IP tunnel MTU

Mon Oct 29 14:54:46 UTC 2012

The core issue here is TCP MSS. PMTUD is a dynamic process for
adjusting MSS, but requires that ICMP be permitted to negotiate the
connection.  The realistic alternative, in a world that filters all
ICMP traffic, is to manually rewrite the MSS.  In IOS this can be
achieved via "ip tcp adjust-mss" and on Linux-based systems, netfilter
can be used to adjust MSS for example.

Keep in mind that the MSS will be smaller than your MTU.
Consider the following example:

 ip mtu 1480
 ip tcp adjust-mss 1440
 tunnel mode ipip

IP packets have 20 bytes of overhead, leaving 1480 bytes for data.  So
for an IP-in-IP tunnel, you'd set your MTU of your tunnel interface to
1480.  Subtract another 20 bytes for the tunneled IP header and 20
bytes (typical) for your TCP header and you're left with 1440 bytes
for data in a TCP connection.  So in this case we write the MSS as
1440.

I use IP-in-IP as an example because it's simple.  GRE tunnels can be
a little more complex.  While the GRE header is typically 4 bytes, it
can grow up to 16 bytes depending on options used.

So for a typical GRE tunnel (4 byte header), you would subtract 20
bytes for the IP header and 4 bytes for the GRE header from your base
MTU of 1500.  This would mean an MTU of 1476, and a TCP MMS of 1436.

Keep in mind that a TCP header can be up to 60 bytes in length, so you
may want to go higher than the typical 20 bytes for your MSS if you're
seeing problems.

On Tue, Oct 23, 2012 at 10:07 AM, Templin, Fred L
<Fred.L.Templin at boeing.com> wrote:
> Hi Roland,
>
>> -----Original Message-----
>> From: Dobbins, Roland [mailto:rdobbins at arbor.net]
>> Sent: Monday, October 22, 2012 6:49 PM
>> To: NANOG list
>> Subject: Re: IP tunnel MTU
>>
>>
>> On Oct 23, 2012, at 5:24 AM, Templin, Fred L wrote:
>>
>> > Since tunnels always reduce the effective MTU seen by data packets due
>> to the encapsulation overhead, the only two ways to accommodate
>> > the tunnel MTU is either through the use of path MTU discovery or
>> through fragmentation and reassembly.
>>
>> Actually, you can set your tunnel MTU manually.
>>
>> For example, the typical MTU folks set for a GRE tunnel is 1476.
>
> Yes; I was aware of this. But, what I want to get to is
> setting the tunnel MTU to infinity.
>
>> This isn't a new issue; it's been around ever since tunneling technologies
>> have been around, and tons have been written on this topic.  Look at your
>> various router/switch vendor Web sites, archives of this list and others,
>> etc.
>
> Sure. I've written a fair amount about it too over the span
> of the last ten years. What is new is that there is now a
> solution near at hand.
>
>> So, it's been known about, dealt with, and documented for a long time.  In
>> terms of doing something about it, the answer there is a) to allow the
>> requisite ICMP for PMTU-D to work to/through any networks within your span
>> of administrative control and b)
>
> That does you no good if there is some other network further
> beyond your span of administrative control that does not allow
> the ICMP PTBs through. And, studies have shown this to be the
> case in a non-trivial number of instances.
>
>> b) adjusting your own tunnel MTUs to
>> appropriate values based upon experimentation.
>
> Adjust it down to what? 1280? Then, if your tunnel with the
> adjusted MTU enters another tunnel with its own adjusted MTU
> there is an MTU underflow that might not get reported if the
> ICMP PTB messages are lost. An alternative is to use IP
> fragmentation, but recent studies have shown that more and
> more operators are unconditionally dropping IPv6 fragments
> and IPv4 fragmentation is not an option due to wrapping IDs
> at high data rates.
>
> Nested tunnels-within-tunnels occur in operational scenarios
> more and more, and adjusting the MTU for only one tunnel in
> the nesting does you no good if there are other tunnels that
> adjust their own MTUs.
>
>> Enterprise endpoint networks are notorious for blocking *all* ICMP (as
>> well as TCP/53 DNS) at their edges due to 'security' misinformation
>> propagated by Confused Information Systems Security Professionals and
>> their ilk.  Be sure that your own network policies aren't part of the
>> problem affecting your userbase, as well as anyone else with a need to
>> communicate with properties on your network via tunnels.
>
> Again, all an operator can control is that which is within their
> own administrative domain. That does no good for ICMPs that are
> lost beyond their administrative domain.
>
> Thanks - Fred
> fred.l.templin at boeing.com
>
>> -----------------------------------------------------------------------
>> Roland Dobbins <rdobbins at arbor.net> // <http://www.arbornetworks.com>
>>
>>         Luck is the residue of opportunity and design.
>>
>>                      -- John Milton
>>
>
>

-- 
Ray Patrick Soucy
Network Engineer
University of Maine System

T: 207-561-3526
F: 207-561-3531

MaineREN, Maine's Research and Education Network
www.maineren.net