Re: [Ntp] CLOCK_TAI (was NTPv5: big picture)

[FUSTE Emmanuel <[e-mail]>]

Le 06/01/2021 à 15:29, Martin Burnicki a écrit :
> FUSTE Emmanuel wrote:
>> Le 06/01/2021 à 11:58, Martin Burnicki a écrit :
>>> Or you would have to provide (and maintain) an NTP server implementation
>>> that supports both v4 and v5, depending on the version of the request.
>> As the vast majority of today ones which still support v1/v2/v3/v4,
>> being based on ntpd,  and v3/v4 for most of the others.
> I don't thing NTPv1 and v2 are really supported by current
> implementations, maybe except for a v2 version code in the mode 6 packets.
>
> If I remember correctly, NTPv2 wasn't even widespread, but some fields
> of the packet were changed in v3. When that happened, NTPv2 wasn't very
> widespread, either.
>
> The only thing a v4 server has to do to when it receives a v3 request is
> to copy the v3 version code into the reply packet instead of the native
> v4 code. That's it.
>
> In a compatible approach you could set up the reply packet as usually,
> and only append extension fields with more detailed information if the
> protocol version sent by the client is v5.
>
> A possible approach could be that
>
> - a v5 client could send an initial v5 request
>
> - when the server sees v5 it replies with a base packet and one or more
> extension fields indicating its features/capabilities
>
> - the client could send further requests with extension fields
> requesting for additional information
>
> - a v4 client would simply send and receive a v4 packet without the
> extensions
Ok we mostly agree.
You advocate for a v5 format which permit to do less implementation work 
in a dual v4/v5 server.
And I prefer a departure from UTC like timestamps which imply a little 
more bit of work on the server side for v4 clients as the base packet 
will not have the same meaning and will contain some "mandatory" extensions.

> I think whether hardware timestamping is supported or not isn't related
> to the packet format. Even NTPv4 can use timestamping in the kernel
> driver, if the OS supports it.
Yes but as the common time-stamping hardware only has and support only 
one time-stamping clock and is responsible to put this clock value in 
the outbound packet, format matter.
You could not share the same interface with hardware time-stamping 
between NTPv4 and PTP for this reason unless you are on custom or very 
uncommon hardware.
For the receiving part it does not matter as the timestamp is purely a 
local measure which could be converted to whatever you want.

Emmanuel.