Re: [RFC PATCH] x86/insn: support decode MOVSXD instruction for MMIO

[Tom Lendacky]

On 3/31/23 10:25, Kirill A. Shutemov wrote:
> On Fri, Mar 31, 2023 at 09:33:31AM -0500, Tom Lendacky wrote:
> > On 3/31/23 09:09, Kirill A. Shutemov wrote:
> > > On Fri, Mar 31, 2023 at 08:40:30AM -0500, Tom Lendacky wrote:
> > > > On 3/31/23 05:06, Kirill A. Shutemov wrote:
> > > > > On Fri, Mar 31, 2023 at 08:49:48AM +0000, David Laight wrote:
> > > > > > From: Wu Zongyong
> > > > > > > Sent: 31 March 2023 03:24
> > > > > > >
> > > > > > > On Thu, Mar 30, 2023 at 03:39:51PM +0300, kirill.shutemov@xxxxxxxxxxxxxxx wrote:
> > > > > > > > On Wed, Mar 29, 2023 at 10:59:37AM +0800, Wu Zongyong wrote:
> > > > > > > > > It seems MOVSXD which opcode is 0x63 is not handled, support
> > > > > > > > > to decode it in insn_decode_mmio().
> > > > > > > >
> > > > > > > > Do you have a particular user in mind?
> > > > > > > To be honest, I don't find a specific user which uses the MOVSXD.
> > > > > > >
> > > > > > > But both Intel and AMD's instructions reference contains MOVSXD and lots
> > > > > > > of MOVSXD instructions occur when I "objdump -S vmlinux", so I think it
> > > > > > > may be useful to support it in insn_decode_mmio().
> > > > > > >
> > > > > > > Are there some special consideration about this instruction?
> > > > > >
> > > > > > It is a sign-extending memory read (32bit to 64bit).
> > > > > > You pretty much never want to do that to a device register.
> > > > > > Also kernel code should be using readl() (etc) which do
> > > > > > unsigned reads.
> > > > > > So they should never happen for mmio.
> > > > > >
> > > > > > Of course, if you mmap() PCIe space directly into a program's
> > > > > > address space anything might happen ...
> > > > >
> > > > > There are two users of the interface: TDX and SEV. TDX doesn't allow
> > > > > userspace MMIO. SEV *seems* allows it, but I am not sure how it is safe.
> > > > >
> > > > > Tom?
> > > >
> > > > The insn_decode_mmio() function is only called by the SEV/TDX related code
> > > > and is specifically MMIO oriented. As David said, this instruction is likely
> > > > not being used for that in the kernel. If we come across a case where this
> > > > is used, we can look at how it is being used in that situation and it can be
> > > > addressed then.
> > >
> > > I was asking if SEV supports userspace MMIO. And if yes, how do you make
> > > it safe?
> >
> > No, SEV doesn't support userspace MMIO.
>
> But where do you filter out userspace MMIO? AFAICS, it goes straight from
> from #VC to insn_decode_mmio(). Hm?

The userspace mapping would have the encryption bit set and MMIO to 
encrypted memory is detected and not allowed.

Thanks,
Tom

>