Re: [PATCH nft 1/2] src: add dscp support

[Pablo Neira Ayuso <pablo@netfilter.org>]

On Wed, Nov 25, 2015 at 11:23:23PM +0000, Patrick McHardy wrote:
> On 25.11, Pablo Neira Ayuso wrote:
> > On Wed, Nov 25, 2015 at 08:45:13PM +0000, Patrick McHardy wrote:
> > > On 25.11, Pablo Neira Ayuso wrote:
> > > > On Wed, Nov 25, 2015 at 08:01:25PM +0000, Patrick McHardy wrote:
> > > > > On 25.11, Pablo Neira Ayuso wrote:
> > > > > >  #define TYPE_MAX		(__TYPE_MAX - 1)
> > > > > > diff --git a/include/proto.h b/include/proto.h
> > > > > > index a43bf98..41af0c1 100644
> > > > > > --- a/include/proto.h
> > > > > > +++ b/include/proto.h
> > > > > > @@ -176,7 +176,7 @@ enum ip_hdr_fields {
> > > > > >  	IPHDR_INVALID,
> > > > > >  	IPHDR_VERSION,
> > > > > >  	IPHDR_HDRLENGTH,
> > > > > > -	IPHDR_TOS,
> > > > > > +	IPHDR_DSCP,
> > > > > 
> > > > > Shouldn't we keep both? I think it should work to do that or at least not be
> > > > > difficult to make it work.
> > > > 
> > > > Last time we discussed this, people arised concerns related to the
> > > > fact that this is obsolete, and nobody should be using this these
> > > > days.
> > > 
> > > I don't really buy into these arguments. We support lots of obsolete things,
> > > the fact is that even though it is obsoleted, it still exists and for old
> > > devices might even be the only thing they support.
> > 
> > ToS breaks useful things like ECN, and the more I keep reading docs on
> > the Internet, the more problem I have to see how the user can benefit
> > from this.
> 
> We *match* on ToS, that cannot possibly break anything. Also I'm unsure how
> this could break ECN even otherwise, ToS does not even use the ECN bits.

ToS bits overlap with ECN bits, from that original ToS 8 bit-field now
we use 6 bit for DSCP and 2 bits for ECN.

> Its quite simple, if the user has old devices that set ToS values, he will
> be able to match on that without manually converting it to DSCP values.
> Given that our current tos definition is also not to practical for that
> since its too broad, I don't really care that much, although I think it
> should rather be fixed than simply thrown out.

People that designed DSCP and ECN did not care about having some
reasonable backward compatible behaviour wrt. ToS. They just changed
the semantics of those bits long time ago.

I can explore keeping this backward if you like, we can probably
accept ToS from the parser, then map it to DSCP, but that will no
achieve what the user expected on the network. I'm usually reticent to
break old stuff, but in this case I would skip.

Cheers,
Pablo