Re: possible deadlock in get_user_pages_unlocked

[Eric Biggers]

On Fri, Feb 09, 2018 at 07:19:25PM -0800, Eric Biggers wrote:
> Hi Al,
> 
> On Sat, Feb 10, 2018 at 01:36:40AM +0000, Al Viro wrote:
> > On Fri, Feb 02, 2018 at 09:57:27AM +0100, Dmitry Vyukov wrote:
> > 
> > > syzbot tests for up to 5 minutes. However, if there is a race involved
> > > then you may need more time because the crash is probabilistic.
> > > But from what I see most of the time, if one can't reproduce it
> > > easily, it's usually due to some differences in setup that just don't
> > > allow the crash to happen at all.
> > > FWIW syzbot re-runs each reproducer on a freshly booted dedicated VM
> > > and what it provided is the kernel output it got during run of the
> > > provided program. So we have reasonably high assurance that this
> > > reproducer worked in at least one setup.
> > 
> > Could you guys check if the following fixes the reproducer?
> > 
> > diff --git a/mm/gup.c b/mm/gup.c
> > index 61015793f952..058a9a8e4e2e 100644
> > --- a/mm/gup.c
> > +++ b/mm/gup.c
> > @@ -861,6 +861,9 @@ static __always_inline long __get_user_pages_locked(struct task_struct *tsk,
> >  		BUG_ON(*locked != 1);
> >  	}
> >  
> > +	if (flags & FOLL_NOWAIT)
> > +		locked = NULL;
> > +
> >  	if (pages)
> >  		flags |= FOLL_GET;
> >  
> 
> Yes that fixes the reproducer for me.
> 

Just to follow up on this: it seems that Al's suggested fix didn't go anywhere,
but someone else eventually ran into this bug (which was a real deadlock) and a
slightly different fix was merged, commit 96312e61282ae.  It fixes the
reproducer for me too.  Telling syzbot so that it can close the bug:

#syz fix: mm/gup.c: teach get_user_pages_unlocked to handle FOLL_NOWAIT

- Eric