Re: [PATCH] x86/mm: Rewrite sme_populate_pgd() in a more sensible way

[Tom Lendacky]

On 12/4/2017 8:57 AM, Kirill A. Shutemov wrote:
> On Mon, Dec 04, 2017 at 08:19:11AM -0600, Tom Lendacky wrote:
> > On 12/4/2017 5:23 AM, Kirill A. Shutemov wrote:
> > > sme_populate_pgd() open-codes a lot of things that are not needed to be
> > > open-coded.
> > >
> > > Let's rewrite it in a more stream-lined way.
> > >
> > > This would also buy us boot-time switching between support between
> > > paging modes, when rest of the pieces will be upstream.
> >
> > Hi Kirill,
> >
> > Unfortunately, some of these can't be changed.  The use of p4d_offset(),
> > pud_offset(), etc., use non-identity mapped virtual addresses which cause
> > failures at this point of the boot process.
>
> Wat? Virtual address is virtual address. p?d_offset() doesn't care about
> what mapping you're using.

Yes it does.  For example, pmd_offset() issues a pud_page_addr() call,
which does a __va() returning a non-identity mapped address (0xffff88...). 
 Only identity mapped virtual addresses have been setup at this point, so
the use of that virtual address panics the kernel.

Thanks,
Tom

> > Also, calls such as __p4d(), __pud(), etc., are part of the paravirt
> > support and can't be used yet, either.
>
> Yeah, I missed this. native_make_p?d() has to be used instead.
>
> > I can take a closer look at some of the others (p*d_none() and
> > p*d_large()) which make use of the native_ macros, but my worry would be
> > that these get changed in the future to the non-native calls and then
> > boot failures occur.
>
> If you want to avoid paravirt altogher for whole compilation unit, one
> more option would be to put #undef CONFIG_PARAVIRT before all includes.
> That's hack, but it works. We already use this in arch/x86/boot/compressed
> code.