git: PF - Force 'sloppy' when establishing conflicting state
Matthew Dillon
dillon at crater.dragonflybsd.org
Tue Sep 2 21:30:39 PDT 2014
commit 9d0d81f910550be540a72ce262aee108ae351eb5
Author: Matthew Dillon <dillon at apollo.backplane.com>
Date: Tue Sep 2 21:19:15 2014 -0700
PF - Force 'sloppy' when establishing conflicting state
* Check whether a PASS IN or PASS OUT conflicts with established translation
state in the opposite direction. When this situation is detected, one
of the PASS rules can establish state (and with recent SMP work, both
PASS rules will establish state). This causes problems because the PASS
rules may only see one direction of the connection due to the RDR or NAT.
If strict TCP sequence space checking occurs the PASS state can generate
RSTs.
To fix this we force the SLOPPY flag to be set for any PASS state being
established in the face of a conflict against a translation rule. This
allows packet flow to short-cut through the state table and is preferable
to disallowing establishment of the state because that would force a
full rules scan (and repeated conflict/failure) for every packet.
History
* PASS IN and PASS OUT rules can interfere with a RDR rule when strict
sequence space tests are made for established TCP state.
In pre-SMP PF, including in FreeBSD and probably also in other BSDs,
two active states are generally established, one for RDR and one for the
PASS IN. PF attempts to establish state for the PASS OUT but hits a
conflict against the established RDR state and fails.
However, in this situation no short-cut state is established in one
direction and ALL packets that would have matched the failed PASS OUT
will cause a full rules scan.
* With the SMP work, the PASS OUT conflict was no longer detected because
the RDR state is on a different RBTREE than the PASS OUT state. This
allowed conflicting state for both PASS IN and PASS OUT to be established.
* Conflicting state in either direction is capable of generating spurious
RSTs against the RDR rule. One direction for sure, the other will generate
RSTs but possibly be obscured by the translation rule (that is, the RST
ends up going somewhere unexpected), so the RDR rule still works. But
the problem remains.
Summary of changes:
sys/net/pf/pf.c | 167 ++++++++++++++++++++++++++++++++++++----
sys/net/pf/pfvar.h | 3 +
usr.sbin/pfctl/pf_print_state.c | 2 +
3 files changed, 157 insertions(+), 15 deletions(-)
http://gitweb.dragonflybsd.org/dragonfly.git/commitdiff/9d0d81f910550be540a72ce262aee108ae351eb5
--
DragonFly BSD source repository
More information about the Commits
mailing list