Re: [PATCH] mm: Fix invalid page pointer returned with FOLL_PIN gups

[Peter Xu <peterx@redhat.com>]

On Thu, Jan 27, 2022 at 11:25:38AM -0400, Jason Gunthorpe wrote:
> On Thu, Jan 27, 2022 at 05:19:56PM +0800, Peter Xu wrote:
> 
> > > > > diff --git a/mm/gup.c b/mm/gup.c
> > > > > index f0af462ac1e2..8ebc04058e97 100644
> > > > > +++ b/mm/gup.c
> > > > > @@ -440,7 +440,7 @@ static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
> > > > >   		pte_t *pte, unsigned int flags)
> > > > >   {
> > > > >   	/* No page to get reference */
> > > > > -	if (flags & FOLL_GET)
> > > > > +	if (flags & (FOLL_GET | FOLL_PIN))
> > > > >   		return -EFAULT;
> > > > 
> > > > Yes. This clearly fixes the problem that the patch describes, and also
> > > > clearly matches up with the Fixes tag. So that's correct.
> > > 
> > > It is a really confusing though, why not just always return -EEXIST
> > > here?
> > 
> > Because in current code GUP handles -EEXIST and -EFAULT differently?
> 
> That has nothing to do with here. We shouldn't be deciding what the
> top layer does way down here. Return the correct error code for what
> was discovered at this layer the upper loop should make the decision
> what it should do
> 
> > We do early bail out on -EFAULT.  -EEXIST was first introduced in 2015 from
> > Kirill for not failing some mlock() or mmap(MAP_POPULATE) on dax (1027e4436b6).
> > Then in 2017 it got used again with pud-sized thp (a00cc7d9dd93d) on dax too.
> > They seem to service the same goal and it seems to be designed that -EEXIST
> > shouldn't fail GUP immediately.
> 
> It must fail GUP immeidately if there is a pages list.

Right, but my point is we don't have an user at all for follow_page_mask()
returning -EEXIST with a **page which is non-NULL.  Or did I miss it?

> 
> Callers that want an early failure must pass in NULL for pages, it is
> just that simple. It has nothing to do with the FOLL flags.
> 
> A WARN_ON would be appropriate to compare the FOLL flags against the
> pages. eg FOLL_GET without a pages is nonsense and should be
> immediately aborted. On the other hand, we avoid this by construction
> internal to gup.c

We have something like that already, although it's only a VM_BUG_ON() not a
BUG_ON() or WARN_ON() at the entry of __get_user_pages():

	VM_BUG_ON(!!pages != !!(gup_flags & (FOLL_GET | FOLL_PIN)));

> 
> > > > Here, however, I think we need to consider this a little more carefully,
> > > > and attempt to actually fix up this case. It is never going to be OK
> > > > here, to return a **pages array that has these little landmines of
> > > > potentially uninitialized pointers. And so continuing on *at all* seems
> > > > very wrong.
> > > 
> > > Indeed, it should just be like this:
> > > 
> > > @@ -1182,6 +1182,10 @@ static long __get_user_pages(struct mm_struct *mm,
> > >                          * Proper page table entry exists, but no corresponding
> > >                          * struct page.
> > >                          */
> > > +                       if (pages) {
> > > +                               page = ERR_PTR(-EFAULT);
> > > +                               goto out;
> > > +                       }
> > >                         goto next_page;
> > >                 } else if (IS_ERR(page)) {
> > >                         ret = PTR_ERR(page);
> > 
> > IIUC not failing -EEXIST immediately seems to be what we want.
> 
> Which is what this does, for the only case it is acceptable - a null
> page list.
> 
> > From that POV, WARN_ON_ONCE() helps better on exposing an illegal return of
> > -EEXIST (as mentioned in the commit message) than the -EFAULT convertion, IMHO.
> 
> Again, that is upside down, -EEXIST should not be a illegal return. It
> should be valid, have a defined meaning 'the vaddr exists but has no
> struct page' and the top loop, and only the top loop, makes the
> decision what to do about it.

I believe this works too and I think I get your point, but as stated above it's
just not used yet so the path is not useful to any real code path.

Especially with above VM_BUG_ON() it means if we'll go into the "if (pages)" we
should have already triggered the VM_BUG_ON() condition when entering the function.

Thanks,

-- 
Peter Xu