[Interest] Help with QSslSocket race condition?

[Henry Skoglund]

On 23/10/2018 20.43, Israel Brewster wrote:
> I am using QSslSockets with a QTcpServer to create a simple 
> client/server program (is there a higher level API I could use 
> instead?). I have subclassed QTcpServer to create my server side sockets 
> as QSslSockets and call startServerEncryption(), but otherwise it is 
> just a standard QTcpServer.
> 
> When a socket first connects to the server, I make the following connection:
> 
> connect(clientConnection, SIGNAL(disconnected()), 
> clientConnection, SLOT(deleteLater()));
> 
> where clientConnection is the QSslSocket created by the QTcpServer. 
> Later, when using the socket (again on the server), I have the following 
> function to wait for data from the client:
> 
>   ////////////////////////////////////////////////////////////
> /// \brief EZPCore::waitForResponseReady
> /// \param socket
> /// \param msec
> /// \return true=data available, false=timeout/error
> ///
> /// Documentation for QAbstractSocket "waitForReadyRead" says
> /// "This function may fail randomly on Windows. Consider using the event loop and the readyRead()
> /// signal if your software will run on Windows."
> /// so that's what this function does.
> bool EZPCore::waitForResponseReady(QSslSocket *socket,int msec){
>    
>    if(!socket || !socket->isValid() || !socket->isEncrypted() || socket->state()!=QAbstractSocket::ConnectedState)
>        
>    return false; //there is something wrong with the socket, no point in waiting.
> 
>      if(socket->bytesAvailable()>0)
>          return true;//success! Bytes are available!
> 
>      QEventLoop wait;  //don't block other operations while waiting
>      QTimer waitTimeout; //only wait a set length of time.
>      waitTimeout.setSingleShot(true);
> 
>      connect(&waitTimeout,&QTimer::timeout,[&wait]{wait.exit(-1);});
>    
>    connect(socket,&QAbstractSocket::disconnected,[&wait]{wait.exit(-2);}); //catch if the socket closes while waiting.
>    
>    connect(socket,&QAbstractSocket::readyRead,[&wait]{wait.exit();}); //zero exit
> 
>      waitTimeout.start(msec);
>      int result=wait.exec();
>    
>    waitTimeout.stop(); //stop the timer if it is still running (probably unneeded?)
> 
>      if(result<0)
>          return false; //not connected or timeout
> 
>      //should only get here if result is 0, indicating readyRead signal 
> received.
> 
>      //if we have bytes available, then return success (0), even if
>      //we had a timeout
>      if(socket->bytesAvailable()>0) //<--------We frequently crash here, 
> UNLESS I don't call "deleteLater()" on disconnect???
>          return true;
>      return false; //no bytes available
> }
> 
> many times this works flawlessly. However,  I am frequently getting a 
> crash when I call bytesAvailable() on the socket at the end of above 
> function - a crash that is resolved if I don't make the connection to 
> "deleteLater" when the socket is established. This tells me that the 
> socket is getting deleted during this event loop. What I don't 
> understand is why?
> 
> - When I come into this function, I check for 
> QAbstractSocket::ConnectedState on the socket, and return immediately if 
> not connected, which should mean that the socket is still connected when 
> this function starts.
> - I connect the "disconnected" signal of this function up to exit the 
> event loop with a non-zero status should the socket disconnect during 
> the event loop (entirely possible, of course)
> - The function returns immediately if the event loop exits with a 
> non-zero status, and the only way for the event loop to exit with a zero 
> status is for readyRead to be emitted.
> 
> As such, the only way for the crashing line of code to be executed 
> should be for a) the socket to be in a connected state (otherwise the 
> function wouldn't run), b) disconnected signal to NOT to be emitted 
> during the event loop (so still connected), and c) the readyRead signal 
> to be emitted. However, the crashing would indicate that the socket *is* 
> being deleted during the event loop, which indicates that the socket is 
> disconnecting. And yet I still get a readyRead signal?
> 
> Incidentally, If I connect the destroyed signal to the event loop exit, 
> I *do* catch that. So I can see the socket being destroyed, but not 
> being disconnected. So obviously I have an architecture issue here: the 
> socket is closing and being destroyed while I am still trying to read 
> from it. How can I fix this?
> 

Hi, it could be that a new connection arrives not inside Qt's normal 
event loop but in your wait.exec() call, i.e. some kind of recursion 
could occur, I mean EZPCore::waitForResponseReady() calls wait.exec() 
which calls another EZPCore::waitForResponseReady()...

The above is just a guess, but why not let Qt run its vanilla event loop 
and refactor your code into 2 separate functions, one that takes care of 
the disconnected signal and one that handles the readyRead signal. 
Neither of them should require an eventloop inside them.

If you need to keep track of timeouts, you could use the QTimer but 
connected to a separate function. Or just let the TCP protocol run its 
course and disconnect automatically for you...

Rgrds Henry