net/socket/tcp_socket_libevent.cc

   1 // Copyright 2013 The Chromium Authors. All rights reserved.
   2 // Use of this source code is governed by a BSD-style license that can be
   3 // found in the LICENSE file.
   4
   5 #include "net/socket/tcp_socket.h"
   6
   7 #include <errno.h>
   8 #include <netinet/tcp.h>
   9 #include <sys/socket.h>
  10
  11 #include "base/bind.h"
  12 #include "base/logging.h"
  13 #include "base/metrics/histogram.h"
  14 #include "base/metrics/stats_counters.h"
  15 #include "base/posix/eintr_wrapper.h"
  16 #include "base/task_runner_util.h"
  17 #include "base/threading/worker_pool.h"
  18 #include "net/base/address_list.h"
  19 #include "net/base/connection_type_histograms.h"
  20 #include "net/base/io_buffer.h"
  21 #include "net/base/ip_endpoint.h"
  22 #include "net/base/net_errors.h"
  23 #include "net/base/net_util.h"
  24 #include "net/base/network_activity_monitor.h"
  25 #include "net/base/network_change_notifier.h"
  26 #include "net/socket/socket_libevent.h"
  27 #include "net/socket/socket_net_log_params.h"
  28
  29 // If we don't have a definition for TCPI_OPT_SYN_DATA, create one.
  30 #ifndef TCPI_OPT_SYN_DATA
  31 #define TCPI_OPT_SYN_DATA 32
  32 #endif
  33
  34 namespace net {
  35
  36 namespace {
  37
  38 // True if OS supports TCP FastOpen.
  39 bool g_tcp_fastopen_supported = false;
  40 // True if TCP FastOpen is user-enabled for all connections.
  41 // TODO(jri): Change global variable to param in HttpNetworkSession::Params.
  42 bool g_tcp_fastopen_user_enabled = false;
  43 // True if TCP FastOpen connect-with-write has failed at least once.
  44 bool g_tcp_fastopen_has_failed = false;
  45
  46 // SetTCPNoDelay turns on/off buffering in the kernel. By default, TCP sockets
  47 // will wait up to 200ms for more data to complete a packet before transmitting.
  48 // After calling this function, the kernel will not wait. See TCP_NODELAY in
  49 // `man 7 tcp`.
  50 bool SetTCPNoDelay(int fd, bool no_delay) {
  51   int on = no_delay ? 1 : 0;
  52   int error = setsockopt(fd, IPPROTO_TCP, TCP_NODELAY, &on, sizeof(on));
  53   return error == 0;
  54 }
  55
  56 // SetTCPKeepAlive sets SO_KEEPALIVE.
  57 bool SetTCPKeepAlive(int fd, bool enable, int delay) {
  58   int on = enable ? 1 : 0;
  59   if (setsockopt(fd, SOL_SOCKET, SO_KEEPALIVE, &on, sizeof(on))) {
  60     PLOG(ERROR) << "Failed to set SO_KEEPALIVE on fd: " << fd;
  61     return false;
  62   }
  63
  64   // If we disabled TCP keep alive, our work is done here.
  65   if (!enable)
  66     return true;
  67
  68 #if defined(OS_LINUX) || defined(OS_ANDROID)
  69   // Set seconds until first TCP keep alive.
  70   if (setsockopt(fd, SOL_TCP, TCP_KEEPIDLE, &delay, sizeof(delay))) {
  71     PLOG(ERROR) << "Failed to set TCP_KEEPIDLE on fd: " << fd;
  72     return false;
  73   }
  74   // Set seconds between TCP keep alives.
  75   if (setsockopt(fd, SOL_TCP, TCP_KEEPINTVL, &delay, sizeof(delay))) {
  76     PLOG(ERROR) << "Failed to set TCP_KEEPINTVL on fd: " << fd;
  77     return false;
  78   }
  79 #endif
  80   return true;
  81 }
  82
  83 #if defined(OS_LINUX) || defined(OS_ANDROID)
  84 // Checks if the kernel supports TCP FastOpen.
  85 bool SystemSupportsTCPFastOpen() {
  86   const base::FilePath::CharType kTCPFastOpenProcFilePath[] =
  87       "/proc/sys/net/ipv4/tcp_fastopen";
  88   std::string system_supports_tcp_fastopen;
  89   if (!base::ReadFileToString(base::FilePath(kTCPFastOpenProcFilePath),
  90                               &system_supports_tcp_fastopen)) {
  91     return false;
  92   }
  93   // The read from /proc should return '1' if TCP FastOpen is enabled in the OS.
  94   if (system_supports_tcp_fastopen.empty() ||
  95       (system_supports_tcp_fastopen[0] != '1')) {
  96     return false;
  97   }
  98   return true;
  99 }
 100
 101 void RegisterTCPFastOpenIntentAndSupport(bool user_enabled,
 102                                          bool system_supported) {
 103   g_tcp_fastopen_supported = system_supported;
 104   g_tcp_fastopen_user_enabled = user_enabled;
 105 }
 106 #endif
 107
 108 }  // namespace
 109
 110 //-----------------------------------------------------------------------------
 111
 112 bool IsTCPFastOpenSupported() {
 113   return g_tcp_fastopen_supported;
 114 }
 115
 116 bool IsTCPFastOpenUserEnabled() {
 117   return g_tcp_fastopen_user_enabled;
 118 }
 119
 120 // This is asynchronous because it needs to do file IO, and it isn't allowed to
 121 // do that on the IO thread.
 122 void CheckSupportAndMaybeEnableTCPFastOpen(bool user_enabled) {
 123 #if defined(OS_LINUX) || defined(OS_ANDROID)
 124   base::PostTaskAndReplyWithResult(
 125       base::WorkerPool::GetTaskRunner(/*task_is_slow=*/false).get(),
 126       FROM_HERE,
 127       base::Bind(SystemSupportsTCPFastOpen),
 128       base::Bind(RegisterTCPFastOpenIntentAndSupport, user_enabled));
 129 #endif
 130 }
 131
 132 TCPSocketLibevent::TCPSocketLibevent(NetLog* net_log,
 133                                      const NetLog::Source& source)
 134     : use_tcp_fastopen_(false),
 135       tcp_fastopen_write_attempted_(false),
 136       tcp_fastopen_connected_(false),
 137       tcp_fastopen_status_(TCP_FASTOPEN_STATUS_UNKNOWN),
 138       logging_multiple_connect_attempts_(false),
 139       net_log_(BoundNetLog::Make(net_log, NetLog::SOURCE_SOCKET)) {
 140   net_log_.BeginEvent(NetLog::TYPE_SOCKET_ALIVE,
 141                       source.ToEventParametersCallback());
 142 }
 143
 144 TCPSocketLibevent::~TCPSocketLibevent() {
 145   net_log_.EndEvent(NetLog::TYPE_SOCKET_ALIVE);
 146   Close();
 147 }
 148
 149 int TCPSocketLibevent::Open(AddressFamily family) {
 150   DCHECK(!socket_);
 151   socket_.reset(new SocketLibevent);
 152   int rv = socket_->Open(ConvertAddressFamily(family));
 153   if (rv != OK)
 154     socket_.reset();
 155   return rv;
 156 }
 157
 158 int TCPSocketLibevent::AdoptConnectedSocket(int socket_fd,
 159                                             const IPEndPoint& peer_address) {
 160   DCHECK(!socket_);
 161
 162   SockaddrStorage storage;
 163   if (!peer_address.ToSockAddr(storage.addr, &storage.addr_len) &&
 164       // For backward compatibility, allows the empty address.
 165       !(peer_address == IPEndPoint())) {
 166     return ERR_ADDRESS_INVALID;
 167   }
 168
 169   socket_.reset(new SocketLibevent);
 170   int rv = socket_->AdoptConnectedSocket(socket_fd, storage);
 171   if (rv != OK)
 172     socket_.reset();
 173   return rv;
 174 }
 175
 176 int TCPSocketLibevent::Bind(const IPEndPoint& address) {
 177   DCHECK(socket_);
 178
 179   SockaddrStorage storage;
 180   if (!address.ToSockAddr(storage.addr, &storage.addr_len))
 181     return ERR_ADDRESS_INVALID;
 182
 183   return socket_->Bind(storage);
 184 }
 185
 186 int TCPSocketLibevent::Listen(int backlog) {
 187   DCHECK(socket_);
 188   return socket_->Listen(backlog);
 189 }
 190
 191 int TCPSocketLibevent::Accept(scoped_ptr<TCPSocketLibevent>* tcp_socket,
 192                               IPEndPoint* address,
 193                               const CompletionCallback& callback) {
 194   DCHECK(tcp_socket);
 195   DCHECK(!callback.is_null());
 196   DCHECK(socket_);
 197   DCHECK(!accept_socket_);
 198
 199   net_log_.BeginEvent(NetLog::TYPE_TCP_ACCEPT);
 200
 201   int rv = socket_->Accept(
 202       &accept_socket_,
 203       base::Bind(&TCPSocketLibevent::AcceptCompleted,
 204                  base::Unretained(this), tcp_socket, address, callback));
 205   if (rv != ERR_IO_PENDING)
 206     rv = HandleAcceptCompleted(tcp_socket, address, rv);
 207   return rv;
 208 }
 209
 210 int TCPSocketLibevent::Connect(const IPEndPoint& address,
 211                                const CompletionCallback& callback) {
 212   DCHECK(socket_);
 213
 214   if (!logging_multiple_connect_attempts_)
 215     LogConnectBegin(AddressList(address));
 216
 217   net_log_.BeginEvent(NetLog::TYPE_TCP_CONNECT_ATTEMPT,
 218                       CreateNetLogIPEndPointCallback(&address));
 219
 220   SockaddrStorage storage;
 221   if (!address.ToSockAddr(storage.addr, &storage.addr_len))
 222     return ERR_ADDRESS_INVALID;
 223
 224   if (use_tcp_fastopen_) {
 225     // With TCP FastOpen, we pretend that the socket is connected.
 226     DCHECK(!tcp_fastopen_write_attempted_);
 227     socket_->SetPeerAddress(storage);
 228     return OK;
 229   }
 230
 231   int rv = socket_->Connect(storage,
 232                             base::Bind(&TCPSocketLibevent::ConnectCompleted,
 233                                        base::Unretained(this), callback));
 234   if (rv != ERR_IO_PENDING)
 235     rv = HandleConnectCompleted(rv);
 236   return rv;
 237 }
 238
 239 bool TCPSocketLibevent::IsConnected() const {
 240   if (!socket_)
 241     return false;
 242
 243   if (use_tcp_fastopen_ && !tcp_fastopen_write_attempted_ &&
 244       socket_->HasPeerAddress()) {
 245     // With TCP FastOpen, we pretend that the socket is connected.
 246     // This allows GetPeerAddress() to return peer_address_.
 247     return true;
 248   }
 249
 250   return socket_->IsConnected();
 251 }
 252
 253 bool TCPSocketLibevent::IsConnectedAndIdle() const {
 254   // TODO(wtc): should we also handle the TCP FastOpen case here,
 255   // as we do in IsConnected()?
 256   return socket_ && socket_->IsConnectedAndIdle();
 257 }
 258
 259 int TCPSocketLibevent::Read(IOBuffer* buf,
 260                             int buf_len,
 261                             const CompletionCallback& callback) {
 262   DCHECK(socket_);
 263   DCHECK(!callback.is_null());
 264
 265   int rv = socket_->Read(
 266       buf, buf_len,
 267       base::Bind(&TCPSocketLibevent::ReadCompleted,
 268                  // Grab a reference to |buf| so that ReadCompleted() can still
 269                  // use it when Read() completes, as otherwise, this transfers
 270                  // ownership of buf to socket.
 271                  base::Unretained(this), make_scoped_refptr(buf), callback));
 272   if (rv != ERR_IO_PENDING)
 273     rv = HandleReadCompleted(buf, rv);
 274   return rv;
 275 }
 276
 277 int TCPSocketLibevent::Write(IOBuffer* buf,
 278                              int buf_len,
 279                              const CompletionCallback& callback) {
 280   DCHECK(socket_);
 281   DCHECK(!callback.is_null());
 282
 283   CompletionCallback write_callback =
 284       base::Bind(&TCPSocketLibevent::WriteCompleted,
 285                  // Grab a reference to |buf| so that WriteCompleted() can still
 286                  // use it when Write() completes, as otherwise, this transfers
 287                  // ownership of buf to socket.
 288                  base::Unretained(this), make_scoped_refptr(buf), callback);
 289   int rv;
 290
 291   if (use_tcp_fastopen_ && !tcp_fastopen_write_attempted_) {
 292     rv = TcpFastOpenWrite(buf, buf_len, write_callback);
 293   } else {
 294     rv = socket_->Write(buf, buf_len, write_callback);
 295   }
 296
 297   if (rv != ERR_IO_PENDING)
 298     rv = HandleWriteCompleted(buf, rv);
 299   return rv;
 300 }
 301
 302 int TCPSocketLibevent::GetLocalAddress(IPEndPoint* address) const {
 303   DCHECK(address);
 304
 305   if (!socket_)
 306     return ERR_SOCKET_NOT_CONNECTED;
 307
 308   SockaddrStorage storage;
 309   int rv = socket_->GetLocalAddress(&storage);
 310   if (rv != OK)
 311     return rv;
 312
 313   if (!address->FromSockAddr(storage.addr, storage.addr_len))
 314     return ERR_ADDRESS_INVALID;
 315
 316   return OK;
 317 }
 318
 319 int TCPSocketLibevent::GetPeerAddress(IPEndPoint* address) const {
 320   DCHECK(address);
 321
 322   if (!IsConnected())
 323     return ERR_SOCKET_NOT_CONNECTED;
 324
 325   SockaddrStorage storage;
 326   int rv = socket_->GetPeerAddress(&storage);
 327   if (rv != OK)
 328     return rv;
 329
 330   if (!address->FromSockAddr(storage.addr, storage.addr_len))
 331     return ERR_ADDRESS_INVALID;
 332
 333   return OK;
 334 }
 335
 336 int TCPSocketLibevent::SetDefaultOptionsForServer() {
 337   DCHECK(socket_);
 338   return SetAddressReuse(true);
 339 }
 340
 341 void TCPSocketLibevent::SetDefaultOptionsForClient() {
 342   DCHECK(socket_);
 343
 344   // This mirrors the behaviour on Windows. See the comment in
 345   // tcp_socket_win.cc after searching for "NODELAY".
 346   // If SetTCPNoDelay fails, we don't care.
 347   SetTCPNoDelay(socket_->socket_fd(), true);
 348
 349   // TCP keep alive wakes up the radio, which is expensive on mobile. Do not
 350   // enable it there. It's useful to prevent TCP middleboxes from timing out
 351   // connection mappings. Packets for timed out connection mappings at
 352   // middleboxes will either lead to:
 353   // a) Middleboxes sending TCP RSTs. It's up to higher layers to check for this
 354   // and retry. The HTTP network transaction code does this.
 355   // b) Middleboxes just drop the unrecognized TCP packet. This leads to the TCP
 356   // stack retransmitting packets per TCP stack retransmission timeouts, which
 357   // are very high (on the order of seconds). Given the number of
 358   // retransmissions required before killing the connection, this can lead to
 359   // tens of seconds or even minutes of delay, depending on OS.
 360 #if !defined(OS_ANDROID) && !defined(OS_IOS)
 361   const int kTCPKeepAliveSeconds = 45;
 362
 363   SetTCPKeepAlive(socket_->socket_fd(), true, kTCPKeepAliveSeconds);
 364 #endif
 365 }
 366
 367 int TCPSocketLibevent::SetAddressReuse(bool allow) {
 368   DCHECK(socket_);
 369
 370   // SO_REUSEADDR is useful for server sockets to bind to a recently unbound
 371   // port. When a socket is closed, the end point changes its state to TIME_WAIT
 372   // and wait for 2 MSL (maximum segment lifetime) to ensure the remote peer
 373   // acknowledges its closure. For server sockets, it is usually safe to
 374   // bind to a TIME_WAIT end point immediately, which is a widely adopted
 375   // behavior.
 376   //
 377   // Note that on *nix, SO_REUSEADDR does not enable the TCP socket to bind to
 378   // an end point that is already bound by another socket. To do that one must
 379   // set SO_REUSEPORT instead. This option is not provided on Linux prior
 380   // to 3.9.
 381   //
 382   // SO_REUSEPORT is provided in MacOS X and iOS.
 383   int boolean_value = allow ? 1 : 0;
 384   int rv = setsockopt(socket_->socket_fd(), SOL_SOCKET, SO_REUSEADDR,
 385                       &boolean_value, sizeof(boolean_value));
 386   if (rv < 0)
 387     return MapSystemError(errno);
 388   return OK;
 389 }
 390
 391 int TCPSocketLibevent::SetReceiveBufferSize(int32 size) {
 392   DCHECK(socket_);
 393   int rv = setsockopt(socket_->socket_fd(), SOL_SOCKET, SO_RCVBUF,
 394                       reinterpret_cast<const char*>(&size), sizeof(size));
 395   return (rv == 0) ? OK : MapSystemError(errno);
 396 }
 397
 398 int TCPSocketLibevent::SetSendBufferSize(int32 size) {
 399   DCHECK(socket_);
 400   int rv = setsockopt(socket_->socket_fd(), SOL_SOCKET, SO_SNDBUF,
 401                       reinterpret_cast<const char*>(&size), sizeof(size));
 402   return (rv == 0) ? OK : MapSystemError(errno);
 403 }
 404
 405 bool TCPSocketLibevent::SetKeepAlive(bool enable, int delay) {
 406   DCHECK(socket_);
 407   return SetTCPKeepAlive(socket_->socket_fd(), enable, delay);
 408 }
 409
 410 bool TCPSocketLibevent::SetNoDelay(bool no_delay) {
 411   DCHECK(socket_);
 412   return SetTCPNoDelay(socket_->socket_fd(), no_delay);
 413 }
 414
 415 void TCPSocketLibevent::Close() {
 416   socket_.reset();
 417
 418   // Record and reset TCP FastOpen state.
 419   if (tcp_fastopen_write_attempted_ ||
 420       tcp_fastopen_status_ == TCP_FASTOPEN_PREVIOUSLY_FAILED) {
 421     UMA_HISTOGRAM_ENUMERATION("Net.TcpFastOpenSocketConnection",
 422                               tcp_fastopen_status_, TCP_FASTOPEN_MAX_VALUE);
 423   }
 424   use_tcp_fastopen_ = false;
 425   tcp_fastopen_connected_ = false;
 426   tcp_fastopen_write_attempted_ = false;
 427   tcp_fastopen_status_ = TCP_FASTOPEN_STATUS_UNKNOWN;
 428 }
 429
 430 bool TCPSocketLibevent::UsingTCPFastOpen() const {
 431   return use_tcp_fastopen_;
 432 }
 433
 434 void TCPSocketLibevent::EnableTCPFastOpenIfSupported() {
 435   if (!IsTCPFastOpenSupported())
 436     return;
 437
 438   // Do not enable TCP FastOpen if it had previously failed.
 439   // This check conservatively avoids middleboxes that may blackhole
 440   // TCP FastOpen SYN+Data packets; on such a failure, subsequent sockets
 441   // should not use TCP FastOpen.
 442   if(!g_tcp_fastopen_has_failed)
 443     use_tcp_fastopen_ = true;
 444   else
 445     tcp_fastopen_status_ = TCP_FASTOPEN_PREVIOUSLY_FAILED;
 446 }
 447
 448 bool TCPSocketLibevent::IsValid() const {
 449   return socket_ != NULL && socket_->socket_fd() != kInvalidSocket;
 450 }
 451
 452 void TCPSocketLibevent::StartLoggingMultipleConnectAttempts(
 453     const AddressList& addresses) {
 454   if (!logging_multiple_connect_attempts_) {
 455     logging_multiple_connect_attempts_ = true;
 456     LogConnectBegin(addresses);
 457   } else {
 458     NOTREACHED();
 459   }
 460 }
 461
 462 void TCPSocketLibevent::EndLoggingMultipleConnectAttempts(int net_error) {
 463   if (logging_multiple_connect_attempts_) {
 464     LogConnectEnd(net_error);
 465     logging_multiple_connect_attempts_ = false;
 466   } else {
 467     NOTREACHED();
 468   }
 469 }
 470
 471 void TCPSocketLibevent::AcceptCompleted(
 472     scoped_ptr<TCPSocketLibevent>* tcp_socket,
 473     IPEndPoint* address,
 474     const CompletionCallback& callback,
 475     int rv) {
 476   DCHECK_NE(ERR_IO_PENDING, rv);
 477   callback.Run(HandleAcceptCompleted(tcp_socket, address, rv));
 478 }
 479
 480 int TCPSocketLibevent::HandleAcceptCompleted(
 481     scoped_ptr<TCPSocketLibevent>* tcp_socket,
 482     IPEndPoint* address,
 483     int rv) {
 484   if (rv == OK)
 485     rv = BuildTcpSocketLibevent(tcp_socket, address);
 486
 487   if (rv == OK) {
 488     net_log_.EndEvent(NetLog::TYPE_TCP_ACCEPT,
 489                       CreateNetLogIPEndPointCallback(address));
 490   } else {
 491     net_log_.EndEventWithNetErrorCode(NetLog::TYPE_TCP_ACCEPT, rv);
 492   }
 493
 494   return rv;
 495 }
 496
 497 int TCPSocketLibevent::BuildTcpSocketLibevent(
 498     scoped_ptr<TCPSocketLibevent>* tcp_socket,
 499     IPEndPoint* address) {
 500   DCHECK(accept_socket_);
 501
 502   SockaddrStorage storage;
 503   if (accept_socket_->GetPeerAddress(&storage) != OK ||
 504       !address->FromSockAddr(storage.addr, storage.addr_len)) {
 505     accept_socket_.reset();
 506     return ERR_ADDRESS_INVALID;
 507   }
 508
 509   tcp_socket->reset(new TCPSocketLibevent(net_log_.net_log(),
 510                                           net_log_.source()));
 511   (*tcp_socket)->socket_.reset(accept_socket_.release());
 512   return OK;
 513 }
 514
 515 void TCPSocketLibevent::ConnectCompleted(const CompletionCallback& callback,
 516                                          int rv) const {
 517   DCHECK_NE(ERR_IO_PENDING, rv);
 518   callback.Run(HandleConnectCompleted(rv));
 519 }
 520
 521 int TCPSocketLibevent::HandleConnectCompleted(int rv) const {
 522   // Log the end of this attempt (and any OS error it threw).
 523   if (rv != OK) {
 524     net_log_.EndEvent(NetLog::TYPE_TCP_CONNECT_ATTEMPT,
 525                       NetLog::IntegerCallback("os_error", errno));
 526   } else {
 527     net_log_.EndEvent(NetLog::TYPE_TCP_CONNECT_ATTEMPT);
 528   }
 529
 530   // Give a more specific error when the user is offline.
 531   if (rv == ERR_ADDRESS_UNREACHABLE && NetworkChangeNotifier::IsOffline())
 532     rv = ERR_INTERNET_DISCONNECTED;
 533
 534   if (!logging_multiple_connect_attempts_)
 535     LogConnectEnd(rv);
 536
 537   return rv;
 538 }
 539
 540 void TCPSocketLibevent::LogConnectBegin(const AddressList& addresses) const {
 541   base::StatsCounter connects("tcp.connect");
 542   connects.Increment();
 543
 544   net_log_.BeginEvent(NetLog::TYPE_TCP_CONNECT,
 545                       addresses.CreateNetLogCallback());
 546 }
 547
 548 void TCPSocketLibevent::LogConnectEnd(int net_error) const {
 549   if (net_error != OK) {
 550     net_log_.EndEventWithNetErrorCode(NetLog::TYPE_TCP_CONNECT, net_error);
 551     return;
 552   }
 553
 554   UpdateConnectionTypeHistograms(CONNECTION_ANY);
 555
 556   SockaddrStorage storage;
 557   int rv = socket_->GetLocalAddress(&storage);
 558   if (rv != OK) {
 559     PLOG(ERROR) << "GetLocalAddress() [rv: " << rv << "] error: ";
 560     NOTREACHED();
 561     net_log_.EndEventWithNetErrorCode(NetLog::TYPE_TCP_CONNECT, rv);
 562     return;
 563   }
 564
 565   net_log_.EndEvent(NetLog::TYPE_TCP_CONNECT,
 566                     CreateNetLogSourceAddressCallback(storage.addr,
 567                                                       storage.addr_len));
 568 }
 569
 570 void TCPSocketLibevent::ReadCompleted(const scoped_refptr<IOBuffer>& buf,
 571                                       const CompletionCallback& callback,
 572                                       int rv) {
 573   DCHECK_NE(ERR_IO_PENDING, rv);
 574   callback.Run(HandleReadCompleted(buf.get(), rv));
 575 }
 576
 577 int TCPSocketLibevent::HandleReadCompleted(IOBuffer* buf, int rv) {
 578   if (tcp_fastopen_write_attempted_ && !tcp_fastopen_connected_) {
 579     // A TCP FastOpen connect-with-write was attempted. This read was a
 580     // subsequent read, which either succeeded or failed. If the read
 581     // succeeded, the socket is considered connected via TCP FastOpen.
 582     // If the read failed, TCP FastOpen is (conservatively) turned off for all
 583     // subsequent connections. TCP FastOpen status is recorded in both cases.
 584     // TODO (jri): This currently results in conservative behavior, where TCP
 585     // FastOpen is turned off on _any_ error. Implement optimizations,
 586     // such as turning off TCP FastOpen on more specific errors, and
 587     // re-attempting TCP FastOpen after a certain amount of time has passed.
 588     if (rv >= 0)
 589       tcp_fastopen_connected_ = true;
 590     else
 591       g_tcp_fastopen_has_failed = true;
 592     UpdateTCPFastOpenStatusAfterRead();
 593   }
 594
 595   if (rv < 0) {
 596     net_log_.AddEvent(NetLog::TYPE_SOCKET_READ_ERROR,
 597                       CreateNetLogSocketErrorCallback(rv, errno));
 598     return rv;
 599   }
 600
 601   base::StatsCounter read_bytes("tcp.read_bytes");
 602   read_bytes.Add(rv);
 603   net_log_.AddByteTransferEvent(NetLog::TYPE_SOCKET_BYTES_RECEIVED, rv,
 604                                 buf->data());
 605   NetworkActivityMonitor::GetInstance()->IncrementBytesReceived(rv);
 606
 607   return rv;
 608 }
 609
 610 void TCPSocketLibevent::WriteCompleted(const scoped_refptr<IOBuffer>& buf,
 611                                        const CompletionCallback& callback,
 612                                        int rv) {
 613   DCHECK_NE(ERR_IO_PENDING, rv);
 614   callback.Run(HandleWriteCompleted(buf.get(), rv));
 615 }
 616
 617 int TCPSocketLibevent::HandleWriteCompleted(IOBuffer* buf, int rv) {
 618   if (rv < 0) {
 619     if (tcp_fastopen_write_attempted_ && !tcp_fastopen_connected_) {
 620       // TCP FastOpen connect-with-write was attempted, and the write failed
 621       // for unknown reasons. Record status and (conservatively) turn off
 622       // TCP FastOpen for all subsequent connections.
 623       // TODO (jri): This currently results in conservative behavior, where TCP
 624       // FastOpen is turned off on _any_ error. Implement optimizations,
 625       // such as turning off TCP FastOpen on more specific errors, and
 626       // re-attempting TCP FastOpen after a certain amount of time has passed.
 627       tcp_fastopen_status_ = TCP_FASTOPEN_ERROR;
 628       g_tcp_fastopen_has_failed = true;
 629     }
 630     net_log_.AddEvent(NetLog::TYPE_SOCKET_WRITE_ERROR,
 631                       CreateNetLogSocketErrorCallback(rv, errno));
 632     return rv;
 633   }
 634
 635   base::StatsCounter write_bytes("tcp.write_bytes");
 636   write_bytes.Add(rv);
 637   net_log_.AddByteTransferEvent(NetLog::TYPE_SOCKET_BYTES_SENT, rv,
 638                                 buf->data());
 639   NetworkActivityMonitor::GetInstance()->IncrementBytesSent(rv);
 640   return rv;
 641 }
 642
 643 int TCPSocketLibevent::TcpFastOpenWrite(
 644     IOBuffer* buf,
 645     int buf_len,
 646     const CompletionCallback& callback) {
 647   SockaddrStorage storage;
 648   int rv = socket_->GetPeerAddress(&storage);
 649   if (rv != OK)
 650     return rv;
 651
 652   int flags = 0x20000000;  // Magic flag to enable TCP_FASTOPEN.
 653 #if defined(OS_LINUX) || defined(OS_ANDROID)
 654   // sendto() will fail with EPIPE when the system doesn't implement TCP
 655   // FastOpen, and with EOPNOTSUPP when the system implements TCP FastOpen
 656   // but it is disabled. Theoretically these shouldn't happen
 657   // since the caller should check for system support on startup, but
 658   // users may dynamically disable TCP FastOpen via sysctl.
 659   flags |= MSG_NOSIGNAL;
 660 #endif // defined(OS_LINUX) || defined(OS_ANDROID)
 661   rv = HANDLE_EINTR(sendto(socket_->socket_fd(),
 662                            buf->data(),
 663                            buf_len,
 664                            flags,
 665                            storage.addr,
 666                            storage.addr_len));
 667   tcp_fastopen_write_attempted_ = true;
 668
 669   if (rv >= 0) {
 670     tcp_fastopen_status_ = TCP_FASTOPEN_FAST_CONNECT_RETURN;
 671     return rv;
 672   }
 673
 674   DCHECK_NE(EPIPE, errno);
 675
 676   // If errno == EINPROGRESS, that means the kernel didn't have a cookie
 677   // and would block. The kernel is internally doing a connect() though.
 678   // Remap EINPROGRESS to EAGAIN so we treat this the same as our other
 679   // asynchronous cases. Note that the user buffer has not been copied to
 680   // kernel space.
 681   if (errno == EINPROGRESS) {
 682     rv = ERR_IO_PENDING;
 683   } else {
 684     rv = MapSystemError(errno);
 685   }
 686
 687   if (rv != ERR_IO_PENDING) {
 688     // TCP FastOpen connect-with-write was attempted, and the write failed
 689     // since TCP FastOpen was not implemented or disabled in the OS.
 690     // Record status and turn off TCP FastOpen for all subsequent connections.
 691     // TODO (jri): This is almost certainly too conservative, since it blanket
 692     // turns off TCP FastOpen on any write error. Two things need to be done
 693     // here: (i) record a histogram of write errors; in particular, record
 694     // occurrences of EOPNOTSUPP and EPIPE, and (ii) afterwards, consider
 695     // turning off TCP FastOpen on more specific errors.
 696     tcp_fastopen_status_ = TCP_FASTOPEN_ERROR;
 697     g_tcp_fastopen_has_failed = true;
 698     return rv;
 699   }
 700
 701   tcp_fastopen_status_ = TCP_FASTOPEN_SLOW_CONNECT_RETURN;
 702   return socket_->WaitForWrite(buf, buf_len, callback);
 703 }
 704
 705 void TCPSocketLibevent::UpdateTCPFastOpenStatusAfterRead() {
 706   DCHECK(tcp_fastopen_status_ == TCP_FASTOPEN_FAST_CONNECT_RETURN ||
 707          tcp_fastopen_status_ == TCP_FASTOPEN_SLOW_CONNECT_RETURN);
 708
 709   if (tcp_fastopen_write_attempted_ && !tcp_fastopen_connected_) {
 710     // TCP FastOpen connect-with-write was attempted, and failed.
 711     tcp_fastopen_status_ =
 712         (tcp_fastopen_status_ == TCP_FASTOPEN_FAST_CONNECT_RETURN ?
 713             TCP_FASTOPEN_FAST_CONNECT_READ_FAILED :
 714             TCP_FASTOPEN_SLOW_CONNECT_READ_FAILED);
 715     return;
 716   }
 717
 718   bool getsockopt_success = false;
 719   bool server_acked_data = false;
 720 #if defined(TCP_INFO)
 721   // Probe to see the if the socket used TCP FastOpen.
 722   tcp_info info;
 723   socklen_t info_len = sizeof(tcp_info);
 724   getsockopt_success = getsockopt(socket_->socket_fd(), IPPROTO_TCP, TCP_INFO,
 725                                   &info, &info_len) == 0 &&
 726                        info_len == sizeof(tcp_info);
 727   server_acked_data = getsockopt_success &&
 728                       (info.tcpi_options & TCPI_OPT_SYN_DATA);
 729 #endif
 730
 731   if (getsockopt_success) {
 732     if (tcp_fastopen_status_ == TCP_FASTOPEN_FAST_CONNECT_RETURN) {
 733       tcp_fastopen_status_ = (server_acked_data ?
 734                               TCP_FASTOPEN_SYN_DATA_ACK :
 735                               TCP_FASTOPEN_SYN_DATA_NACK);
 736     } else {
 737       tcp_fastopen_status_ = (server_acked_data ?
 738                               TCP_FASTOPEN_NO_SYN_DATA_ACK :
 739                               TCP_FASTOPEN_NO_SYN_DATA_NACK);
 740     }
 741   } else {
 742     tcp_fastopen_status_ =
 743         (tcp_fastopen_status_ == TCP_FASTOPEN_FAST_CONNECT_RETURN ?
 744          TCP_FASTOPEN_SYN_DATA_GETSOCKOPT_FAILED :
 745          TCP_FASTOPEN_NO_SYN_DATA_GETSOCKOPT_FAILED);
 746   }
 747 }
 748
 749 }  // namespace net