eternalwrt-mt798x/target/linux/generic/pending-6.18/795-09-net-ethernet-mtk_ppe-offload-flows-to-MxL862xx-switc.patch

From 031ec521141aaa7fe13a937e0f942319f501cc4d Mon Sep 17 00:00:00 2001
From: Daniel Golle <daniel@makrotopia.org>
Date: Thu, 23 Apr 2026 15:24:32 +0100
Subject: [PATCH] net: ethernet: mtk_ppe: offload flows to MxL862xx switches in
 tag_8021q mode

Extend the PPE flow-offload output path to handle egress through
an MxL862xx switch running in tag_8021q mode
(DSA_TAG_PROTO_MXL862_8021Q), in addition to the existing support
for MediaTek's native 8-byte DSA tag (DSA_TAG_PROTO_MTK).

In tag_8021q mode MxL862xx expects every ingress frame on the CPU port
to carry an outer 802.1Q tag whose VID identifies the target user
port; the switch's per-port egress catchall rule then strips that
outer tag. The VID to use is the one produced by
dsa_tag_8021q_standalone_vid(dp), which bakes the tag_8021q reserved
marker (RSV = GENMASK(11, 10)), the DSA tree's switch ID, and the user
port's dp->index into the low 12 bits of the VID.

Change mtk_flow_get_dsa_port() to return, alongside dp->index, the VID
the PPE should push on the offloaded frame, in a new push_vid
out-parameter:

  - DSA_TAG_PROTO_MTK               -> push_vid = 0;
                                       caller uses mtk_foe_entry_set_dsa()
                                       (magic etype) as before.
  - DSA_TAG_PROTO_MXL862_8021Q      -> push_vid = standalone VID;
                                       caller pushes it as a VLAN via
                                       mtk_foe_entry_set_vlan().

The switch statement is intentionally scoped: other tag_8021q-based
taggers (sja1105, ocelot_8021q, ...) use different tag layers
(CTAG-in-range rather than Q-in-Q) and/or bridge-scoped VIDs and are
not interchangeable here.

The prerequisite reorder commit ("mtk_ppe_offload: set output device
before VLAN/PPPoE push") put mtk_flow_set_output_device() ahead of the
user VLAN push loop, so the outer MxL862xx VID lands in vlan1 and a
subsequent user VLAN stacks into vlan2. Queue selection stays with the
DSA queue map introduced in the preceding patch.

Signed-off-by: Daniel Golle <daniel@makrotopia.org>
---
 .../net/ethernet/mediatek/mtk_ppe_offload.c   | 54 +++++++++++++++++--
 1 file changed, 49 insertions(+), 5 deletions(-)

--- a/drivers/net/ethernet/mediatek/mtk_ppe_offload.c
+++ b/drivers/net/ethernet/mediatek/mtk_ppe_offload.c
@@ -10,6 +10,7 @@
 #include <net/flow_offload.h>
 #include <net/pkt_cls.h>
 #include <net/dsa.h>
+#include <linux/dsa/8021q.h>
 #include "mtk_eth_soc.h"
 #include "mtk_wed.h"
 
@@ -166,8 +167,25 @@ mtk_flow_mangle_ipv4(const struct flow_a
 	return 0;
 }
 
+/*
+ * Inspect *dev as a DSA user netdev.  On success, substitute *dev with
+ * the conduit netdev, return the DSA user port's dp->index, and in
+ * *push_vid hand back either 0 (native MTK tag; the PPE injects its
+ * magic etype later via mtk_foe_entry_set_dsa()) or the outer 802.1Q
+ * VID the PPE should push on the egressing frame (tag_8021q taggers
+ * that use an outer Q-in-Q tag with a standalone per-port VID, i.e.
+ * the MxL862xx tag_8021q tagger).
+ *
+ * push_vid may be NULL for callers that only want the idev->conduit
+ * substitution (e.g. ingress-side ppe_idx selection, where no VID
+ * is to be pushed on the flow entry).
+ *
+ * Other tag_8021q-based taggers (sja1105, ocelot_8021q, ...) use
+ * different tag layers and/or bridge-scoped VIDs; do not extend the
+ * switch below to cover them without per-tagger review.
+ */
 static int
-mtk_flow_get_dsa_port(struct net_device **dev)
+mtk_flow_get_dsa_port(struct net_device **dev, u16 *push_vid)
 {
 #if IS_ENABLED(CONFIG_NET_DSA)
 	struct dsa_port *dp;
@@ -176,8 +194,18 @@ mtk_flow_get_dsa_port(struct net_device
 	if (IS_ERR(dp))
 		return -ENODEV;
 
-	if (dp->cpu_dp->tag_ops->proto != DSA_TAG_PROTO_MTK)
+	switch (dp->cpu_dp->tag_ops->proto) {
+	case DSA_TAG_PROTO_MTK:
+		if (push_vid)
+			*push_vid = 0;
+		break;
+	case DSA_TAG_PROTO_MXL862_8021Q:
+		if (push_vid)
+			*push_vid = dsa_tag_8021q_standalone_vid(dp);
+		break;
+	default:
 		return -ENODEV;
+	}
 
 	*dev = dsa_port_to_conduit(dp);
 
@@ -195,6 +223,7 @@ mtk_flow_set_output_device(struct mtk_et
 	struct mtk_wdma_info info = {};
 	struct mtk_mac *mac;
 	int pse_port, dsa_port, queue;
+	u16 push_vid = 0;
 
 	if (mtk_flow_get_wdma_info(dev, dest_mac, &info) == 0) {
 		mtk_foe_entry_set_wdma(eth, foe, info.wdma_idx, info.queue,
@@ -220,7 +249,7 @@ mtk_flow_set_output_device(struct mtk_et
 		goto out;
 	}
 
-	dsa_port = mtk_flow_get_dsa_port(&dev);
+	dsa_port = mtk_flow_get_dsa_port(&dev, &push_vid);
 
 	if (dev == eth->netdev[0])
 		pse_port = PSE_GDM1_PORT;
@@ -231,8 +260,12 @@ mtk_flow_set_output_device(struct mtk_et
 	else
 		return -EOPNOTSUPP;
 
-	if (dsa_port >= 0)
-		mtk_foe_entry_set_dsa(eth, foe, dsa_port);
+	if (dsa_port >= 0) {
+		if (push_vid)
+			mtk_foe_entry_set_vlan(eth, foe, push_vid);
+		else
+			mtk_foe_entry_set_dsa(eth, foe, dsa_port);
+	}
 
 	if (dsa_port >= 0 && dsa_port < MTK_DSA_USER_PORT_MAX) {
 		mac = netdev_priv(dev);
@@ -275,6 +308,17 @@ mtk_flow_offload_replace(struct mtk_eth
 		flow_rule_match_meta(rule, &match);
 		if (mtk_is_netsys_v2_or_greater(eth)) {
 			idev = __dev_get_by_index(&init_net, match.key->ingress_ifindex);
+			/*
+			 * If idev is a DSA user netdev, substitute it with
+			 * its conduit so the valid-idev check below passes
+			 * and ppe_index picks up the conduit's PPE engine.
+			 * Without this, upstream flows from a DSA user port
+			 * (e.g. MxL862xx lanN) install on ppe[0] while the
+			 * conduit's GDMA routes ingress through its own
+			 * ppe[mac->ppe_idx]; the lookup misses and the
+			 * flow never binds in HW.
+			 */
+			mtk_flow_get_dsa_port(&idev, NULL);
 			if (idev && idev->netdev_ops == eth->netdev[0]->netdev_ops) {
 				struct mtk_mac *mac = netdev_priv(idev);