Re: [PATCH] x86: do not overrun page table ranges in gup

[Nick Piggin <nickpiggin@yahoo.com.au>]

On Tuesday 29 July 2008 10:33, Linus Torvalds wrote:
> On Tue, 29 Jul 2008, Johannes Weiner wrote:
> > Actually, I think the prettier fix would be to just establish that
> > garuantee:
> >
> > --- a/arch/x86/mm/gup.c
> > +++ b/arch/x86/mm/gup.c
> > @@ -223,7 +223,7 @@ int get_user_pages_fast(unsigned long start, int
> > nr_pages, int write, struct page **pages)
> >  {
> >  	struct mm_struct *mm = current->mm;
> > -	unsigned long end = start + (nr_pages << PAGE_SHIFT);
> > +	unsigned long end = PAGE_ALIGN(start + (nr_pages << PAGE_SHIFT));
>
> Umm. 'end' is guaranteed to be page-aligned if 'start' is.
>
> So if this makes a difference, that implies that _start_ isn't
> page-aligned, and then you when you add PAGE_SIZE to 'addr', you are going
> to miss 'end' again.
>
> So no, the right fix would be to align 'start' first, which means that
> everything else (including 'end') will be page-aligned. Aligning just one
> or the other is very very wrong.
>
> But yeah, this looks like a nasty bug. It's also sad that the code
> that _should_ be architecture-independent, isn't - because every
> architecture defines the _whole_ "get_user_pages_fast()", even though part
> of it is very much arch-independent (the whole alignment/access_ok part).

I guess when we get a couple more architectures implementing it, we
should split that into a little helper perhaps. I just don't know
quite how it is going to pan out.