On Fri, Nov 16, 2012 at 12:10:39PM -0800, David Rientjes wrote:
> On Fri, 16 Nov 2012, Kirill A. Shutemov wrote:
> 
> > > > > > diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> > > > > > index d767a7c..05490b3 100644
> > > > > > --- a/mm/huge_memory.c
> > > > > > +++ b/mm/huge_memory.c
> > > > > > @@ -1259,6 +1259,8 @@ int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
> > > > > >  		pmd_t entry;
> > > > > >  		entry = pmdp_get_and_clear(mm, addr, pmd);
> > > > > >  		entry = pmd_modify(entry, newprot);
> > > > > > +		if (is_huge_zero_pmd(entry))
> > > > > > +			entry = pmd_wrprotect(entry);
> > > > > >  		set_pmd_at(mm, addr, pmd, entry);
> > > > > >  		spin_unlock(&vma->vm_mm->page_table_lock);
> > > > > >  		ret = 1;
> > > > > 
> > > > > Nack, this should be handled in pmd_modify().
> > > > 
> > > > I disagree. It means we will have to enable hzp per arch. Bad idea.
> > > > 
> > > 
> > > pmd_modify() only exists for those architectures with thp support already, 
> > > so you've already implicitly enabled for all the necessary architectures 
> > > with your patchset.
> > 
> > Now we have huge zero page fully implemented inside mm/huge_memory.c. Push
> > this logic to pmd_modify() means we expose hzp implementation details to
> > arch code. Looks ugly for me.
> > 
> 
> So you are suggesting that anybody who ever does pmd_modify() in the 
> future is responsible for knowing about the zero page and to protect 
> against giving it write permission in the calling code??

Looks like we don't need the patch at all.

IIUC, if you ask for PROT_WRITE vm_get_page_prot() will translate it to
_PAGE_COPY or similar and you'll only get the page writable on pagefault.

Could anybody confirm that it's correct?

-- 
 Kirill A. Shutemov