On Fri, Mar 13, 2020 at 9:27 AM Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> On Fri, 13 Mar 2020 09:17:22 +0530 Souptick Joarder <jrdr.linux@xxxxxxxxx> wrote:
>
> > On Fri, Mar 13, 2020 at 8:28 AM Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote:
> > >
> > > On Fri, 13 Mar 2020 07:41:00 +0530 Souptick Joarder <jrdr.linux@xxxxxxxxx> wrote:
> > >
> > > > walk_page_range() already has a check for lockdep_assert_held().
> > > > So additional check for lockdep_assert_held() can be removed from
> > > > hmm_range_fault().
> > > >
> > > > ...
> > > >
> > > > --- a/mm/hmm.c
> > > > +++ b/mm/hmm.c
> > > > @@ -681,7 +681,6 @@ long hmm_range_fault(struct hmm_range *range, unsigned int flags)
> > > > struct mm_struct *mm = range->notifier->mm;
> > > > int ret;
> > > >
> > > > - lockdep_assert_held(&mm->mmap_sem);
> > > >
> > > > do {
> > > > /* If range is no longer valid force retry. */
> > >
> > > It isn't very obvious that hmm_range_fault() is and will only be called
> > > from walk_page_range() (is it?)
> > >
> >
> > Sorry Andrew, didn't get this part ?
> > * hmm_range_fault() is and will only be called
> > from walk_page_range() (is it?) *
>
> The patch assumes that hmm_range_fault() will only ever be called via
> walk_page_range(). How do we know this is the case? And that it
> always will be the case?
>
Ahh, Sorry, I think change log creates confusion.
The patch assumes that walk_page_range() is called from hmm_range_fault().
currently there are two caller for hmm_range_fault().
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c, line 859
drivers/gpu/drm/nouveau/nouveau_svm.c, line 544
in both case, * &mm->mmap_sem * lock is taken before calling hmm_range_fault().
Now inside hmm_range_fault() there is a check for
lockdep_assert_held(&mm->mmap_sem)
and again inside loop walk_page_range() is called which cross check
same lockdep_assert_held().
So the idea is to remove the first check
lockdep_assert_held(&mm->mmap_sem) in hmm_range_fault().
