Re: [PATCH] Detect early free of a live mm

[Mark Rutland]

On Thu, Mar 01, 2018 at 01:40:24PM +0100, Michal Hocko wrote:
> On Thu 01-03-18 11:22:37, Mark Rutland wrote:
> > On Thu, Mar 01, 2018 at 10:35:22AM +0100, Michal Hocko wrote:
> > > On Wed 28-02-18 12:18:10, Mark Rutland wrote:
> > > > Ugh, I messed up Peter's email when sending this out. please
> > > > s/infraded/infradead/ if replying to the first mail.
> > > > 
> > > > Sorry about that.
> > > > 
> > > > Mark.
> > > > 
> > > > On Wed, Feb 28, 2018 at 12:14:58PM +0000, Mark Rutland wrote:
> > > > > KASAN splats indicate that in some cases we free a live mm, then
> > > > > continue to access it, with potentially disastrous results. This is
> > > > > likely due to a mismatched mmdrop() somewhere in the kernel, but so far
> > > > > the culprit remains elusive.
> > > > > 
> > > > > Let's have __mmdrop() verify that the mm isn't live for the current
> > > > > task, similar to the existing check for init_mm. This way, we can catch
> > > > > this class of issue earlier, and without requiring KASAN.
> > > 
> > > Wouldn't it be better to check the mm_users count instead?
> > > {VM_}BUG_ON(atomic_read(&mm->mm_users));
> > 
> > Perhaps, but that won't catch a mismatched mmput(), which could
> > decrement mm_users (and mm_count) to zero early, before current->mm is
> > cleared in exit_mm().
> 
> true
> 
> > Locally, I'm testing with:
> > 
> >   BUG_ON(mm == current->mm);
> >   BUG_ON(mm == current->active_mm);
> >   BUG_ON(refcount_read(&mm->mm_users) != 0);
> >   BUG_ON(refcount_read(&mm->mm_count) != 0);
> > 
> > ... so as to also catch an early free from another thread.
> 
> I would be careful about active_mm. At least {un}use_mm does some tricks
> with it.

If that can leave mm == current->active_mm during __mmdrop(), I think
that would be a bug in its own right, since that would leave
current->active_mm pointing at a freed mm.

AFAICT, use_mm() does the usual mm_count balancing when flipping
active_mm (i.e. we mmgrab() the new mm, make it active_mm, then mmdrop()
the old active_mm), so even if the mmdrop() ends up freeing the old
active_mm, that shouldn't trigger the BUG_ON(mm == current->active_mm).

In unuse_mm() we don't poke active_mm at all, and leave that to the
usual context_switch() + finish_task_switch() grab/drop.

> > > Relying on current->mm resetting works currently but this is quite a
> > > subtle dependency.
> > 
> > Do you mean because it only catches an early free in a thread using that
> > mm, or is there another subtlety you had in mind?
> 
> No, I was more thinking about future changes when we stop clearing
> tsk->mm at exit. This is rather unlikely to be honest because that
> requires more changes.

Ok.

Thanks,
Mark.