On Fri, Jan 28, 2022 at 3:09 PM Benjamin Kaduk <[email protected]> wrote:

> On Fri, Jan 28, 2022 at 02:55:57PM -0800, David Schinazi wrote:
> > Thank you for your review Ben. Responses inline.
> >
> > David
> >
> > On Fri, Jan 28, 2022 at 2:23 PM Benjamin Kaduk via Datatracker <
> > [email protected]> wrote:
> >
> > > ----------------------------------------------------------------------
> > > DISCUSS:
> > > ----------------------------------------------------------------------
> > >
> > > Section 5 refers to a "max_packet_size" transport parameter but I do
> not
> > > see that parameter defined in the registry or RFC 9000.
> > > It seems that a transport parameter of that name was present in earlier
> > > versions of draft-ietf-quic-transport, but got renamed to
> > > max_udp_payload_size in the -28, so hopefully this is just a trivial
> > > rename.
> > >
> >
> > You're absolutely right, we missed the rename and will fix this by
> > landing your PR #76.
>
> Excellent, I'm glad this was the easy option, and as you note in the
> follow-up, it's fixed in the editor's copy already, so I'll go clear now.
>
> > ----------------------------------------------------------------------
> > > COMMENT:
> > > ----------------------------------------------------------------------
> > >
> > > I put some editorial suggestions (including the presumed resolution of
> the
> > > DISCUSS) on github at https://github.com/quicwg/datagram/pull/76 .
> > >
> >
> > Thanks, let's discuss those on the PR.
> >
> > Section 2
> > >
> > >    *  QUIC uses a more nuanced loss recovery mechanism than the DTLS
> > >       handshake, which has a basic packet loss retransmission timer.
> > >
> > > This is true of DTLS 1.2 and prior versions, which technically is right
> > > now the current version of DTLS.  However, it's not quite true of DTLS
> > > 1.3, which includes an explicit ACK message to supplement the
> > > retransmission timer.  DTLS 1.3 stands a pretty decent chance of being
> > > published as an RFC prior to this document (per ekr, it should have the
> > > last technical changes from the WG finalized this weekend and then go
> into
> > > the "real" AUTH48 state), so I think we ought to speak to the
> mechanisms
> > > of DTLS 1.3 here.
> > >
> >
> > How about we just remove "which has a basic packet loss retransmission
> > timer"?
> > I think it's safe to say that QUIC congestion control is more nuanced
> than
> > that of
> > any version of DTLS, including DTLS 1.3.
>
> That seems reasonable.  I agree that QUIC CC is clearly more nuanced than
> DTLS of any version.  And the value of going into any level of detail about
> what DTLS actually does provide seems pretty minimal.
>

Thanks. I've now made that change to the editor's copy.

David

> Section 3
> > >
> > >    For most uses of DATAGRAM frames, it is RECOMMENDED to send a value
> > >    of 65535 in the max_datagram_frame_size transport parameter to
> > >    indicate that this endpoint will accept any DATAGRAM frame that fits
> > >    inside a QUIC packet.
> > >
> > > It's interesting to compare this to the RFC 9000 max_udp_payload_size
> > > default of 65527, the maximum permitted UDP payload.  Indeed, the QUIC
> > > 1-RTT packet header does not even contain a length field that would
> limit
> > > the frame size.  So I'm not entirely sure what motivates the 65535
> value
> > > specifically.  (I do see the subsequent discussion about how there are
> > > other factors, including max_packet_size/max_udp_payload_size, that can
> > > further limit what is usable.)
> > >
> >
> > We picked that value because it means "don't limit the length" and
> doesn't
> > involve thinking about header sizes or doing math.
>
> Okay.  I can't fault that, even if it seems that since we're bounded by
> max_udp_payload_size we could just as easily reuse the 65527 value.
>
> Thanks for the quic(k) turnaround!
>
> -Ben
>

Reply via email to