Re: [PATCH v2 28/32] mm/memory: allow pte_offset_map[_lock]() to fail

[Hugh Dickins]

On Fri, 9 Jun 2023, Andrew Morton wrote:
> On Thu, 8 Jun 2023 18:43:38 -0700 (PDT) Hugh Dickins <hughd@xxxxxxxxxx> wrote:
> 
> > copy_pte_range(): use pte_offset_map_nolock(), and allow for it to fail;
> > but with a comment on some further assumptions that are being made there.
> > 
> > zap_pte_range() and zap_pmd_range(): adjust their interaction so that
> > a pte_offset_map_lock() failure in zap_pte_range() leads to a retry in
> > zap_pmd_range(); remove call to pmd_none_or_trans_huge_or_clear_bad().
> > 
> > Allow pte_offset_map_lock() to fail in many functions.  Update comment
> > on calling pte_alloc() in do_anonymous_page().  Remove redundant calls
> > to pmd_trans_unstable(), pmd_devmap_trans_unstable(), pmd_none() and
> > pmd_bad(); but leave pmd_none_or_clear_bad() calls in free_pmd_range()
> > and copy_pmd_range(), those do simplify the next level down.
> > 
> > ...
> >
> > @@ -3728,11 +3737,9 @@ vm_fault_t do_swap_page(struct vm_fault *vmf)
> >  			vmf->page = pfn_swap_entry_to_page(entry);
> >  			vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
> >  					vmf->address, &vmf->ptl);
> > -			if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
> > -				spin_unlock(vmf->ptl);
> > -				goto out;
> > -			}
> > -
> > +			if (unlikely(!vmf->pte ||
> > +				     !pte_same(*vmf->pte, vmf->orig_pte)))
> > +				goto unlock;
> >  			/*
> >  			 * Get a page reference while we know the page can't be
> >  			 * freed.
> 
> This hunk falls afoul of
> https://lkml.kernel.org/r/20230602092949.545577-5-ryan.roberts@xxxxxxx.
> 
> I did this:
> 
> @@ -3729,7 +3738,8 @@ vm_fault_t do_swap_page(struct vm_fault
>  			vmf->page = pfn_swap_entry_to_page(entry);
>  			vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
>  					vmf->address, &vmf->ptl);
> -			if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
> +			if (unlikely(!vmf->pte ||
> +				     !pte_same(*vmf->pte, vmf->orig_pte)))
>  				goto unlock;
>  
>  			/*

Yes, that's exactly right: thanks, Andrew.

Hugh