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

[Tom Lendacky]

On 12/4/2017 12:50 PM, Kirill A. Shutemov wrote:
> On Mon, Dec 04, 2017 at 12:33:01PM -0600, Tom Lendacky wrote:
> > On 12/4/2017 10:34 AM, Kirill A. Shutemov wrote:
> > > On Mon, Dec 04, 2017 at 04:00:26PM +0000, Tom Lendacky wrote:
> > > > 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.
> > >
> > > Stupid me. You are right.
> > >
> > > What about something like this:
> > >
> > > diff --git a/arch/x86/mm/mem_encrypt.c b/arch/x86/mm/mem_encrypt.c
> > > index d9a9e9fc75dd..65e0d68f863f 100644
> > > --- a/arch/x86/mm/mem_encrypt.c
> > > +++ b/arch/x86/mm/mem_encrypt.c
> > > @@ -12,6 +12,23 @@
> > >    #define DISABLE_BRANCH_PROFILING
> > > +/*
> > > + * Since we're dealing with identity mappings, physical and virtual
> > > + * addresses are the same, so override these defines which are ultimately
> > > + * used by the headers in misc.h.
> > > + */
> > > +#define __pa(x)  ((unsigned long)(x))
> > > +#define __va(x)  ((void *)((unsigned long)(x)))
> >
> > No, you can't do this.  There are routines in this file that are called
> > after the kernel has switched to its standard virtual address map where
> > this definition of __va() will likely cause a failure.
>
> Let's than split it up into separate compilation unit.
>
> > > +/*
> > > + * Special hack: we have to be careful, because no indirections are
> > > + * allowed here, and paravirt_ops is a kind of one. As it will only run in
> > > + * baremetal anyway, we just keep it from happening. (This list needs to
> > > + * be extended when new paravirt and debugging variants are added.)
> > > + */
> > > +#undef CONFIG_PARAVIRT
> > > +#undef CONFIG_PARAVIRT_SPINLOCKS
> >
> > I'd really, really like to avoid doing something like this.
>
> Any other proposals?
>
> Current code is way too hairy and hard to modify.

I'd like to hear some other opinions, but if there are none, I don't see
an issue with splitting this up, I guess.  I just dislike doing some of
these hacks, but if it makes things cleaner and simpler to understand,
then I guess I can't really object.

Thanks,
Tom

>