[PATCH v1 3/4] arm64: Fix the page leak in pud/pmd_set_huge

[Mark Rutland]

On Wed, Mar 14, 2018 at 04:57:29PM +0530, Chintan Pandya wrote:
> 
> 
> On 3/14/2018 4:23 PM, Mark Rutland wrote:
> > On Wed, Mar 14, 2018 at 02:18:24PM +0530, Chintan Pandya wrote:
> > > While setting huge page, we need to take care of
> > > previously existing next level mapping. Since,
> > > we are going to overrite previous mapping, the
> > > only reference to next level page table will get
> > > lost and the next level page table will be zombie,
> > > occupying space forever. So, free it before
> > > overriding.
> > 
> > > @@ -939,6 +940,9 @@ int pud_set_huge(pud_t *pudp, phys_addr_t phys, pgprot_t prot)
> > >   		return 0;
> > >   	BUG_ON(phys & ~PUD_MASK);
> > > +	if (pud_val(*pud) && !pud_huge(*pud))
> > > +		free_page((unsigned long)__va(pud_val(*pud)));
> > > +
> > >   	set_pud(pudp, pfn_pud(__phys_to_pfn(phys), sect_prot));
> > >   	return 1;
> > >   }
> > > @@ -953,6 +957,9 @@ int pmd_set_huge(pmd_t *pmdp, phys_addr_t phys, pgprot_t prot)
> > >   		return 0;
> > >   	BUG_ON(phys & ~PMD_MASK);
> > > +	if (pmd_val(*pmd) && !pmd_huge(*pmd))
> > > +		free_page((unsigned long)__va(pmd_val(*pmd)));
> > > +
> > 
> > As Marc noted, (assuming the subsequent revert is applied) in both of
> > these cases, these tables are still live, and thus it is not safe to
> > free them.
> > 
> > Consider that immediately after freeing the pages, they may get
> > re-allocated elsewhere, where they may be modified. If this happens
> > before TLB invalidation, page table walks may allocate junk into TLBs,
> > resulting in a number of problems.
> Ah okay. What about this sequence,
> 1) I store old PMD/PUD values
> 2) Update the PMD/PUD with section mapping
> 3) Invalidate TLB
> 4) Then free the *leaked* page

You must invalidate the TLB *before* setting the new entry:

1) store the old entry value
2) clear the entry
3) invalidate the TLB

... then you can either:

4) update the entry
5) free the old table

... or:

4) free the old table
5) update the entry

Thanks,
Mark.