Re: Possible software flow problem in serial_core

[Peter Hurley <peter@hurleysoftware.com>]

[ +cc linux-serial ]

On 03/20/2014 09:44 AM, Bollinger, Seth wrote:
> Hello All,
>
> We’ve recently run into a possible problem with software flow control
> handling in the serial_core layer.
>
> Here is the scenario:
>
> 1. Transmit from uart to remote device
> 2. Remote device sends us an XOFF
> 3. The tty layer receives the XOFF
> 4. stop_tty() - The uart transmitter is stopped (ops->stop_tx) just as the
> serial_core ring is cleared (this could trap a few bytes in the fifo).
> 5. Remote device sends us an XON
> 6. The tty layer receives the XON
> 7. start_tty() - However, the serial_core ring is empty, so the call to
> start the uart transmitter (uart_start:ops->start_tx) is skipped
>
> Any window, however small, could leave bytes stuck in the transmitter
> forever -- particularly if there will be no further transmission until
> receiving a response.
>
> I can't find any functionality in the drivers that accounts for this
> possibility.  Can you help me find how Linux serial drivers manage this
> eventuality?
>
> Thanks,
>
> Seth

What hardware is this?

A 16550 transmitter doesn't shut-off at uart->stop_tx(), so the remaining
data in the transmitter FIFO is still transmitted.

Some hardware has (relatively) big FIFOs and allows the transmitter to be
shutoff, but these should unconditionally re-enable the transmitter without
clearing the xmit fifo.

Are you seeing lost data? Maybe it's because soft flow-control is to
slow to shutoff the sender?

Regards,
Peter Hurley