On 10/02, Andrii Nakryiko wrote:
> On Wed, Oct 2, 2019 at 6:43 PM Stanislav Fomichev <sdf@xxxxxxxxxxx> wrote:
> >
> > On 10/02, Andrii Nakryiko wrote:
> > > On Wed, Oct 2, 2019 at 10:35 AM Stanislav Fomichev <sdf@xxxxxxxxxx> wrote:
> > > >
> > > > Always use init_net flow dissector BPF program if it's attached and fall
> > > > back to the per-net namespace one. Also, deny installing new programs if
> > > > there is already one attached to the root namespace.
> > > > Users can still detach their BPF programs, but can't attach any
> > > > new ones (-EPERM).
>
> I find this quite confusing for users, honestly. If there is no root
> namespace dissector we'll successfully attach per-net ones and they
> will be working fine. That some process will attach root one and all
> the previously successfully working ones will suddenly "break" without
> users potentially not realizing why. I bet this will be hair-pulling
> investigation for someone. Furthermore, if root net dissector is
> already attached, all subsequent attachment will now start failing.
The idea is that if sysadmin decides to use system-wide dissector it would
be attached from the init scripts/systemd early in the boot process.
So the users in your example would always get EPERM/EBUSY/EXIST.
I don't really see a realistic use-case where root and non-root
namespaces attach/detach flow dissector programs at non-boot
time (or why non-root containers could have BPF dissector and root
could have C dissector; multi-nic machine?).
But I totally see your point about confusion. See below.
> I'm not sure what's the better behavior here is, but maybe at least
> forcibly detach already attached ones, so when someone goes and tries
> to investigate, they will see that their BPF program is not attached
> anymore. Printing dmesg warning would be hugely useful here as well.
We can do for_each_net and detach non-root ones; that sounds
feasible and may avoid the confusion (at least when you query
non-root ns to see if the prog is still there, you get a valid
indication that it's not).
> Alternatively, if there is any per-net dissector attached, we might
> disallow root net dissector to be installed. Sort of "too late to the
> party" way, but at least not surprising to successfully installed
> dissectors.
We can do this as well.
> Thoughts?
Let me try to implement both of your suggestions and see which one makes
more sense. I'm leaning towards the later (simple check to see if
any non-root ns has the prog attached).
I'll follow up with a v2 if all goes well.
> > > > Cc: Petar Penkov <ppenkov@xxxxxxxxxx>
> > > > Signed-off-by: Stanislav Fomichev <sdf@xxxxxxxxxx>
> > > > ---
> > > > Documentation/bpf/prog_flow_dissector.rst | 3 +++
> > > > net/core/flow_dissector.c | 11 ++++++++++-
> > > > 2 files changed, 13 insertions(+), 1 deletion(-)
> > > >
> > > > diff --git a/Documentation/bpf/prog_flow_dissector.rst b/Documentation/bpf/prog_flow_dissector.rst
> > > > index a78bf036cadd..4d86780ab0f1 100644
> > > > --- a/Documentation/bpf/prog_flow_dissector.rst
> > > > +++ b/Documentation/bpf/prog_flow_dissector.rst
> > > > @@ -142,3 +142,6 @@ BPF flow dissector doesn't support exporting all the metadata that in-kernel
> > > > C-based implementation can export. Notable example is single VLAN (802.1Q)
> > > > and double VLAN (802.1AD) tags. Please refer to the ``struct bpf_flow_keys``
> > > > for a set of information that's currently can be exported from the BPF context.
> > > > +
> > > > +When BPF flow dissector is attached to the root network namespace (machine-wide
> > > > +policy), users can't override it in their child network namespaces.
> > > > diff --git a/net/core/flow_dissector.c b/net/core/flow_dissector.c
> > > > index 7c09d87d3269..494e2016fe84 100644
> > > > --- a/net/core/flow_dissector.c
> > > > +++ b/net/core/flow_dissector.c
> > > > @@ -115,6 +115,11 @@ int skb_flow_dissector_bpf_prog_attach(const union bpf_attr *attr,
> > > > struct bpf_prog *attached;
> > > > struct net *net;
> > > >
> > > > + if (rcu_access_pointer(init_net.flow_dissector_prog)) {
> > > > + /* Can't override root flow dissector program */
> > > > + return -EPERM;
> > > > + }
> > >
> > > This is racy, shouldn't this be checked after grabbing a lock below?
> > What kind of race do you have in mind?
>
> I was thinking about the case of two competing attaches for root
> init_net, but it seems like we will double-check again under lock, so
> this is fine as is.
>
> >
> > Even if I put this check under the mutex, it's still possible that if
> > two cpus concurrently start attaching flow dissector programs (i.e. call
> > sys_bpf(BPF_PROG_ATTACH)) at the same time (one to root ns, the other
> > to non-root ns), the cpu that is attaching to non-root can grab mutex first,
> > pass all the checks and attach the prog (higher frequency, tubo boost, etc).
> >
> > The mutex is there to protect only against concurrent attaches to the
> > _same_ netns. For the sake of simplicity we have a global one instead
> > of a mutex per net-ns.
> >
> > So I'd rather not grab the mutex and keep it simple. Even in there is a
> > race, in __skb_flow_dissect we always check init_net first.
> >
> > > > +
> > > > net = current->nsproxy->net_ns;
> > > > mutex_lock(&flow_dissector_mutex);
> > > > attached = rcu_dereference_protected(net->flow_dissector_prog,
> > > > @@ -910,7 +915,11 @@ bool __skb_flow_dissect(const struct net *net,
> > > > WARN_ON_ONCE(!net);
> > > > if (net) {
> > > > rcu_read_lock();
> > > > - attached = rcu_dereference(net->flow_dissector_prog);
> > > > + attached =
> > > > + rcu_dereference(init_net.flow_dissector_prog);
> > > > +
> > > > + if (!attached)
> > > > + attached = rcu_dereference(net->flow_dissector_prog);
> > > >
> > > > if (attached) {
> > > > struct bpf_flow_keys flow_keys;
> > > > --
> > > > 2.23.0.444.g18eeb5a265-goog
> > > >
