Re: [mpls] [bier] The first nibble issue associated with MPLS encapsulation

[Eric C Rosen <[e-mail]>]

On 4/14/2016 6:06 AM, Alexander Vainshtein wrote:
> Stewart and all,
> I concur with Stewart that there is a strong case for 0 in the first nibble for all non-IP flows.
>
> As for the need for sub-typing:
> AFAIK quite a few implementations (including some HW-based packet processors) treat 0 in the first nibble after the label stack as an indication of an Ethernet PW.
>
> Some of them go as far as to hash on the assumed L2 headers for ECMP. This causes serious problems, e.g., with the TDM PWs that could be reordered if handled by such packet processors in transit LSRs.
>
> This makes quite a case for sub-typing IMO regardless of BIER.
> At the same time, it seems that all the bits in CW structure are used - at least for some PW types in some cases.

It seems to me that Sasha's reasoning supports the conclusion that it is 
best to avoid 0 in the first nibble (even if this is not the conclusion 
he drew!)  I don't think sub-typing is a solution, because it doesn't 
offer any protection against the behavior of existing hardware.

On 4/14/2016 5:31 AM, Stewart Bryant wrote:
> The nibble value  is recorded in the IP types registry
> and any wish to take another value really needs to be discussed
> with the INT area.

Are you referring to the IP Version Number registry?  (I am not aware of 
an "IP types registry".)   We're not asking for any modification to 
that, since we're not doing another version of IP. I don't see any need 
to get into  one of those long and non-productive arguments that the INT 
area seems to relish.

On 4/13/2016 11:07 PM, Xuxiaohu wrote:
> In fact, there is an attempt to have a registry for the first nibble (a.k.a., MPLS payload type field).

Hopefully that attempt won't succeed!

On 4/14/2016 7:06 AM, Loa Andersson wrote:
> While I can live with 0x0000, 0x0010 or 0x0101, RFC 4928 actually says:
>
>    It is REQUIRED, however, that applications depend upon in-order
>    packet delivery restrict the first nibble values to 0x0 and 0x1.
>
> If that is what we want for bier, there is a case to use 0x0 or 0x1 for
> bier- 

Strictly speaking, RFC 4928 needs to be updated, as the reasoning given 
there does not really support the above "requirement".  But the effort 
to do that hardly seems worthwhile.  RFC 4928 already makes clear that 
the above requirement is protecting against the risk that IPv5 ECMP will 
affect some MPLS packets inappropriately. No one seems to think that 
that is a real risk.