Re: [PATCHv7 00/14] mm, x86/cc: Implement support for unaccepted memory

[Kirill A. Shutemov]

On Mon, Jun 27, 2022 at 01:54:45PM +0200, Ard Biesheuvel wrote:
> On Mon, 27 Jun 2022 at 13:30, Kirill A. Shutemov
> <kirill.shutemov@xxxxxxxxxxxxxxx> wrote:
> >
> > On Fri, Jun 24, 2022 at 10:37:10AM -0600, Peter Gonda wrote:
> > > On Tue, Jun 14, 2022 at 6:03 AM Kirill A. Shutemov
> > > <kirill.shutemov@xxxxxxxxxxxxxxx> wrote:
> > > >
> > > > UEFI Specification version 2.9 introduces the concept of memory
> > > > acceptance: some Virtual Machine platforms, such as Intel TDX or AMD
> > > > SEV-SNP, requiring memory to be accepted before it can be used by the
> > > > guest. Accepting happens via a protocol specific for the Virtual
> > > > Machine platform.
> > > >
> > > > Accepting memory is costly and it makes VMM allocate memory for the
> > > > accepted guest physical address range. It's better to postpone memory
> > > > acceptance until memory is needed. It lowers boot time and reduces
> > > > memory overhead.
> > > >
> > > > The kernel needs to know what memory has been accepted. Firmware
> > > > communicates this information via memory map: a new memory type --
> > > > EFI_UNACCEPTED_MEMORY -- indicates such memory.
> > > >
> > > > Range-based tracking works fine for firmware, but it gets bulky for
> > > > the kernel: e820 has to be modified on every page acceptance. It leads
> > > > to table fragmentation, but there's a limited number of entries in the
> > > > e820 table
> > > >
> > > > Another option is to mark such memory as usable in e820 and track if the
> > > > range has been accepted in a bitmap. One bit in the bitmap represents
> > > > 2MiB in the address space: one 4k page is enough to track 64GiB or
> > > > physical address space.
> > > >
> > > > In the worst-case scenario -- a huge hole in the middle of the
> > > > address space -- It needs 256MiB to handle 4PiB of the address
> > > > space.
> > > >
> > > > Any unaccepted memory that is not aligned to 2M gets accepted upfront.
> > > >
> > > > The approach lowers boot time substantially. Boot to shell is ~2.5x
> > > > faster for 4G TDX VM and ~4x faster for 64G.
> > > >
> > > > TDX-specific code isolated from the core of unaccepted memory support. It
> > > > supposed to help to plug-in different implementation of unaccepted memory
> > > > such as SEV-SNP.
> > > >
> > > > The tree can be found here:
> > > >
> > > > https://github.com/intel/tdx.git guest-unaccepted-memory
> > >
> > > Hi Kirill,
> > >
> > > I have a couple questions about this feature mainly about how cloud
> > > customers can use this, I assume since this is a confidential compute
> > > feature a large number of the users of these patches will be cloud
> > > customers using TDX and SNP. One issue I see with these patches is how
> > > do we as a cloud provider know whether a customer's linux image
> > > supports this feature, if the image doesn't have these patches UEFI
> > > needs to fully validate the memory, if the image does we can use this
> > > new protocol. In GCE we supply our VMs with a version of the EDK2 FW
> > > and the customer doesn't input into which UEFI we run, as far as I can
> > > tell from the Azure SNP VM documentation it seems very similar. We
> > > need to somehow tell our UEFI in the VM what to do based on the image.
> > > The current way I can see to solve this issue would be to have our
> > > customers give us metadata about their VM's image but this seems kinda
> > > burdensome on our customers (I assume we'll have more features which
> > > both UEFI and kernel need to both support inorder to be turned on like
> > > this one) and error-prone, if a customer incorrectly labels their
> > > image it may fail to boot.. Has there been any discussion about how to
> > > solve this? My naive thoughts were what if UEFI and Kernel had some
> > > sort of feature negotiation. Maybe that could happen via an extension
> > > to exit boot services or a UEFI runtime driver, I'm not sure what's
> > > best here just some ideas.
> >
> > Just as an idea, we can put info into UTS_VERSION which can be read from
> > the built bzImage. We have info on SMP and preeption there already.
> >
> 
> Instead of hacking this into the binary, couldn't we define a protocol
> that the kernel will call from the EFI stub (before EBS()) to identify
> itself as an image that understands unaccepted memory, and knows how
> to deal with it?
> 
> That way, the firmware can accept all the memory on behalf of the OS
> at ExitBootServices() time, unless the OS has indicated there is no
> need to do so.

I agree it would be better. But I think it would require change to EFI
spec, no?

-- 
 Kirill A. Shutemov