Re: [RFC PATCH net-next] netfilter: nf_conntrack: add ct expression support for netdev egress chains

From: Hari Chandrakanthan

Date: Mon Oct 12 2026 - 08:21:47 EST

On 10/6/2026 5:09 PM, Florian Westphal wrote:

Hari Chandrakanthan<hari.chandrakanthan@xxxxxxxxxxxxxxxx> wrote:

Add support for using the ct expression in nftables netdev egress chains.
This enables QoS policy enforcement at the netdev egress hook by
allowing ct operations such as copying connmark to packet mark.

Ok so far.

Add an explicit NFPROTO_NETDEV case in nf_ct_netns_get() and
nf_ct_netns_put() that enables conntrack for IPv4, IPv6 and bridge when a
ct expression is added to a netdev chain.

Why? The egress chain is passive (its a 'read property off ct'). why
does it have to turn on conntrack, let alone for bridge too?

I agree that enabling inet/bridge conntrack solely for netdev/egress would be unnecessary overhead.
To address this, as suggested by Pablo, we are considering enabling conntrack lookup/creation at
netdev/ingress. This would allow netdev chains to access conntrack state without requiring conntrack
to be enabled for the inet or bridge families, avoiding unnecessary overhead for netdev-only use cases.
Kindly share your thoughts on this approach.

Restrict ct expression use in the netdev family to egress hooks only, as
the connection entry is not yet available at ingress.

This makes sense.

Sharing this change as an RFC, to get feedback. The patch has been tested
by configuring nft rules at netdev egress hook to set ct mark and copy
ct mark into skb->mark. Also, the patch is validated at netdev ingress to
ensure the nft rule with ct mark set action is rejected.

Signed-off-by: Hari Chandrakanthan<hari.chandrakanthan@xxxxxxxxxxxxxxxx>
---
net/netfilter/nf_conntrack_proto.c | 26 ++++++++++++++++++++++++++
net/netfilter/nft_ct.c | 22 ++++++++++++++++++++++
2 files changed, 48 insertions(+)

diff --git a/net/netfilter/nf_conntrack_proto.c b/net/netfilter/nf_conntrack_proto.c
index 7a40e4e0e33e..b8ee46262901 100644
--- a/net/netfilter/nf_conntrack_proto.c
+++ b/net/netfilter/nf_conntrack_proto.c
@@ -587,11 +587,36 @@ static int nf_ct_netns_inet_get(struct net *net)
int nf_ct_netns_get(struct net *net, u8 nfproto)
{
int err;
+ bool bridge_acquired = false;
switch (nfproto) {
case NFPROTO_INET:
err = nf_ct_netns_inet_get(net);
break;
+ case NFPROTO_NETDEV:
+ err = nf_ct_netns_do_get(net, NFPROTO_BRIDGE);
+ if (err < 0) {
+ mutex_lock(&nf_ct_proto_mutex);
+ if (nf_ct_bridge_info) {
+ /* Module present but hook registration failed.*/
+ mutex_unlock(&nf_ct_proto_mutex);
+ return err;
+ }
+ mutex_unlock(&nf_ct_proto_mutex);
+ /* Bridge module absent, netdev egress handles routed
+ * traffic too, bridge conntrack is only needed for
+ * bridged frames.
+ */
+ } else {
+ bridge_acquired = true;
+ }
+ err = nf_ct_netns_inet_get(net);
+ if (err < 0) {
+ if (bridge_acquired)
+ nf_ct_netns_put(net, NFPROTO_BRIDGE);
+ return err;
+ }
+ break;

I don't understand the need for this. Conntrack needs to be enabled to
track, but your use case makes no sense if conntrack isn't being used
already.

As mentioned in the response to your earlier question, we are looking at enabling conntrack
creation/lookup at netdev/ingress instead. This would establish the ct entry earlier in the packet path,
making it available at egress without requiring the netdev family rule to initialize conntrack for inet and bridge.
The NFPROTO_NETDEV case in nf_ct_netns_get() would then be simplified accordingly.

diff --git a/net/netfilter/nft_ct.c b/net/netfilter/nft_ct.c
index 3c4c2faa7398..e90e73475b0c 100644
--- a/net/netfilter/nft_ct.c
+++ b/net/netfilter/nft_ct.c
@@ -649,6 +649,15 @@ static void nft_ct_get_destroy(const struct nft_ctx *ctx,
nf_ct_netns_put(ctx->net, ctx->family);
}
+static int nft_ct_validate(const struct nft_ctx *ctx,
+ const struct nft_expr *expr)
+{
+ if (ctx->family != NFPROTO_NETDEV)
+ return 0;
+
+ return nft_chain_validate_hooks(ctx->chain, 1 << NF_NETDEV_EGRESS);
+}
+
static void nft_ct_set_destroy(const struct nft_ctx *ctx,
const struct nft_expr *expr)
{
@@ -732,6 +741,7 @@ static const struct nft_expr_ops nft_ct_get_ops = {
.init = nft_ct_get_init,
.destroy = nft_ct_get_destroy,
.dump = nft_ct_get_dump,
+ .validate = nft_ct_validate,
};

OK.

#ifdef CONFIG_MITIGATION_RETPOLINE
@@ -742,6 +752,7 @@ static const struct nft_expr_ops nft_ct_get_fast_ops = {
.init = nft_ct_get_init,
.destroy = nft_ct_get_destroy,
.dump = nft_ct_get_dump,
+ .validate = nft_ct_validate,
};
#endif

OK.

@@ -752,9 +763,19 @@ static const struct nft_expr_ops nft_ct_set_ops = {
.init = nft_ct_set_init,
.destroy = nft_ct_set_destroy,
.dump = nft_ct_set_dump,
+ .validate = nft_ct_validate,
};

Not sure. Whats the use case to set connmark, labels etc. at netdev
egress stage? I think use case was to READ those at egress stage
to set skb->mark etc?

One use case for setting connmark at netdev/egress is when flow classification itself is performed only at egress.

For example, in the WiFi QoS (SCS) use case, the Access Point can now classify the downlink flows
destined for a client at the WLAN netdevice/egress hook.

Classifying traffic at the netdev/egress limits the operation to packets transmitted through that interface,
without impacting traffic for other interfaces in the system.

The resulting QoS classification metadata can be stored in connmark at the time of processing the first packet,
allowing subsequent packets belonging to the same flow to reuse the previously determined classification state,
without affecting traffic on other interfaces.

#ifdef CONFIG_NF_CONNTRACK_ZONES
+static int nft_ct_set_zone_validate(const struct nft_ctx *ctx,
+ const struct nft_expr *expr)
+{
+ if (ctx->family == NFPROTO_NETDEV)
+ return -EOPNOTSUPP;
+
+ return 0;
+}
+
static const struct nft_expr_ops nft_ct_set_zone_ops = {
.type = &nft_ct_type,
.size = NFT_EXPR_SIZE(sizeof(struct nft_ct)),
@@ -762,6 +783,7 @@ static const struct nft_expr_ops nft_ct_set_zone_ops = {
.init = nft_ct_set_init,
.destroy = nft_ct_set_destroy,
.dump = nft_ct_set_dump,
+ .validate = nft_ct_set_zone_validate,

Makes sense to reject it.