kern/164402: [pf] pf crashes with a particular set of rules when first matching packet arrives

[Ermal Luçi]

On Tue, Apr 17, 2012 at 10:38 AM, Ermal Luçi <eri at freebsd.org> wrote:
> 2012/4/17 Gleb Smirnoff <glebius at freebsd.org>:
>> On Tue, Apr 17, 2012 at 10:06:15AM +0200, Ermal Lu?i wrote:
>> E> 2012/4/16 Gleb Smirnoff <glebius at freebsd.org>:
>> E> > On Sun, Apr 15, 2012 at 12:00:21PM +0000, Gleb Smirnoff wrote:
>> E> > T>  On Sun, Apr 15, 2012 at 11:10:03AM +0000, Gleb Smirnoff wrote:
>> E> > T>  T>    I have a vague suspicion on what is happening. Your description of
>> E> > T>  T>  the problem looks like if a packet processing in the kernel has entered
>> E> > T>  T>  an endless loop.
>> E> > T>  T>
>> E> > T>  T>    Looking at pf_route() I see such possibility. From OpenBSD we have
>> E> > T>  T>  this protection against endless looping:
>> E> > T>  T>
>> E> > T>  T>          if ((*m)->m_pkthdr.pf.routed++ > 3) {
>> E> > T>  T>                  m0 = *m;
>> E> > T>  T>                  *m = NULL;
>> E> > T>  T>                  goto bad;
>> E> > T>  T>          }
>> E> > T>  T>
>> E> > T>  T>  In our code this transforms to:
>> E> > T>  T>
>> E> > T>  T>          if (pd->pf_mtag->routed++ > 3) {
>> E> > T>  T>                  m0 = *m;
>> E> > T>  T>                  *m = NULL;
>> E> > T>  T>                  goto bad;
>> E> > T>  T>          }
>> E> > T>  T>
>> E> > T>  T>  The root difference between storing the tag on mbuf and on pfdesc
>> E> > T>  T>  is that we lose pfdesc, and thus the tag, when we enter pf_test()
>> E> > T>  T>  recursively. And pf_route() does this recursion:
>> E> > T>  T>
>> E> > T>  T>          if (oifp != ifp) {
>> E> > T>  T>                  if (pf_test(PF_OUT, ifp, &m0, NULL) != PF_PASS) {
>> E> > T>  T>                          goto bad;
>> E> > T>  T>          ....
>> E> > T>
>> E> > T>  On second look I see that my suspicion may not be true. In the
>> E> > T>  beginning of pf_test() we do pf_get_mtag() which preserves already
>> E> > T>  present tag if there is one.
>> E> >
>> E> > Further investigation showed that problem exist when route applied
>> E> > ends in lo0, and packet passes to if_simloop(). There all mtags are
>> E> > stripped from the mbuf, including the pf mtag. Then packet is again
>> E> > processed by ip_input() again entering pf(4), if it again matches
>> E> > a routing rule, then we got an endless loop.
>> E> >
>> E> > We can try to fix this applying MTAG_PERSISTENT to the pf(4) tag id.
>> E>
>> E> That seems like the best fix for this case.
>>
>> In this case crash or freeze is fixed, but still packet is dropped. Example
>> of such rule:
>>
>> pass in on igb0 fastroute proto tcp from any to $localip
>>
>> Anyway, dropping packets is much better than crashing.

To be more explicit, this breaks functionality.
Since as i said the firewall has already taken descision and the state
keeping for sure will drop this matching tcp packets.
You will not see it for other protos that do not have the state
transition like tcp though your statistics will be wrong.
This is not justifible and its better to crash :)

Though as i said the skip firewall flag seems more sensible.
>>
>
> Actually after some coffee :) i think its better marking the packet
> with M_SKIP_FIREWALL since
> it has already taken its decision on this packet.
>
> The simloop consumers seem to be just facilities of how things work
> from what i can see.
>
> So just delivering the packet by sending skipping the firewalls seems
> more sensibile!
>
> Though the persistent case for the tags should be revisited since it
> may fix some other issues with pf(4) tags, and some others.
>
>> --
>> Totus tuus, Glebius.
>
>
>
> --
> Ermal



-- 
Ermal