Re: [Linux PATCH] Fix to hugepages to work around new PWT handling

[Dave McCracken <dcm@mccr.org>]

On Thursday, June 24, 2010, Jeremy Fitzhardinge wrote:
> > Um, this is the upper level code.  The entire purpose of make_huge_pte is
> > to  construct a present huge pte from page and pgprot. The problem is
> > that the original code makes the pte, then sets the present bit via
> > pte_mkhuge().  This means the Xen-specific macro that triggers on
> > present is misled and doesn't do the pfn_to_mfn().  Without this patch
> > hugepages is handing pfns to the hypervisor to map instead of mfns.
> >
> >   
> 
> In principle, setting present should cause the pte to be converted from
> pfn to mfn, but I don't think that ever happens with normal ptes (since
> non-present ptes contain swap info).  But I don't see where a huge pte
> gets present set; pte_mkhuge itself doesn't do anything except set PSE.

Wow.  I just dug through the code.  The landscape has sure changed since the 
last time I followed this path.

It used to be that vma->vm_page_prot only contained the various read/write 
flags for that vma.  At that time pte_mkhuge() did in fact add _PAGE_PRESENT|
_PAGE_PSE to the pte.

Now it appears that vma->vm_page_prot does include _PAGE_PRESENT in all its 
various states.  So this part of the patch is in fact unnecessary.

It's what I get for not rechecking my facts to be sure they haven't changed.  
Sorry.

Dave McCracken
Oracle Corp.