On Thu, Oct 31, 2019 at 11:39:58PM +0100, Daniel Borkmann wrote:

On 10/31/19 2:41 PM, Lukas Wunner wrote:

quoted

Commit e687ad60af09 ("netfilter: add netfilter ingress hook after
handle_ing() under unique static key") introduced the ability to
classify packets on ingress.

Allow the same on egress.

The need for this arose because I had to filter egress packets which do
not match a specific ethertype.  The most common solution appears to be
to enslave the interface to a bridge and use ebtables, but that's
cumbersome to configure and comes with a (small) performance penalty.
An alternative approach is tc, but that doesn't afford equivalent
matching options as netfilter.

Hmm, have you tried tc BPF on the egress hook (via sch_cls_act -> cls_bpf)?

There's another reason I chose netfilter over tc:  I need to activate the
filter from a kernel module, hence need an in-kernel (rather than user space)
API.

netfilter provides that via nf_register_net_hook(), I couldn't find
anything similar for tc.  And an egress netfilter hook seemed like
an obvious missing feature given the presence of an ingress hook.

The module I need this for is out-of-tree:
https://github.com/RevolutionPi/piControl/commit/da199ccd2099

In my experience the argument that a feature is needed for an out-of-tree
module holds zero value upstream.  If there's no in-tree user, the feature
isn't merged, I've seen this more than enough.  Which is why I didn't mention
it in the first place.

For our use case I wouldn't even need the nft user space support which I
posted separately, I just implemented it for completeness and to increase
acceptability of the present series.

Thanks,

Lukas