Re: [PATCHv7 12/19] x86/mm: Adjust virtual address space layout in early boot.

From: Kirill A. Shutemov
Date: Thu Sep 28 2017 - 10:28:17 EST


On Thu, Sep 28, 2017 at 03:38:38PM +0200, Ingo Molnar wrote:
>
> * Kirill A. Shutemov <kirill@xxxxxxxxxxxxx> wrote:
>
> > On Thu, Sep 28, 2017 at 10:31:55AM +0200, Ingo Molnar wrote:
> > >
> > > * Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx> wrote:
> > >
> > > > We need to adjust virtual address space to support switching between
> > > > paging modes.
> > > >
> > > > The adjustment happens in __startup_64().
> > >
> > > > +#ifdef CONFIG_X86_5LEVEL
> > > > + if (__read_cr4() & X86_CR4_LA57) {
> > > > + pgtable_l5_enabled = 1;
> > > > + pgdir_shift = 48;
> > > > + ptrs_per_p4d = 512;
> > > > + }
> > > > +#endif
> > >
> > > So CR4 really sucks as a parameter passing interface - was it us who enabled LA57
> > > in the early boot code, right? Couldn't we add a flag which gets set there, or
> > > something?
> >
> > It's not necessary that we enabled LA57. At least I tried to write code
> > that doesn't assume this. We enable it if bootloader haven't done this
> > already for us.
> >
> > What is so awful about using CR4 as passing interface? It's one-time
> > check, so performance shouldn't be an issue.
>
> As a starter, this code is in generic x86 code [choose_random_location()], is this
> CR4 bit known to AMD as well and is it guaranteed to be sane across all x86 CPUs?
> I don't think so.

It's architectural thing, so it's consistent across all x86
implementations.

> CR4 is a poor interface to pass CPU features through. Generaly we try to enumerate
> CPU features via CPUID, and/or enable synthetic CPU features in certain cases, and
> work from there.

Okay, has_cpuflag(X86_FEATURE_LA57) seems would do.

--
Kirill A. Shutemov