Re: [PATCH v8 15/16] x86/virt/tdx: Flush cache in kexec() when TDX is enabled

[Huang, Kai]

On Tue, 2023-01-10 at 16:30 -0800, Hansen, Dave wrote:
> On 1/10/23 16:13, Huang, Kai wrote:
> > On Tue, 2023-01-10 at 07:27 -0800, Dave Hansen wrote:
> ...
> > > Think about it this way: kexec() is modifying persistent (across kexec)
> > > state to get the system ready for the new kernel.  The caches are
> > > persistent state.  Devices have persistent state.  Memory state persists
> > > across kexec().  The memory integrity metadata persists.
> > > 
> > > What persistent state does a conversion to KeyID-0 affect?  It resets
> > > the integrity metadata and the memory contents.
> > > 
> > > Kexec leaves memory contents in place and doesn't zero them, so memory
> > > contents don't matter.  The integrity metadata also doesn't matter
> > > because the memory will be used as KeyID-0 and that KeyID doesn't read
> > > the integrity metadata.
> > 
> > Right.  So I guess we just need to call out the new kernel will use memory as
> > KeyID-0?
> 
> Not even that.
> 
> Say the new kernel wanted to use the memory as KeyID-3.  What would it
> do?  It would *ASSUME* that the memory *WASN'T* KeyID-3.  It would
> convert it to KeyID-3.  That conversion would work from *any* KeyID.
> 
> So:
> 
> 	KeyID-0: OK, because it has no integrity enforcement
> 	KeyID-1: OK, new kernel will convert the page
> 	KeyID-2: OK, new kernel will convert the page
> 	...
> 	KeyID-$MAX: OK, new kernel will convert the page
> 
> So, "OK" everywhere.  Nothing to do... anywhere.
> 
> Either I'm totally missing how this works, or you're desperately trying
> to make this more complicated than it is.
> 

You are right.  The page conversion must do MOVDIR64B first even converting the
page from KeyID 0.  I was wrongly thinking when converting from KeyID 0 we don't
need to do MOVDIR64B.  My bad.

Sorry for the noise.  Thanks for your time.  I'll remove all those staff.