Re: [PATCH] mm/pgtable: return null if no ptl in __pte_offset_map_lock

[José Pekkarinen]

On 2023-11-15 16:19, Matthew Wilcox wrote:
> On Wed, Nov 15, 2023 at 08:55:05AM +0200, José Pekkarinen wrote:
> > Documentation of __pte_offset_map_lock suggest there is situations 
> > where
>
> You should have cc'd Hugh who changed all this code recently.

    Hi,

    Sorry, he seems to be missing if I run get_maintainer.pl:

$ ./scripts/get_maintainer.pl include/linux/mm.h
Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> (maintainer:MEMORY MANAGEMENT)
linux-mm@xxxxxxxxx (open list:MEMORY MANAGEMENT)
linux-kernel@xxxxxxxxxxxxxxx (open list)

> > a pmd may not have a corresponding page table, in which case it should
> > return NULL without changing ptlp. Syzbot found its ways to produce a
> > NULL dereference in the function showing this case. This patch will
> > provide the exit path suggested if this unlikely situation turns up. 
> > The
> > output of the kasan null-ptr-report follows:
>
> There's no need to include all this nonsense in the changelog.

    No problem, we can clean the patch if we find there is something
worth upstreaming.

> >  spin_lock include/linux/spinlock.h:351 [inline]
> >  __pte_offset_map_lock+0x154/0x360 mm/pgtable-generic.c:373
> >  pte_offset_map_lock include/linux/mm.h:2939 [inline]
> >  filemap_map_pages+0x698/0x11f0 mm/filemap.c:3582
>
> This was the only interesting part.
>
> > +++ b/include/linux/mm.h
> > @@ -2854,7 +2854,7 @@ void ptlock_free(struct ptdesc *ptdesc);
> >
> >  static inline spinlock_t *ptlock_ptr(struct ptdesc *ptdesc)
> >  {
> > -	return ptdesc->ptl;
> > +	return (likely(ptdesc)) ? ptdesc->ptl : NULL;
> >  }
>
> I don't think we should be changing ptlock_ptr().

    This is where the null ptr dereference originates, so the only
alternative I can think of is to protect the life cycle of the ptdesc
to prevent it to die between the pte check and the spin_unlock of
__pte_offset_map_lock. Would that work for you?

> > +++ b/mm/pgtable-generic.c
> > @@ -370,6 +370,8 @@ pte_t *__pte_offset_map_lock(struct mm_struct *mm, 
> > pmd_t *pmd,
> >  	if (unlikely(!pte))
> >  		return pte;
> >  	ptl = pte_lockptr(mm, &pmdval);
> > +	if (unlikely(!ptl))
> > +		return NULL;
> >  	spin_lock(ptl);
>
> I don't understand how this could possibly solve the problem.  If 
> there's
> no PTE level, then __pte_offset_map() should return NULL and we'd 
> already
> return due to the check for !pte.

    I tested the syzbot reproducer in x86 and it doesn't produce this 
kasan
report anymore.

    José.