And here's something i just tested on net-next that
fixes this for me.

cheers,
jamal

On Wed, 2010-03-31 at 12:38 -0400, jamal wrote:

This may be oversight in current implementation and possibly
nobody has needed it before - hence it is not functional.

I want to have a drop-all policy on a per-interface level
for incoming packets and add exceptions as i need them.
[using the flow table is cheap if you have xfrm built in].
i.e something along the lines of:

#eth0, wild-card drop all
ip xfrm policy add src 0.0.0.0/0 dst 0.0.0.0/0 dev eth0 \
       dir in ptype main action block priority $SOME-HIGH-value
#eth0, exception
ip xfrm policy add blah blah dev eth0 \
dir in ptype main action allow priority $SOME-small-value
#eth1, wild-card drop all
ip xfrm policy add src 0.0.0.0/0 dst 0.0.0.0/0 dev eth1 \
       dir in ptype main action block priority $SOME-HIGH-value
#eth1 exception ...

The problem is this works as long as i dont specify an interface.
i.e, this would work in the in-direction:

ip xfrm policy add src 0.0.0.0/0 dst 0.0.0.0/0 \
        dir in ptype main action block priority $SOME-HIGH-value

This would not work:
ip xfrm policy add src 0.0.0.0/0 dst 0.0.0.0/0 dev eth0 \
       dir in ptype main action block priority $SOME-HIGH-value


The checks in the selector matching is the culprit, example for v4:

__xfrm4_selector_match(struct xfrm_selector *sel, struct flowi *fl)
{
        return  .... &&
                .... &&
                (fl->oif == sel->ifindex || !sel->ifindex);
}

i.e in the second case i have a non-zero sel->ifindex but
a zero fl->oif; so it wont match.

One approach to fix this is to pass the direction then i can do
in the function call, then i can do something along the lines of
matching if:
(fl_dir == FLOW_DIR_IN && (fl->iif == sel->ifindex || !sel->ifindex) ||
(fl->oif == sel->ifindex || !sel->ifindex);

Is there any reason the selector matching only assumes fl->oif?
Are there any unforeseen issues/breakages if i added a check for the
above.

cheers,
jamal