Re: [RFC PATCH v2] x86/sev: enforce RIP-relative accesses in early SEV/SME code

[Kevin Loughlin]

On Mon, Jan 15, 2024 at 2:12 AM Kirill A. Shutemov
<kirill.shutemov@xxxxxxxxxxxxxxx> wrote:
>
> On Fri, Jan 12, 2024 at 10:29:36AM -0800, Kevin Loughlin wrote:
> >
> > Per my tests, yes we can; I replaced the fixup_*() functions with
> > GET_RIP_RELATIVE_PTR()/PTR_TO_RIP_RELATIVE_PTR(), and guests with and
> > without SEV, SEV-ES, and SEV-SNP all successfully booted under both
> > clang and gcc builds.
>
> BTW, do we need both macros? Caller can do &var, right?

While I don't think the caller doing "&var" would work without passing
it as a separate argument like `GET_RIP_RELATIVE_PTR(var, &var)` (as
we would still need the original var's string name in the macro for
the inline assembly `#var(%%rip)`), we should nonetheless be able to
merge both into a single macro with one "var" argument. Specifically,
the only current difference between the macros is the input operand
constraint, and GET_RIP_RELATIVE_PTR()'s constraint will work for
both. I will make this change in v3.

> If we are okay with single macro, maybe rename it to RIP_RELATIVE_PTR().

With the merge into a single macro (and upon thinking about the
macro's behavior), I have a slight preference for
`RIP_RELATIVE_ADDR()` in v3 because it makes it clearer that the macro
behaves like the address-of operator "&" (just guaranteeing the use of
RIP-relative addressing to obtain the address). However, I'm happy to
go with RIP_RELATIVE_PTR() if you feel that's better.

> One other thing: I see you sprinkle casts to for every use of the macros.
> But why? void* can cast to any other pointer without explicit casting.

You're right; the casting is unnecessary. I'll eliminate it in v3. Thanks.

> > On Fri, Jan 12, 2024 at 4:17 AM Kirill A. Shutemov
> > <kirill.shutemov@xxxxxxxxxxxxxxx> wrote:
> > >
> > > Also, is there any reason why GET_RIP_RELATIVE_PTR() and
> > > PTR_TO_RIP_RELATIVE_PTR() have to be macros? Inline functions would be
> > > cleaner.
> >
> > I used macros because we need to use both the global variable itself
> > and the global variable's string name (obtained via #var in the macro)
> > in the inline assembly. As a secondary reason, the macro also avoids
> > the need to provide separate functions for each type of variable for
> > which we'd like to get RIP-relative pointers (ex: u64, unsigned int,
> > unsigned long, etc.).
>
> If we do it only on pointers, wouldn't void * -> void * be enough?

Only using pointers would indeed eliminate the secondary factor as a
reason to use macros. However, the primary motivation for a macro
would remain: we still need the string name of the variable for the
inline assembly.