sctp doesn't honour net.ipv6.bindv6only

sctp doesn't honour net.ipv6.bindv6only

[David Laight]

The sctp code doesn't use sk->sk_ipv6only (which is initialised
from net.ipv6.bindv6only) but instead uses its own flag
sp->v4mapped which always defaults to 1.

There may also be an expectation that
  [gs]etsockopt(sctp_fd, IPPROTO_IPV6, IPV6_V6ONLY,...)
will access the flag that sctp uses internally.
(Matching TCP and UDP.)

Patch is trivial (remembering the inverted state)
but I don't have a clean enough source tree :-(

	David

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

Re: sctp doesn't honour net.ipv6.bindv6only

[Marcelo Ricardo Leitner]

On Tue, May 19, 2020 at 10:47:17AM +0000, David Laight wrote:
> The sctp code doesn't use sk->sk_ipv6only (which is initialised
> from net.ipv6.bindv6only) but instead uses its own flag

It actually does, via [__]ipv6_only_sock() calls since 7dab83de50c7
("sctp: Support ipv6only AF_INET6 sockets.").

> sp->v4mapped which always defaults to 1.
> 
> There may also be an expectation that
>   [gs]etsockopt(sctp_fd, IPPROTO_IPV6, IPV6_V6ONLY,...)
> will access the flag that sctp uses internally.
> (Matching TCP and UDP.)

My understanding is that these are slightly different.

v4mapped, if false, will allow the socket to deal with both address
types, without mapping. If true, it will map v4 into v6.
v6only, if false, it will do mapping for tcp/udp, but sctp won't use
it. If true, it will deny using v4, which is complementary to v4mapped
for sctp.

Did I miss anything?
 
  Marcelo

RE: sctp doesn't honour net.ipv6.bindv6only

[David Laight]

From: Marcelo Ricardo Leitner
> Sent: 19 May 2020 20:47
> 
> On Tue, May 19, 2020 at 10:47:17AM +0000, David Laight wrote:
> > The sctp code doesn't use sk->sk_ipv6only (which is initialised
> > from net.ipv6.bindv6only) but instead uses its own flag
> 
> It actually does, via [__]ipv6_only_sock() calls since 7dab83de50c7
> ("sctp: Support ipv6only AF_INET6 sockets.").
> 
> > sp->v4mapped which always defaults to 1.
> >
> > There may also be an expectation that
> >   [gs]etsockopt(sctp_fd, IPPROTO_IPV6, IPV6_V6ONLY,...)
> > will access the flag that sctp uses internally.
> > (Matching TCP and UDP.)
> 
> My understanding is that these are slightly different.
> 
> v4mapped, if false, will allow the socket to deal with both address
> types, without mapping. If true, it will map v4 into v6.
> v6only, if false, it will do mapping for tcp/udp, but sctp won't use
> it. If true, it will deny using v4, which is complementary to v4mapped
> for sctp.
> 
> Did I miss anything?

Possibly I did, I wasn't looking much beyond the [sg]etsockopt code.
Although our code supports SCTP/IPv6 and I have tested it a bit
I don't think any of our customers use it (yet).
We default to creating IPv6 listening sockets but all the connections
are IPv4.

I think I'm still confused though:

IIRC v6only (mainly) affects listening sockets.
If 0 (the default on Linux) an IPv4 connection will 'attach to' an
IPv6 socket and the application will see v4mapped addresses [1].
If 1 the application needs to create two separate sockets to receive
both IPv4 and IPV6 connections.

I can't see how SCTP would be any different to TCP and UDP.
It can't make any sense to dual-home with a mixture of IPv4/6 addresses.

So does v4mapped just control the format of the addresses on the socket
interface when an IPv4 connection is using an IPv6 socket? 

[1] Actually, thinking further I can't remember whether this is true.
All our code allows for v4mapped addresses and decodes them for printing.

	David

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