Re: [PATCH v5 6/6] x86/tdx: Implement MOVS for MMIO

[Kirill A. Shutemov]

On Wed, Aug 28, 2024 at 12:44:36PM +0200, Alexey Gladkov wrote:
> From: "Alexey Gladkov (Intel)" <legion@xxxxxxxxxx>
> 
> Add emulation of the MOVS instruction on MMIO regions. MOVS emulation
> consists of dividing it into a series of read and write operations,
> which in turn will be validated separately.
> 
> This implementation is based on the same principle as in SEV. It splits
> MOVS into separate read and write operations, which in turn can cause
> nested #VEs depending on which of the arguments caused the first #VE.
> 
> The difference with the SEV implementation is the execution context. SEV
> code is executed in atomic context. Exception handler in TDX is executed
> with interrupts enabled. That's why the approach to locking is
> different. In TDX, mmap_lock is taken to verify and emulate the
> instruction.
> 
> Another difference is how the read and write instructions are executed
> for MOVS emulation. While in SEV each read/write operation returns to
> user space, in TDX these operations are performed from the kernel
> context.

It looks like SEV only returns to userspace to retry the instruction after
stepping on failed __get_user()/__put_user(), unrolling back to
vc_raw_handle_exception() and handling page fault there.

But I'm not sure what happens with #VC inside vc_read_mem() and
vc_write_mem(). Can the #VC exception be nested? Tom?

-- 
  Kiryl Shutsemau / Kirill A. Shutemov