[Cake] act_connmark + dscp

[Toke Høiland-Jørgensen]

Kevin Darbyshire-Bryant <kevin at darbyshire-bryant.me.uk> writes:

> Before I go too far down this road (and to avoid the horror of
> actually trying to code it) here’s what I’m trying to achieve.
>
>
> act_connmark + dscp is designed to copy a DSCP code to/from conntrack mark.  It uses 8 bits of the mark field, currently the most significant byte.
>
> Bits 31-26: DSCP
> Bit 25: Spare/Future
> Bit 24: Valid DSCP set
>
> The valid bit is set when the ‘getdscp’ function has written a DSCP
> value into the conntrack (& hence skb) mark. This allows us & other
> skb->mark/ct->mark applications (eg iptables, cake qdisc) to know that
> a DSCP value has been placed in the field. We cannot simply use a
> non-zero DSCP because zero is a valid DSCP.

If someone installs the action, the field is supposedly always copied;
so why do we need another flag?

Instead, I'd suggest adding user options with mask and shift to make it
possible to configure where in the mark field the values are stored...

> ’setdscp’ restores the DSCP field from the stored DSCP if valid bit is set.
>
>
> Modes:
>
> getdscp - If ct->mark doesn’t contain a valid DSCP, copy the DSCP
> field from the current skb (if valid for that skb ie. ip4/ip6) and set
> the valid bit.
>
> bleach/aka wash - As part of the getdscp routine, optionally null out
> the DSCP bits. We need a bleach routine because we need to see the
> DSCP bits intact so we can copy them to the mark. If cake washes them
> then we don’t get to see them, defeating the entire point.
>
> setdscp - if skb->mark valid bit set, copy the skb->mark stored DSCP
> to the diffserv bits of the current skb (if appropriate)

TC actions are executed before the qdisc, so I don't think bleach/wash
is really needed? Also, it can already be achieved with the pedit
action. So I think 'get' and 'set' are enough.

> I’ve attached a *VERY* rough starting point for how far I’ve got - not
> even compiled yet… I doubt it does. Not a natural (kernel) programmer
> so everything is a struggle.
>
> Two things I’m falling over:
>
> 1) Really not sure how to get parameters in to set the
> ‘getdscp/setdscp/wash’ options.

You'll need to add new netlink parameter values and parse them like we
do in cake (after the existing struct-copy-based option passing). I
think a single option that can be set to *either* "get" *or* "set" will
be simpler, then you don't have to worry about what happens if someone
enables both.

> 2) Is it possible to have different tc actions on an interface and a
> mirred fed ifb interface? That could be a real killer actually. We
> need to do different things on egress vs ingress.

Yes, you can chain actions, and they are per-interface.

> 3) (Spanish Inquisition moment) The original connmark code goes and
> does a nf_ct_get_tuplepr if the nf_ct_get fails.  Why would the
> nf_ct_get fail?

Presumably because nf_ct_get just reads a field in the skb; which may
not have been set yet in some cases, necessitating a full lookup.

-Toke