Chris Wright wrote:
>
> +void __devinit machine_specific_modify_cpu_capabilities(struct cpuinfo_x86 *c)
> +{
> + clear_bit(X86_FEATURE_VME, c->x86_capability);
> + clear_bit(X86_FEATURE_DE, c->x86_capability);
> + clear_bit(X86_FEATURE_PSE, c->x86_capability);
> + clear_bit(X86_FEATURE_PGE, c->x86_capability);
> + clear_bit(X86_FEATURE_SEP, c->x86_capability);
> + clear_bit(X86_FEATURE_MWAIT, c->x86_capability);
> + if (!(xen_start_info->flags & SIF_PRIVILEGED))
> + clear_bit(X86_FEATURE_MTRR, c->x86_capability);
> + c->hlt_works_ok = 0;
>
That's pretty heinous. Suppose Xen decides to start supporting any of
these features. Now, you need to change all of the Xen modified kernels
to remove pieces of this.
You've effectively introduced completely arbitrary, but most
importantly, static constraints into Linux based on what features Xen
currently supports. This is not productive on either end.
This is the entire point of the VMI interface. You can arbitrarily
enable and disable these features in a hidden layer, call it VMI, call
it Xen-bridge, call it hypercall page, whatever. Using this layer, you
can automatically abstract away exactly this type of nasty, deliberately
deprecating extraneous code. We use this feature exactly to strip away
these bits, completely hidden from the guest, by "trapping" the CPUID
instruction in the VMI layer. The VMI interface as it stands today is
unimportant - it has and will continue to change to accommodate Xen.
But the principles of it are important for maintainability, flexibility,
and future compatibility. One of these principles is to avoid
unbalanced changes like this to the guest kernel.
And we certainly support VME, DE, PSE, PGE, and SEP features, and really
would not like them disabled unconditionally in a virtual environment.
Zach
