GPF when doing jail -r, possibly an use-after-free

[Mikolaj Golub]

On Sat, 7 Jul 2012 20:38:23 +0000 Bjoern A. Zeeb wrote:

 BAZ> On 6. Jul 2012, at 05:53 , Mikolaj Golub wrote:

 >> 
 >> On Thu, 5 Jul 2012 20:21:53 +0000 Bjoern A. Zeeb wrote:
 >> 
 >> BAZ> On 5. Jul 2012, at 19:53 , Mikolaj Golub wrote:
 >> 
 >>>> 
 >>>> On Thu, 05 Jul 2012 12:18:20 -0700 Xin Li wrote:
 >>>> 
 >>>> XL> Hi, Mikolaj,
 >>>> 
 >>>> XL> On 07/04/12 00:00, Mikolaj Golub wrote:
 >>>>>> Is this observed after destroying epair? There is an issue with
 >>>>>> epair: on destroy, when epair_clone_destroy() calls
 >>>>>> ether_ifdetach() for its second half it does not switch to its vnet
 >>>>>> and if_detach_internal() can't find the interface and just returns.
 >>>>>> As a result V_ifnet list is left with dead reference.
 >>>> 
 >>>> XL> Yes.
 >>>> 
 >>>>>> http://lists.freebsd.org/pipermail/freebsd-virtualization/2011-January/000628.html
 >>>>>> 
 >>>>>> Here is an updated patch against CURRENT:
 >>>>>> 
 >>>>>> http://people.freebsd.org/~trociny/if_epair.c.epair_clone_destroy.1.patch
 >>>> 
 >>>> XL> Your
 >>>>>> 
 >>>> XL> patch did fixed the problem, thanks!  Are you going to commit it
 >>>> XL> against -HEAD and then MFC after a while?
 >>>> 
 >>>> I would like Bjoern review it before me committing, or at least tell he does
 >>>> not mind, if he does not have time to review -)
 >> 
 >> BAZ> To me the patch looks wrong; I am wondering if someone broke some other central
 >> BAZ> assumptions but given I cannot currently spend time on this and if it fixes things
 >> BAZ> feel free to go ahead.
 >> 
 >> If you told what looks wrong I could try to dig at that direction and might be
 >> back with a better solution, instead of committing a presumably wrong fix.

 BAZ> sorry;  vnet.c:vnet_destroy() should dtrt already wrt to interfaces moving to parents
 BAZ> and being detached.

But this is not the issue with vnet_destroy() not moving interfaces to
parents. It does move them. It is with destroying epair. When epair that has
its ends in different vnets is destroyed, and ether_ifdetach() is called for
the second end without switching to its vnet it fails to find the iface in the
wrong vnet and just silently returns (which I think is also wrong:
if_detach_internal() should panic here). As a result the pointer is not
removed from vnet ifnet list. And later, when someone is traversing this list
and tries to access the pointer (this is often vnet_destroy(), which is
usually called after removing interfaces, but might be e.g. ifconfig) dead
pointer dereference occurs.

My patch just makes epair_clone_destroy() switch to the proper vnet before
calling ether_ifdetach().

Or have I missed your point?

-- 
Mikolaj Golub