PF doesn't work with changed interfaces names.

[Daniel Hartmeier]

On Wed, Aug 24, 2005 at 05:09:14PM +0200, Pawel Jakub Dawidek wrote:

> When we change interface name with:
> 
> 	# ifconfig fxp0 name net0
> 
> and we add a firewall rule, restart pf, remove the rule, restart pf, we got:
> 
> Fatal trap 12: page fault while in kernel mode

The rule might have created an interface-bound state entry on fxp0. I
don't know off-hand how 'ifconfig name' interacts with pf_if.c pfi_*()
functions, but if it destroys the kif object of fxp0 (and creates a new
one for net0), there might be a problem in pf_if.c pfi_maybe_destroy()

#ifdef __FreeBSD__
        if ((p->pfik_flags & (PFI_IFLAG_ATTACHED | PFI_IFLAG_GROUP)) ||
            ((p->pfik_rules > 0 || p->pfik_states > 0) &&
             (p->pfik_flags & PFI_IFLAG_PLACEHOLDER) == 0))
#else
        if ((p->pfik_flags & (PFI_IFLAG_ATTACHED | PFI_IFLAG_GROUP)) ||
            p->pfik_rules > 0 || p->pfik_states > 0)
#endif
                return (0);

The non-FreeBSD version strictly returns when the pfi_kif object still
contains state entries, but the FreeBSD version might be free'ing the
object when it still contains state entries. Those state entries point
back to the pfi_kif object that contains them. If this happens, you
might see exactly the crash you describe, i.e. pf_state_compare_*() then
tries to access the no-longer-existing pfi_kif object to traverse state
entries in there, accessing invalid memory.

I have to study the use of PFI_IFLAG_PLACEHOLDER more, maybe Max has an
idea what goes wrong there on interface name changes (ifconfig name)...

As a short-term workaround, I think disabling pf and flusing the state
table (pfctl -d; pfctl -Fs) before the ifconfig invokation would prevent
the panic.

Daniel