System.Net.Socket: TSocket.Accept not behaving correctly in Linux ?

[optimax 2]

Hi,

I have a Tcp daemon thread using TSocket from System.Net.Socket (cross platform), listening for incoming TCP connection and creating client threads when clients connect.

It works just fine when compiled and run under Windows. I can connect with a Windows TCP client, a Linux TCP client, or even a simple Linux telnet session, and handle the client communication in the client thread.

When compiled and run under Linux, it seems that TSocket.Accept always returns nil, even when a connection is successfully established via telnet or other client. Hence the code that creates the client threads is never executed....
In Linux, my logs show an infinite series of 'Daemon - Socket.Accept with timeout 500ms' followed by 'Daemon - No new connection after 500ms timeout', although a connection can be successfully established.

However, without a client thread and a reference to the client socket, there isn't much I can do to handle the client communication...

Has anyone seen this behavior when compiling a TCP server for Linux based on System.Net.Socket ?

Any hint or help would be very much appreciated.

procedure TTcpDaemon.Execute;
var
  LConnectionSocket: TSocket;
  LConnectionThread: TTcpConnectionThread;
begin
  FServerSocket := TSocket.Create(TSocketType.TCP, TEncoding.UTF8);
  FServerSocket.Listen(FIP, '', FPort);
  Log('Dameon started (IP: ' + FIP + ' Port: ' + IntToStr(FPort) + ')');
  Log('Daemon - ConnectionThreadCount: ' + IntToStr(GetConnectionThreadCount));

  while not Terminated do
    begin
      try
        Log('Daemon - Socket.Accept with timeout 500ms');
        LConnectionSocket := FServerSocket.Accept(500);
        if Assigned(LConnectionSocket) then
          begin
            Log('Daemon - New connection - ' + LConnectionSocket.RemoteAddress + ':' + IntToStr(LConnectionSocket.RemotePort));
            LConnectionThread := TTcpConnectionThread.Create(LConnectionSocket);
            FThreadList.Add(LConnectionThread);
            Log('Daemon - Adding new connection thread - ConnectionThreadCount: ' + IntToStr(GetConnectionThreadCount));
            Log('Daemon - Starting new connection thread');
            LConnectionThread.Start;
          end
        else
          Log('Daemon - No new connection after 500ms timeout');
      except
        on E: Exception do
          Log(Self.ClassName + ' Exception (' + E.ClassName + '): ' + E.Message);
      end;
    end;


  TerminateAllThreads;
  FServerSocket.Close(True);
end;

 

[optimax 2]

So, I found some interesting details, which all seem to point to a buggy implementation of TSocket.Accept when targeting Linux.

 

Instead of

LConnectionSocket := FServerSocket.Accept(500);
if Assigned(LConnectionSocket) then
  begin
    ...

I wrote the following code:

LReadFds := TFDSet.Create(Self);
LWaitResult := FServerSocket.Select(ReadFds, nil, nil, 500);
if LWaitResult := TWaitResult.wrSignaled then
  begin
    LConnectionSocket := FServerSocket.Accept;
    if Assigned(LConnectionSocket) then
     begin
       Log('Daemon - New connection - ' + LConnectionSocket.RemoteAddress + ':' + IntToStr(LConnectionSocket.RemotePort));
       LConnectionThread := TTCpConnectionThread.Create(LConnectionSocket);
       ...

... and it works, both when compiled for Windows and Linux !

 

The difference being that TSocket.Accept is now called without a parameter (infinite wait, but returns immediately because a client connection is pending), this pointed to the problematic implementation of TSocket.Accept with a timeout:

  if not (TSocketState.Listening in FState) then
    raise ESocketError.CreateRes(@sSocketNotListening);
  Result := nil;
  if Timeout <> INFINITE then
  begin
    FD_ZERO(FDR);
    _FD_SET(FSocket, FDR);
    time.tv_sec := Timeout div 1000;
    time.tv_usec := (Timeout mod 1000) * 1000;
    Res := socketselect(1, @FDR, nil, nil, @time);  // << this is where things go wrong when compiled for Linux
    CheckSocketResult(Res, 'select');
    if Res = 0 then
      Exit;
  end;
  Len := SizeOf(Addr);
  ClientHandle := acceptsocket(FSocket, Addr, Len);
  CheckSocketResult(ClientHandle, 'accept');
  Result := GetClientSocket(ClientHandle, Addr);

The problem is the call to socketselect(1, @FDR, nil, nil, @time). This code works fine in Windows because the first hardcoded parameter (nfds = 1) is ignored. However, in Linux, this parameter must be the set to the highest-numbered file descriptor in any of the three sets + 1. So for Linux, this implementation is telling socketselect() to only look at file descriptor 0, which is almost certainly not what the socket is set to.

 

This is however correctly implemented in TSocket.Select (I am not sure why TSocket.Accept does not simply call TSocket.Select, rather than re-implementing the same functionality at low level):

  ReadFds := GetFds(CheckRead);
  WriteFds := GetFds(CheckWrite);
  ErrorFds := GetFds(CheckError);
  if Microseconds >= 0 then
  begin
    time.tv_sec := Microseconds div 1000000;
    time.tv_usec := Microseconds mod 1000000;
    TimePtr := @time;
  end else
    TimePtr := nil;
  Res := socketselect(Max(GetMaxFds(CheckRead), Max(GetMaxFds(CheckWrite), GetMaxFds(CheckError))) + 1, ReadFds, WriteFds, ErrorFds, PTimeVal(TimePtr));
  CheckSocketResult(Res, 'select');
  if Res = 0 then
    Result := TWaitResult.wrTimeout
  else
    Result := TWaitResult.wrSignaled;

More information about the low level call to select() can be found here:
https://stackoverflow.com/questions/2008059/socket-select-works-in-windows-and-times-out-in-linux

 

I guess the next step is to report this as a bug to Embarcadero...

 

[Dmitry Arefiev 101]

This is known issue:

https://quality.embarcadero.com/browse/RSP-19708

[optimax 2]

Thanks for the link. I just added a comment to RSP-19708 with the details of the code for TSocket.Accept and TSocket.Select, showing the error in the implementation of TSocket.Accept and the correct implementation of TSocket.Select.

Edited May 25, 2020 by optimax