Re: [PATCH 3/3] conf: Be more precise about minimum MTUs

[Stefano Brivio <sbrivio@redhat.com>]

On Thu, 20 Feb 2025 14:55:30 +1100
David Gibson <david@gibson.dropbear.id.au> wrote:

> On Wed, Feb 19, 2025 at 06:37:28AM +0100, Stefano Brivio wrote:
> > On Wed, 19 Feb 2025 14:14:29 +1100
> > David Gibson <david@gibson.dropbear.id.au> wrote:
> >   
> > > Currently we reject the -m option if given a value less than ETH_MAX_MTU  
> > 
> > ETH_MIN_MTU
> >   
> > > (68).  That define is derived from the kernel, but its name is misleading:
> > > it doesn't really have anything to do with Ethernet per se, but is rather
> > > the minimum payload any L2 link must be able to handle in order to carry
> > > IPv4.  
> > 
> > Yes, that should be IPV4_MIN_MTU instead, but it was only added as
> > recently as 4.14 kernels, so I opted for ETH_MIN_MTU. A misnomer as you
> > pointed out, but safe.  
> 
> Ah, thanks, I hadn't realised that newer kernels had better named
> constants.  When I respin I'll use matching names.
> 
> > > For IPv6, it's not sufficient: that requires an MTU of at least
> > > 1280.
> > > 
> > > Furthermore, the value of 68 is the minimum IP *fragment* size the link
> > > must be able to carry.  Since we don't support IP fragmentation, it's not
> > > sufficient for us.  Instead we should clamp the MTU to 576 for IPv4 - the
> > > minimum IP datagram size that all hosts must be able to accept.  
> > 
> > First off, the only assumption in RFC 791 terms we can _perhaps_ make is
> > that we are some kind of "module" (also called "node", could be host or
> > router), not a (full) host. Maybe not even a module. So, with that
> > regard, we don't need to be prepared to _accept_ (for ourselves as
> > destination) any particular datagram size.
> > 
> > Second, even if all hosts need to be able to accept 576-byte datagrams,
> > that doesn't mean that all links need to be able to carry them. The MTU
> > refers _to the link_, not to what a host is able to accept.  
> 
> Ah... yes.  I was thinking that that requirement implied that a link
> which can't fragment was useless if it couldn't carry 576-byte
> datagrams, but thinking over your examples here I realise I was
> mistaken.
> 
> > And that's the reason why you can set 68 bytes as MTU on most network
> > interfaces on Linux. We set sub-576 values ourselves in tests:
> > 
> > $ grep -rn "mtu 256" *
> > passt_tcp:95:guest	ip link set dev __IFNAME__ mtu 256
> > passt_vu_tcp:95:guest	ip link set dev __IFNAME__ mtu 256
> > 
> > That is, indeed, all hosts (not "modules") need to be able to accept
> > (not "forward") datagram sizes of at least 576 bytes... but that's only
> > assuming you can deliver those datagrams to them.
> > 
> > This is not just a theoretical matter. As late as 2018, I was made
> > aware of a setup with several (local!) nodes with links between them
> > having ~380 bytes as MTU.
> > 
> > Sure enough, the reason why I know about this was an issue coming from
> > the same flawed assumption made in kernel commit c9fefa08190f
> > ("ip6_tunnel: get the min mtu properly in ip6_tnl_xmit"), and fixed by
> > 82a40777de12 ("ip6_tunnel: use the right value for ipv4 min mtu check
> > in ip6_tnl_xmit").
> > 
> > See also commit b4331a681822 ("vti6: Change minimum MTU to IPV4_MIN_MTU,
> > vti6 can carry IPv4 too") on the subject of what links can carry vs.
> > what endpoints should be able to forward.
> >   
> > > Move the verification of the MTU's lower bound to logic specific to the IP
> > > versions and correct those errors.
> > > 
> > > Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
> > > ---
> > >  conf.c | 20 +++++++++++++++-----
> > >  ip.h   |  7 +++++++
> > >  util.h |  3 ---
> > >  3 files changed, 22 insertions(+), 8 deletions(-)
> > > 
> > > diff --git a/conf.c b/conf.c
> > > index c5ee07b0..e127acc1 100644
> > > --- a/conf.c
> > > +++ b/conf.c
> > > @@ -1663,9 +1663,9 @@ void conf(struct ctx *c, int argc, char **argv)
> > >  			if (errno || *e)
> > >  				die("Invalid MTU: %s", optarg);
> > >  
> > > -			if (mtu && (mtu < ETH_MIN_MTU || mtu > ETH_MAX_MTU)) {
> > > -				die("MTU %lu out of range (%u..%u)", mtu,
> > > -				    ETH_MIN_MTU, ETH_MAX_MTU);
> > > +			if (mtu > ETH_MAX_MTU) {
> > > +				die("MTU %lu too large (max %u)",
> > > +				    mtu, ETH_MAX_MTU);
> > >  			}
> > >  
> > >  			c->mtu = mtu;
> > > @@ -1838,10 +1838,20 @@ void conf(struct ctx *c, int argc, char **argv)
> > >  	log_conf_parsed = true;		/* Stop printing everything */
> > >  
> > >  	nl_sock_init(c, false);
> > > -	if (!v6_only)
> > > +	if (!v6_only) {
> > > +		if (c->mtu < IPV4_MINMAX_DATAGRAM) {  
> > 
> > Now, if you want to make this symmetric with the IPv6 case, we could
> > also move this here... it just unnecessarily adds lines of code, and
> > this function is already (necessarily) rather long.  
> 
> Sorry, I'm not following what change you're suggesting (or discussing?).

The exact change I quoted: moving the check on the minimum MTU to here:

		if (c->mtu < IPV4_MINMAX_DATAGRAM) {  

compared to doing it earlier in conf().

> > > +			die("MTU %"PRIu16" is too small for IPv4 (minimum %u)",
> > > +			    c->mtu, IPV4_MINMAX_DATAGRAM);
> > > +		}
> > >  		c->ifi4 = conf_ip4(ifi4, &c->ip4);
> > > -	if (!v4_only)
> > > +	}
> > > +	if (!v4_only) {
> > > +		if (c->mtu < IPV6_MIN_MTU) {
> > > +			die("MTU %"PRIu16" is too small for IPv6 (minimum %u)",
> > > +			    c->mtu, IPV6_MIN_MTU);  
> > 
> > Does the fact that we don't disable IPv6 imply that IPv6 must be
> > working at all times? In my opinion not.
> > 
> > It's also rather convenient to be able to specify '-m 200' (for
> > whatever test) without having to give '-4' explicitly.
> >   
> > >From a functionality perspective, I think warn() would be a better  
> > choice.  
> 
> warn() and disable the relevant protocol.  That makes sense, I'll make
> that change.

I don't think it makes sense to disable IPv4, highlighting quote:

> > Does the fact that we don't disable IPv6 imply that IPv6 must be
> > working at all times? In my opinion not.

...you can advertise a small MTU for whatever reason. The guest might
configure it or not. The guest might change it later on. We have no way
to re-enable IPv6 once it's disabled, though.

So let's just do what the user says, I would suggest, and warn them
that it *might* not work. There is zero functionality gained by
disabling IPv6.

-- 
Stefano