Of course, if Alan is OK with this, I'm more than OK too. :-)

It may well be better.

(But the polling routines would need to clear all interrupts, not
just rx/tx. For example, if the controller indicated some error, and
nobody clears it, then we'll start reentering infinitely.)

For a lot of devices and platforms you'd probably mask them instead ?

quoted

If you use a clear_irqs callback, you can drop characters if
one arrives between the last character buffer read and calling
clear_irqs.

Only if we call clear_irqs() after reading the characters, but we do
it before. So if new characters are available, we will reenter NMI,
which is OK.

Recursively or not... again you get platform specific magic in places
we don't want.

In the driver poll method for most uarts is going to be at least buried
in what is usually arch specific uarts.