Re: [tsvwg] UDP options and header-data split (zero copy)

[Joseph Touch <[e-mail]>]

> On Aug 1, 2021, at 6:45 PM, Tom Herbert <[e-mail]> wrote:
> 
> 
> 
> On Sun, Aug 1, 2021, 6:13 PM Joseph Touch <[e-mail] <mailto:[e-mail]>> wrote:
> 
> RFC 9000 talks about DF being set for IPv4, but that’s the default for IPv6 (no on-path fragmentation).
> 
> It does say "no fragmentation at the IP layer”, which presumably also prohibits IP source fragmentation, though it’s not at all clear why, other than to avoid black-holing through NAT. DF being set in IPv4 is critical to RFC1191 path MTU discovery, but that does not prohibit use of source fragmentation as long as DF is set. 
> 
> Path MTU discovery doesn't work if the source fragments.

It does - it depends on what the transport wants to do.

> For instance, if a host fragments a 1500 byte packet into three packets of 500 bytes then we're probing the path with 500 bytes not 1500 which is what the transport protocol expects.

And if the packets get through, then there’s no issue. If they fail, the source will get back a PTB with the size that works at the IP layer, which informs transport.

So it will work fine - IF the user wants to let IP do source fragmentation. The only requirement is that the IP set DF, which is always implicit for IPv6.

> However, that does not prevent use of UDP fragmentation - which would both be opaque to avoiding IP fragmentation and QUIC, as well as would traverse NATs because of the replicated UDP header in UDP fragments.
> 
> UDP fragmentation similarly breaks pmtud.

It would be a layer that adapts to it in lieu of the user - if that’s what the user wants (i.e., if they ask for UDP fragmentation to be used).

Joe