Re: [PATCH] x86/mm: fix regression with huge pages on PAE

From: "Kirill A. Shutemov" <kirill@shutemov.name>
To: Ingo Molnar <mingo@kernel.org>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>,
	Borislav Petkov <bp@alien8.de>,
	hpa@zytor.com, tglx@linutronix.de, mingo@redhat.com,
	akpm@linux-foundation.org, linux-mm@kvack.org,
	linux-kernel@vger.kernel.org, x86@kernel.org, jgross@suse.com,
	konrad.wilk@oracle.com, elliott@hpe.com,
	boris.ostrovsky@oracle.com, Toshi Kani <toshi.kani@hpe.com>,
	Linus Torvalds <torvalds@linux-foundation.org>
Subject: Re: [PATCH] x86/mm: fix regression with huge pages on PAE
Date: Thu, 12 Nov 2015 11:00:19 +0200	[thread overview]
Message-ID: <20151112090018.GA22481@node.shutemov.name> (raw)
In-Reply-To: <20151112085418.GA18963@gmail.com>

On Thu, Nov 12, 2015 at 09:54:18AM +0100, Ingo Molnar wrote:
> 
> * Kirill A. Shutemov <kirill.shutemov@linux.intel.com> wrote:
> 
> > diff --git a/arch/x86/include/asm/page_types.h b/arch/x86/include/asm/page_types.h
> > index c5b7fb2774d0..cc071c6f7d4d 100644
> > --- a/arch/x86/include/asm/page_types.h
> > +++ b/arch/x86/include/asm/page_types.h
> > @@ -9,19 +9,21 @@
> >  #define PAGE_SIZE	(_AC(1,UL) << PAGE_SHIFT)
> >  #define PAGE_MASK	(~(PAGE_SIZE-1))
> >  
> > +#define PMD_PAGE_SIZE		(_AC(1, UL) << PMD_SHIFT)
> > +#define PMD_PAGE_MASK		(~(PMD_PAGE_SIZE-1))
> > +
> > +#define PUD_PAGE_SIZE		(_AC(1, UL) << PUD_SHIFT)
> > +#define PUD_PAGE_MASK		(~(PUD_PAGE_SIZE-1))
> > +
> >  #define __PHYSICAL_MASK		((phys_addr_t)((1ULL << __PHYSICAL_MASK_SHIFT) - 1))
> >  #define __VIRTUAL_MASK		((1UL << __VIRTUAL_MASK_SHIFT) - 1)
> >  
> > -/* Cast PAGE_MASK to a signed type so that it is sign-extended if
> > +/* Cast *PAGE_MASK to a signed type so that it is sign-extended if
> >     virtual addresses are 32-bits but physical addresses are larger
> >     (ie, 32-bit PAE). */
> >  #define PHYSICAL_PAGE_MASK	(((signed long)PAGE_MASK) & __PHYSICAL_MASK)
> > -
> > -#define PMD_PAGE_SIZE		(_AC(1, UL) << PMD_SHIFT)
> > -#define PMD_PAGE_MASK		(~(PMD_PAGE_SIZE-1))
> > -
> > -#define PUD_PAGE_SIZE		(_AC(1, UL) << PUD_SHIFT)
> > -#define PUD_PAGE_MASK		(~(PUD_PAGE_SIZE-1))
> > +#define PHYSICAL_PMD_PAGE_MASK	(((signed long)PMD_PAGE_MASK) & __PHYSICAL_MASK)
> > +#define PHYSICAL_PUD_PAGE_MASK	(((signed long)PUD_PAGE_MASK) & __PHYSICAL_MASK)
> 
> that's a really odd way of writing it, 'long' is signed by default ...

See the comment above (it was there before the patch). 'signed' can be
considered as documentation -- we want sign-extension here.

> There seems to be 150+ such cases in the kernel source though - weird.
> 
> More importantly, how does this improve things on 32-bit PAE kernels? If I follow 
> the values correctly then PMD_PAGE_MASK is 'UL' i.e. 32-bit:
> 
> > +#define PMD_PAGE_SIZE                (_AC(1, UL) << PMD_SHIFT)
> > +#define PMD_PAGE_MASK                (~(PMD_PAGE_SIZE-1))
> 
> thus PHYSICAL_PMD_PAGE_MASK is 32-bit too:
> 
> > +#define PHYSICAL_PMD_PAGE_MASK       (((signed long)PMD_PAGE_MASK) & __PHYSICAL_MASK)
> 
> so how is the bug fixed?

Again, see the comment.
I've checked that it generates correct value (using kernel/bounds.c).

-- 
 Kirill A. Shutemov