Alan Cox wrote:
> It would be nice not to export it at all or to protect it, paravirt_ops
> is a rootkit authors dream ticket. I'm opposed to paravirt_ops until it
> is properly protected, its an unpleasantly large security target if not.
>
Do you have an example of an attack which would become significantly
easier with pv_ops in use? I agree it might make a juicy target, but
surely it is just a matter of degree given that any attacker who can get
to pv_ops can do pretty much anything else.
> It would be a lot safer if we could have the struct paravirt_ops in
> protected read-only const memory space, set it up in the core kernel
> early on in boot when we play "guess todays hypervisor" and then make
> sure it stays in read only (even to kernel) space.
>
Yes, I'd thought about doing something like that, but as Arjan pointed
out, nothing is actually read-only in the kernel when using a 2M
mapping. It's also ameliorated by the fact that some of the entrypoints
are never used at runtime, because the code has been patched inline (but
I don't think it would ever be desirable to patch every entrypoint,
since some are just not worth the effort, complexity or obfuscation
which result from patching).
J
