Re: [PATCH RESEND v6 02/16] mm/gup: Fix __get_user_pages() on fault retry of hugetlb

[Peter Xu]

On Mon, Mar 02, 2020 at 08:02:34PM +0100, David Hildenbrand wrote:
> On 20.02.20 16:53, Peter Xu wrote:
> > When follow_hugetlb_page() returns with *locked==0, it means we've got
> > a VM_FAULT_RETRY within the fauling process and we've released the
> > mmap_sem.  When that happens, we should stop and bail out.
> > 
> > Signed-off-by: Peter Xu <peterx@xxxxxxxxxx>
> > ---
> >  mm/gup.c | 10 ++++++++++
> >  1 file changed, 10 insertions(+)
> > 
> > diff --git a/mm/gup.c b/mm/gup.c
> > index 1b4411bd0042..76cb420c0fb7 100644
> > --- a/mm/gup.c
> > +++ b/mm/gup.c
> > @@ -849,6 +849,16 @@ static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
> >  				i = follow_hugetlb_page(mm, vma, pages, vmas,
> >  						&start, &nr_pages, i,
> >  						gup_flags, locked);
> > +				if (locked && *locked == 0) {
> > +					/*
> > +					 * We've got a VM_FAULT_RETRY
> > +					 * and we've lost mmap_sem.
> > +					 * We must stop here.
> > +					 */
> > +					BUG_ON(gup_flags & FOLL_NOWAIT);
> > +					BUG_ON(ret != 0);
> 
> Can we be sure ret is really set to != 0 at this point? At least,
> reading the code this is not clear to me.

Here I wanted to make sure ret is zero (it's BUG_ON, not assert).

"ret" is the fallback return value only if error happens when i==0.
Here we want to make sure even if no page is pinned we'll return zero
gracefully when VM_FAULT_RETRY happened when following the hugetlb
pages.

> 
> Shouldn't we set "ret = i" and assert that i is an error (e.g., EBUSY?).
> Or set -EBUSY explicitly?

No.  Here "i" could only be either positive (when we've got some pages
pinned no matter where), or zero (when follow_hugetlb_page released
the mmap_sem on the first page that it wants to pin).  So imo "i"
should never be negative instead.

Thanks,

-- 
Peter Xu