Re: [PATCH bpf-next 0/3] Introduce pinnable bpf_link kernel abstraction

From: Alexei Starovoitov <alexei.starovoitov@gmail.com>
To: Daniel Borkmann <daniel@iogearbox.net>
Cc: "Toke Høiland-Jørgensen" <toke@redhat.com>,
	"Alexei Starovoitov" <ast@fb.com>,
	"Andrii Nakryiko" <andrii.nakryiko@gmail.com>,
	"Andrii Nakryiko" <andriin@fb.com>, bpf <bpf@vger.kernel.org>,
	Networking <netdev@vger.kernel.org>,
	"Kernel Team" <kernel-team@fb.com>
Subject: Re: [PATCH bpf-next 0/3] Introduce pinnable bpf_link kernel abstraction
Date: Thu, 5 Mar 2020 14:50:28 -0800	[thread overview]
Message-ID: <20200305225027.nazs3gtlcw3gjjvn@ast-mbp> (raw)
In-Reply-To: <473a3e8a-03ea-636c-f054-3c960bf0fdbd@iogearbox.net>

On Thu, Mar 05, 2020 at 11:34:18PM +0100, Daniel Borkmann wrote:
> On 3/5/20 5:34 PM, Alexei Starovoitov wrote:
> > On Thu, Mar 05, 2020 at 11:37:11AM +0100, Toke Høiland-Jørgensen wrote:
> > > Alexei Starovoitov <alexei.starovoitov@gmail.com> writes:
> > > > On Wed, Mar 04, 2020 at 08:47:44AM +0100, Toke Høiland-Jørgensen wrote:
> [...]
> > > Anyway, what I was trying to express:
> > > 
> > > > Still that doesn't mean that pinned link is 'immutable'.
> > > 
> > > I don't mean 'immutable' in the sense that it cannot be removed ever.
> > > Just that we may end up in a situation where an application can see a
> > > netdev with an XDP program attached, has the right privileges to modify
> > > it, but can't because it can't find the pinned bpf_link. Right? Or am I
> > > misunderstanding your proposal?
> > > 
> > > Amending my example from before, this could happen by:
> > > 
> > > 1. Someone attaches a program to eth0, and pins the bpf_link to
> > >     /sys/fs/bpf/myprog
> > > 
> > > 2. eth0 is moved to a different namespace which mounts a new sysfs at
> > >     /sys
> > > 
> > > 3. Inside that namespace, /sys/fs/bpf/myprog is no longer accessible, so
> > >     xdp-loader can't get access to the original bpf_link; but the XDP
> > >     program is still attached to eth0.
> > 
> > The key to decide is whether moving netdev across netns should be allowed
> > when xdp attached. I think it should be denied. Even when legacy xdp
> > program is attached, since it will confuse user space managing part.
> 
> There are perfectly valid use cases where this is done already today (minus
> bpf_link), for example, consider an orchestrator that is setting up the BPF
> program on the device, moving to the newly created application pod during
> the CNI call in k8s, such that the new pod does not have the /sys/fs/bpf/
> mount instance and if unprivileged cannot remove the BPF prog from the dev
> either. We do something like this in case of ipvlan, meaning, we attach a
> rootlet prog that calls into single slot of a tail call map, move it to the
> application pod, and only out of Cilium's own pod and it's pod-local bpf fs
> instance we manage the pinned tail call map to update the main programs in
> that single slot w/o having to switch any netns later on.

Right. You mentioned this use case before, but I managed to forget about it.
Totally makes sense for prog to stay attached to netdev when it's moved.
I think pod manager would also prefer that pod is not able to replace
xdp prog from inside the container. It sounds to me that steps 1,2,3 above
is exactly the desired behavior. Otherwise what stops some application
that started in a pod to override it?