Actually, I started working on these lines, mostly to learn how things
work. I'm not all that far though, and I'm sure someone
experienced in the code could be as far as I am in a day. At this point
I'm almost ready to conclude that this needs to be done using userfs
if the kernel can't page.
My concept was:
- use swap space for storage. This is like Sun's approach, in that
one device can be used for both purposes, saving disk space as a
side benefit.
- It can grow and shrink. Unused swap space is available for
tempfs, and vice versa.
- If done in userfs, then read/write from the swap would be automatic.
- It would bypass the buffer cache. Not too important to the plan,
but by using mm so that recently accessed pages are in memory,
this is similar to what the buffer cache gives anyway.