Re: [tip:sched/core] sched/core: Fix a potential double-fetch bug in sched_copy_attr()

[Thomas Gleixner]

On Mon, 28 Jan 2019, Peter Zijlstra wrote:
> On Sun, Jan 27, 2019 at 12:04:44PM +0100, Thomas Gleixner wrote:
> > On Mon, 21 Jan 2019, tip-bot for Kangjie Lu wrote:
> > > diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> > > index a674c7db2f29..d4d3514c4fe9 100644
> > > --- a/kernel/sched/core.c
> > > +++ b/kernel/sched/core.c
> > > @@ -4499,6 +4499,9 @@ static int sched_copy_attr(struct sched_attr __user *uattr, struct sched_attr *a
> > >  	if (ret)
> > >  		return -EFAULT;
> > >  
> > > +	/* In case attr->size was changed by user-space: */
> > > +	attr->size = size;
> > > +
> > 
> > Just when pondering to send that to Linus, I tried to write up a concise
> > summary for this which made me look at the patch.
> > 
> > If the size changed, then its clear that user space fiddled with the date
> > between the size fetch and the full copy from user. So why restoring the
> > size instead of doing the obvious:
> > 
> >    	 if (attr->size != size)
> > 	 	return -ECRAP;
> > 
> > Hmm?
> 
> Sure; but if we do that we should also change perf_copy_attr() which has
> the exact same thing.

Yes please.

The point is that by default the data passed to a function (and it does not
matter whether it's a syscall) by pointer is immutable. There is exactly
ONE syscall which is specifically designed to deal with mutable data and
that's a constant source of headache ....

Kangjie is right that all double fetch operations like the one in
sched_copy_attr() are prone to concurrent modification problems. But then
we really should say NO instead of silently trying to fix things up. I
personally would even kill the process immediately, no matter whether the
corruption is caused by malicious intent or by sheer stupidity.

Thanks,

	tglx