Re: [ntpwg] The next step of draft-ietf-ntp-checksum-trailer

[Harlan Stenn <[e-mail]>]

Miroslav Lichvar writes:
> On Mon, Mar 07, 2016 at 12:35:56PM +0000, Harlan Stenn wrote:
> > Originally, if there was an extension field there MUST be a MAC.
> 
> Right. That requirement was not necessary and it was later removed.

Really?  Please say more.  Where can I learn about this?

> > Tal wrrote Errata 3627 (or whatever) and removed that requirement.
> 
> Yes, but why is that a problem?

Because one can legislate the value of pi to be 3 since that would
make so many things easier.  That doesn't mean it's going to work in the
real world.

>> MAC protection is good.  NTS uses a 'single' Field Type, but the first
>> few packets have no MAC protection.  Therefore, the first few packets of
>> an NTS exchange arguably MUST have MAC protection outside the NTS
>> packet
> 
> With what key would that MAC be generated? Unless I misunderstood how
> NTS is supposed to work, the first few packets are simply not
> authenticated and can't be used for anything except initialization of
> the NTS association.

Symmetric key would be an OK choice here.  Preferably using the "MAC in
an EF".

>> We're getting ready to publish a "MAC in an Extension Field", and
>> arguably this is how it should have been originally done.  So in keeping
>> with the above logic, there are 2 pieces of information we want to see
>> for an extension field.  One is "SHOULD or MAY have MAC protection after
>> this field," and the other is "This extension field provides MAC
>> protection."
> 
> What will the receiver of this information do with it?

Whatever it wants.  This goes to giving the client basic information
that is useful in deciding what to do in case the client  gets EFs it
doesn't understand.

> The client must
> have its own policy on what extension fields has to be covered with
> MAC in order to accept the packet. The server can't dictate that.

Of course.

> Normally, the client will require all fields to be covered by MAC if
> authentication is enabled.

Yes.

So where is the confusion here?

>>>> - Use a flag bit in the Field Type that says "My extension field does
>>>> not require a following MAC."
>>> 
>>> Meaning the field is not related to authentication?
>> 
>> Meaning "My extension field does not require a MAC after it".  This is
>> the "MAY" case as opposed to the "SHOULD" case.  An example of this is
>> Tal's checksum complement proposal, as currently written.  If ntpd sends
>> out an unauthenticated packet, the FPGA engine is free to rewrite the
>> transmit timestamp, but since the OS has already computed a UDP
>> checksum, if the transmit timestamp in the packet is changed it's almost
>> certain that a checksum complement will be needed to make sure the
>> existing UDP checksum will still be correct.  That's what Tal's proposal
>> is all about, and it's the first time we've seen a proposal for an
>> extension field where one MUST NOT have a MAC at the end of the packet.
> 
> Ok, but that information is already included in the specification. The
> client or server that generates that field needs to know that.

The >sender< needs to know.  The >recipient< might be running old
software, current software, or software that's aware of these new EFs.

This all goes to what happens down the road when a newer sender is
talking to an older recipient.

If a recipient gets data it does not understand, it may benefit from a
hint as to what's going on with that unknown EF.

>>> The SHA1 MAC is 24 octets, that's why there is a requirement for the
>>> last extension field to be at least 28 octets.
>> 
>> 4 byte keyid plus 24 octets.  So 28 bytes is clearly potentially
>> ambiguous.
> 
> A SHA1 MAC has 160 bits for the digest and 32 bit for the key ID.
> That's 192 bits or 24 octets total.

OK.  What I remember is hearing that we were seeing 3 sizes of old-style
MACs, and they were 20, 24, and 28 bytes' long.  I'll see what I can
find out.

>> Extension fields of 4, 8, 12, or 16 bytes are also unambiguous.  Why
>> prohibit these lengths?
> 
> Because MAC in a packet from an old client could be parsed as one or
> multiple extension fields.

An old client will ONLY send autokey.  So this is a non-issue, right?

It's pretty easy to parse an EF and make sure it's sane.  If the current
spot in the extended payload parses as a good EF, then one can do a
quick check on any following data in the payload and get a good idea as
to what's going on.  At worst, you have a packet that CORRECTLY parses
as both "a bunch of EFs" and "an old-style MAC in a packet that either
has no extension field or is an autokey packet".  If only one of those
interpretations is correct, then we can dance on the head of the
false-negative or false-positive pin.

Remember, this is *only* a POTENTIAL problem when mixing new and old
implementations.  In that case folks may need to administratively
segregate these machines.  We can certainly solve this problem in NTPv5.

>>>> - a MAC consists of a 4-byte keyid and a payload.  If there is no
>>>> autokey extension field, the keyid MUST be <65535, and we SHOULD have
>>> 
>>> Isn't that restriction for keyid just for the autokey RFC? There are
>>> NTP implementations that allow any keyid with symmetric keys.
>> 
>> Originally:
>> 
>> - any extension field meant there MUST be a MAC.
>> - the only defined extention field as been autokey.
>> - if no extension field was present, then no autokey, and a MAC meant
>>   symmetric key was being used.  keyid < 65535.
>> - if an extension field was present, the autokey was being used, there
>>   MUST be a MAC, and  keyid > 65535.
> 
> That applies to the reference implementation, but not necessarily for
> other NTPv4 implementations. They can parse extension fields, but they
> don't have to know anything about autokey and can authenticate with
> MACs using key IDs above 65535.

The RFCs have clear requirements on the keyids.  If an implementation
chooses to be non-compliant that's the implementer's choice.  An
implementation that makes this choice has no leg to stand on if it
doesn't interoperate with compliant implementations.

And in this case it's not really that bad either - this older and
non-compliant implementation simply has to see if the 32-bit keyid is a
valid key in order to disambiguate the packet.

But we're back to the point where if you're running NTP networks where
you have both old and new NTP instances, you may need to have a way to
identify which machines are "old" and which are "new" and communicate
accordingly.

H
_______________________________________________
ntpwg mailing list
[e-mail]
http://lists.ntp.org/listinfo/ntpwg