Re: [PATCH v5 11/12] x86/paravirt: switch functions with custom code to ALTERNATIVE

[Jürgen Groß]

On 08.03.21 19:30, Borislav Petkov wrote:
> On Mon, Mar 08, 2021 at 01:28:43PM +0100, Juergen Gross wrote:
> > diff --git a/arch/x86/include/asm/paravirt.h b/arch/x86/include/asm/paravirt.h
> > index 36cd71fa097f..04b3067f31b5 100644
> > --- a/arch/x86/include/asm/paravirt.h
> > +++ b/arch/x86/include/asm/paravirt.h
> > @@ -137,7 +137,8 @@ static inline void write_cr0(unsigned long x)
> >   
> >   static inline unsigned long read_cr2(void)
> >   {
> > -	return PVOP_CALLEE0(unsigned long, mmu.read_cr2);
> > +	return PVOP_ALT_CALLEE0(unsigned long, mmu.read_cr2,
> > +				"mov %%cr2, %%rax;", ~X86_FEATURE_XENPV);
>
> Just some cursory poking first - indepth review later.
>
> Do I see this correctly that the negated feature can be expressed with, to use
> this example here:
>
> 	ALTERNATIVE_TERNARY(mmu.read_cr2, X86_FEATURE_XENPV, "", "mov %%cr2, %%rax;");
>
> ?

No.

This would leave the Xen-pv case with a nop, while we need it to call
mmu.read_cr2().

In the Xen-pv case there must be _no_ alternative patching in order to
have the paravirt patching do its patching (indirect->direct call).

This is exactly the reason why I need to "not feature".

The only other solution I can think of would be a "split static_call"
handling using ALTERNATIVE_TERNARY():

ALTERNATIVE_TERNARY(initial_static_call(mmu.read_cr2),
                    X86_FEATURE_XENPV,
                    final_static_call(mmu.read_cr2),
                    "mov %%cr2, %%rax;");

with initial_static_call() doing an indirect call, while
final_static_call() would do a direct call.

Not sure we really want that.


Juergen

Attachment: OpenPGP_0xB0DE9DD628BF132F.asc
Description: application/pgp-keys

Attachment: OpenPGP_signature
Description: OpenPGP digital signature