Re: Fwd: New Version Notification for draft-hinden-6man-hbh-processing-01.txt

[Brian E Carpenter <[e-mail]>]

Tom,
On 11-Jun-21 10:47, Tom Herbert wrote:
> On Thu, Jun 10, 2021 at 3:20 PM Brian E Carpenter
> <[e-mail]> wrote:
>>
>> On 11-Jun-21 03:07, Tom Herbert wrote:
>>> On Thu, Jun 10, 2021 at 12:41 AM Fernando Gont
>>> <fernando.gont=[e-mail]> wrote:
>> ...
>> <big snip>
>> ...
>>> If you are pursuing that path, you should also limite the overall IPv6
>>> header chain length. That;s probably even more important than limiting
>>> the number of EHs or number of HbH options.
>>>
>>>> Fernando,
>>>
>>>> That limit and other pertinent ones are defined in RFC8504 section 5.3
>>>> and RFC8883. The limits in RFC8504 are specified for hosts, but it
>>>> would be straightforward to apply them to routers  where the behavior
>>>> when a limit is exceeded would be different-- end hosts should drop
>>>> packets that exceed the limit, intermediate nodes should stop parsing
>>>> and forward the packet. If a node decides to drop the packet then it
>>>> can send an RFC8883 ICMP error to inform the sender what the exceeded
>>>> limit was. Recommended defaults can be provided for limits; e.g. in
>>>> RFC8504 the limit of HBH or DestOpts is eight meaning that receivers
>>>> should support up to eight options and senders should be able to send
>>>> up to eight options with reasonable confidence their packets won't be
>>>> dropped. A similar default could be established for length of IP
>>>> header chain (for hosts this isn't immediately necessary since hosts,
>>>> unlike routers, usually don't have a concept of fixed sized parsing
>>>> buffer holding headers, although with hardware acceleration that limit
>>>> might be more applicable to set in the host).
>>
>> Which all, IMHO, shows that the Internet cannot ever be assumed to be
>> transparent to extension headers of any kind; either we accept that
>> extensions are confined to limited domains or we need a probing process
>> before using them, to determine whether they can survive the trip.
>>
>> Not much has changed (and it's exactly the same for IPv4 options): unusable
>> across the Internet.
> 
> Brian,
> 
> By that same logic, IPv6 itself is unusable across the Internet since
> not all Internet paths support IPv6 :-). 

Yes, I overstated my case a bit, but I do find that it is necessary
to be realistic in our designs, so I agree with what you say below.

   Brian
 
> For that matter, probably any
> protocol, with the possible exception of plain TCP/IPv4 without
> options, is "unusable" if the definition of usable is that there is a
> 100% probability that packets of a protocol will reach their
> destination.
> 
> It seems like this is an exercise of making protocols ever more
> usable-- either by fallbacks (e.g. Happy Eyeballs) as you mentioned,
> but also by clarifications to the protocol requirements to elicit
> practical implementation. For instance, RFC8200 allowing nodes to
> completely ignore HBH options instead of requiring every single node
> in the path to process was a big step in the right direction.
> Specifying a reasonable default limit for the number of options that
> should be processed or length of the IPv6 header chain that is
> supported would be another step. Note these things won't get us to
> 100% support, but they could increase reachability and hence increase
> utility of the protocols.
> 
> Tom
> 
>>
>>    Brian