On Tue, Apr 14, 2020 at 04:06:20PM -0300, Jason Gunthorpe wrote:
> On Tue, Apr 14, 2020 at 11:58:29AM -0700, Ira Weiny wrote:
> > On Tue, Apr 14, 2020 at 01:32:34PM -0300, Jason Gunthorpe wrote:
> > > On Tue, Apr 14, 2020 at 05:10:01PM +0200, Alexander Gordeev wrote:
> > > > Commit 0005d20 ("mm/gup: Move page table entry dereference
> > > > into helper function") wrapped access to page table entries
> > > > larger than sizeof(long) into a race-aware accessor. One of
> > > > the two dereferences in gup_fast path was however overlooked.
> > > >
> > > > CC: Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx>
> > > > CC: linux-mm@xxxxxxxxx
> > > > Signed-off-by: Alexander Gordeev <agordeev@xxxxxxxxxxxxx>
> > > > mm/gup.c | 2 +-
> > > > 1 file changed, 1 insertion(+), 1 deletion(-)
> > > >
> > > > diff --git a/mm/gup.c b/mm/gup.c
> > > > index d53f7dd..eceb98b 100644
> > > > +++ b/mm/gup.c
> > > > @@ -2208,7 +2208,7 @@ static int gup_pte_range(pmd_t pmd, unsigned long addr, unsigned long end,
> > > > if (!head)
> > > > goto pte_unmap;
> > > >
> > > > - if (unlikely(pte_val(pte) != pte_val(*ptep))) {
> > > > + if (unlikely(pte_val(pte) != pte_val(gup_get_pte(ptep)))) {
> > >
> > > It doesn't seem like this needs the special helper as it is just
> > > checking that the pte hasn't changed, it doesn't need to be read
> > > exactly.
> > >
> > > But it probably should technically still be a READ_ONCE. Although I
> > > think the atomic inside try_grab_compound_head prevents any real
> > > problems.
> >
> > I think we should go for consistency here and use the helper function.
>
> It seems quite expensive to do two more unncessary barriers..
But won't a failure to read the 'real' pte result in falling back to GUP slow?
Not sure which is worse?
And most arch's don't suffer from this...
Ira
