Re: [PATCH 05/31] mm/filemap: allow pte_offset_map_lock() to fail

[Qi Zheng]

On 2023/5/24 10:35, Hugh Dickins wrote:
> On Mon, 22 May 2023, Qi Zheng wrote:
> > On 2023/5/22 12:54, Hugh Dickins wrote:
> > > filemap_map_pages() allow pte_offset_map_lock() to fail; and remove the
> > > pmd_devmap_trans_unstable() check from filemap_map_pmd(), which can safely
> > > return to filemap_map_pages() and let pte_offset_map_lock() discover that.
> > >
> > > Signed-off-by: Hugh Dickins <hughd@xxxxxxxxxx>
> > > ---
> > >    mm/filemap.c | 12 +++++-------
> > >    1 file changed, 5 insertions(+), 7 deletions(-)
> > >
> > > diff --git a/mm/filemap.c b/mm/filemap.c
> > > index 28b42ee848a4..9e129ad43e0d 100644
> > > --- a/mm/filemap.c
> > > +++ b/mm/filemap.c
> > > @@ -3408,13 +3408,6 @@ static bool filemap_map_pmd(struct vm_fault *vmf,
> > > struct folio *folio,
> > >     if (pmd_none(*vmf->pmd))
> > >      pmd_install(mm, vmf->pmd, &vmf->prealloc_pte);
> > >    -	/* See comment in handle_pte_fault() */
> > > -	if (pmd_devmap_trans_unstable(vmf->pmd)) {
> > > -		folio_unlock(folio);
> > > -		folio_put(folio);
> > > -		return true;
> > > -	}
> > > -
> > >    	return false;
> > >    }
> > >    
> > > @@ -3501,6 +3494,11 @@ vm_fault_t filemap_map_pages(struct vm_fault *vmf,
> > >    
> > >     addr = vma->vm_start + ((start_pgoff - vma->vm_pgoff) << PAGE_SHIFT);
> > >     vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, addr, &vmf->ptl);
> > > +	if (!vmf->pte) {
> > > +		folio_unlock(folio);
> > > +		folio_put(folio);
> >
> > Missing:
> > 		ret = VM_FAULT_NOPAGE;
> > ?
>
> No, not missed.  Here ret is 0, which leads do_read_fault() to try
> __do_fault() afterwards.  Whereas VM_FAULT_NOPAGE would send it back
> to userspace to retry the whole fault.  Either will work, but I think
> the intention of VM_FAULT_NOPAGE here in filemap_map_pages() is to say
> "the page you want is now inserted", which is probably not the case.

Got it. Thanks for the explanation.

> Hugh

--
Thanks,
Qi