Re: Strange problem with SCTP+IPv6

From: Corey Minyard <minyard@acm.org>
To: David Laight <David.Laight@ACULAB.COM>
Cc: 'Marcelo Ricardo Leitner' <marcelo.leitner@gmail.com>,
	Michael Tuexen <Michael.Tuexen@lurchi.franken.de>,
	Xin Long <lucien.xin@gmail.com>,
	Vlad Yasevich <vyasevich@gmail.com>,
	Neil Horman <nhorman@tuxdriver.com>,
	"linux-sctp@vger.kernel.org" <linux-sctp@vger.kernel.org>,
	LKML <linux-kernel@vger.kernel.org>
Subject: Re: Strange problem with SCTP+IPv6
Date: Tue, 23 Jun 2020 16:17:56 +0000	[thread overview]
Message-ID: <20200623161756.GE3235@minyard.net> (raw)
In-Reply-To: <c1121947c9a94703b4ab6dc434a7c3f8@AcuMS.aculab.com>

On Tue, Jun 23, 2020 at 01:17:28PM +0000, David Laight wrote:
> From: Marcelo Ricardo Leitner
> > Sent: 22 June 2020 19:33
> > On Mon, Jun 22, 2020 at 08:01:24PM +0200, Michael Tuexen wrote:
> > > > On 22. Jun 2020, at 18:57, Corey Minyard <minyard@acm.org> wrote:
> > > >
> > > > On Mon, Jun 22, 2020 at 08:01:23PM +0800, Xin Long wrote:
> > > >> On Sun, Jun 21, 2020 at 11:56 PM Corey Minyard <minyard@acm.org> wrote:
> > > >>>
> > > >>> I've stumbled upon a strange problem with SCTP and IPv6.  If I create an
> > > >>> sctp listening socket on :: and set the IPV6_V6ONLY socket option on it,
> > > >>> then I make a connection to it using ::1, the connection will drop after
> > > >>> 2.5 seconds with an ECONNRESET error.
> > > >>>
> > > >>> It only happens on SCTP, it doesn't have the issue if you connect to a
> > > >>> full IPv6 address instead of ::1, and it doesn't happen if you don't
> > > >>> set IPV6_V6ONLY.  I have verified current end of tree kernel.org.
> > > >>> I tried on an ARM system and x86_64.
> > > >>>
> > > >>> I haven't dug into the kernel to see if I could find anything yet, but I
> > > >>> thought I would go ahead and report it.  I am attaching a reproducer.
> > > >>> Basically, compile the following code:
> > > >> The code only set IPV6_V6ONLY on server side, so the client side will
> > > >> still bind all the local ipv4 addresses (as you didn't call bind() to
> > > >> bind any specific addresses ). Then after the connection is created,
> > > >> the client will send HB on the v4 paths to the server. The server
> > > >> will abort the connection, as it can't support v4.
> > > >>
> > > >> So you can work around it by either:
> > > >>
> > > >>  - set IPV6_V6ONLY on client side.
> > > >>
> > > >> or
> > > >>
> > > >>  - bind to the specific v6 addresses on the client side.
> > > >>
> > > >> I don't see RFC said something about this.
> > > >> So it may not be a good idea to change the current behaviour
> > > >> to not establish the connection in this case, which may cause regression.
> > > >
> > > > Ok, I understand this.  It's a little strange, but I see why it works
> > > > this way.
> > > I don't. I would expect it to work as I described in my email.
> > > Could someone explain me how and why it is behaving different from
> > > my expectation?
> > 
> > It looks like a bug to me. Testing with this test app here, I can see
> > the INIT_ACK being sent with a bunch of ipv4 addresses in it and
> > that's unexpected for a v6only socket. As is, it's the server saying
> > "I'm available at these other addresses too, but not."
> 
> Does it even make sense to mix IPv4 and IPv6 addresses on the same
> connection?
> I don't remember ever seeing both types of address in a message,
> but may not have looked.

That's an interesting question.  Do the RFCs say anything?  I would
assume it was ok unless ipv6only was set.

> 
> I also wonder whether the connection should be dropped for an error
> response on a path that has never been validated.

That actually bothered me a bit more.  Shouldn't it stay up if any path
is up?  That's kind of the whole point of multihoming.

> 
> OTOH the whole 'multi-homing' part of SCTP sucks.

I don't think so.

> The IP addresses a server needs to bind to depend on where the
> incoming connection will come from.
> A local connection may be able to use a 192.168.x.x address
> but a remote connection must not - as it may be defined locally
> at the remote system.
> But both connections can come into the public (routable) address.
> We have to tell customers to explicitly configure the local IP
> addresses - which means the application has to know what they are.
> Fortunately these apps are pretty static - usually M3UA.

Umm, no,  If you have a private address, it better be behind a firewall,
and the firewall should handle rewriting the packet to fix the addresses.

It doesn't appear that Linux netfilter does this.  There is a TODO in
the code for this.  But that's how it *should* work.

-corey

> 
> 	David
> 
> -
> Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
> Registration No: 1397386 (Wales)
> 