Re: netpfil with if_output and ip(6)_output

On 14.11.2016 16:13, Franco Fichtner wrote:
> Hi Andrey,
> 
>> On 14 Nov 2016, at 1:55 PM, Andrey V. Elsukov <bu7cher_at_yandex.ru> wrote:
>>
>> I have some thought related to your proposal.
>> What you think if we will introduce new KPI to work with fwd_tags?
>> With such KPI we can make fwd_tags opaque for PFIL consumers and handle
>> tags identically in all *proto*_output() routines.
>>
>> For first glance I can propose the following:
>>
>> /* ip_var.h */
>> #define	IP_HAS_NEXTHOP(m)	((m)->m_flags & M_IP_NEXTHOP)
>> int ip_set_fwdtag(struct mbuf *m, struct sockaddr_in *dst,
>>    u_short ifidx);
>> int ip_get_fwdtag(struct mbuf *m, struct sockaddr_in *dst,
>>    u_short *ifidx);
>> void ip_flush_fwdtag(struct mbuf *m);
>>
>>
>> /* ip6_var.h */
>> #define	IP6_HAS_NEXTHOP(m)	((m)->m_flags & M_IP6_NEXTHOP)
>> int ip6_set_fwdtag(struct mbuf *m, struct sockaddr_in6 *dst,
>>    u_short ifidx);
>> int ip6_get_fwdtag(struct mbuf *m, struct sockaddr_in6 *dst,
>>    u_short *ifidx);
>> void ip6_flush_fwdtag(struct mbuf *m);
> 
> This looks reasonable, thank you.  How would we proceed with the
> implementation / porting of ipfw to the new framework?
> 
>> Since I'm not quite aware how PF handles PACKET_TAG_IPFORWARD tags, you
>> can modify this to fully cover its needs.
> 
> Right now, it doesn't, which is the original problem: PACKET_TAG_IPFORWARD
> aligns well with netpfil, now we only need pf to do the same (and
> maybe the new ipfw net64 code at some point as well).

nat64 does direct if_output only when undocumented
IPFIREWALL_NAT64_DIRECT_OUTPUT option was used. In other cases (by
default) translated packet will be scheduled via netisr.

Also, do not forget that we have special handling for fwd_tags in
udp/tcp output path.

> ipfw only requires the forwarding address, pf will want to select
> the outgoing interface in the same step so the KPI already covers
> that.

Yes, we get both ifindex and destination address from one call.
In future ipfw(4) also can be extended to use ifindex.

> As a note, set_fwdtag will have to be able to (safely) rewrite the
> internal structure, in case one filter overwrites the decision of
> the other, or we call flush + set again?

In my meaning it should remove old tag and set new one. This can be
optimized for cases when we have tag with nexthop for the same address
family.

> Another small oddity is the flipping of M_FASTFWD_OURS which is
> write only for forwarding, but overwriting with another all requires
> to unset it.  Should M_FASTFWD_OURS set/unset be part of the KBI as
> well?

It looks like M_FASTWFD_OURS also can be covered by this KPI.

-- 
WBR, Andrey V. Elsukov