In the ancient days of V7 UNIX, this problem came up in a related way:
processes engaged in DMA I/O were unkillable and unswappable (after all, it
would be bad for the DMA to complete into memory that doesn't belong to the
initiator any more) by dint of having their scheduler priority set below
PZERO (not to be confused with "nice"). Literally, the signal stuff
wouldn't deliver any signal to such a process, and the swapper stuff passed
over such processes - they were "inelegible" for swapping.

I would be surprised if this mechanism were not still in place.

	Erik <fair@clock.org>