Connecting Linux to hypervisors [LWN.net]

Connecting Linux to hypervisors

Posted Aug 12, 2006 17:27 UTC (Sat) by giraffedata (subscriber, #1954)
Parent article: Connecting Linux to hypervisors

The hypervisor ROM thing doesn't make much sense as described. The hypervisor ROM shouldn't be loaded into the kernel by Linux and shouldn't be code maintained by Linux developers. It should exist permanently in the virtual machine's address space -- that's what ROM means. It should be totally out of the control of Linux developers and under the control of the hypervisor developers, and the stability of the interface would flow directly from that fact. Just like traditional ISA BIOS.

Ordinarily, hypervisors just offer new instructions instead of memory you can branch to (i.e. an instruction causes an interrupt which hypervisor code that is invisible to Linux handles), but I suppose hypervisor ROM might be faster or more convenient.

(Log in to post comments)

Connecting Linux to hypervisors

Posted Aug 23, 2006 14:49 UTC (Wed) by Duncan (guest, #6647) [Link]

> [Hypervisor ROM should be j]ust like traditional
> ISA BIOS[, thus not controlled by Linux devs].

You are looking at it from the perspective of guest OS. What about when
Linux is the host OS? That's what the debate is about here.

An entirely user-mode host is slower than a host built with a cooperating
kernel that has exposed certain bits (like interrupt control) directly to
the host application. What is being debated here is what that exposed
interface should look like from the kernel as host side, and whether it
will be nailed hard and fast like most regular user mode interfaces, or
specifically allowed to change, as can most of the kernel other than the
user mode interfaces.

IOW, from the kernel as host perspective, Linux /is/ the hardware-like
hypervisor ROM, with Linux developers therefore responsible for developing
and maintaining that interface. Will it be set in stone as the regular
user interface, or specifically allowed to change, as a regular kernel
interface like that exposed to kernel modules?

Duncan