Re: [PATCH v8 8.2/9] x86/virt/tdx: Reduce TDMR's reserved areas by using CMRs to find memory holes

["Huang, Kai" <kai.huang@intel.com>]

On Mon, 2024-12-09 at 18:46 -0800, Dan Williams wrote:
> Huang, Kai wrote:
> [..]
> > > So in the end, I buy that the CMR's have something to offer here. But I
> > > think that "why" I mentioned above casts doubt on whether
> > > for_each_mem_pfn_range() is the right primitive on which to build the
> > > TDX memblocks in the first place.
> > 
> > We can change to just use CMRs as TDX memory blocks, i.e., always cover all CMRs
> > in TDMRs, but this will have much wider impact.
> > 
> > The main concern is the PAMT allocation: PAMT is allocated from page allocator,
> > but the CMRs -- the RAM as defined by the platform and the TDX module - - can
> > cover more, and sometimes much more, regions than the regions end up to the page
> > allocator.
> > 
> > E.g., today we can use 'memmap=' to reserve part of memory for other purpose. 
> > And in the future CMRs may cover CXL memory regions which could be much larger
> > IIUC.
> 
> Please do not point to memmap= as a reason to complicate the TDX
> initialization. memmap= is a debug / expert feature where the user gets
> to keep all the pieces if they get it wrong.
> 
> Please do not point to theoretical CXL futures as a reason not to do the
> right thing in TDX initialization. The CXL TEE Security Protocol makes
> CXL memory indistinguishable from DDR. It is layering violation for TDX
> module initialization to add complexity and policy assumptions as if it
> knows better than the published CMRs what memory resources are available
> in the platform.

OK.  Thanks for all the feedback.

> 
> Please drop my reviewed-by on this patch until we have a solution and a
> simple story for the recently discovered problems that CMR enumeration
> solves. This includes reserve-area population and disambiguating
> reserve-area enumeration from late-to-online memory resources.

OK.  I would like to get the solution agreed.  IIUC you prefer to just using
CMRs to build TDX-usable memory regions when constructing TDMRs.  Please let me
know if I have misunderstanding.

Hi Dave,

Please let me know if you are OK with this solution?