Re: [PATCH] mm: mmu_notifier: re-fix freed page still mapped insecondary MMU

[Robin Holt]

On Tue, Apr 16, 2013 at 06:26:36PM +0800, Xiao Guangrong wrote:
> On 04/16/2013 05:31 PM, Robin Holt wrote:
> > On Tue, Apr 16, 2013 at 02:39:49PM +0800, Xiao Guangrong wrote:
> >> The commit 751efd8610d3 (mmu_notifier_unregister NULL Pointer deref
> >> and multiple ->release()) breaks the fix:
> >>     3ad3d901bbcfb15a5e4690e55350db0899095a68
> >>     (mm: mmu_notifier: fix freed page still mapped in secondary MMU)
> > 
> > Can you describe how the page is still mapped?  I thought I had all
> > cases covered.  Whichever call hits first, I thought we had one callout
> > to the registered notifiers.  Are you saying we need multiple callouts?
> 
> No.
> 
> You patch did this:
> 
>                 hlist_del_init_rcu(&mn->hlist);    1 <======
> +               spin_unlock(&mm->mmu_notifier_mm->lock);
> +
> +               /*
> +                * Clear sptes. (see 'release' description in mmu_notifier.h)
> +                */
> +               if (mn->ops->release)
> +                       mn->ops->release(mn, mm);    2 <======
> +
> +               spin_lock(&mm->mmu_notifier_mm->lock);
> 
> At point 1, you delete the notify, but the page is still on LRU. Other
> cpu can reclaim the page but without call ->invalid_page().
> 
> At point 2, you call ->release(), the secondary MMU make page Accessed/Dirty
> but that page has already been on the free-list of page-alloctor.

That expectation on srcu _REALLY_ needs to be documented better.
Maybe I missed it in the comments, but there is an expectation beyond
the synchronize_srcu().  This code has been extremely poorly described
and I think it is time we fix that up.

I do see that in comments for mmu_notifier_unregister, there is an
expectation upon already having all the spte's removed prior to making
this call.  I think that is also a stale comment as it mentions a lock
which I am not sure ever really existed.

> > Also, shouldn't you be asking for a revert commit and then supply a
> > subsequent commit for the real fix?  I thought that was the process for
> > doing a revert.
> 
> Can not do that pure reversion since your patch moved hlist_for_each_entry_rcu
> which has been modified now.
> 
> Should i do pure-eversion + hlist_for_each_entry_rcu update first?

Let's not go off without considering this first.

It looks like what we really need to do is ensure there is a method
for ensuring that the mmu_notifier remains on the list while callouts
invalidate_page() callouts are being made and also a means of ensuring
that only one ->release() callout is made.

First, is it the case that when kvm calls mmu_notifier_unregister(),
it has already cleared the spte's?  (what does spte stand for anyway)?
If so, then we really need to close the hole in __mmu_notifier_release().
I think we would need to modify code in both _unregister and _release,
but the issue is really _release.


I need to get ready and drive into work.  If you want to float something
out there, that is fine.  Otherwise, I will try to work something up
when I get to the office.

Thanks,
Robin
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/