POSIX only: Move client socket functionality from TCPClientSocket into TCPSocket.

net/socket/tcp_socket_libevent.cc

Index: net/socket/tcp_socket_libevent.cc
diff --git a/net/socket/tcp_socket_libevent.cc b/net/socket/tcp_socket_libevent.cc
index 486113374963df9eb1f36cbaaee715ff7c737c57..140cd83f9d11fc4b5085f28a1bad2721ee1582cc 100644
--- a/net/socket/tcp_socket_libevent.cc
+++ b/net/socket/tcp_socket_libevent.cc
@@ -2,11 +2,12 @@
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
-#include "net/socket/tcp_socket_libevent.h"
+#include "net/socket/tcp_socket.h"

[wtc, 2013/09/13 22:22:44]

This should include tcp_socket_libevent.h because this file is named
tcp_socket_libevent.cc.

[yzshen1, 2013/09/13 23:52:40]

This is intended, because we need IsTCPFastOpenEnabled() from tcp_socket.h.

The previous code tcp_client_socket_libevent.cc included tcp_client_socket.h for
the same reason.


On 2013/09/13 22:22:44, wtc wrote:

 
 #include <errno.h>
 #include <fcntl.h>
 #include <netdb.h>
+#include <netinet/tcp.h>
 #include <sys/socket.h>
 
 #include "build/build_config.h"
@@ -15,30 +16,146 @@
 #include <netinet/in.h>

[wtc, 2013/09/13 22:22:44]

I believe this file can only be used on OS_POSIX. Some of the C system headers
we include above only exist on OS_POSIX. So this #if defined(OS_POSIX) check
seems unnecessary.

[yzshen1, 2013/09/13 23:52:40]

Done.

 #endif
 
+#include "base/callback_helpers.h"
 #include "base/logging.h"
+#include "base/metrics/histogram.h"
+#include "base/metrics/stats_counters.h"
 #include "base/posix/eintr_wrapper.h"
+#include "net/base/address_list.h"
+#include "net/base/connection_type_histograms.h"
+#include "net/base/io_buffer.h"
 #include "net/base/ip_endpoint.h"
 #include "net/base/net_errors.h"
 #include "net/base/net_util.h"
-#include "net/socket/socket_descriptor.h"
+#include "net/base/network_change_notifier.h"
 #include "net/socket/socket_net_log_params.h"
 
+// If we don't have a definition for TCPI_OPT_SYN_DATA, create one.
+#ifndef TCPI_OPT_SYN_DATA
+#define TCPI_OPT_SYN_DATA 32
+#endif
+
 namespace net {
 
+namespace {
+
+const int kTCPKeepAliveSeconds = 45;
+
+// SetTCPNoDelay turns on/off buffering in the kernel. By default, TCP sockets
+// will wait up to 200ms for more data to complete a packet before transmitting.
+// After calling this function, the kernel will not wait. See TCP_NODELAY in
+// `man 7 tcp`.
+bool SetTCPNoDelay(int fd, bool no_delay) {
+  int on = no_delay ? 1 : 0;
+  int error = setsockopt(fd, IPPROTO_TCP, TCP_NODELAY, &on, sizeof(on));
+  return error == 0;
+}
+
+// SetTCPKeepAlive sets SO_KEEPALIVE.
+bool SetTCPKeepAlive(int fd, bool enable, int delay) {
+  int on = enable ? 1 : 0;
+  if (setsockopt(fd, SOL_SOCKET, SO_KEEPALIVE, &on, sizeof(on))) {
+    PLOG(ERROR) << "Failed to set SO_KEEPALIVE on fd: " << fd;
+    return false;
+  }
+#if defined(OS_LINUX) || defined(OS_ANDROID)
+  // Set seconds until first TCP keep alive.
+  if (setsockopt(fd, SOL_TCP, TCP_KEEPIDLE, &delay, sizeof(delay))) {
+    PLOG(ERROR) << "Failed to set TCP_KEEPIDLE on fd: " << fd;
+    return false;
+  }
+  // Set seconds between TCP keep alives.
+  if (setsockopt(fd, SOL_TCP, TCP_KEEPINTVL, &delay, sizeof(delay))) {
+    PLOG(ERROR) << "Failed to set TCP_KEEPINTVL on fd: " << fd;
+    return false;
+  }
+#endif
+  return true;
+}
+
+int MapConnectError(int os_error) {
+  switch (os_error) {
+    case EACCES:
+      return ERR_NETWORK_ACCESS_DENIED;
+    case ETIMEDOUT:
+      return ERR_CONNECTION_TIMED_OUT;
+    default: {
+      int net_error = MapSystemError(os_error);
+      if (net_error == ERR_FAILED)
+        return ERR_CONNECTION_FAILED;  // More specific than ERR_FAILED.
+
+      // Give a more specific error when the user is offline.
+      if (net_error == ERR_ADDRESS_UNREACHABLE &&
+          NetworkChangeNotifier::IsOffline()) {
+        return ERR_INTERNET_DISCONNECTED;
+      }
+      return net_error;
+    }
+  }
+}
+
+}  // namespace
+
+//-----------------------------------------------------------------------------
+
+TCPSocketLibevent::Watcher::Watcher(
+    const base::Closure& read_ready_callback,
+    const base::Closure& write_ready_callback)
+    : read_ready_callback_(read_ready_callback),
+      write_ready_callback_(write_ready_callback) {
+}
+
+TCPSocketLibevent::Watcher::~Watcher() {
+}
+
+void TCPSocketLibevent::Watcher::OnFileCanReadWithoutBlocking(int /* fd */) {
+  if (!read_ready_callback_.is_null())
+    read_ready_callback_.Run();

[wtc, 2013/09/13 22:22:44]

Do we need to reset read_ready_callback_ and write_ready_callback_ after
running them?

[yzshen1, 2013/09/13 23:52:40]

|read_ready_callback_| here is a closure to forward the notification to a
certain method of TCPSocketLibevent. It never changes during the lifetime of
TCPSocketLibevent. So we shouldn't reset it. (same for |write_ready_callback_|.)

+  else
+    NOTREACHED();
+}
+
+void TCPSocketLibevent::Watcher::OnFileCanWriteWithoutBlocking(int /* fd */) {
+  if (!write_ready_callback_.is_null())
+    write_ready_callback_.Run();
+  else
+    NOTREACHED();
+}
+
 TCPSocketLibevent::TCPSocketLibevent(NetLog* net_log,
                                      const NetLog::Source& source)
     : socket_(kInvalidSocket),
+      accept_watcher_(base::Bind(&TCPSocketLibevent::DidCompleteAccept,
+                                 base::Unretained(this)),
+                      base::Closure()),
       accept_socket_(NULL),
       accept_address_(NULL),
+      read_watcher_(base::Bind(&TCPSocketLibevent::DidCompleteRead,
+                               base::Unretained(this)),
+                    base::Closure()),
+      write_watcher_(base::Closure(),
+                     base::Bind(&TCPSocketLibevent::DidCompleteConnectOrWrite,
+                                base::Unretained(this))),
+      read_buf_len_(0),
+      write_buf_len_(0),
+      use_tcp_fastopen_(IsTCPFastOpenEnabled()),
+      tcp_fastopen_connected_(false),
+      fast_open_status_(FAST_OPEN_STATUS_UNKNOWN),
+      waiting_connect_(false),
+      connect_os_error_(0),
+      logging_multiple_connect_attempts_(false),
       net_log_(BoundNetLog::Make(net_log, NetLog::SOURCE_SOCKET)) {
   net_log_.BeginEvent(NetLog::TYPE_SOCKET_ALIVE,
                       source.ToEventParametersCallback());
 }
 
 TCPSocketLibevent::~TCPSocketLibevent() {
-  if (socket_ != kInvalidSocket)
-    Close();
   net_log_.EndEvent(NetLog::TYPE_SOCKET_ALIVE);
+  if (tcp_fastopen_connected_) {
+    UMA_HISTOGRAM_ENUMERATION("Net.TcpFastOpenSocketConnection",
+                              fast_open_status_, FAST_OPEN_MAX_VALUE);
+  }
+  Close();

[wtc, 2013/09/13 22:22:44]

It may be better to call Close() before logging events and updating histograms.
The original code does this.

[yzshen1, 2013/09/13 23:52:40]

That will reset |tcp_fastopen_connected_| and cause the previous if-block to
never run.

On 2013/09/13 22:22:44, wtc wrote:

[wtc, 2013/09/18 23:26:17]

I see. So the original code worked because it never reset
tcp_fastopen_connected_ to false, right?

[yzshen1, 2013/09/18 23:40:35]

Right. It is never set to false again.

 }
 
 int TCPSocketLibevent::Open(AddressFamily family) {
@@ -61,7 +178,8 @@ int TCPSocketLibevent::Open(AddressFamily family) {
   return OK;
 }
 
-int TCPSocketLibevent::Adopt(int socket) {
+int TCPSocketLibevent::AdoptConnectedSocket(int socket,
+                                            const IPEndPoint& peer_address) {
   DCHECK(CalledOnValidThread());
   DCHECK_EQ(socket_, kInvalidSocket);
 
@@ -73,16 +191,9 @@ int TCPSocketLibevent::Adopt(int socket) {
     return result;
   }
 
-  return OK;
-}
-
-int TCPSocketLibevent::Release() {
-  DCHECK(CalledOnValidThread());
-  DCHECK(accept_callback_.is_null());
+  peer_address_.reset(new IPEndPoint(peer_address));
 
-  int result = socket_;
-  socket_ = kInvalidSocket;
-  return result;
+  return OK;
 }
 
 int TCPSocketLibevent::Bind(const IPEndPoint& address) {
@@ -102,19 +213,6 @@ int TCPSocketLibevent::Bind(const IPEndPoint& address) {
   return OK;
 }
 
-int TCPSocketLibevent::GetLocalAddress(IPEndPoint* address) const {
-  DCHECK(CalledOnValidThread());
-  DCHECK(address);
-
-  SockaddrStorage storage;
-  if (getsockname(socket_, storage.addr, &storage.addr_len) < 0)
-    return MapSystemError(errno);
-  if (!address->FromSockAddr(storage.addr, storage.addr_len))
-    return ERR_FAILED;
-
-  return OK;
-}
-
 int TCPSocketLibevent::Listen(int backlog) {
   DCHECK(CalledOnValidThread());
   DCHECK_GT(backlog, 0);
@@ -145,7 +243,7 @@ int TCPSocketLibevent::Accept(scoped_ptr<TCPSocketLibevent>* socket,
   if (result == ERR_IO_PENDING) {
     if (!base::MessageLoopForIO::current()->WatchFileDescriptor(
             socket_, true, base::MessageLoopForIO::WATCH_READ,
-            &accept_socket_watcher_, this)) {
+            &accept_socket_watcher_, &accept_watcher_)) {
       PLOG(ERROR) << "WatchFileDescriptor failed on read";
       return MapSystemError(errno);
     }
@@ -158,11 +256,201 @@ int TCPSocketLibevent::Accept(scoped_ptr<TCPSocketLibevent>* socket,
   return result;
 }
 
+int TCPSocketLibevent::Connect(const IPEndPoint& address,
+                               const CompletionCallback& callback) {
+  DCHECK(CalledOnValidThread());
+  DCHECK_NE(socket_, kInvalidSocket);
+  DCHECK(!waiting_connect_);
+
+  // |peer_address_| will be non-NULL if Connect() has been called. Unless
+  // Close() is called to reset the internal state, a second call to Connect()
+  // is not allowed.
+  // Please note that we enforce this even if the previous Connect() has
+  // completed and failed. Connecting the same |socket_| again after a

[wtc, 2013/09/13 22:22:44]

Nit: "Please note that we enforce this ... and failed." can be removed because
it applies only to Windows.

[yzshen1, 2013/09/13 23:52:40]

I thought this sentence also applies here? 
Maybe I have misunderstood something?

I changed it to "Please note that we don't allow a second Connect() even if the
previous Connect() has failed". Does it sound better?

[wtc, 2013/09/18 23:26:17]

Your comment is fine. Thanks.

(Since this won't work on POSIX, I thought the comment wasn't
necessary.)

+  // connection attempt failed results in unspecified behavior according to
+  // POSIX.
+  DCHECK(!peer_address_);
+
+  if (!logging_multiple_connect_attempts_)
+    LogConnectBegin(AddressList(address));
+
+  peer_address_.reset(new IPEndPoint(address));
+
+  int rv = DoConnect();
+  if (rv == ERR_IO_PENDING) {
+    // Synchronous operation not supported.
+    DCHECK(!callback.is_null());
+    write_callback_ = callback;
+    waiting_connect_ = true;
+  } else {
+    DoConnectComplete(rv);
+  }
+
+  return rv;
+}
+
+bool TCPSocketLibevent::IsConnected() const {
+  DCHECK(CalledOnValidThread());
+
+  if (socket_ == kInvalidSocket || waiting_connect_)
+    return false;
+
+  if (use_tcp_fastopen_ && !tcp_fastopen_connected_ && peer_address_) {

[wtc, 2013/09/13 22:22:44]

Do we need to check peer_address_ in this conditional expression?

[yzshen1, 2013/09/13 23:52:40]

Yes. The following is one example:

- |use_tcp_fastopen_| is always true if fast open is enabled.
- the user hasn't called Connect(), |tcp_fastopen_connected_| == false.

In this case, we don't want IsConnected() to return true.


On 2013/09/13 22:22:44, wtc wrote:

+    // With TCP FastOpen, we pretend that the socket is connected.
+    // This allows GetPeerAddress() to return peer_address_.
+    return true;
+  }
+
+  // Check if connection is alive.
+  char c;
+  int rv = HANDLE_EINTR(recv(socket_, &c, 1, MSG_PEEK));
+  if (rv == 0)
+    return false;
+  if (rv == -1 && errno != EAGAIN && errno != EWOULDBLOCK)
+    return false;
+
+  return true;
+}
+
+bool TCPSocketLibevent::IsConnectedAndIdle() const {
+  DCHECK(CalledOnValidThread());
+
+  if (socket_ == kInvalidSocket || waiting_connect_)
+    return false;
+
+  // TODO(wtc): should we also handle the TCP FastOpen case here,
+  // as we do in IsConnected()?
+
+  // Check if connection is alive and we haven't received any data
+  // unexpectedly.
+  char c;
+  int rv = HANDLE_EINTR(recv(socket_, &c, 1, MSG_PEEK));
+  if (rv >= 0)
+    return false;
+  if (errno != EAGAIN && errno != EWOULDBLOCK)
+    return false;
+
+  return true;
+}
+
+int TCPSocketLibevent::Read(IOBuffer* buf,
+                            int buf_len,
+                            const CompletionCallback& callback) {
+  DCHECK(CalledOnValidThread());
+  DCHECK_NE(kInvalidSocket, socket_);
+  DCHECK(!waiting_connect_);
+  DCHECK(read_callback_.is_null());
+  // Synchronous operation not supported
+  DCHECK(!callback.is_null());
+  DCHECK_GT(buf_len, 0);
+
+  int nread = HANDLE_EINTR(read(socket_, buf->data(), buf_len));
+  if (nread >= 0) {
+    base::StatsCounter read_bytes("tcp.read_bytes");
+    read_bytes.Add(nread);
+    net_log_.AddByteTransferEvent(NetLog::TYPE_SOCKET_BYTES_RECEIVED, nread,
+                                  buf->data());
+    RecordFastOpenStatus();
+    return nread;
+  }
+  if (errno != EAGAIN && errno != EWOULDBLOCK) {
+    int net_error = MapSystemError(errno);
+    net_log_.AddEvent(NetLog::TYPE_SOCKET_READ_ERROR,
+                      CreateNetLogSocketErrorCallback(net_error, errno));
+    return net_error;
+  }
+
+  if (!base::MessageLoopForIO::current()->WatchFileDescriptor(
+          socket_, true, base::MessageLoopForIO::WATCH_READ,
+          &read_socket_watcher_, &read_watcher_)) {
+    DVLOG(1) << "WatchFileDescriptor failed on read, errno " << errno;
+    return MapSystemError(errno);
+  }
+
+  read_buf_ = buf;
+  read_buf_len_ = buf_len;
+  read_callback_ = callback;
+  return ERR_IO_PENDING;
+}
+
+int TCPSocketLibevent::Write(IOBuffer* buf,
+                             int buf_len,
+                             const CompletionCallback& callback) {
+  DCHECK(CalledOnValidThread());
+  DCHECK_NE(kInvalidSocket, socket_);
+  DCHECK(!waiting_connect_);
+  DCHECK(write_callback_.is_null());
+  // Synchronous operation not supported
+  DCHECK(!callback.is_null());
+  DCHECK_GT(buf_len, 0);
+
+  int nwrite = InternalWrite(buf, buf_len);
+  if (nwrite >= 0) {
+    base::StatsCounter write_bytes("tcp.write_bytes");
+    write_bytes.Add(nwrite);
+    net_log_.AddByteTransferEvent(NetLog::TYPE_SOCKET_BYTES_SENT, nwrite,
+                                  buf->data());
+    return nwrite;
+  }
+  if (errno != EAGAIN && errno != EWOULDBLOCK) {
+    int net_error = MapSystemError(errno);
+    net_log_.AddEvent(NetLog::TYPE_SOCKET_WRITE_ERROR,
+                      CreateNetLogSocketErrorCallback(net_error, errno));
+    return net_error;
+  }
+
+  if (!base::MessageLoopForIO::current()->WatchFileDescriptor(
+          socket_, true, base::MessageLoopForIO::WATCH_WRITE,
+          &write_socket_watcher_, &write_watcher_)) {
+    DVLOG(1) << "WatchFileDescriptor failed on write, errno " << errno;
+    return MapSystemError(errno);
+  }
+
+  write_buf_ = buf;
+  write_buf_len_ = buf_len;
+  write_callback_ = callback;
+  return ERR_IO_PENDING;
+}
+
+int TCPSocketLibevent::GetLocalAddress(IPEndPoint* address) const {
+  DCHECK(CalledOnValidThread());
+  DCHECK(address);

[wtc, 2013/09/13 22:22:44]

We probably should do
  if (socket_ == kInvalidSocket)
    return ERR_SOCKET_NOT_CONNECTED;
to match the original code.

Note: it is possible that the getsockname call will fail with the same error.

[yzshen1, 2013/09/13 23:52:40]

IMHO, ERR_SOCKET_NOT_CONNECTED is not accurate.
For example, GetLocalAddress() can be used after Bind() but before Connect().
It is not a method related to the connection status.

Removing this part matches the behavior of tcp_socket_win.cc as well. What do
you think?


On 2013/09/13 22:22:44, wtc wrote:

[wtc, 2013/09/18 23:26:17]

Yes, this is fine.

If socket_ is kInvalidSocket (-1), I think
getsockname(socket_, ...) will fail with EBADF, which we
map to ERR_INVALID_HANDLE. That's a reasonable error code.

+
+  SockaddrStorage storage;
+  if (getsockname(socket_, storage.addr, &storage.addr_len) < 0)
+    return MapSystemError(errno);
+  if (!address->FromSockAddr(storage.addr, storage.addr_len))
+    return ERR_FAILED;

[wtc, 2013/09/13 22:22:44]

Please use a better error code. (You fixed this in tcp_socket_win.cc.)

[yzshen1, 2013/09/13 23:52:40]

Done. I used ERR_ADDRESS_INVALID.

+
+  return OK;
+}
+
+int TCPSocketLibevent::GetPeerAddress(IPEndPoint* address) const {
+  DCHECK(CalledOnValidThread());
+  DCHECK(address);
+  if (!IsConnected())
+    return ERR_SOCKET_NOT_CONNECTED;
+  *address = *peer_address_;
+  return OK;
+}
+
 int TCPSocketLibevent::SetDefaultOptionsForServer() {
+  DCHECK(CalledOnValidThread());
   return SetAddressReuse(true);
 }
 
+void TCPSocketLibevent::SetDefaultOptionsForClient() {
+  DCHECK(CalledOnValidThread());
+
+  // This mirrors the behaviour on Windows. See the comment in
+  // tcp_socket_win.cc after searching for "NODELAY".
+  SetTCPNoDelay(socket_, true);  // If SetTCPNoDelay fails, we don't care.
+  SetTCPKeepAlive(socket_, true, kTCPKeepAliveSeconds);
+}
+
 int TCPSocketLibevent::SetAddressReuse(bool allow) {
+  DCHECK(CalledOnValidThread());
+
   // SO_REUSEADDR is useful for server sockets to bind to a recently unbound
   // port. When a socket is closed, the end point changes its state to TIME_WAIT
   // and wait for 2 MSL (maximum segment lifetime) to ensure the remote peer
@@ -184,14 +472,96 @@ int TCPSocketLibevent::SetAddressReuse(bool allow) {
   return OK;
 }
 
+bool TCPSocketLibevent::SetReceiveBufferSize(int32 size) {
+  DCHECK(CalledOnValidThread());
+  int rv = setsockopt(socket_, SOL_SOCKET, SO_RCVBUF,
+      reinterpret_cast<const char*>(&size),
+      sizeof(size));
+  DCHECK(!rv) << "Could not set socket receive buffer size: " << errno;
+  return rv == 0;
+}
+
+bool TCPSocketLibevent::SetSendBufferSize(int32 size) {
+  DCHECK(CalledOnValidThread());
+  int rv = setsockopt(socket_, SOL_SOCKET, SO_SNDBUF,
+      reinterpret_cast<const char*>(&size),
+      sizeof(size));
+  DCHECK(!rv) << "Could not set socket send buffer size: " << errno;
+  return rv == 0;
+}
+
+bool TCPSocketLibevent::SetKeepAlive(bool enable, int delay) {
+  DCHECK(CalledOnValidThread());
+  return SetTCPKeepAlive(socket_, enable, delay);
+}
+
+bool TCPSocketLibevent::SetNoDelay(bool no_delay) {
+  DCHECK(CalledOnValidThread());
+  return SetTCPNoDelay(socket_, no_delay);
+}
+
 void TCPSocketLibevent::Close() {
+  DCHECK(CalledOnValidThread());
+
   if (socket_ != kInvalidSocket) {
-    bool ok = accept_socket_watcher_.StopWatchingFileDescriptor();
-    DCHECK(ok);
     if (HANDLE_EINTR(close(socket_)) < 0)
       PLOG(ERROR) << "close";
     socket_ = kInvalidSocket;
   }
+
+  bool ok = accept_socket_watcher_.StopWatchingFileDescriptor();
+  DCHECK(ok);
+  ok = read_socket_watcher_.StopWatchingFileDescriptor();
+  DCHECK(ok);
+  ok = write_socket_watcher_.StopWatchingFileDescriptor();
+  DCHECK(ok);
+
+  if (!accept_callback_.is_null()) {
+    accept_socket_ = NULL;
+    accept_address_ = NULL;
+    accept_callback_.Reset();
+  }
+
+  if (!read_callback_.is_null()) {
+    read_buf_ = NULL;
+    read_buf_len_ = 0;
+    read_callback_.Reset();
+  }
+
+  if (!write_callback_.is_null()) {
+    write_buf_ = NULL;
+    write_buf_len_ = 0;
+    write_callback_.Reset();
+  }
+
+  tcp_fastopen_connected_ = false;
+  fast_open_status_ = FAST_OPEN_STATUS_UNKNOWN;
+  waiting_connect_ = false;
+  peer_address_.reset();
+  connect_os_error_ = 0;
+}
+
+bool TCPSocketLibevent::UsingTCPFastOpen() const {
+  return use_tcp_fastopen_;
+}
+
+void TCPSocketLibevent::StartLoggingMultipleConnectAttempts(
+    const AddressList& addresses) {
+  if (!logging_multiple_connect_attempts_) {
+    logging_multiple_connect_attempts_ = true;
+    LogConnectBegin(addresses);
+  } else {
+    NOTREACHED();
+  }
+}
+
+void TCPSocketLibevent::EndLoggingMultipleConnectAttempts(int net_error) {
+  if (logging_multiple_connect_attempts_) {
+    LogConnectEnd(net_error);
+    logging_multiple_connect_attempts_ = false;
+  } else {
+    NOTREACHED();
+  }
 }
 
 int TCPSocketLibevent::AcceptInternal(scoped_ptr<TCPSocketLibevent>* socket,
@@ -217,7 +587,7 @@ int TCPSocketLibevent::AcceptInternal(scoped_ptr<TCPSocketLibevent>* socket,
   }
   scoped_ptr<TCPSocketLibevent> tcp_socket(new TCPSocketLibevent(
       net_log_.net_log(), net_log_.source()));
-  int adopt_result = tcp_socket->Adopt(new_socket);
+  int adopt_result = tcp_socket->AdoptConnectedSocket(new_socket, ip_end_point);
   if (adopt_result != OK) {
     net_log_.EndEventWithNetErrorCode(NetLog::TYPE_TCP_ACCEPT, adopt_result);
     return adopt_result;
@@ -229,7 +599,191 @@ int TCPSocketLibevent::AcceptInternal(scoped_ptr<TCPSocketLibevent>* socket,
   return OK;
 }
 
-void TCPSocketLibevent::OnFileCanReadWithoutBlocking(int fd) {
+int TCPSocketLibevent::DoConnect() {
+  DCHECK_EQ(0, connect_os_error_);
+
+  net_log_.BeginEvent(NetLog::TYPE_TCP_CONNECT_ATTEMPT,
+                      CreateNetLogIPEndPointCallback(peer_address_.get()));
+
+  // Connect the socket.
+  if (!use_tcp_fastopen_) {
+    SockaddrStorage storage;
+    if (!peer_address_->ToSockAddr(storage.addr, &storage.addr_len))
+      return ERR_INVALID_ARGUMENT;
+
+    if (!HANDLE_EINTR(connect(socket_, storage.addr, storage.addr_len))) {
+      // Connected without waiting!
+      return OK;
+    }
+  } else {
+    // With TCP FastOpen, we pretend that the socket is connected.
+    DCHECK(!tcp_fastopen_connected_);
+    return OK;
+  }
+
+  // Check if the connect() failed synchronously.
+  connect_os_error_ = errno;
+  if (connect_os_error_ != EINPROGRESS)
+    return MapConnectError(connect_os_error_);
+
+  // Otherwise the connect() is going to complete asynchronously, so watch
+  // for its completion.
+  if (!base::MessageLoopForIO::current()->WatchFileDescriptor(
+          socket_, true, base::MessageLoopForIO::WATCH_WRITE,
+          &write_socket_watcher_, &write_watcher_)) {
+    connect_os_error_ = errno;
+    DVLOG(1) << "WatchFileDescriptor failed: " << connect_os_error_;
+    return MapSystemError(connect_os_error_);
+  }
+
+  return ERR_IO_PENDING;
+}
+
+void TCPSocketLibevent::DoConnectComplete(int result) {
+  // Log the end of this attempt (and any OS error it threw).
+  int os_error = connect_os_error_;
+  connect_os_error_ = 0;
+  if (result != OK) {
+    net_log_.EndEvent(NetLog::TYPE_TCP_CONNECT_ATTEMPT,
+                      NetLog::IntegerCallback("os_error", os_error));
+  } else {
+    net_log_.EndEvent(NetLog::TYPE_TCP_CONNECT_ATTEMPT);
+  }
+
+  if (!logging_multiple_connect_attempts_)
+    LogConnectEnd(result);
+}
+
+void TCPSocketLibevent::LogConnectBegin(const AddressList& addresses) {
+  base::StatsCounter connects("tcp.connect");
+  connects.Increment();
+
+  net_log_.BeginEvent(NetLog::TYPE_TCP_CONNECT,
+                      addresses.CreateNetLogCallback());
+}
+
+void TCPSocketLibevent::LogConnectEnd(int net_error) {
+  if (net_error == OK)
+    UpdateConnectionTypeHistograms(CONNECTION_ANY);
+
+  if (net_error != OK) {
+    net_log_.EndEventWithNetErrorCode(NetLog::TYPE_TCP_CONNECT, net_error);
+    return;
+  }
+
+  SockaddrStorage storage;
+  int rv = getsockname(socket_, storage.addr, &storage.addr_len);
+  if (rv != 0) {
+    PLOG(ERROR) << "getsockname() [rv: " << rv << "] error: ";
+    NOTREACHED();
+    net_log_.EndEventWithNetErrorCode(NetLog::TYPE_TCP_CONNECT, rv);
+    return;
+  }
+
+  net_log_.EndEvent(NetLog::TYPE_TCP_CONNECT,
+                    CreateNetLogSourceAddressCallback(storage.addr,
+                                                      storage.addr_len));
+}
+
+void TCPSocketLibevent::DidCompleteRead() {
+  RecordFastOpenStatus();
+  if (read_callback_.is_null())

[wtc, 2013/09/13 22:22:44]

Do you know why read_callback_ can be null? (I know this is_null() check
comes from the original code.)

[yzshen1, 2013/09/13 23:52:40]

I don't know any scenario in which read_callback_ can be null when
DidCompleteRead() is called.
I just migrated it from the original code.

Do you suggest me to remove it or we just stay at the safe side for now?

On 2013/09/13 22:22:44, wtc wrote:

[wtc, 2013/09/18 23:26:17]

No change needed. I was just curious if you understand why.

+    return;
+
+  int bytes_transferred;
+  bytes_transferred = HANDLE_EINTR(read(socket_, read_buf_->data(),
+                                        read_buf_len_));
+
+  int result;
+  if (bytes_transferred >= 0) {
+    result = bytes_transferred;
+    base::StatsCounter read_bytes("tcp.read_bytes");
+    read_bytes.Add(bytes_transferred);
+    net_log_.AddByteTransferEvent(NetLog::TYPE_SOCKET_BYTES_RECEIVED, result,
+                                  read_buf_->data());
+  } else {
+    result = MapSystemError(errno);
+    if (result != ERR_IO_PENDING) {
+      net_log_.AddEvent(NetLog::TYPE_SOCKET_READ_ERROR,
+                        CreateNetLogSocketErrorCallback(result, errno));
+    }
+  }
+
+  if (result != ERR_IO_PENDING) {
+    read_buf_ = NULL;
+    read_buf_len_ = 0;
+    bool ok = read_socket_watcher_.StopWatchingFileDescriptor();
+    DCHECK(ok);
+    base::ResetAndReturn(&read_callback_).Run(result);
+  }
+}
+
+void TCPSocketLibevent::DidCompleteWrite() {
+  DCHECK(!write_callback_.is_null());
+
+  int bytes_transferred;
+  bytes_transferred = HANDLE_EINTR(write(socket_, write_buf_->data(),
+                                         write_buf_len_));
+
+  int result;
+  if (bytes_transferred >= 0) {
+    result = bytes_transferred;
+    base::StatsCounter write_bytes("tcp.write_bytes");
+    write_bytes.Add(bytes_transferred);
+    net_log_.AddByteTransferEvent(NetLog::TYPE_SOCKET_BYTES_SENT, result,
+                                  write_buf_->data());
+  } else {
+    result = MapSystemError(errno);
+    if (result != ERR_IO_PENDING) {
+      net_log_.AddEvent(NetLog::TYPE_SOCKET_WRITE_ERROR,
+                        CreateNetLogSocketErrorCallback(result, errno));
+    }
+  }
+
+  if (result != ERR_IO_PENDING) {
+    write_buf_ = NULL;
+    write_buf_len_ = 0;
+    write_socket_watcher_.StopWatchingFileDescriptor();
+    base::ResetAndReturn(&write_callback_).Run(result);
+  }
+}
+
+void TCPSocketLibevent::DidCompleteConnect() {
+  DCHECK(waiting_connect_);
+
+  // Get the error that connect() completed with.
+  int os_error = 0;
+  socklen_t len = sizeof(os_error);
+  if (getsockopt(socket_, SOL_SOCKET, SO_ERROR, &os_error, &len) < 0)
+    os_error = errno;
+
+  int result = OK;
+  // TODO(eroman): Is this check really necessary?
+  if (os_error == EINPROGRESS || os_error == EALREADY) {

[wtc, 2013/09/13 22:22:44]

Let's remove this check and the TODO comment.

EINPROGRESS and EALREADY can only be returned by connect(). So we won't get
these two error codes from the getsockopt(...SO_ERROR...) call above.

[yzshen1, 2013/09/13 23:52:40]

Done.

+    NOTREACHED();  // This indicates a bug in libevent or our code.
+    result = ERR_UNEXPECTED;
+  } else {
+    result = MapConnectError(os_error);
+  }
+
+  connect_os_error_ = os_error;
+  if (result != ERR_IO_PENDING) {
+    DoConnectComplete(result);
+    waiting_connect_ = false;
+    write_socket_watcher_.StopWatchingFileDescriptor();
+    base::ResetAndReturn(&write_callback_).Run(result);
+  }
+}
+
+void TCPSocketLibevent::DidCompleteConnectOrWrite() {
+  if (waiting_connect_) {
+    DidCompleteConnect();
+  } else if (!write_callback_.is_null()) {

[wtc, 2013/09/13 22:22:44]

(I know this is_null() check comes from the original code.)

I suggest we just call DidCompleteWrite, and let DidCompleteWrite check
write_callback_.is_null() just like DidCompleteRead checks
read_callback_.is_null().

[yzshen1, 2013/09/13 23:52:40]

Done.

+    DidCompleteWrite();
+  }
+}
+
+void TCPSocketLibevent::DidCompleteAccept() {
   DCHECK(CalledOnValidThread());
 
   int result = AcceptInternal(accept_socket_, accept_address_);
@@ -244,8 +798,85 @@ void TCPSocketLibevent::OnFileCanReadWithoutBlocking(int fd) {
   }
 }
 
-void TCPSocketLibevent::OnFileCanWriteWithoutBlocking(int fd) {
-  NOTREACHED();
+int TCPSocketLibevent::InternalWrite(IOBuffer* buf, int buf_len) {
+  int nwrite;
+  if (use_tcp_fastopen_ && !tcp_fastopen_connected_) {
+    SockaddrStorage storage;
+    if (!peer_address_->ToSockAddr(storage.addr, &storage.addr_len)) {
+      errno = EINVAL;
+      return -1;
+    }
+
+    int flags = 0x20000000; // Magic flag to enable TCP_FASTOPEN.
+#if defined(OS_LINUX)
+    // sendto() will fail with EPIPE when the system doesn't support TCP Fast
+    // Open. Theoretically that shouldn't happen since the caller should check
+    // for system support on startup, but users may dynamically disable TCP Fast
+    // Open via sysctl.
+    flags |= MSG_NOSIGNAL;
+#endif // defined(OS_LINUX)
+    nwrite = HANDLE_EINTR(sendto(socket_,
+                                 buf->data(),
+                                 buf_len,
+                                 flags,
+                                 storage.addr,
+                                 storage.addr_len));
+    tcp_fastopen_connected_ = true;
+
+    if (nwrite < 0) {
+      DCHECK_NE(EPIPE, errno);
+
+      // If errno == EINPROGRESS, that means the kernel didn't have a cookie
+      // and would block. The kernel is internally doing a connect() though.
+      // Remap EINPROGRESS to EAGAIN so we treat this the same as our other
+      // asynchronous cases. Note that the user buffer has not been copied to
+      // kernel space.
+      if (errno == EINPROGRESS) {
+        errno = EAGAIN;
+        fast_open_status_ = FAST_OPEN_SLOW_CONNECT_RETURN;
+      } else {
+        fast_open_status_ = FAST_OPEN_ERROR;
+      }
+    } else {
+      fast_open_status_ = FAST_OPEN_FAST_CONNECT_RETURN;
+    }
+  } else {
+    nwrite = HANDLE_EINTR(write(socket_, buf->data(), buf_len));
+  }
+  return nwrite;
+}
+
+void TCPSocketLibevent::RecordFastOpenStatus() {
+  if (use_tcp_fastopen_ &&
+      (fast_open_status_ == FAST_OPEN_FAST_CONNECT_RETURN ||
+       fast_open_status_ == FAST_OPEN_SLOW_CONNECT_RETURN)) {
+    DCHECK_NE(FAST_OPEN_STATUS_UNKNOWN, fast_open_status_);
+    bool getsockopt_success(false);
+    bool server_acked_data(false);
+#if defined(TCP_INFO)
+    // Probe to see the if the socket used TCP Fast Open.
+    tcp_info info;
+    socklen_t info_len = sizeof(tcp_info);
+    getsockopt_success =
+        getsockopt(socket_, IPPROTO_TCP, TCP_INFO, &info, &info_len) == 0 &&
+        info_len == sizeof(tcp_info);
+    server_acked_data = getsockopt_success &&
+        (info.tcpi_options & TCPI_OPT_SYN_DATA);
+#endif
+    if (getsockopt_success) {
+      if (fast_open_status_ == FAST_OPEN_FAST_CONNECT_RETURN) {
+        fast_open_status_ = (server_acked_data ? FAST_OPEN_SYN_DATA_ACK :
+                             FAST_OPEN_SYN_DATA_NACK);
+      } else {
+        fast_open_status_ = (server_acked_data ? FAST_OPEN_NO_SYN_DATA_ACK :
+                             FAST_OPEN_NO_SYN_DATA_NACK);
+      }
+    } else {
+      fast_open_status_ = (fast_open_status_ == FAST_OPEN_FAST_CONNECT_RETURN ?
+                           FAST_OPEN_SYN_DATA_FAILED :
+                           FAST_OPEN_NO_SYN_DATA_FAILED);
+    }
+  }
 }
 
 }  // namespace net