Peter Chubb wrote:

In a hosted VMM like LinuxOnLinux or UML, context switch time can be a
major problem (as mmap when repeated for each guest page frame takes a
long time).  One solution is to allow the host kernel to keep a cache of
address space contexts, and switch between them in a single
operation. 
  

Other VMMs which have a large usermode component, like lguest and kvm, 
do maintain two address spaces mapping the same set of pages.  But 
unlike UML (and I guess LoL), the guest mappings are not represented as 
VMAs, but just as a raw processor pagetable.  They need some special 
switcher code to go into that state, so it doesn't look like this would 
be terribly useful for them.

Am I right in presuming that this is really only useful for VMMs which 
want to use mmap/mprotect/munmap for the virtual MMU implementation?

It might be interesting if the two cases could be unified in some way, 
so that the VMMs could use a common usermode mechanism to achieve the 
same end, which is what Carsten was proposing.  But its not obvious to 
me how much common mechanism can be pulled out, since its a pretty 
deeply architecture-specific operation.

    J