Fix: RX Stall may occur locking the system

[ithinuel]

A the end of a writing operation with large amount of data, the tx fifo may get filled and then process enters the wait_idle method.

However, while the TX fifo gets emptied, the RX fifo gets filled and it may cause an RX stall, effectively stopping the state machine, and preventing the idle condition from ever raising.

[9names]

Lgtm! Nice explanation of the problem

[jannic]

Definitely a good catch!

What I don't like is that the new code just drops RX data, but does that unreliably: It depends on timing whether all incoming data is dropped before the function returns. For a public function I'd search for a better solution which either doesn't drop RX (return some error instead and let the caller handle it), or always drops all RX data. (The latter can't be 100% deterministic as it can't prevent new RX from appearing between the last check and the function return.)

For this library, I'd say this is fine: There are only two callers of this function, one clears RX anyways, and the other one is in some error branch on finish.

Regarding the error case: If one transaction ends with an error in finish, a subsequent call to setup could panic because of this code:

        // flush read fifo
        assert!(self.rx.read().is_none(), "rx FIFO shall be empty");

Shouldnt setup actually flush the RX FIFO instead of panicing?

In any case: This is independent from this pull request → LGTM.