Xenomai might be fast enough with some coprocessing added if necessary. It makes your RT threads have essentially "normal" POSIX RT api, but runs them on separate kernel that can respond to things like interrupts much, much faster (including stable timeslicing)