TCP: timing may cause socket to enter invalid state which cannot reconnect

Libraries
1

I’m running FreeRTOS in virtual environment that creates unusual timing issues which is probably why I ran into this but I believe this could also happen on real hardware.

The scenario is that I lose a connection and I try to shutdown the socket and reconnect. This fails because the socket is already in the wrong state and FreeRTOS_shutdown returns pdFREERTOS_ERRNO_EOPNOTSUPP causing the AWS stack to abort execution and failing to reconnect later as the the socket was not fully shutdown.

The reason pdFREERTOS_ERRNO_EOPNOTSUPP is returned is because the socket that loses connection can have its state change away from established before FreeRTOS_shutdown is called.

    BaseType_t FreeRTOS_shutdown( Socket_t xSocket,
                                  BaseType_t xHow )
    {
        FreeRTOS_Socket_t * pxSocket = ( FreeRTOS_Socket_t * ) xSocket;
        BaseType_t xResult;

        if( prvValidSocket( pxSocket, FREERTOS_IPPROTO_TCP, pdTRUE ) == pdFALSE )
        {
            /*_RB_ Is this comment correct?  The socket is not of a type that
             * supports the listen() operation. */
            xResult = -pdFREERTOS_ERRNO_EOPNOTSUPP;
        }
        else if( pxSocket->u.xTCP.ucTCPState != ( uint8_t ) eESTABLISHED )
        {
            /*_RB_ Is this comment correct?  The socket is not of a type that
             * supports the listen() operation. */
            xResult = -pdFREERTOS_ERRNO_EOPNOTSUPP;
        }

The condition else if( pxSocket->u.xTCP.ucTCPState != ( uint8_t ) eESTABLISHED ) is entered.

The comment inside the condition makes no sense to me. It is fine for the FreeRTOS_shutdown to not do anything in this case but it should return a more useful error, something like maybe pdFREERTOS_ERRNO_ENOTCONN. This way the stack can disregard the error and interpret it as shutdown not needed so I can continue safely instead of aborting. Right now SOCKETS_Shutdown in lib/amazon_freertos/libraries/abstractions/secure_sockets/freertos_plus_tcp/iot_secure_sockets.c has no way to distinguish between the two and passes an error further up the stack causing execution to be aborted leaving the socket in unusable state.

2

Apologies for the delay in your post showing up - it went into moderation. It will take a short while to investigate what you are reporting - thanks for the detail.

3

Hello, I just replied to your post reporting an issue on github.

Does that answer your question?

4

To add to what @htibosch has said, if SOCKETS_shutdown returns a zero value then it means that the shutdown is in progress. It is a stream of events that need to happen before the connection actually closes. Thus, to actively shutdown a connection, one must use code similar to what Hein added to the issue. I am adding his code here (with a minor change) for simplicity:

TickType_t xTicksToWait = MAX_TIME_TO_WAIT;
TimeOut_t xTimeOut;

FreeRTOS_shutdown( xSocket, FREERTOS_SHUT_RDWR );
vTaskSetTimeOutState( &xTimeOut );

while( xTaskCheckForTimeOut( &xTimeOut, &xTicksToWait ) == pdFALSE )
{
    int32_t rc = FreeRTOS_recv( xSocket )
    if( ( rc < 0 ) && ( rc != -pdFREERTOS_ERRNO_EAGAIN ) )
    {
        /* The TCP connection is broken. */
        break;
    }
}

In the above example, you can replace FreeRTOS_xxx() calls with the corresponding SOCKETS_xxx() calls.

To actually close a socket, one then needs to call SOCKETS_close()/FreeRTOS_closesocket().

Let me know if that helps. :slight_smile:

5

Thank you @kanherea for making the example complete. It must be said that the code assumes that the call to FreeRTOS_recv() is blocking, i.e. FREERTOS_SO_RCVTIMEO has been set to a non-zero value.

I just found out that @paul-szczepanek-arm is using older code. Today’s version is like this:

if( prvValidSocket( pxSocket, FREERTOS_IPPROTO_TCP, pdTRUE ) == pdFALSE )
{
    /*_RB_ Is this comment correct?  The socket is not of a type that
     * supports the listen() operation. */
    xResult = -pdFREERTOS_ERRNO_EOPNOTSUPP;
}
else if( pxSocket->u.xTCP.ucTCPState != ( uint8_t ) eESTABLISHED )
{
    /* The socket is not connected. */
    xResult = -pdFREERTOS_ERRNO_ENOTCONN;
}
else

The comment in the first if statement can be extended to :

/* Either "pxSocket" is not a valid pointer, or the
* socket is not bound, or it is not a TCP socket. */

But once again: FreeRTOS_shutdown()should only be called on a connected socket. The shutting down is not immediate, it takes time to exchange a FIN packets.
During the shutting down, FreeRTOS_recv() might still return data received from the peer.

6

Ah, didn’t realise this change has already been made, I guess we came to the same conclusion independently. Apologies this escaped my attention.

It’s not my code that was failing, it’s the AWS ota agent code. Thank you.