net/socket/tcp_socket_win.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 #include "net/socket/tcp_socket_win.h"
   7
   8 #include <mstcpip.h>
   9
  10 #include "base/callback_helpers.h"
  11 #include "base/logging.h"
  12 #include "base/profiler/scoped_tracker.h"
  13 #include "base/win/windows_version.h"
  14 #include "net/base/address_list.h"
  15 #include "net/base/connection_type_histograms.h"
  16 #include "net/base/io_buffer.h"
  17 #include "net/base/ip_endpoint.h"
  18 #include "net/base/net_errors.h"
  19 #include "net/base/net_util.h"
  20 #include "net/base/network_activity_monitor.h"
  21 #include "net/base/network_change_notifier.h"
  22 #include "net/base/winsock_init.h"
  23 #include "net/base/winsock_util.h"
  24 #include "net/socket/socket_descriptor.h"
  25 #include "net/socket/socket_net_log_params.h"
  26
  27 namespace net {
  28
  29 namespace {
  30
  31 const int kTCPKeepAliveSeconds = 45;
  32
  33 int SetSocketReceiveBufferSize(SOCKET socket, int32 size) {
  34   int rv = setsockopt(socket, SOL_SOCKET, SO_RCVBUF,
  35                       reinterpret_cast<const char*>(&size), sizeof(size));
  36   int net_error = (rv == 0) ? OK : MapSystemError(WSAGetLastError());
  37   DCHECK(!rv) << "Could not set socket receive buffer size: " << net_error;
  38   return net_error;
  39 }
  40
  41 int SetSocketSendBufferSize(SOCKET socket, int32 size) {
  42   int rv = setsockopt(socket, SOL_SOCKET, SO_SNDBUF,
  43                       reinterpret_cast<const char*>(&size), sizeof(size));
  44   int net_error = (rv == 0) ? OK : MapSystemError(WSAGetLastError());
  45   DCHECK(!rv) << "Could not set socket send buffer size: " << net_error;
  46   return net_error;
  47 }
  48
  49 // Disable Nagle.
  50 // The Nagle implementation on windows is governed by RFC 896.  The idea
  51 // behind Nagle is to reduce small packets on the network.  When Nagle is
  52 // enabled, if a partial packet has been sent, the TCP stack will disallow
  53 // further *partial* packets until an ACK has been received from the other
  54 // side.  Good applications should always strive to send as much data as
  55 // possible and avoid partial-packet sends.  However, in most real world
  56 // applications, there are edge cases where this does not happen, and two
  57 // partial packets may be sent back to back.  For a browser, it is NEVER
  58 // a benefit to delay for an RTT before the second packet is sent.
  59 //
  60 // As a practical example in Chromium today, consider the case of a small
  61 // POST.  I have verified this:
  62 //     Client writes 649 bytes of header  (partial packet #1)
  63 //     Client writes 50 bytes of POST data (partial packet #2)
  64 // In the above example, with Nagle, a RTT delay is inserted between these
  65 // two sends due to nagle.  RTTs can easily be 100ms or more.  The best
  66 // fix is to make sure that for POSTing data, we write as much data as
  67 // possible and minimize partial packets.  We will fix that.  But disabling
  68 // Nagle also ensure we don't run into this delay in other edge cases.
  69 // See also:
  70 //    http://technet.microsoft.com/en-us/library/bb726981.aspx
  71 bool DisableNagle(SOCKET socket, bool disable) {
  72   BOOL val = disable ? TRUE : FALSE;
  73   int rv = setsockopt(socket, IPPROTO_TCP, TCP_NODELAY,
  74                       reinterpret_cast<const char*>(&val),
  75                       sizeof(val));
  76   DCHECK(!rv) << "Could not disable nagle";
  77   return rv == 0;
  78 }
  79
  80 // Enable TCP Keep-Alive to prevent NAT routers from timing out TCP
  81 // connections. See http://crbug.com/27400 for details.
  82 bool SetTCPKeepAlive(SOCKET socket, BOOL enable, int delay_secs) {
  83   unsigned delay = delay_secs * 1000;
  84   struct tcp_keepalive keepalive_vals = {
  85       enable ? 1u : 0u,  // TCP keep-alive on.
  86       delay,  // Delay seconds before sending first TCP keep-alive packet.
  87       delay,  // Delay seconds between sending TCP keep-alive packets.
  88   };
  89   DWORD bytes_returned = 0xABAB;
  90   int rv = WSAIoctl(socket, SIO_KEEPALIVE_VALS, &keepalive_vals,
  91                     sizeof(keepalive_vals), NULL, 0,
  92                     &bytes_returned, NULL, NULL);
  93   DCHECK(!rv) << "Could not enable TCP Keep-Alive for socket: " << socket
  94               << " [error: " << WSAGetLastError() << "].";
  95
  96   // Disregard any failure in disabling nagle or enabling TCP Keep-Alive.
  97   return rv == 0;
  98 }
  99
 100 int MapConnectError(int os_error) {
 101   switch (os_error) {
 102     // connect fails with WSAEACCES when Windows Firewall blocks the
 103     // connection.
 104     case WSAEACCES:
 105       return ERR_NETWORK_ACCESS_DENIED;
 106     case WSAETIMEDOUT:
 107       return ERR_CONNECTION_TIMED_OUT;
 108     default: {
 109       int net_error = MapSystemError(os_error);
 110       if (net_error == ERR_FAILED)
 111         return ERR_CONNECTION_FAILED;  // More specific than ERR_FAILED.
 112
 113       // Give a more specific error when the user is offline.
 114       if (net_error == ERR_ADDRESS_UNREACHABLE &&
 115           NetworkChangeNotifier::IsOffline()) {
 116         return ERR_INTERNET_DISCONNECTED;
 117       }
 118
 119       return net_error;
 120     }
 121   }
 122 }
 123
 124 }  // namespace
 125
 126 //-----------------------------------------------------------------------------
 127
 128 // Nothing to do for Windows since it doesn't support TCP FastOpen.
 129 // TODO(jri): Remove these along with the corresponding global variables.
 130 bool IsTCPFastOpenSupported() { return false; }
 131 bool IsTCPFastOpenUserEnabled() { return false; }
 132 void CheckSupportAndMaybeEnableTCPFastOpen(bool user_enabled) {}
 133
 134 // This class encapsulates all the state that has to be preserved as long as
 135 // there is a network IO operation in progress. If the owner TCPSocketWin is
 136 // destroyed while an operation is in progress, the Core is detached and it
 137 // lives until the operation completes and the OS doesn't reference any resource
 138 // declared on this class anymore.
 139 class TCPSocketWin::Core : public base::RefCounted<Core> {
 140  public:
 141   explicit Core(TCPSocketWin* socket);
 142
 143   // Start watching for the end of a read or write operation.
 144   void WatchForRead();
 145   void WatchForWrite();
 146
 147   // The TCPSocketWin is going away.
 148   void Detach() { socket_ = NULL; }
 149
 150   // The separate OVERLAPPED variables for asynchronous operation.
 151   // |read_overlapped_| is used for both Connect() and Read().
 152   // |write_overlapped_| is only used for Write();
 153   OVERLAPPED read_overlapped_;
 154   OVERLAPPED write_overlapped_;
 155
 156   // The buffers used in Read() and Write().
 157   scoped_refptr<IOBuffer> read_iobuffer_;
 158   scoped_refptr<IOBuffer> write_iobuffer_;
 159   int read_buffer_length_;
 160   int write_buffer_length_;
 161
 162   bool non_blocking_reads_initialized_;
 163
 164  private:
 165   friend class base::RefCounted<Core>;
 166
 167   class ReadDelegate : public base::win::ObjectWatcher::Delegate {
 168    public:
 169     explicit ReadDelegate(Core* core) : core_(core) {}
 170     ~ReadDelegate() override {}
 171
 172     // base::ObjectWatcher::Delegate methods:
 173     void OnObjectSignaled(HANDLE object) override;
 174
 175    private:
 176     Core* const core_;
 177   };
 178
 179   class WriteDelegate : public base::win::ObjectWatcher::Delegate {
 180    public:
 181     explicit WriteDelegate(Core* core) : core_(core) {}
 182     ~WriteDelegate() override {}
 183
 184     // base::ObjectWatcher::Delegate methods:
 185     void OnObjectSignaled(HANDLE object) override;
 186
 187    private:
 188     Core* const core_;
 189   };
 190
 191   ~Core();
 192
 193   // The socket that created this object.
 194   TCPSocketWin* socket_;
 195
 196   // |reader_| handles the signals from |read_watcher_|.
 197   ReadDelegate reader_;
 198   // |writer_| handles the signals from |write_watcher_|.
 199   WriteDelegate writer_;
 200
 201   // |read_watcher_| watches for events from Connect() and Read().
 202   base::win::ObjectWatcher read_watcher_;
 203   // |write_watcher_| watches for events from Write();
 204   base::win::ObjectWatcher write_watcher_;
 205
 206   DISALLOW_COPY_AND_ASSIGN(Core);
 207 };
 208
 209 TCPSocketWin::Core::Core(TCPSocketWin* socket)
 210     : read_buffer_length_(0),
 211       write_buffer_length_(0),
 212       non_blocking_reads_initialized_(false),
 213       socket_(socket),
 214       reader_(this),
 215       writer_(this) {
 216   memset(&read_overlapped_, 0, sizeof(read_overlapped_));
 217   memset(&write_overlapped_, 0, sizeof(write_overlapped_));
 218
 219   read_overlapped_.hEvent = WSACreateEvent();
 220   write_overlapped_.hEvent = WSACreateEvent();
 221 }
 222
 223 TCPSocketWin::Core::~Core() {
 224   // Make sure the message loop is not watching this object anymore.
 225   read_watcher_.StopWatching();
 226   write_watcher_.StopWatching();
 227
 228   WSACloseEvent(read_overlapped_.hEvent);
 229   memset(&read_overlapped_, 0xaf, sizeof(read_overlapped_));
 230   WSACloseEvent(write_overlapped_.hEvent);
 231   memset(&write_overlapped_, 0xaf, sizeof(write_overlapped_));
 232 }
 233
 234 void TCPSocketWin::Core::WatchForRead() {
 235   // We grab an extra reference because there is an IO operation in progress.
 236   // Balanced in ReadDelegate::OnObjectSignaled().
 237   AddRef();
 238   read_watcher_.StartWatching(read_overlapped_.hEvent, &reader_);
 239 }
 240
 241 void TCPSocketWin::Core::WatchForWrite() {
 242   // We grab an extra reference because there is an IO operation in progress.
 243   // Balanced in WriteDelegate::OnObjectSignaled().
 244   AddRef();
 245   write_watcher_.StartWatching(write_overlapped_.hEvent, &writer_);
 246 }
 247
 248 void TCPSocketWin::Core::ReadDelegate::OnObjectSignaled(HANDLE object) {
 249   DCHECK_EQ(object, core_->read_overlapped_.hEvent);
 250   if (core_->socket_) {
 251     if (core_->socket_->waiting_connect_)
 252       core_->socket_->DidCompleteConnect();
 253     else
 254       core_->socket_->DidSignalRead();
 255   }
 256
 257   core_->Release();
 258 }
 259
 260 void TCPSocketWin::Core::WriteDelegate::OnObjectSignaled(
 261     HANDLE object) {
 262   DCHECK_EQ(object, core_->write_overlapped_.hEvent);
 263   if (core_->socket_)
 264     core_->socket_->DidCompleteWrite();
 265
 266   core_->Release();
 267 }
 268
 269 //-----------------------------------------------------------------------------
 270
 271 TCPSocketWin::TCPSocketWin(net::NetLog* net_log,
 272                            const net::NetLog::Source& source)
 273     : socket_(INVALID_SOCKET),
 274       accept_event_(WSA_INVALID_EVENT),
 275       accept_socket_(NULL),
 276       accept_address_(NULL),
 277       waiting_connect_(false),
 278       waiting_read_(false),
 279       waiting_write_(false),
 280       connect_os_error_(0),
 281       logging_multiple_connect_attempts_(false),
 282       net_log_(BoundNetLog::Make(net_log, NetLog::SOURCE_SOCKET)) {
 283   net_log_.BeginEvent(NetLog::TYPE_SOCKET_ALIVE,
 284                       source.ToEventParametersCallback());
 285   EnsureWinsockInit();
 286 }
 287
 288 TCPSocketWin::~TCPSocketWin() {
 289   Close();
 290   net_log_.EndEvent(NetLog::TYPE_SOCKET_ALIVE);
 291 }
 292
 293 int TCPSocketWin::Open(AddressFamily family) {
 294   DCHECK(CalledOnValidThread());
 295   DCHECK_EQ(socket_, INVALID_SOCKET);
 296
 297   socket_ = CreatePlatformSocket(ConvertAddressFamily(family), SOCK_STREAM,
 298                                  IPPROTO_TCP);
 299   if (socket_ == INVALID_SOCKET) {
 300     PLOG(ERROR) << "CreatePlatformSocket() returned an error";
 301     return MapSystemError(WSAGetLastError());
 302   }
 303
 304   if (SetNonBlocking(socket_)) {
 305     int result = MapSystemError(WSAGetLastError());
 306     Close();
 307     return result;
 308   }
 309
 310   return OK;
 311 }
 312
 313 int TCPSocketWin::AdoptConnectedSocket(SOCKET socket,
 314                                        const IPEndPoint& peer_address) {
 315   DCHECK(CalledOnValidThread());
 316   DCHECK_EQ(socket_, INVALID_SOCKET);
 317   DCHECK(!core_.get());
 318
 319   socket_ = socket;
 320
 321   if (SetNonBlocking(socket_)) {
 322     int result = MapSystemError(WSAGetLastError());
 323     Close();
 324     return result;
 325   }
 326
 327   core_ = new Core(this);
 328   peer_address_.reset(new IPEndPoint(peer_address));
 329
 330   return OK;
 331 }
 332
 333 int TCPSocketWin::AdoptListenSocket(SOCKET socket) {
 334   DCHECK(CalledOnValidThread());
 335   DCHECK_EQ(socket_, INVALID_SOCKET);
 336
 337   socket_ = socket;
 338
 339   if (SetNonBlocking(socket_)) {
 340     int result = MapSystemError(WSAGetLastError());
 341     Close();
 342     return result;
 343   }
 344
 345   // |core_| is not needed for sockets that are used to accept connections.
 346   // The operation here is more like Open but with an existing socket.
 347
 348   return OK;
 349 }
 350
 351 int TCPSocketWin::Bind(const IPEndPoint& address) {
 352   DCHECK(CalledOnValidThread());
 353   DCHECK_NE(socket_, INVALID_SOCKET);
 354
 355   SockaddrStorage storage;
 356   if (!address.ToSockAddr(storage.addr, &storage.addr_len))
 357     return ERR_ADDRESS_INVALID;
 358
 359   int result = bind(socket_, storage.addr, storage.addr_len);
 360   if (result < 0) {
 361     PLOG(ERROR) << "bind() returned an error";
 362     return MapSystemError(WSAGetLastError());
 363   }
 364
 365   return OK;
 366 }
 367
 368 int TCPSocketWin::Listen(int backlog) {
 369   DCHECK(CalledOnValidThread());
 370   DCHECK_GT(backlog, 0);
 371   DCHECK_NE(socket_, INVALID_SOCKET);
 372   DCHECK_EQ(accept_event_, WSA_INVALID_EVENT);
 373
 374   accept_event_ = WSACreateEvent();
 375   if (accept_event_ == WSA_INVALID_EVENT) {
 376     PLOG(ERROR) << "WSACreateEvent()";
 377     return MapSystemError(WSAGetLastError());
 378   }
 379
 380   int result = listen(socket_, backlog);
 381   if (result < 0) {
 382     PLOG(ERROR) << "listen() returned an error";
 383     return MapSystemError(WSAGetLastError());
 384   }
 385
 386   return OK;
 387 }
 388
 389 int TCPSocketWin::Accept(scoped_ptr<TCPSocketWin>* socket,
 390                          IPEndPoint* address,
 391                          const CompletionCallback& callback) {
 392   DCHECK(CalledOnValidThread());
 393   DCHECK(socket);
 394   DCHECK(address);
 395   DCHECK(!callback.is_null());
 396   DCHECK(accept_callback_.is_null());
 397
 398   net_log_.BeginEvent(NetLog::TYPE_TCP_ACCEPT);
 399
 400   int result = AcceptInternal(socket, address);
 401
 402   if (result == ERR_IO_PENDING) {
 403     // Start watching.
 404     WSAEventSelect(socket_, accept_event_, FD_ACCEPT);
 405     accept_watcher_.StartWatching(accept_event_, this);
 406
 407     accept_socket_ = socket;
 408     accept_address_ = address;
 409     accept_callback_ = callback;
 410   }
 411
 412   return result;
 413 }
 414
 415 int TCPSocketWin::Connect(const IPEndPoint& address,
 416                           const CompletionCallback& callback) {
 417   DCHECK(CalledOnValidThread());
 418   DCHECK_NE(socket_, INVALID_SOCKET);
 419   DCHECK(!waiting_connect_);
 420
 421   // |peer_address_| and |core_| will be non-NULL if Connect() has been called.
 422   // Unless Close() is called to reset the internal state, a second call to
 423   // Connect() is not allowed.
 424   // Please note that we enforce this even if the previous Connect() has
 425   // completed and failed. Although it is allowed to connect the same |socket_|
 426   // again after a connection attempt failed on Windows, it results in
 427   // unspecified behavior according to POSIX. Therefore, we make it behave in
 428   // the same way as TCPSocketPosix.
 429   DCHECK(!peer_address_ && !core_.get());
 430
 431   if (!logging_multiple_connect_attempts_)
 432     LogConnectBegin(AddressList(address));
 433
 434   peer_address_.reset(new IPEndPoint(address));
 435
 436   int rv = DoConnect();
 437   if (rv == ERR_IO_PENDING) {
 438     // Synchronous operation not supported.
 439     DCHECK(!callback.is_null());
 440     read_callback_ = callback;
 441     waiting_connect_ = true;
 442   } else {
 443     DoConnectComplete(rv);
 444   }
 445
 446   return rv;
 447 }
 448
 449 bool TCPSocketWin::IsConnected() const {
 450   DCHECK(CalledOnValidThread());
 451
 452   if (socket_ == INVALID_SOCKET || waiting_connect_)
 453     return false;
 454
 455   if (waiting_read_)
 456     return true;
 457
 458   // Check if connection is alive.
 459   char c;
 460   int rv = recv(socket_, &c, 1, MSG_PEEK);
 461   if (rv == 0)
 462     return false;
 463   if (rv == SOCKET_ERROR && WSAGetLastError() != WSAEWOULDBLOCK)
 464     return false;
 465
 466   return true;
 467 }
 468
 469 bool TCPSocketWin::IsConnectedAndIdle() const {
 470   DCHECK(CalledOnValidThread());
 471
 472   if (socket_ == INVALID_SOCKET || waiting_connect_)
 473     return false;
 474
 475   if (waiting_read_)
 476     return true;
 477
 478   // Check if connection is alive and we haven't received any data
 479   // unexpectedly.
 480   char c;
 481   int rv = recv(socket_, &c, 1, MSG_PEEK);
 482   if (rv >= 0)
 483     return false;
 484   if (WSAGetLastError() != WSAEWOULDBLOCK)
 485     return false;
 486
 487   return true;
 488 }
 489
 490 int TCPSocketWin::Read(IOBuffer* buf,
 491                        int buf_len,
 492                        const CompletionCallback& callback) {
 493   DCHECK(CalledOnValidThread());
 494   DCHECK_NE(socket_, INVALID_SOCKET);
 495   DCHECK(!waiting_read_);
 496   CHECK(read_callback_.is_null());
 497   DCHECK(!core_->read_iobuffer_.get());
 498
 499   return DoRead(buf, buf_len, callback);
 500 }
 501
 502 int TCPSocketWin::Write(IOBuffer* buf,
 503                         int buf_len,
 504                         const CompletionCallback& callback) {
 505   DCHECK(CalledOnValidThread());
 506   DCHECK_NE(socket_, INVALID_SOCKET);
 507   DCHECK(!waiting_write_);
 508   CHECK(write_callback_.is_null());
 509   DCHECK_GT(buf_len, 0);
 510   DCHECK(!core_->write_iobuffer_.get());
 511
 512   WSABUF write_buffer;
 513   write_buffer.len = buf_len;
 514   write_buffer.buf = buf->data();
 515
 516   // TODO(wtc): Remove the assertion after enough testing.
 517   AssertEventNotSignaled(core_->write_overlapped_.hEvent);
 518   DWORD num;
 519   int rv = WSASend(socket_, &write_buffer, 1, &num, 0,
 520                    &core_->write_overlapped_, NULL);
 521   if (rv == 0) {
 522     if (ResetEventIfSignaled(core_->write_overlapped_.hEvent)) {
 523       rv = static_cast<int>(num);
 524       if (rv > buf_len || rv < 0) {
 525         // It seems that some winsock interceptors report that more was written
 526         // than was available. Treat this as an error.  http://crbug.com/27870
 527         LOG(ERROR) << "Detected broken LSP: Asked to write " << buf_len
 528                    << " bytes, but " << rv << " bytes reported.";
 529         return ERR_WINSOCK_UNEXPECTED_WRITTEN_BYTES;
 530       }
 531       net_log_.AddByteTransferEvent(NetLog::TYPE_SOCKET_BYTES_SENT, rv,
 532                                     buf->data());
 533       NetworkActivityMonitor::GetInstance()->IncrementBytesSent(rv);
 534       return rv;
 535     }
 536   } else {
 537     int os_error = WSAGetLastError();
 538     if (os_error != WSA_IO_PENDING) {
 539       int net_error = MapSystemError(os_error);
 540       net_log_.AddEvent(NetLog::TYPE_SOCKET_WRITE_ERROR,
 541                         CreateNetLogSocketErrorCallback(net_error, os_error));
 542       return net_error;
 543     }
 544   }
 545   waiting_write_ = true;
 546   write_callback_ = callback;
 547   core_->write_iobuffer_ = buf;
 548   core_->write_buffer_length_ = buf_len;
 549   core_->WatchForWrite();
 550   return ERR_IO_PENDING;
 551 }
 552
 553 int TCPSocketWin::GetLocalAddress(IPEndPoint* address) const {
 554   DCHECK(CalledOnValidThread());
 555   DCHECK(address);
 556
 557   SockaddrStorage storage;
 558   if (getsockname(socket_, storage.addr, &storage.addr_len))
 559     return MapSystemError(WSAGetLastError());
 560   if (!address->FromSockAddr(storage.addr, storage.addr_len))
 561     return ERR_ADDRESS_INVALID;
 562
 563   return OK;
 564 }
 565
 566 int TCPSocketWin::GetPeerAddress(IPEndPoint* address) const {
 567   DCHECK(CalledOnValidThread());
 568   DCHECK(address);
 569   if (!IsConnected())
 570     return ERR_SOCKET_NOT_CONNECTED;
 571   *address = *peer_address_;
 572   return OK;
 573 }
 574
 575 int TCPSocketWin::SetDefaultOptionsForServer() {
 576   return SetExclusiveAddrUse();
 577 }
 578
 579 void TCPSocketWin::SetDefaultOptionsForClient() {
 580   // Increase the socket buffer sizes from the default sizes for WinXP.  In
 581   // performance testing, there is substantial benefit by increasing from 8KB
 582   // to 64KB.
 583   // See also:
 584   //    http://support.microsoft.com/kb/823764/EN-US
 585   // On Vista, if we manually set these sizes, Vista turns off its receive
 586   // window auto-tuning feature.
 587   //    http://blogs.msdn.com/wndp/archive/2006/05/05/Winhec-blog-tcpip-2.aspx
 588   // Since Vista's auto-tune is better than any static value we can could set,
 589   // only change these on pre-vista machines.
 590   if (base::win::GetVersion() < base::win::VERSION_VISTA) {
 591     const int32 kSocketBufferSize = 64 * 1024;
 592     SetSocketReceiveBufferSize(socket_, kSocketBufferSize);
 593     SetSocketSendBufferSize(socket_, kSocketBufferSize);
 594   }
 595
 596   DisableNagle(socket_, true);
 597   SetTCPKeepAlive(socket_, true, kTCPKeepAliveSeconds);
 598 }
 599
 600 int TCPSocketWin::SetExclusiveAddrUse() {
 601   // On Windows, a bound end point can be hijacked by another process by
 602   // setting SO_REUSEADDR. Therefore a Windows-only option SO_EXCLUSIVEADDRUSE
 603   // was introduced in Windows NT 4.0 SP4. If the socket that is bound to the
 604   // end point has SO_EXCLUSIVEADDRUSE enabled, it is not possible for another
 605   // socket to forcibly bind to the end point until the end point is unbound.
 606   // It is recommend that all server applications must use SO_EXCLUSIVEADDRUSE.
 607   // MSDN: http://goo.gl/M6fjQ.
 608   //
 609   // Unlike on *nix, on Windows a TCP server socket can always bind to an end
 610   // point in TIME_WAIT state without setting SO_REUSEADDR, therefore it is not
 611   // needed here.
 612   //
 613   // SO_EXCLUSIVEADDRUSE will prevent a TCP client socket from binding to an end
 614   // point in TIME_WAIT status. It does not have this effect for a TCP server
 615   // socket.
 616
 617   BOOL true_value = 1;
 618   int rv = setsockopt(socket_, SOL_SOCKET, SO_EXCLUSIVEADDRUSE,
 619                       reinterpret_cast<const char*>(&true_value),
 620                       sizeof(true_value));
 621   if (rv < 0)
 622     return MapSystemError(errno);
 623   return OK;
 624 }
 625
 626 int TCPSocketWin::SetReceiveBufferSize(int32 size) {
 627   DCHECK(CalledOnValidThread());
 628   return SetSocketReceiveBufferSize(socket_, size);
 629 }
 630
 631 int TCPSocketWin::SetSendBufferSize(int32 size) {
 632   DCHECK(CalledOnValidThread());
 633   return SetSocketSendBufferSize(socket_, size);
 634 }
 635
 636 bool TCPSocketWin::SetKeepAlive(bool enable, int delay) {
 637   return SetTCPKeepAlive(socket_, enable, delay);
 638 }
 639
 640 bool TCPSocketWin::SetNoDelay(bool no_delay) {
 641   return DisableNagle(socket_, no_delay);
 642 }
 643
 644 void TCPSocketWin::Close() {
 645   DCHECK(CalledOnValidThread());
 646
 647   if (socket_ != INVALID_SOCKET) {
 648     // Only log the close event if there's actually a socket to close.
 649     net_log_.AddEvent(NetLog::EventType::TYPE_SOCKET_CLOSED);
 650
 651     // Note: don't use CancelIo to cancel pending IO because it doesn't work
 652     // when there is a Winsock layered service provider.
 653
 654     // In most socket implementations, closing a socket results in a graceful
 655     // connection shutdown, but in Winsock we have to call shutdown explicitly.
 656     // See the MSDN page "Graceful Shutdown, Linger Options, and Socket Closure"
 657     // at http://msdn.microsoft.com/en-us/library/ms738547.aspx
 658     shutdown(socket_, SD_SEND);
 659
 660     // This cancels any pending IO.
 661     if (closesocket(socket_) < 0)
 662       PLOG(ERROR) << "closesocket";
 663     socket_ = INVALID_SOCKET;
 664   }
 665
 666   if (!accept_callback_.is_null()) {
 667     accept_watcher_.StopWatching();
 668     accept_socket_ = NULL;
 669     accept_address_ = NULL;
 670     accept_callback_.Reset();
 671   }
 672
 673   if (accept_event_) {
 674     WSACloseEvent(accept_event_);
 675     accept_event_ = WSA_INVALID_EVENT;
 676   }
 677
 678   if (core_.get()) {
 679     if (waiting_connect_) {
 680       // We closed the socket, so this notification will never come.
 681       // From MSDN' WSAEventSelect documentation:
 682       // "Closing a socket with closesocket also cancels the association and
 683       // selection of network events specified in WSAEventSelect for the
 684       // socket".
 685       core_->Release();
 686     }
 687     core_->Detach();
 688     core_ = NULL;
 689   }
 690
 691   waiting_connect_ = false;
 692   waiting_read_ = false;
 693   waiting_write_ = false;
 694
 695   read_callback_.Reset();
 696   write_callback_.Reset();
 697   peer_address_.reset();
 698   connect_os_error_ = 0;
 699 }
 700
 701 void TCPSocketWin::StartLoggingMultipleConnectAttempts(
 702     const AddressList& addresses) {
 703   if (!logging_multiple_connect_attempts_) {
 704     logging_multiple_connect_attempts_ = true;
 705     LogConnectBegin(addresses);
 706   } else {
 707     NOTREACHED();
 708   }
 709 }
 710
 711 void TCPSocketWin::EndLoggingMultipleConnectAttempts(int net_error) {
 712   if (logging_multiple_connect_attempts_) {
 713     LogConnectEnd(net_error);
 714     logging_multiple_connect_attempts_ = false;
 715   } else {
 716     NOTREACHED();
 717   }
 718 }
 719
 720 int TCPSocketWin::AcceptInternal(scoped_ptr<TCPSocketWin>* socket,
 721                                  IPEndPoint* address) {
 722   SockaddrStorage storage;
 723   int new_socket = accept(socket_, storage.addr, &storage.addr_len);
 724   if (new_socket < 0) {
 725     int net_error = MapSystemError(WSAGetLastError());
 726     if (net_error != ERR_IO_PENDING)
 727       net_log_.EndEventWithNetErrorCode(NetLog::TYPE_TCP_ACCEPT, net_error);
 728     return net_error;
 729   }
 730
 731   IPEndPoint ip_end_point;
 732   if (!ip_end_point.FromSockAddr(storage.addr, storage.addr_len)) {
 733     NOTREACHED();
 734     if (closesocket(new_socket) < 0)
 735       PLOG(ERROR) << "closesocket";
 736     int net_error = ERR_ADDRESS_INVALID;
 737     net_log_.EndEventWithNetErrorCode(NetLog::TYPE_TCP_ACCEPT, net_error);
 738     return net_error;
 739   }
 740   scoped_ptr<TCPSocketWin> tcp_socket(new TCPSocketWin(
 741       net_log_.net_log(), net_log_.source()));
 742   int adopt_result = tcp_socket->AdoptConnectedSocket(new_socket, ip_end_point);
 743   if (adopt_result != OK) {
 744     net_log_.EndEventWithNetErrorCode(NetLog::TYPE_TCP_ACCEPT, adopt_result);
 745     return adopt_result;
 746   }
 747   *socket = tcp_socket.Pass();
 748   *address = ip_end_point;
 749   net_log_.EndEvent(NetLog::TYPE_TCP_ACCEPT,
 750                     CreateNetLogIPEndPointCallback(&ip_end_point));
 751   return OK;
 752 }
 753
 754 void TCPSocketWin::OnObjectSignaled(HANDLE object) {
 755   WSANETWORKEVENTS ev;
 756   if (WSAEnumNetworkEvents(socket_, accept_event_, &ev) == SOCKET_ERROR) {
 757     PLOG(ERROR) << "WSAEnumNetworkEvents()";
 758     return;
 759   }
 760
 761   if (ev.lNetworkEvents & FD_ACCEPT) {
 762     int result = AcceptInternal(accept_socket_, accept_address_);
 763     if (result != ERR_IO_PENDING) {
 764       accept_socket_ = NULL;
 765       accept_address_ = NULL;
 766       base::ResetAndReturn(&accept_callback_).Run(result);
 767     }
 768   } else {
 769     // This happens when a client opens a connection and closes it before we
 770     // have a chance to accept it.
 771     DCHECK(ev.lNetworkEvents == 0);
 772
 773     // Start watching the next FD_ACCEPT event.
 774     WSAEventSelect(socket_, accept_event_, FD_ACCEPT);
 775     accept_watcher_.StartWatching(accept_event_, this);
 776   }
 777 }
 778
 779 int TCPSocketWin::DoConnect() {
 780   DCHECK_EQ(connect_os_error_, 0);
 781   DCHECK(!core_.get());
 782
 783   net_log_.BeginEvent(NetLog::TYPE_TCP_CONNECT_ATTEMPT,
 784                       CreateNetLogIPEndPointCallback(peer_address_.get()));
 785
 786   core_ = new Core(this);
 787
 788   // WSAEventSelect sets the socket to non-blocking mode as a side effect.
 789   // Our connect() and recv() calls require that the socket be non-blocking.
 790   WSAEventSelect(socket_, core_->read_overlapped_.hEvent, FD_CONNECT);
 791
 792   SockaddrStorage storage;
 793   if (!peer_address_->ToSockAddr(storage.addr, &storage.addr_len))
 794     return ERR_ADDRESS_INVALID;
 795
 796   int result;
 797   {
 798     // TODO(ricea): Remove ScopedTracker below once crbug.com/436634 is fixed.
 799     tracked_objects::ScopedTracker tracking_profile(
 800         FROM_HERE_WITH_EXPLICIT_FUNCTION("436634 connect()"));
 801     result = connect(socket_, storage.addr, storage.addr_len);
 802   }
 803
 804   if (!result) {
 805     // Connected without waiting!
 806     //
 807     // The MSDN page for connect says:
 808     //   With a nonblocking socket, the connection attempt cannot be completed
 809     //   immediately. In this case, connect will return SOCKET_ERROR, and
 810     //   WSAGetLastError will return WSAEWOULDBLOCK.
 811     // which implies that for a nonblocking socket, connect never returns 0.
 812     // It's not documented whether the event object will be signaled or not
 813     // if connect does return 0.  So the code below is essentially dead code
 814     // and we don't know if it's correct.
 815     NOTREACHED();
 816
 817     if (ResetEventIfSignaled(core_->read_overlapped_.hEvent))
 818       return OK;
 819   } else {
 820     int os_error = WSAGetLastError();
 821     if (os_error != WSAEWOULDBLOCK) {
 822       LOG(ERROR) << "connect failed: " << os_error;
 823       connect_os_error_ = os_error;
 824       int rv = MapConnectError(os_error);
 825       CHECK_NE(ERR_IO_PENDING, rv);
 826       return rv;
 827     }
 828   }
 829
 830   // TODO(ricea): Remove ScopedTracker below once crbug.com/436634 is fixed.
 831   tracked_objects::ScopedTracker tracking_profile(
 832       FROM_HERE_WITH_EXPLICIT_FUNCTION("436634 WatchForRead()"));
 833
 834   core_->WatchForRead();
 835   return ERR_IO_PENDING;
 836 }
 837
 838 void TCPSocketWin::DoConnectComplete(int result) {
 839   // Log the end of this attempt (and any OS error it threw).
 840   int os_error = connect_os_error_;
 841   connect_os_error_ = 0;
 842   if (result != OK) {
 843     net_log_.EndEvent(NetLog::TYPE_TCP_CONNECT_ATTEMPT,
 844                       NetLog::IntegerCallback("os_error", os_error));
 845   } else {
 846     net_log_.EndEvent(NetLog::TYPE_TCP_CONNECT_ATTEMPT);
 847   }
 848
 849   if (!logging_multiple_connect_attempts_)
 850     LogConnectEnd(result);
 851 }
 852
 853 void TCPSocketWin::LogConnectBegin(const AddressList& addresses) {
 854   net_log_.BeginEvent(NetLog::TYPE_TCP_CONNECT,
 855                       addresses.CreateNetLogCallback());
 856 }
 857
 858 void TCPSocketWin::LogConnectEnd(int net_error) {
 859   if (net_error == OK)
 860     UpdateConnectionTypeHistograms(CONNECTION_ANY);
 861
 862   if (net_error != OK) {
 863     net_log_.EndEventWithNetErrorCode(NetLog::TYPE_TCP_CONNECT, net_error);
 864     return;
 865   }
 866
 867   struct sockaddr_storage source_address;
 868   socklen_t addrlen = sizeof(source_address);
 869   int rv = getsockname(
 870       socket_, reinterpret_cast<struct sockaddr*>(&source_address), &addrlen);
 871   if (rv != 0) {
 872     LOG(ERROR) << "getsockname() [rv: " << rv
 873                << "] error: " << WSAGetLastError();
 874     NOTREACHED();
 875     net_log_.EndEventWithNetErrorCode(NetLog::TYPE_TCP_CONNECT, rv);
 876     return;
 877   }
 878
 879   net_log_.EndEvent(
 880       NetLog::TYPE_TCP_CONNECT,
 881       CreateNetLogSourceAddressCallback(
 882           reinterpret_cast<const struct sockaddr*>(&source_address),
 883           sizeof(source_address)));
 884 }
 885
 886 int TCPSocketWin::DoRead(IOBuffer* buf, int buf_len,
 887                          const CompletionCallback& callback) {
 888   if (!core_->non_blocking_reads_initialized_) {
 889     WSAEventSelect(socket_, core_->read_overlapped_.hEvent,
 890                    FD_READ | FD_CLOSE);
 891     core_->non_blocking_reads_initialized_ = true;
 892   }
 893   int rv = recv(socket_, buf->data(), buf_len, 0);
 894   if (rv == SOCKET_ERROR) {
 895     int os_error = WSAGetLastError();
 896     if (os_error != WSAEWOULDBLOCK) {
 897       int net_error = MapSystemError(os_error);
 898       net_log_.AddEvent(
 899           NetLog::TYPE_SOCKET_READ_ERROR,
 900           CreateNetLogSocketErrorCallback(net_error, os_error));
 901       return net_error;
 902     }
 903   } else {
 904     net_log_.AddByteTransferEvent(NetLog::TYPE_SOCKET_BYTES_RECEIVED, rv,
 905                                   buf->data());
 906     NetworkActivityMonitor::GetInstance()->IncrementBytesReceived(rv);
 907     return rv;
 908   }
 909
 910   waiting_read_ = true;
 911   read_callback_ = callback;
 912   core_->read_iobuffer_ = buf;
 913   core_->read_buffer_length_ = buf_len;
 914   core_->WatchForRead();
 915   return ERR_IO_PENDING;
 916 }
 917
 918 void TCPSocketWin::DidCompleteConnect() {
 919   DCHECK(waiting_connect_);
 920   DCHECK(!read_callback_.is_null());
 921   int result;
 922
 923   WSANETWORKEVENTS events;
 924   int rv;
 925   {
 926     // TODO(pkasting): Remove ScopedTracker below once crbug.com/462784 is
 927     // fixed.
 928     tracked_objects::ScopedTracker tracking_profile1(
 929         FROM_HERE_WITH_EXPLICIT_FUNCTION(
 930             "462784 TCPSocketWin::DidCompleteConnect -> WSAEnumNetworkEvents"));
 931     rv = WSAEnumNetworkEvents(socket_, core_->read_overlapped_.hEvent, &events);
 932   }
 933   int os_error = 0;
 934   if (rv == SOCKET_ERROR) {
 935     NOTREACHED();
 936     os_error = WSAGetLastError();
 937     result = MapSystemError(os_error);
 938   } else if (events.lNetworkEvents & FD_CONNECT) {
 939     os_error = events.iErrorCode[FD_CONNECT_BIT];
 940     result = MapConnectError(os_error);
 941   } else {
 942     NOTREACHED();
 943     result = ERR_UNEXPECTED;
 944   }
 945
 946   connect_os_error_ = os_error;
 947   DoConnectComplete(result);
 948   waiting_connect_ = false;
 949
 950   // TODO(pkasting): Remove ScopedTracker below once crbug.com/462784 is fixed.
 951   tracked_objects::ScopedTracker tracking_profile4(
 952       FROM_HERE_WITH_EXPLICIT_FUNCTION(
 953           "462784 TCPSocketWin::DidCompleteConnect -> read_callback_"));
 954   DCHECK_NE(result, ERR_IO_PENDING);
 955   base::ResetAndReturn(&read_callback_).Run(result);
 956 }
 957
 958 void TCPSocketWin::DidCompleteWrite() {
 959   DCHECK(waiting_write_);
 960   DCHECK(!write_callback_.is_null());
 961
 962   DWORD num_bytes, flags;
 963   BOOL ok = WSAGetOverlappedResult(socket_, &core_->write_overlapped_,
 964                                    &num_bytes, FALSE, &flags);
 965   WSAResetEvent(core_->write_overlapped_.hEvent);
 966   waiting_write_ = false;
 967   int rv;
 968   if (!ok) {
 969     int os_error = WSAGetLastError();
 970     rv = MapSystemError(os_error);
 971     net_log_.AddEvent(NetLog::TYPE_SOCKET_WRITE_ERROR,
 972                       CreateNetLogSocketErrorCallback(rv, os_error));
 973   } else {
 974     rv = static_cast<int>(num_bytes);
 975     if (rv > core_->write_buffer_length_ || rv < 0) {
 976       // It seems that some winsock interceptors report that more was written
 977       // than was available. Treat this as an error.  http://crbug.com/27870
 978       LOG(ERROR) << "Detected broken LSP: Asked to write "
 979                  << core_->write_buffer_length_ << " bytes, but " << rv
 980                  << " bytes reported.";
 981       rv = ERR_WINSOCK_UNEXPECTED_WRITTEN_BYTES;
 982     } else {
 983       net_log_.AddByteTransferEvent(NetLog::TYPE_SOCKET_BYTES_SENT, num_bytes,
 984                                     core_->write_iobuffer_->data());
 985       NetworkActivityMonitor::GetInstance()->IncrementBytesSent(num_bytes);
 986     }
 987   }
 988
 989   core_->write_iobuffer_ = NULL;
 990
 991   DCHECK_NE(rv, ERR_IO_PENDING);
 992   base::ResetAndReturn(&write_callback_).Run(rv);
 993 }
 994
 995 void TCPSocketWin::DidSignalRead() {
 996   DCHECK(waiting_read_);
 997   DCHECK(!read_callback_.is_null());
 998
 999   int os_error = 0;
1000   WSANETWORKEVENTS network_events;
1001   int rv = WSAEnumNetworkEvents(socket_, core_->read_overlapped_.hEvent,
1002                                 &network_events);
1003   if (rv == SOCKET_ERROR) {
1004     os_error = WSAGetLastError();
1005     rv = MapSystemError(os_error);
1006   } else if (network_events.lNetworkEvents) {
1007     // TODO(pkasting): Remove ScopedTracker below once crbug.com/462778 is
1008     // fixed.
1009     tracked_objects::ScopedTracker tracking_profile2(
1010         FROM_HERE_WITH_EXPLICIT_FUNCTION(
1011             "462778 TCPSocketWin::DidSignalRead -> DoRead"));
1012     DCHECK_EQ(network_events.lNetworkEvents & ~(FD_READ | FD_CLOSE), 0);
1013     // If network_events.lNetworkEvents is FD_CLOSE and
1014     // network_events.iErrorCode[FD_CLOSE_BIT] is 0, it is a graceful
1015     // connection closure. It is tempting to directly set rv to 0 in
1016     // this case, but the MSDN pages for WSAEventSelect and
1017     // WSAAsyncSelect recommend we still call DoRead():
1018     //   FD_CLOSE should only be posted after all data is read from a
1019     //   socket, but an application should check for remaining data upon
1020     //   receipt of FD_CLOSE to avoid any possibility of losing data.
1021     //
1022     // If network_events.iErrorCode[FD_READ_BIT] or
1023     // network_events.iErrorCode[FD_CLOSE_BIT] is nonzero, still call
1024     // DoRead() because recv() reports a more accurate error code
1025     // (WSAECONNRESET vs. WSAECONNABORTED) when the connection was
1026     // reset.
1027     rv = DoRead(core_->read_iobuffer_.get(), core_->read_buffer_length_,
1028                 read_callback_);
1029     if (rv == ERR_IO_PENDING)
1030       return;
1031   } else {
1032     // This may happen because Read() may succeed synchronously and
1033     // consume all the received data without resetting the event object.
1034     core_->WatchForRead();
1035     return;
1036   }
1037
1038   waiting_read_ = false;
1039   core_->read_iobuffer_ = NULL;
1040   core_->read_buffer_length_ = 0;
1041
1042   DCHECK_NE(rv, ERR_IO_PENDING);
1043   base::ResetAndReturn(&read_callback_).Run(rv);
1044 }
1045
1046 bool TCPSocketWin::GetEstimatedRoundTripTime(base::TimeDelta* out_rtt) const {
1047   DCHECK(out_rtt);
1048   // TODO(bmcquade): Consider implementing using
1049   // GetPerTcpConnectionEStats/GetPerTcp6ConnectionEStats.
1050   return false;
1051 }
1052
1053 }  // namespace net