Re: [PATCH] tcp: probe for SO_PEEK_OFF both in tcpv4 and tcp6

[Stefano Brivio <sbrivio@redhat.com>]

On Sat, 20 Jul 2024 10:46:07 -0400
Jon Maloy <jmaloy@redhat.com> wrote:

> My first approach to this was to condition the use of SO_PEEK_OFF with 
> tcpv4, e.g., basically
> a test like if (v4 && peek_offset_cap) {...} everywhere, but then I made 
> an interesting discovery.
> 
> It turns out that, unless the ´-4' option is explicitly given on the 
> command line, all sockets are
> v6, even those that are later used as v4 sockets.

Not necessarily: if IPv6 support is disabled for other reasons, such as
the fact that we didn't find any routable IPv6 address, sockets will
also be IPv4-only. See tcp_sock_init().

> So, the set_peek_off() 
> call failed even
> for supposedly v4 sockets.
> 
> I checked this by adding a printout to the tcp_listen_handler(), and 
> noticed that all returns from
> the accept4() call goes into the AF_INET6 branch, even when the client 
> (iperf3) call is using an IPv4 address.
> During traffic, the very same socket is marked as v4 in the tcp_tap_conn 
> structure, and this seems to
> have worked just fine until I added the set_peek_offset call().
> 
> I believe this is an issue that has been introduced during the last 
> months, since I didn't start
> using the ´-4' option consistently until some months ago, and then it 
> worked.

That's not actually an issue, it's intended, see commit 8e914238b6de
("tcp: Use dual stack sockets for port forwarding when possible").

Could it be that you enabled IPv6 on your setup meanwhile, and that's
why you see this now? I think I tested your changes on an isolated
IPv4-only setup, as I didn't run a kernel version with SO_PEEK_OFF
support at the time.

-- 
Stefano