Linux 1.3.(some high numbers) kernels have fairly good real time performance. Applications can use POSIX real time scheduling with absolute priorities higher than any process. I ran the realtime test programs associated with some program (don't remember the name) for POSIX realtime process testing and noted that the longest time that the kernal locked out the realtime application never exceeded 135 microseconds on my Pentium 100. I assume this means that the longest kernel call tested didn't exceed that number...some other cases might go longer.
I would like to run a test where a realtime process ran on a precision timed interrupt at the same time the overall Linux kernel performed it full range of functions. This realtime process sole job would be to measure interrupt latency and histogram them, probably through the /proc filesyste. My learning curve for this task would be quite steep but if someone would like to take on this task for a little education, I'd be interested in the results.