Re: [tip:efi/core] x86/efi: Unmap EFI boot services code/data regions from efi_pgd

[Ard Biesheuvel]

On Mon, 17 Dec 2018 at 20:48, Prakhya, Sai Praneeth
<sai.praneeth.prakhya@xxxxxxxxx> wrote:
>
> > > > > Hi Thomas and Ingo,
> > > > >
> > > > > I recently noticed that the below commits [1] and [2] are broken
> > > > > when kernel command line argument "efi=old_map" is passed. Sorry!
> > > > > I missed to test this condition prior to sending these patches to mailing list.
> > > > > I am working on a fix and will send it to mailing list as soon as it's ready.
> > > > >
> > > >
> > > > Could you elaborate on the problem please?
> > >
> > > Sure! My bad..
> > >
> > > Little bit of history here:
> > > Boris with this patch set [1] introduced statically mapping EFI
> > > Runtime Services at -4G and also introduced "efi=old_map" to return to
> > > previous EFI functionality which used ioremap and __va(pa).
> > >
> > > [3] and [4] are links to old_map_region()
> > >
> > > The commit 08cfb38f3ef4 ("x86/efi: Unmap EFI boot services code/data
> > > regions from efi_pgd"), unmaps EFI boot services code/data regions
> > > *only* from efi_pgd but efi=old_map maps EFI boot services code/data
> > > regions into swapper_pgd. Also, efi=old_map  uses either
> > > ioremap() or __va(md->phys_addr) to map EFI runtime/boot time services and
> > doesn't use kernel_map_pages_in_pgd().
> > >
> > > So, we need a different unmapping routine to unmap EFI boot services
> > > code/data regions from swapper_pgd if they were mapped using efi=old_map.
> > >
> >
> > For the short term, could we simply make your changes dependent on efi !=
> > old_map? That gives us some time to fix the old_map case properly.
>
> Yes, I think we could and it should work but I hesitated to propose it because
> (as you already noted) it's a short term fix and not the right fix.
>

What is the status here?

> Alternatively, we could also evaluate if we need to support efi=old_map case going further.
> I thought dropping it would be a bad idea because it changes kernel user visible interface
> (because it's a kernel command line argument) and never brought it up.
> Not sure what Thomas, Ingo or Linus might think about dropping a kernel command line
> argument.
>

Dropping a command line argument is not a problem in itself, unless
anyone is actively using it :-)

As far as I can tell, the SGI x86 UV platforms still rely on this, so
we're stuck with it for the foreseeable future.

This means we need a fixes that makes your unmapping code conditional
on !old_memmap. Do you have an ETA for that?