Re: Non-Last Small IPv6 Fragments

[Ole Troan <[e-mail]>]

> On 14 Jan 2019, at 21:40, Erik Kline <[e-mail]> wrote:
> 
>> On Sun, 13 Jan 2019 at 22:24, Fernando Gont <[e-mail]> wrote:
>> 
>> Hi, Erik,
>> 
>>> On 14/1/19 02:19, Erik Kline wrote:
>>> 
>>>> Ole Troan wrote on 13/01/2019 20:03:
>>>>> Let’s fix path MTU discovery. There’s no fix for fragmentation.
>>>> 
>>>> pmtu discovery is hard because it needs a way for an intermediate
>>>    node to be able to signal a transmitting node to dynamically drop
>>>    the MTU if network conditions change.  The only way to reliably work
>>>    around this is to transmit at 1280 and claim implementation /
>>>    operational breakage if this cannot get through.  This, however,
>>>    robs the host devices of the ability to use higher MTUs.
>>> 
>>>    “Fix” as in something different than rfc8201.
>>> 
>>> 
>>> How about a version of the fragment header that:
>>> 
>>>    (a) *always* has the L4 header (or a configurable number of bytes)
>>> in every non-first fragment
>>> 
>>> This could be documented on a per-L4+ basis how much needs to be
>>> included, i.e. we could have one doc for UDP and (say) a separate doc
>>> for information QUIC would need on a per-fragment basis.
>> 
>> If you are requiring that the first frag always contains the L4 ehader,
>> that's already achieved by RFC7112.
> 
> Not what I suggested.  I said every fragment header would include the
> L4 header (or other specifiable bytes).
> 
>> If you really mean that the FH contains the L4 header, then:
>> 
>> * There's the issue of EHs in the fragmentable part
> 
> That would be addressed in an associated L4 doc, and it would probably
> say to ignore non-transport headers.
> 
>> * It may be hard to enforce in the presence of tunnels, unless the
>> tunnel reassembles and refragments where necesary
>> 
>> 
>> Still, that doesn't solve the biggest issues with fragmention:
>> 
>> * a stateful operation for anotherwise stateless protocol
>> 
>> * i'd argue that for the IPv6 case, one of the base reasons for which
>> fragments are dropped is because they employ EHs
>> 
>> 
>> 
>>> This increases the overhead in a given fragment, but also helps to
>>> ensure that (eventually) intermediate systems can examine this field and
>>> preemptively make a drop/no-drop decision.
>> 
>> Then we're back on draft-gont-v6ops-ipv6-ehs-packet-drops
> 
> Understood, but any L3 solution is going to necessitate running into
> this.  UDP trailers is certainly cute (and +1 from me), but every
> transport will have to have some similar fix (or applications will
> have to consider rfc8085#section-3.2 type guidelines).  If there's to
> be an L3 solution to this L3 problem it will face some deployment
> issues but if its usefulness is sufficiently attractive it could find
> reasonably wide-spread support in due time (in the absence of ipng-ng,
> IPv6 is it for the long haul).
> 
> [Silly Idea #1]
> 
> (/me thinks back to spud bof)
> 
> Here's another random suggestion that maybe belongs in panrg or (more
> likely) the dumpster:
> 
> A new hop-by-hop header under "00 1", i.e.
> 
>    00 - skip over this option and continue processing the header
> 
>    1 - Option Data may change en route
> 
> of the form, for example, Option Type 00 1 00001, Opt Data Len 2,
> Option Data == MTU.  Here, the observation is that the main thing we
> have experience with "working" (for the most generous definition of
> "works") in this problemspace in the internet today is TCP MSS
> Clamping.  So let's have an MTU (or MSS) HbH option that can be
> revised monotonically downwards but never lower than 1280.  If it
> helps pad to 8 byte alignment we can have a 2nd 2-byte field that is
> untouched representing the sender's original MTU/MSS.
> 
> In this way, a sender could include an HbH with either:
> 
>    [HbH | 0 | 0x0202 | 2 | (1500=0x5dc) | Pad1 | Pad1]
> 
> or
> 
>    [HbH | 0 | 0x0202 | 4 | (1500=0x5dc) | (1500=0x5dc) ]
> 
> with one of the MTU/MSS fields to be munged along the path.
> 
> If this were the only HbH option included it adds only 8 bytes, fits
> (along with most L4 headers) within a 128 byte hardware classifier,
> and end systems could explore its usefulness via ping.  (Separately,
> there could of course be sockopt/cmsg things to set/get the values.)

https://tools.ietf.org/html/draft-hinden-6man-mtu-option-00


> 
> [Silly Idea #2]
> Separate, even sillier idea: everyone converges on using 16 bits of
> the flow ID to encode the lowest MTU encountered along the path.  Here
> again, TCP MSS clamping style behaviour would be applied to this
> field.