity on at least one of the fds,
  86                 // or 20 ms (about one frame at 50 fps) has elapsed.
  87                 //
  88                 // We could in theory wait forever and rely on wakeup()
  89                 // from add_client_deferred() and add_data_deferred(),
  90                 // but wakeup is a pretty expensive operation, and the
  91                 // two threads might end up fighting over a lock, so it's
  92                 // seemingly (much) more efficient to just have a timeout here.
  93                 int nfds = epoll_pwait(epoll_fd, events, EPOLL_MAX_EVENTS, EPOLL_TIMEOUT_MS, &sigset_without_usr1_block);
  94                 if (nfds == -1 && errno != EINTR) {
  95                         log_perror("epoll_wait");
  96                         exit(1);
  97                 }
  98
  99                 lock_guard<mutex> lock(mu);  // We release the mutex between iterations.
 100
 101                 process_queued_data();
 102
 103                 // Process each client where we have socket activity.
 104                 for (int i = 0; i < nfds; ++i) {
 105                         Client *client = reinterpret_cast<Client *>(events[i].data.ptr);
 106
 107                         if (events[i].events & (EPOLLERR | EPOLLRDHUP | EPOLLHUP)) {
 108                                 close_client(client);
 109                                 continue;
 110                         }
 111
 112                         process_client(client);
 113                 }
 114
 115                 // Process each client where its stream has new data,
 116                 // even if there was no socket activity.
 117                 for (unique_ptr<Stream> &stream : streams) {
 118                         vector<Client *> to_process;
 119                         swap(stream->to_process, to_process);
 120                         for (Client *client : to_process) {
 121                                 process_client(client);
 122                         }
 123                 }
 124
 125                 // Finally, go through each client to see if it's timed out
 126                 // in the READING_REQUEST state. (Seemingly there are clients
 127                 // that can hold sockets up for days at a time without sending
 128                 // anything at all.)
 129                 timespec timeout_time;
 130                 if (clock_gettime(CLOCK_MONOTONIC_COARSE, &timeout_time) == -1) {
 131                         log_perror("clock_gettime(CLOCK_MONOTONIC_COARSE)");
 132                         continue;
 133                 }
 134                 timeout_time.tv_sec -= REQUEST_READ_TIMEOUT_SEC;
 135                 while (!clients_ordered_by_connect_time.empty()) {
 136                         const pair<timespec, int> &connect_time_and_fd = clients_ordered_by_connect_time.front();
 137
 138                         // See if we have reached the end of clients to process.
 139                         if (is_earlier(timeout_time, connect_time_and_fd.first)) {
 140                                 break;
 141                         }
 142
 143                         // If this client doesn't exist anymore, just ignore it
 144                         // (it was deleted earlier).
 145                         auto client_it = clients.find(connect_time_and_fd.second);
 146                         if (client_it == clients.end()) {
 147                                 clients_ordered_by_connect_time.pop();
 148                                 continue;
 149                         }
 150                         Client *client = &client_it->second;
 151                         if (!is_equal(client->connect_time, connect_time_and_fd.first)) {
 152                                 // Another client has taken this fd in the meantime.
 153                                 clients_ordered_by_connect_time.pop();
 154                                 continue;
 155                         }
 156
 157                         if (client->state != Client::READING_REQUEST) {
 158                                 // Only READING_REQUEST can time out.
 159                                 clients_ordered_by_connect_time.pop();
 160                                 continue;
 161                         }
 162
 163                         // OK, it timed out.
 164                         close_client(client);
 165                         clients_ordered_by_connect_time.pop();
 166                 }
 167         }
 168 }
 169
 170 CubemapStateProto Server::serialize()
 171 {
 172         // We don't serialize anything queued, so empty the queues.
 173         process_queued_data();
 174
 175         // Set all clients in a consistent state before serializing
 176         // (ie., they have no remaining lost data). Otherwise, increasing
 177         // the backlog could take clients into a newly valid area of the backlog,
 178         // sending a stream of zeros instead of skipping the data as it should.
 179         //
 180         // TODO: Do this when clients are added back from serialized state instead;
 181         // it would probably be less wasteful.
 182         for (auto &fd_and_client : clients) {
 183                 skip_lost_data(&fd_and_client.second);
 184         }
 185
 186         CubemapStateProto serialized;
 187         for (const auto &fd_and_client : clients) {
 188                 serialized.add_clients()->MergeFrom(fd_and_client.second.serialize());
 189         }
 190         for (unique_ptr<Stream> &stream : streams) {
 191                 serialized.add_streams()->MergeFrom(stream->serialize());
 192         }
 193         return serialized;
 194 }
 195
 196 void Server::add_client_deferred(int sock, Acceptor *acceptor)
 197 {
 198         lock_guard<mutex> lock(queued_clients_mutex);
 199         queued_add_clients.push_back(std::make_pair(sock, acceptor));
 200 }
 201
 202 void Server::add_client(int sock, Acceptor *acceptor)
 203 {
 204         const bool is_tls = acceptor->is_tls();
 205         auto inserted = clients.insert(make_pair(sock, Client(sock)));
 206         assert(inserted.second == true);  // Should not already exist.
 207         Client *client_ptr = &inserted.first->second;
 208
 209         // Connection timestamps must be nondecreasing. I can't find any guarantee
 210         // that even the monotonic clock can't go backwards by a small amount
 211         // (think switching between CPUs with non-synchronized TSCs), so if
 212         // this actually should happen, we hack around it by fudging
 213         // connect_time.
 214         if (!clients_ordered_by_connect_time.empty() &&
 215             is_earlier(client_ptr->connect_time, clients_ordered_by_connect_time.back().first)) {
 216                 client_ptr->connect_time = clients_ordered_by_connect_time.back().first;
 217         }
 218         clients_ordered_by_connect_time.push(make_pair(client_ptr->connect_time, sock));
 219
 220         // Start listening on data from this socket.
 221         epoll_event ev;
 222         if (is_tls) {
 223                 // Even in the initial state (READING_REQUEST), TLS needs to
 224                 // send data for the handshake, and thus might end up needing
 225                 // to know about EPOLLOUT.
 226                 ev.events = EPOLLIN | EPOLLOUT | EPOLLET | EPOLLRDHUP;
 227         } else {
 228                 // EPOLLOUT will be added once we go out of READING_REQUEST.
 229                 ev.events = EPOLLIN | EPOLLET | EPOLLRDHUP;
 230         }
 231         ev.data.ptr = client_ptr;
 232         if (epoll_ctl(epoll_fd, EPOLL_CTL_ADD, sock, &ev) == -1) {
 233                 log_perror("epoll_ctl(EPOLL_CTL_ADD)");
 234                 exit(1);
 235         }
 236
 237         if (is_tls) {
 238                 assert(tls_server_contexts.count(acceptor));
 239                 client_ptr->tls_context = tls_accept(tls_server_contexts[acceptor]);
 240                 if (client_ptr->tls_context == nullptr) {
 241                         log(ERROR, "tls_accept() failed");
 242                         close_client(client_ptr);
 243                         return;
 244                 }
 245                 tls_make_exportable(client_ptr->tls_context, 1);
 246         }
 247
 248         process_client(client_ptr);
 249 }
 250
 251 void Server::add_client_from_serialized(const ClientProto &client)
 252 {
 253         lock_guard<mutex> lock(mu);
 254         Stream *stream;
 255         int stream_index = lookup_stream_by_url(client.url());
 256         if (stream_index == -1) {
 257                 assert(client.state() != Client::SENDING_DATA);
 258                 stream = nullptr;
 259         } else {
 260                 stream = streams[stream_index].get();
 261         }
 262         auto inserted = clients.insert(make_pair(client.sock(), Client(client, stream)));
 263         assert(inserted.second == true);  // Should not already exist.
 264         Client *client_ptr = &inserted.first->second;
 265
 266         // Connection timestamps must be nondecreasing.
 267         assert(clients_ordered_by_connect_time.empty() ||
 268                !is_earlier(client_ptr->connect_time, clients_ordered_by_connect_time.back().first));
 269         clients_ordered_by_connect_time.push(make_pair(client_ptr->connect_time, client.sock()));
 270
 271         // Start listening on data from this socket.
 272         epoll_event ev;
 273         if (client.state() == Client::READING_REQUEST) {
 274                 // See the corresponding comment in Server::add_client().
 275                 if (client.has_tls_context()) {
 276                         ev.events = EPOLLIN | EPOLLOUT | EPOLLET | EPOLLRDHUP;
 277                 } else {
 278                         ev.events = EPOLLIN | EPOLLET | EPOLLRDHUP;
 279                 }
 280         } else {
 281                 // If we don't have more data for this client, we'll be putting it into
 282                 // the sleeping array again soon.
 283                 ev.events = EPOLLOUT | EPOLLET | EPOLLRDHUP;
 284         }
 285         ev.data.ptr = client_ptr;
 286         if (epoll_ctl(epoll_fd, EPOLL_CTL_ADD, client.sock(), &ev) == -1) {
 287                 log_perror("epoll_ctl(EPOLL_CTL_ADD)");
 288                 exit(1);
 289         }
 290
 291         if (client_ptr->state == Client::WAITING_FOR_KEYFRAME ||
 292             client_ptr->state == Client::PREBUFFERING ||
 293             (client_ptr->state == Client::SENDING_DATA &&
 294              client_ptr->stream_pos == client_ptr->stream->bytes_received)) {
 295                 client_ptr->stream->put_client_to_sleep(client_ptr);
 296         } else {
 297                 process_client(client_ptr);
 298         }
 299 }
 300
 301 int Server::lookup_stream_by_url(const string &url) const
 302 {
 303         map<string, int>::const_iterator stream_url_it = stream_url_map.find(url);
 304         if (stream_url_it == stream_url_map.end()) {
 305                 return -1;
 306         }
 307         return stream_url_it->second;
 308 }
 309
 310 int Server::add_stream(const string &url, size_t backlog_size, size_t prebuffering_bytes, Stream::Encoding encoding, Stream::Encoding src_encoding)
 311 {
 312         lock_guard<mutex> lock(mu);
 313         stream_url_map.insert(make_pair(url, streams.size()));
 314         streams.emplace_back(new Stream(url, backlog_size, prebuffering_bytes, encoding, src_encoding));
 315         return streams.size() - 1;
 316 }
 317
 318 int Server::add_stream_from_serialized(const StreamProto &stream, int data_fd)
 319 {
 320         lock_guard<mutex> lock(mu);
 321         stream_url_map.insert(make_pair(stream.url(), streams.size()));
 322         streams.emplace_back(new Stream(stream, data_fd));
 323         return streams.size() - 1;
 324 }
 325
 326 void Server::set_backlog_size(int stream_index, size_t new_size)
 327 {
 328         lock_guard<mutex> lock(mu);
 329         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
 330         streams[stream_index]->set_backlog_size(new_size);
 331 }
 332
 333 void Server::set_prebuffering_bytes(int stream_index, size_t new_amount)
 334 {
 335         lock_guard<mutex> lock(mu);
 336         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
 337         streams[stream_index]->prebuffering_bytes = new_amount;
 338 }
 339
 340 void Server::set_encoding(int stream_index, Stream::Encoding encoding)
 341 {
 342         lock_guard<mutex> lock(mu);
 343         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
 344         streams[stream_index]->encoding = encoding;
 345 }
 346
 347 void Server::set_src_encoding(int stream_index, Stream::Encoding encoding)
 348 {
 349         lock_guard<mutex> lock(mu);
 350         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
 351         streams[stream_index]->src_encoding = encoding;
 352 }
 353
 354 void Server::set_header(int stream_index, const string &http_header, const string &stream_header)
 355 {
 356         lock_guard<mutex> lock(mu);
 357         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
 358         streams[stream_index]->http_header = http_header;
 359
 360         if (stream_header != streams[stream_index]->stream_header) {
 361                 // We cannot start at any of the older starting points anymore,
 362                 // since they'd get the wrong header for the stream (not to mention
 363                 // that a changed header probably means the stream restarted,
 364                 // which means any client starting on the old one would probably
 365                 // stop playing properly at the change point). Next block
 366                 // should be a suitable starting point (if not, something is
 367                 // pretty strange), so it will fill up again soon enough.
 368                 streams[stream_index]->suitable_starting_points.clear();
 369         }
 370         streams[stream_index]->stream_header = stream_header;
 371 }
 372
 373 void Server::set_pacing_rate(int stream_index, uint32_t pacing_rate)
 374 {
 375         lock_guard<mutex> lock(mu);
 376         assert(clients.empty());
 377         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
 378         streams[stream_index]->pacing_rate = pacing_rate;
 379 }
 380
 381 void Server::add_gen204(const std::string &url, const std::string &allow_origin)
 382 {
 383         lock_guard<mutex> lock(mu);
 384         assert(clients.empty());
 385         ping_url_map[url] = allow_origin;
 386 }
 387
 388 void Server::create_tls_context_for_acceptor(const Acceptor *acceptor)
 389 {
 390         assert(acceptor->is_tls());
 391
 392         bool is_server = true;
 393         TLSContext *server_context = tls_create_context(is_server, TLS_V12);
 394
 395         const string &cert = acceptor->get_certificate_chain();
 396         int num_cert = tls_load_certificates(server_context, reinterpret_cast<const unsigned char *>(cert.data()), cert.size());
 397         assert(num_cert > 0);  // Should have been checked by config earlier.
 398
 399         const string &key = acceptor->get_private_key();
 400         int num_key = tls_load_private_key(server_context, reinterpret_cast<const unsigned char *>(key.data()), key.size());
 401         assert(num_key > 0);  // Should have been checked by config earlier.
 402
 403         tls_server_contexts.insert(make_pair(acceptor, server_context));
 404 }
 405
 406 void Server::add_data_deferred(int stream_index, const char *data, size_t bytes, uint16_t metacube_flags)
 407 {
 408         assert(stream_index >= 0 && stream_index < ssize_t(streams.size()));
 409         streams[stream_index]->add_data_deferred(data, bytes, metacube_flags);
 410 }
 411
 412 // See the .h file for postconditions after this function.
 413 void Server::process_client(Client *client)
 414 {
 415         switch (client->state) {
 416         case Client::READING_REQUEST: {
 417                 if (client->tls_context != nullptr) {
 418                         if (send_pending_tls_data(client)) {
 419                                 // send_pending_tls_data() hit postconditions #1 or #4.
 420                                 return;
 421                         }
 422                 }
 423
 424 read_request_again:
 425                 // Try to read more of the request.
 426                 char buf[1024];
 427                 int ret;
 428                 if (client->tls_context == nullptr) {
 429                         ret = read_nontls_data(client, buf, sizeof(buf));
 430                         if (ret == -1) {
 431                                 // read_nontls_data() hit postconditions #1 or #2.
 432                                 return;
 433                         }
 434                 } else {
 435                         ret = read_tls_data(client, buf, sizeof(buf));
 436                         if (ret == -1) {
 437                                 // read_tls_data() hit postconditions #1, #2 or #4.
 438                                 return;
 439                         }
 440                 }
 441
 442                 RequestParseStatus status = wait_for_double_newline(&client->request, buf, ret);
 443
 444                 switch (status) {
 445                 case RP_OUT_OF_SPACE:
 446                         log(WARNING, "[%s] Client sent overlong request!", client->remote_addr.c_str());
 447                         close_client(client);
 448                         return;
 449                 case RP_NOT_FINISHED_YET:
 450                         // OK, we don't have the entire header yet. Fine; we'll get it later.
 451                         // See if there's more data for us.
 452                         goto read_request_again;
 453                 case RP_EXTRA_DATA:
 454                         log(WARNING, "[%s] Junk data after request!", client->remote_addr.c_str());
 455                         close_client(client);
 456                         return;
 457                 case RP_FINISHED:
 458                         break;
 459                 }
 460
 461                 assert(status == RP_FINISHED);
 462
 463                 if (client->tls_context && !client->in_ktls_mode && tls_established(client->tls_context)) {
 464                         // We're ready to enter kTLS mode, unless we still have some
 465                         // handshake data to send (which then must be sent as non-kTLS).
 466                         if (send_pending_tls_data(client)) {
 467                                 // send_pending_tls_data() hit postconditions #1 or #4.
 468                                 return;
 469                         }
 470                         ret = tls_make_ktls(client->tls_context, client->sock);
 471                         if (ret < 0) {
 472                                 log_tls_error("tls_make_ktls", ret);
 473                                 close_client(client);
 474                                 return;
 475                         }
 476                         client->in_ktls_mode = true;
 477                 }
 478
 479                 int error_code = parse_request(client);
 480                 if (error_code == 200) {
 481                         construct_header(client);
 482                 } else if (error_code == 204) {
 483                         construct_204(client);
 484                 } else {
 485                         construct_error(client, error_code);
 486                 }
 487
 488                 // We've changed states, so fall through.
 489                 assert(client->state == Client::SENDING_SHORT_RESPONSE ||
 490                        client->state == Client::SENDING_HEADER);
 491         }
 492         case Client::SENDING_SHORT_RESPONSE:
 493         case Client::SENDING_HEADER: {
 494 sending_header_or_short_response_again:
 495                 int ret;
 496                 do {
 497                         ret = write(client->sock,
 498                                     client->header_or_short_response.data() + client->header_or_short_response_bytes_sent,
 499                                     client->header_or_short_response.size() - client->header_or_short_response_bytes_sent);
 500                 } while (ret == -1 && errno == EINTR);
 501
 502                 if (ret == -1 && errno == EAGAIN) {
 503                         // We're out of socket space, so now we're at the “low edge” of epoll's
 504                         // edge triggering. epoll will tell us when there is more room, so for now,
 505                         // just return.
 506                         // This is postcondition #4.
 507                         return;
 508                 }
 509
 510                 if (ret == -1) {
 511                         // Error! Postcondition #1.
 512                         log_perror("write");
 513                         close_client(client);
 514                         return;
 515                 }
 516
 517                 client->header_or_short_response_bytes_sent += ret;
 518                 assert(client->header_or_short_response_bytes_sent <= client->header_or_short_response.size());
 519
 520                 if (client->header_or_short_response_bytes_sent < client->header_or_short_response.size()) {
 521                         // We haven't sent all yet. Fine; go another round.
 522                         goto sending_header_or_short_response_again;
 523                 }
 524
 525                 // We're done sending the header or error! Clear it to release some memory.
 526                 client->header_or_short_response.clear();
 527
 528                 if (client->state == Client::SENDING_SHORT_RESPONSE) {
 529                         if (more_requests(client)) {
 530                                 // We're done sending the error, but should keep on reading new requests.
 531                                 goto read_request_again;
 532                         } else {
 533                                 // We're done sending the error, so now close.
 534                                 // This is postcondition #1.
 535                                 close_client(client);
 536                         }
 537                         return;
 538                 }
 539
 540                 Stream *stream = client->stream;
 541                 if (client->stream_pos == size_t(-2)) {
 542                         // Start sending from the beginning of the backlog.
 543                         client->stream_pos = min<size_t>(
 544                             stream->bytes_received - stream->backlog_size,
 545                             0);
 546                         client->state = Client::SENDING_DATA;
 547                         goto sending_data;
 548                 } else if (stream->prebuffering_bytes == 0) {
 549                         // Start sending from the first keyframe we get. In other
 550                         // words, we won't send any of the backlog, but we'll start
 551                         // sending immediately as we get the next keyframe block.
 552                         // Note that this is functionally identical to the next if branch,
 553                         // except that we save a binary search.
 554                         client->stream_pos = stream->bytes_received;
 555                         client->state = Client::WAITING_FOR_KEYFRAME;
 556                 } else {
 557                         // We're not going to send anything to the client before we have
 558                         // N bytes. However, this wait might be boring; we can just as well
 559                         // use it to send older data if we have it. We use lower_bound()
 560                         // so that we are conservative and never add extra latency over just
 561                         // waiting (assuming CBR or nearly so); otherwise, we could want e.g.
 562                         // 100 kB prebuffer but end up sending a 10 MB GOP.
 563                         deque<size_t>::const_iterator starting_point_it =
 564                                 lower_bound(stream->suitable_starting_points.begin(),
 565                                             stream->suitable_starting_points.end(),
 566                                             stream->bytes_received - stream->prebuffering_bytes);
 567                         if (starting_point_it == stream->suitable_starting_points.end()) {
 568                                 // None found. Just put us at the end, and then wait for the
 569                                 // first keyframe to appear.
 570                                 client->stream_pos = stream->bytes_received;
 571                                 client->state = Client::WAITING_FOR_KEYFRAME;
 572                         } else {
 573                                 client->stream_pos = *starting_point_it;
 574                                 client->state = Client::PREBUFFERING;
 575                                 goto prebuffering;
 576                         }
 577                 }
 578                 // Fall through.
 579         }
 580         case Client::WAITING_FOR_KEYFRAME: {
 581                 Stream *stream = client->stream;
 582                 if (stream->suitable_starting_points.empty() ||
 583                     client->stream_pos > stream->suitable_starting_points.back()) {
 584                         // We haven't received a keyframe since this stream started waiting,
 585                         // so keep on waiting for one.
 586                         // This is postcondition #3.
 587                         stream->put_client_to_sleep(client);
 588                         return;
 589                 }
 590                 client->stream_pos = stream->suitable_starting_points.back();
 591                 client->state = Client::PREBUFFERING;
 592                 // Fall through.
 593         }
 594         case Client::PREBUFFERING: {
 595 prebuffering:
 596                 Stream *stream = client->stream;
 597                 size_t bytes_to_send = stream->bytes_received - client->stream_pos;
 598                 assert(bytes_to_send <= stream->backlog_size);
 599                 if (bytes_to_send < stream->prebuffering_bytes) {
 600                         // We don't have enough bytes buffered to start this client yet.
 601                         // This is postcondition #3.
 602                         stream->put_client_to_sleep(client);
 603                         return;
 604                 }
 605                 client->state = Client::SENDING_DATA;
 606                 // Fall through.
 607         }
 608         case Client::SENDING_DATA: {
 609 sending_data:
 610                 skip_lost_data(client);
 611                 Stream *stream = client->stream;
 612
 613 sending_data_again:
 614                 size_t bytes_to_send = stream->bytes_received - client->stream_pos;
 615                 assert(bytes_to_send <= stream->backlog_size);
 616                 if (bytes_to_send == 0) {
 617                         return;
 618                 }
 619
 620                 // See if we need to split across the circular buffer.
 621                 bool more_data = false;
 622                 if ((client->stream_pos % stream->backlog_size) + bytes_to_send > stream->backlog_size) {
 623                         bytes_to_send = stream->backlog_size - (client->stream_pos % stream->backlog_size);
 624                         more_data = true;
 625                 }
 626
 627                 ssize_t ret;
 628                 do {
 629                         off_t offset = client->stream_pos % stream->backlog_size;
 630                         ret = sendfile(client->sock, stream->data_fd, &offset, bytes_to_send);
 631                 } while (ret == -1 && errno == EINTR);
 632
 633                 if (ret == -1 && errno == EAGAIN) {
 634                         // We're out of socket space, so return; epoll will wake us up
 635                         // when there is more room.
 636                         // This is postcondition #4.
 637                         return;
 638                 }
 639                 if (ret == -1) {
 640                         // Error, close; postcondition #1.
 641                         log_perror("sendfile");
 642                         close_client(client);
 643                         return;
 644                 }
 645                 client->stream_pos += ret;
 646                 client->bytes_sent += ret;
 647
 648                 if (client->stream_pos == stream->bytes_received) {
 649                         // We don't have any more data for this client, so put it to sleep.
 650                         // This is postcondition #3.
 651                         stream->put_client_to_sleep(client);
 652                 } else if (more_data && size_t(ret) == bytes_to_send) {
 653                         goto sending_data_again;
 654                 }
 655                 break;
 656         }
 657         default:
 658                 assert(false);
 659         }
 660 }
 661
 662 bool Server::send_pending_tls_data(Client *client)
 663 {
 664         // See if there's data from the TLS library to write.
 665         if (client->tls_data_to_send == nullptr) {
 666                 client->tls_data_to_send = tls_get_write_buffer(client->tls_context, &client->tls_data_left_to_send);
 667                 if (client->tls_data_to_send == nullptr) {
 668                         // Really no data to send.
 669                         return false;
 670                 }
 671         }
 672
 673 send_data_again:
 674         int ret;
 675         do {
 676                 ret = write(client->sock, client->tls_data_to_send, client->tls_data_left_to_send);
 677         } while (ret == -1 && errno == EINTR);
 678         assert(ret < 0 || size_t(ret) <= client->tls_data_left_to_send);
 679
 680         if (ret == -1 && errno == EAGAIN) {
 681                 // We're out of socket space, so now we're at the “low edge” of epoll's
 682                 // edge triggering. epoll will tell us when there is more room, so for now,
 683                 // just return.
 684                 // This is postcondition #4.
 685                 return true;
 686         }
 687         if (ret == -1) {
 688                 // Error! Postcondition #1.
 689                 log_perror("write");
 690                 close_client(client);
 691                 return true;
 692         }
 693         if (ret > 0 && size_t(ret) == client->tls_data_left_to_send) {
 694                 // All data has been sent, so we don't need to go to sleep.
 695                 tls_buffer_clear(client->tls_context);
 696                 client->tls_data_to_send = nullptr;
 697                 return false;
 698         }
 699
 700         // More data to send, so try again.
 701         client->tls_data_to_send += ret;
 702         client->tls_data_left_to_send -= ret;
 703         goto send_data_again;
 704 }
 705
 706 int Server::read_nontls_data(Client *client, char *buf, size_t max_size)
 707 {
 708         int ret;
 709         do {
 710                 ret = read(client->sock, buf, max_size);
 711         } while (ret == -1 && errno == EINTR);
 712
 713         if (ret == -1 && errno == EAGAIN) {
 714                 // No more data right now. Nothing to do.
 715                 // This is postcondition #2.
 716                 return -1;
 717         }
 718         if (ret == -1) {
 719                 log_perror("read");
 720                 close_client(client);
 721                 return -1;
 722         }
 723         if (ret == 0) {
 724                 // OK, the socket is closed.
 725                 close_client(client);
 726                 return -1;
 727         }
 728
 729         return ret;
 730 }
 731
 732 int Server::read_tls_data(Client *client, char *buf, size_t max_size)
 733 {
 734 read_again:
 735         int ret;
 736         do {
 737                 ret = read(client->sock, buf, max_size);
 738         } while (ret == -1 && errno == EINTR);
 739
 740         if (ret == -1 && errno == EAGAIN) {
 741                 // No more data right now. Nothing to do.
 742                 // This is postcondition #2.
 743                 return -1;
 744         }
 745         if (ret == -1) {
 746                 log_perror("read");
 747                 close_client(client);
 748                 return -1;
 749         }
 750         if (ret == 0) {
 751                 // OK, the socket is closed.
 752                 close_client(client);
 753                 return -1;
 754         }
 755
 756         // Give it to the TLS library.
 757         int err = tls_consume_stream(client->tls_context, reinterpret_cast<const unsigned char *>(buf), ret, nullptr);
 758         if (err < 0) {
 759                 log_tls_error("tls_consume_stream", err);
 760                 close_client(client);
 761                 return -1;
 762         }
 763         if (err == 0) {
 764                 // Not consumed any data. See if we can read more.
 765                 goto read_again;
 766         }
 767
 768         // Read any decrypted data available for us. (We can reuse buf, since it's free now.)
 769         ret = tls_read(client->tls_context, reinterpret_cast<unsigned char *>(buf), max_size);
 770         if (ret == 0) {
 771                 // No decrypted data for us yet, but there might be some more handshaking
 772                 // to send. Do that if needed, then look for more data.
 773                 if (send_pending_tls_data(client)) {
 774                         // send_pending_tls_data() hit postconditions #1 or #4.
 775                         return -1;
 776                 }
 777                 goto read_again;
 778         }
 779         if (ret < 0) {
 780                 log_tls_error("tls_read", ret);
 781                 close_client(client);
 782                 return -1;
 783         }
 784
 785         assert(ret > 0);
 786         return ret;
 787 }
 788
 789 // See if there's some data we've lost. Ideally, we should drop to a block boundary,
 790 // but resync will be the mux's problem.
 791 void Server::skip_lost_data(Client *client)
 792 {
 793         Stream *stream = client->stream;
 794         if (stream == nullptr) {
 795                 return;
 796         }
 797         size_t bytes_to_send = stream->bytes_received - client->stream_pos;
 798         if (bytes_to_send > stream->backlog_size) {
 799                 size_t bytes_lost = bytes_to_send - stream->backlog_size;
 800                 client->stream_pos = stream->bytes_received - stream->backlog_size;
 801                 client->bytes_lost += bytes_lost;
 802                 ++client->num_loss_events;
 803         }
 804 }
 805
 806 int Server::parse_request(Client *client)
 807 {
 808         vector<string> lines = split_lines(client->request);
 809         client->request.clear();
 810         if (lines.empty()) {
 811                 return 400;  // Bad request (empty).
 812         }
 813
 814         // Parse the headers, for logging purposes.
 815         // TODO: Case-insensitivity.
 816         multimap<string, string> headers = extract_headers(lines, client->remote_addr);
 817         multimap<string, string>::const_iterator referer_it = headers.find("Referer");
 818         if (referer_it != headers.end()) {
 819                 client->referer = referer_it->second;
 820         }
 821         multimap<string, string>::const_iterator user_agent_it = headers.find("User-Agent");
 822         if (user_agent_it != headers.end()) {
 823                 client->user_agent = user_agent_it->second;
 824         }
 825
 826         vector<string> request_tokens = split_tokens(lines[0]);
 827         if (request_tokens.size() < 3) {
 828                 return 400;  // Bad request (empty).
 829         }
 830         if (request_tokens[0] != "GET") {
 831                 return 400;  // Should maybe be 405 instead?
 832         }
 833
 834         string url = request_tokens[1];
 835         client->url = url;
 836         if (url.size() > 8 && url.find("?backlog") == url.size() - 8) {
 837                 client->stream_pos = -2;
 838                 url = url.substr(0, url.size() - 8);
 839         } else {
 840                 client->stream_pos = -1;
 841         }
 842
 843         // Figure out if we're supposed to close the socket after we've delivered the response.
 844         string protocol = request_tokens[2];
 845         if (protocol.find("HTTP/") != 0) {
 846                 return 400;  // Bad request.
 847         }
 848         client->close_after_response = false;
 849         client->http_11 = true;
 850         if (protocol == "HTTP/1.0") {
 851                 // No persistent connections.
 852                 client->close_after_response = true;
 853                 client->http_11 = false;
 854         } else {
 855                 multimap<string, string>::const_iterator connection_it = headers.find("Connection");
 856                 if (connection_it != headers.end() && connection_it->second == "close") {
 857                         client->close_after_response = true;
 858                 }
 859         }
 860
 861         map<string, int>::const_iterator stream_url_map_it = stream_url_map.find(url);
 862         if (stream_url_map_it == stream_url_map.end()) {
 863                 map<string, string>::const_iterator ping_url_map_it = ping_url_map.find(url);
 864                 if (ping_url_map_it == ping_url_map.end()) {
 865                         return 404;  // Not found.
 866                 } else {
 867                         return 204;  // No error.
 868                 }
 869         }
 870
 871         Stream *stream = streams[stream_url_map_it->second].get();
 872         if (stream->http_header.empty()) {
 873                 return 503;  // Service unavailable.
 874         }
 875
 876         // Streams currently never end, so we don't have a content-length,
 877         // and can just as well tell the client it's Connection: close
 878         // (otherwise, we'd have to implement chunking TE for no good reason).
 879         // When we start to support fragments, this will change.
 880         client->close_after_response = true;
 881
 882         client->stream = stream;
 883         if (setsockopt(client->sock, SOL_SOCKET, SO_MAX_PACING_RATE, &client->stream->pacing_rate, sizeof(client->stream->pacing_rate)) == -1) {
 884                 if (client->stream->pacing_rate != ~0U) {
 885                         log_perror("setsockopt(SO_MAX_PACING_RATE)");
 886                 }
 887         }
 888         client->request.clear();
 889
 890         return 200;  // OK!
 891 }
 892
 893 void Server::construct_header(Client *client)
 894 {
 895         Stream *stream = client->stream;
 896         client->header_or_short_response = stream->http_header;
 897         if (client->http_11) {
 898                 assert(client->header_or_short_response.find("HTTP/1.0") == 0);
 899                 client->header_or_short_response[7] = '1';  // Change to HTTP/1.1.
 900                 if (client->close_after_response) {
 901                         client->header_or_short_response.append("Connection: close\r\n");
 902                 }
 903         } else {
 904                 assert(client->close_after_response);
 905         }
 906         if (stream->encoding == Stream::STREAM_ENCODING_RAW) {
 907                 client->header_or_short_response.append("\r\n");
 908         } else if (stream->encoding == Stream::STREAM_ENCODING_METACUBE) {
 909                 client->header_or_short_response.append(
 910                         "Content-encoding: metacube\r\n"
 911                         "\r\n");
 912                 if (!stream->stream_header.empty()) {
 913                         metacube2_block_header hdr;
 914                         memcpy(hdr.sync, METACUBE2_SYNC, sizeof(hdr.sync));
 915                         hdr.size = htonl(stream->stream_header.size());
 916                         hdr.flags = htons(METACUBE_FLAGS_HEADER);
 917                         hdr.csum = htons(metacube2_compute_crc(&hdr));
 918                         client->header_or_short_response.append(
 919                                 string(reinterpret_cast<char *>(&hdr), sizeof(hdr)));
 920                 }
 921         } else {
 922                 assert(false);
 923         }
 924         client->header_or_short_response.append(stream->stream_header);
 925
 926         // Switch states.
 927         client->state = Client::SENDING_HEADER;
 928         change_epoll_events(client, EPOLLOUT | EPOLLET | EPOLLRDHUP);
 929 }
 930
 931 void Server::construct_error(Client *client, int error_code)
 932 {
 933         char error[256];
 934         if (client->http_11 && client->close_after_response) {
 935                 snprintf(error, sizeof(error),
 936                         "HTTP/1.1 %d Error\r\nContent-type: text/plain\r\nConnection: close\r\n\r\nSomething went wrong. Sorry.\r\n",
 937                         error_code);
 938         } else {
 939                 snprintf(error, sizeof(error),
 940                         "HTTP/1.%d %d Error\r\nContent-type: text/plain\r\nContent-length: 30\r\n\r\nSomething went wrong. Sorry.\r\n",
 941                         client->http_11, error_code);
 942         }
 943         client->header_or_short_response = error;
 944
 945         // Switch states.
 946         client->state = Client::SENDING_SHORT_RESPONSE;
 947         change_epoll_events(client, EPOLLOUT | EPOLLET | EPOLLRDHUP);
 948 }
 949
 950 void Server::construct_204(Client *client)
 951 {
 952         map<string, string>::const_iterator ping_url_map_it = ping_url_map.find(client->url);
 953         assert(ping_url_map_it != ping_url_map.end());
 954
 955         if (client->http_11) {
 956                 client->header_or_short_response = "HTTP/1.1 204 No Content\r\n";
 957                 if (client->close_after_response) {
 958                         client->header_or_short_response.append("Connection: close\r\n");
 959                 }
 960         } else {
 961                 client->header_or_short_response = "HTTP/1.0 204 No Content\r\n";
 962                 assert(client->close_after_response);
 963         }
 964         if (!ping_url_map_it->second.empty()) {
 965                 client->header_or_short_response.append("Access-Control-Allow-Origin: ");
 966                 client->header_or_short_response.append(ping_url_map_it->second);
 967         }
 968         client->header_or_short_response.append("\r\n");
 969
 970         // Switch states.
 971         client->state = Client::SENDING_SHORT_RESPONSE;
 972         change_epoll_events(client, EPOLLOUT | EPOLLET | EPOLLRDHUP);
 973 }
 974
 975 template<class T>
 976 void delete_from(vector<T> *v, T elem)
 977 {
 978         typename vector<T>::iterator new_end = remove(v->begin(), v->end(), elem);
 979         v->erase(new_end, v->end());
 980 }
 981
 982 void Server::close_client(Client *client)
 983 {
 984         if (epoll_ctl(epoll_fd, EPOLL_CTL_DEL, client->sock, nullptr) == -1) {
 985                 log_perror("epoll_ctl(EPOLL_CTL_DEL)");
 986                 exit(1);
 987         }
 988
 989         // This client could be sleeping, so we'll need to fix that. (Argh, O(n).)
 990         if (client->stream != nullptr) {
 991                 delete_from(&client->stream->sleeping_clients, client);
 992                 delete_from(&client->stream->to_process, client);
 993         }
 994
 995         if (client->tls_context) {
 996                 tls_destroy_context(client->tls_context);
 997         }
 998
 999         // Log to access_log.
1000         access_log->write(client->get_stats());
1001
1002         // Bye-bye!
1003         safe_close(client->sock);
1004
1005         clients.erase(client->sock);
1006 }
1007
1008 void Server::change_epoll_events(Client *client, uint32_t events)
1009 {
1010         epoll_event ev;
1011         ev.events = events;
1012         ev.data.ptr = client;
1013
1014         if (epoll_ctl(epoll_fd, EPOLL_CTL_MOD, client->sock, &ev) == -1) {
1015                 log_perror("epoll_ctl(EPOLL_CTL_MOD)");
1016                 exit(1);
1017         }
1018 }
1019
1020 bool Server::more_requests(Client *client)
1021 {
1022         if (client->close_after_response) {
1023                 return false;
1024         }
1025
1026         // Switch states and reset the parsers. We don't reset statistics.
1027         client->state = Client::READING_REQUEST;
1028         client->url.clear();
1029         client->stream = NULL;
1030         client->header_or_short_response.clear();
1031         client->header_or_short_response_bytes_sent = 0;
1032
1033         change_epoll_events(client, EPOLLIN | EPOLLET | EPOLLRDHUP);  // No TLS handshake, so no EPOLLOUT needed.
1034
1035         return true;
1036 }
1037
1038 void Server::process_queued_data()
1039 {
1040         {
1041                 lock_guard<mutex> lock(queued_clients_mutex);
1042
1043                 for (const pair<int, Acceptor *> &id_and_acceptor : queued_add_clients) {
1044                         add_client(id_and_acceptor.first, id_and_acceptor.second);
1045                 }
1046                 queued_add_clients.clear();
1047         }
1048
1049         for (unique_ptr<Stream> &stream : streams) {
1050                 stream->process_queued_data();
1051         }
1052 }