Thats an application of the "process can do is own shutdown logic", just
applied to the kernel. You swap everything out (SCO guarantees it will all
fit - we'd have to keep an extra reserve for that as we handle swap
differently). You write out the entire kernel memory image, all the ram
the lot. When you boot you pick up the log, load the entire image except
for a small workspace area you always keep (remember the kernel code
segment is constant), recover the registers and continue. That kind of
trick isnt hard to do (people were doing it with game 'backup' utilities
on 8bit micros).
People take that further at the high end continous running systems as you
install a kernel upgrade, save the entire system state to swap, switch
to the new kernel and recover it - with no down time and not even dropping
network connections.
Alan