The specific example qouted was using the enhanced UART - hence the framing error detection.

Also the protocol had been designed to return only one byte of data, and used auto address mode, so the interrupt would only happen if there was data for that particular address - hence the chance of waiting in the interrupt forever was remote. I guess some kind of timer to avoid that would be good though.

The point is, it was designed to sit there and wait for the data. If there was more than one byte, then exiting the interrupt as soon as possible would be more sensible.