Received-SPF: pass (google.com: domain of linux-nfs-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Content-Type: text/plain; charset="utf-8"
Content-Transfer-Encoding: 7bit
MIME-Version: 1.0
From:   "NeilBrown" <neilb@suse.de>
To:     "Jeff Layton" <jlayton@kernel.org>
Cc:     "Petr Vorel" <pvorel@suse.cz>, ltp@lists.linux.it,
        "Cyril Hrubis" <chrubis@suse.cz>, linux-nfs@vger.kernel.org
Subject: Re: [PATCH 1/1] nfslock01.sh: Don't test on NFS v3 on TCP
In-reply-to: <d441b9f9dfcbb4719d97c7b3b5950dfeeb8913d2.camel@kernel.org>
References: <20230502075921.3614794-1-pvorel@suse.cz>,
 <d441b9f9dfcbb4719d97c7b3b5950dfeeb8913d2.camel@kernel.org>
Date:   Wed, 03 May 2023 07:21:36 +1000
Message-id: <168306249608.19756.8898498995753159325@noble.neil.brown.name>
Precedence: bulk

On Tue, 02 May 2023, Jeff Layton wrote:
> On Tue, 2023-05-02 at 09:59 +0200, Petr Vorel wrote:
> > 
> > Although there is suggestion, how to fix the problem in kernel [2]:
> > 
> >     > Maybe rpcb_create_local() shall detect that it is not in root
> >     > netns, and only try AF_INET connection to > localhost in that case.
> > 
> >     That would be simple and might be sensible.  IF changing the AF_UNIX
> >     path to "/run/rpcbind.sock" isn't sufficient, then testing for the
> >     root_ns is probably the best second option.
> > 
> 
> Was it determined that changing the location of the socket wasn't
> sufficient to fix this? FWIW, My Fedora 38 machine seems to listen on
> that socket already:
> 
>     [Socket]
>     ListenStream=/run/rpcbind.sock

I think the best solution for this problem is to change the kernel to
first try to send to an abstract socket: "\0/run/rpcbind.sock".  Only if
that fails do we try "/run/rpcbind.sock".

We also change rpcbind to listen on both
   ListenStream=@/run/rpcbind.sock
   ListenStream=/run/rpcbind.sock

Abstract sockets are local to a network namespace, while non-abstract
Unix domain sockets are local to a file and so can only be local to a
mount namespace.  We clearly need rpcbind lookup from the kernel to be
netns-local, so abstract is the obvious choice.

NeilBrown