Re: [5.4 PATCH] mm/gup: Do not force a COW break on file-backed memory

[Matthew Wilcox]

On Thu, Dec 02, 2021 at 10:54:48AM -0800, Linus Torvalds wrote:
> On Wed, Dec 1, 2021 at 8:11 PM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
> >
> > The other patch we've been kicking around (and works) is:
> >
> >  static inline bool should_force_cow_break(struct vm_area_struct *vma, unsigned
> > int flags)
> >  {
> > -       return is_cow_mapping(vma->vm_flags) && (flags & FOLL_GET);
> > +       return is_cow_mapping(vma->vm_flags) &&
> > +               (!(vma->vm_flags & VM_DENYWRITE)) && (flags & FOLL_GET);
> >  }
> 
> That patch makes no sense to me.
> 
> It may "work", but it doesn't actually do anything sensible or really
> fix the problem that I can tell.

Oh absolutely, it's semantically nonsense.  The only reason it fixes the
problem is that VM_DENYWRITE VMAs are the only ones considered for the
RO_THP merging, so they're the only ones which we've seen causing a
problem.

> I suspect a real fix would be bigger and more invasive.

Darn.  I was hoping you were going to say something like "The real
problem is follow_trans_huge_pmd() is complete garbage and it should
just do X, Y and Z".  Or "When we force on FOLL_WRITE, we should also
force on FOLL_SPLIT_PMD".

> If the answer is not to backport all the other changes (and they were
> _really_ invasive), I think one answer may be to simply move the
> "should_force_cow_break()" down to below the point where you've looked
> up the page.
> 
> Then you can actually look at "is this a file mapped page", and say
> "if so, that's ok, we can return it as-is".
> 
> Otherwise, you do something like
> 
>         foll_flags |= FOLL_WRITE;
>         free_page(page);
>         goto repeat;
> 
> to repeat the loop (now with FOLL_WRITE).
> 
> So the patch is bigger and more involved, because you would have done
> the page lookup (for reading) and now notice "Oh, I need it for
> writing instead" so you need to undo and re-do).
> 
> But at least - unlike backporting everything else - it would be
> limited to that one __get_user_pages() function.
> 
> Hmm?
> 
> (And you'd need to handle that follow_hugetlb_page() case too), not
> just the follow_page_mask() one)

Thanks, I'll take a look.