In message <20050517.135630.48490311.imp@bsdimp.com>,
Warner Losh writes:

>> Yes; your old patch for i386 SMP effectively collapsed all SPLs to a
>> single level. That way, no driver should interrupt another.
>
>Make careful measurements as to how this impacts interrupt latency.
>Having the serial driver interrupt the audio driver to process
>characters is a good thing (16550's have a much lower latency
>tolerance than do most audio devices).  If too much of the kernel
>blocks interrupts entirely, this can create starvation for things like
>16550 RX FIFOs.

That goes without saying. I've ben dealing with that since DZ-11s and
4856es with 16450s.  Yamamoto-san's patch was for experimentation,
labelled: for someone interested to try).

The biggest --- perhaps the only -- unresolved technical issue from
last time this came up centered around locking and SPL issues.  My
recollection is that there was a near-consensus that we'd need to move
to a total ordering of locks and SPLs; but I may well be misremembering.