[PATCH RFC net-next 0/5] net: ethernet: ti: am65-cpsw: add AF_XDP zero copy support

[PATCH RFC net-next 0/5] net: ethernet: ti: am65-cpsw: add AF_XDP zero copy support

[Roger Quadros]

This series adds AF_XDP zero coppy support to am65-cpsw driver.

Tests were performed on AM62x-sk with xdpsock application [1].

A clear improvement is seen in 64 byte packets on Transmit (txonly)
and receive (rxdrop).
1500 byte test seems to be limited by line rate (1G link) so no
improvement seen there in packet rate. A test on higher speed link
(or PHY-less setup) might be worthwile.

There is some issue during l2fwd with 64 byte packets and benchmark
results show 0. I'm still investigating this issue.

AF_XDP performance using 64 byte packets in Kpps.
Benchmark:	XDP-SKB		XDP-Native	XDP-Native(ZeroCopy)
rxdrop		317		504		824
txonly		400		405		757
l2fwd 		207		264		0

AF_XDP performance using 1500 byte packets in Kpps.
Benchmark:	XDP-SKB		XDP-Native	XDP-Native(ZeroCopy)
rxdrop		82		82		82
txonly		82		82		82
l2fwd 		82		82		82

[1]: https://github.com/xdp-project/bpf-examples/tree/master/AF_XDP-example

To: 

Signed-off-by: Roger Quadros <rogerq@kernel.org>
---
Roger Quadros (5):
      net: ethernet: ti: am65-cpsw: fix BPF Program change on multi-port CPSW
      net: ethernet: ti: am65-cpsw: add XSK pool helpers
      net: ethernet: ti: am65-cpsw: Add AF_XDP zero copy for RX
      net: ethernet: ti: am65-cpsw: Add AF_XDP zero copy for TX
      net: ethernet: ti: am65-cpsw: enable zero copy in XDP features

 drivers/net/ethernet/ti/Makefile         |   2 +-
 drivers/net/ethernet/ti/am65-cpsw-nuss.c | 526 +++++++++++++++++++++++++++----
 drivers/net/ethernet/ti/am65-cpsw-nuss.h |  37 ++-
 drivers/net/ethernet/ti/am65-cpsw-xdp.c  | 155 +++++++++
 4 files changed, 656 insertions(+), 64 deletions(-)
---
base-commit: 9f607dc39b6658ba8ea647bd99725e68c66071b7
change-id: 20250225-am65-cpsw-xdp-zc-2af9e4be1356

Best regards,
-- 
Roger Quadros <rogerq@kernel.org>

[PATCH RFC net-next 1/5] net: ethernet: ti: am65-cpsw: fix BPF Program change on multi-port CPSW

[Roger Quadros]

On a multi-port CPSW system, stopping and starting just one port (ndev)
will not restart the queues if other ports (ndevs) are open.

Instead, check the usage_count variable to know if CPSW is running
and if so restart all the queues.

Signed-off-by: Roger Quadros <rogerq@kernel.org>
---
 drivers/net/ethernet/ti/am65-cpsw-nuss.c | 25 ++++++++++++++++++++-----
 1 file changed, 20 insertions(+), 5 deletions(-)

diff --git a/drivers/net/ethernet/ti/am65-cpsw-nuss.c b/drivers/net/ethernet/ti/am65-cpsw-nuss.c
index 988ce9119306..cd713bb57b91 100644
--- a/drivers/net/ethernet/ti/am65-cpsw-nuss.c
+++ b/drivers/net/ethernet/ti/am65-cpsw-nuss.c
@@ -1924,18 +1924,33 @@ static int am65_cpsw_xdp_prog_setup(struct net_device *ndev,
 				    struct bpf_prog *prog)
 {
 	struct am65_cpsw_port *port = am65_ndev_to_port(ndev);
-	bool running = netif_running(ndev);
+	struct am65_cpsw_common *common = port->common;
+	bool running = !!port->common->usage_count;
 	struct bpf_prog *old_prog;
+	int ret;
 
-	if (running)
-		am65_cpsw_nuss_ndo_slave_stop(ndev);
+	if (running) {
+		/* stop all queues */
+		am65_cpsw_destroy_txqs(common);
+		am65_cpsw_destroy_rxqs(common);
+	}
 
 	old_prog = xchg(&port->xdp_prog, prog);
 	if (old_prog)
 		bpf_prog_put(old_prog);
 
-	if (running)
-		return am65_cpsw_nuss_ndo_slave_open(ndev);
+	if (running) {
+		/* start all queues */
+		ret = am65_cpsw_create_rxqs(common);
+		if (ret)
+			return ret;
+
+		ret = am65_cpsw_create_txqs(common);
+		if (ret) {
+			am65_cpsw_destroy_rxqs(common);
+			return ret;
+		}
+	}
 
 	return 0;
 }

-- 
2.34.1

[PATCH RFC net-next 2/5] net: ethernet: ti: am65-cpsw: add XSK pool helpers

[Roger Quadros]

To prepare for XSK zero copy support, add XSK pool
helpers in a new file am65-cpsw-xdp.c

As queues are shared between ports we can no longer
support the case where zero copy (XSK Pool) is enabled
for the queue on one port but not for other ports.

Current solution is to drop the packet if Zero copy
is not enabled for that port + queue but enabled for
some other port + same queue.

xdp_zc_queues bitmap tracks if queue is setup as XSK
pool and xsk_port_id array tracks which port the XSK
queue is assigned to for zero copy.

Signed-off-by: Roger Quadros <rogerq@kernel.org>
---
 drivers/net/ethernet/ti/Makefile         |   2 +-
 drivers/net/ethernet/ti/am65-cpsw-nuss.c |  21 ++++--
 drivers/net/ethernet/ti/am65-cpsw-nuss.h |  20 +++++
 drivers/net/ethernet/ti/am65-cpsw-xdp.c  | 122 +++++++++++++++++++++++++++++++
 4 files changed, 156 insertions(+), 9 deletions(-)

diff --git a/drivers/net/ethernet/ti/Makefile b/drivers/net/ethernet/ti/Makefile
index cbcf44806924..48d07afe30f9 100644
--- a/drivers/net/ethernet/ti/Makefile
+++ b/drivers/net/ethernet/ti/Makefile
@@ -26,7 +26,7 @@ keystone_netcp_ethss-y := netcp_ethss.o netcp_sgmii.o netcp_xgbepcsr.o cpsw_ale.
 obj-$(CONFIG_TI_K3_CPPI_DESC_POOL) += k3-cppi-desc-pool.o
 
 obj-$(CONFIG_TI_K3_AM65_CPSW_NUSS) += ti-am65-cpsw-nuss.o
-ti-am65-cpsw-nuss-y := am65-cpsw-nuss.o cpsw_sl.o am65-cpsw-ethtool.o cpsw_ale.o
+ti-am65-cpsw-nuss-y := am65-cpsw-nuss.o cpsw_sl.o am65-cpsw-ethtool.o cpsw_ale.o am65-cpsw-xdp.o
 ti-am65-cpsw-nuss-$(CONFIG_TI_AM65_CPSW_QOS) += am65-cpsw-qos.o
 ti-am65-cpsw-nuss-$(CONFIG_TI_K3_AM65_CPSW_SWITCHDEV) += am65-cpsw-switchdev.o
 obj-$(CONFIG_TI_K3_AM65_CPTS) += am65-cpts.o
diff --git a/drivers/net/ethernet/ti/am65-cpsw-nuss.c b/drivers/net/ethernet/ti/am65-cpsw-nuss.c
index cd713bb57b91..a946bcd770c4 100644
--- a/drivers/net/ethernet/ti/am65-cpsw-nuss.c
+++ b/drivers/net/ethernet/ti/am65-cpsw-nuss.c
@@ -58,9 +58,6 @@
 
 #define AM65_CPSW_MAX_PORTS	8
 
-#define AM65_CPSW_MIN_PACKET_SIZE	VLAN_ETH_ZLEN
-#define AM65_CPSW_MAX_PACKET_SIZE	2024
-
 #define AM65_CPSW_REG_CTL		0x004
 #define AM65_CPSW_REG_STAT_PORT_EN	0x014
 #define AM65_CPSW_REG_PTYPE		0x018
@@ -505,7 +502,7 @@ static inline void am65_cpsw_put_page(struct am65_cpsw_rx_flow *flow,
 static void am65_cpsw_nuss_rx_cleanup(void *data, dma_addr_t desc_dma);
 static void am65_cpsw_nuss_tx_cleanup(void *data, dma_addr_t desc_dma);
 
-static void am65_cpsw_destroy_rxq(struct am65_cpsw_common *common, int id)
+void am65_cpsw_destroy_rxq(struct am65_cpsw_common *common, int id)
 {
 	struct am65_cpsw_rx_chn *rx_chn = &common->rx_chns;
 	struct am65_cpsw_rx_flow *flow;
@@ -554,7 +551,7 @@ static void am65_cpsw_destroy_rxqs(struct am65_cpsw_common *common)
 	k3_udma_glue_disable_rx_chn(common->rx_chns.rx_chn);
 }
 
-static int am65_cpsw_create_rxq(struct am65_cpsw_common *common, int id)
+int am65_cpsw_create_rxq(struct am65_cpsw_common *common, int id)
 {
 	struct am65_cpsw_rx_chn *rx_chn = &common->rx_chns;
 	struct page_pool_params pp_params = {
@@ -658,7 +655,7 @@ static int am65_cpsw_create_rxqs(struct am65_cpsw_common *common)
 	return ret;
 }
 
-static void am65_cpsw_destroy_txq(struct am65_cpsw_common *common, int id)
+void am65_cpsw_destroy_txq(struct am65_cpsw_common *common, int id)
 {
 	struct am65_cpsw_tx_chn *tx_chn = &common->tx_chns[id];
 
@@ -692,7 +689,7 @@ static void am65_cpsw_destroy_txqs(struct am65_cpsw_common *common)
 		am65_cpsw_destroy_txq(common, id);
 }
 
-static int am65_cpsw_create_txq(struct am65_cpsw_common *common, int id)
+int am65_cpsw_create_txq(struct am65_cpsw_common *common, int id)
 {
 	struct am65_cpsw_tx_chn *tx_chn = &common->tx_chns[id];
 	int ret;
@@ -1324,7 +1321,7 @@ static int am65_cpsw_nuss_rx_packets(struct am65_cpsw_rx_flow *flow,
 	dma_unmap_single(rx_chn->dma_dev, buf_dma, buf_dma_len, DMA_FROM_DEVICE);
 	k3_cppi_desc_pool_free(rx_chn->desc_pool, desc_rx);
 
-	if (port->xdp_prog) {
+	if (am65_cpsw_xdp_is_enabled(port)) {
 		xdp_init_buff(&xdp, PAGE_SIZE, &port->xdp_rxq[flow->id]);
 		xdp_prepare_buff(&xdp, page_addr, AM65_CPSW_HEADROOM,
 				 pkt_len, false);
@@ -1960,6 +1957,9 @@ static int am65_cpsw_ndo_bpf(struct net_device *ndev, struct netdev_bpf *bpf)
 	switch (bpf->command) {
 	case XDP_SETUP_PROG:
 		return am65_cpsw_xdp_prog_setup(ndev, bpf->prog);
+	case XDP_SETUP_XSK_POOL:
+		return am65_cpsw_xsk_setup_pool(ndev, bpf->xsk.pool,
+						bpf->xsk.queue_id);
 	default:
 		return -EINVAL;
 	}
@@ -3527,7 +3527,12 @@ static int am65_cpsw_nuss_probe(struct platform_device *pdev)
 	common = devm_kzalloc(dev, sizeof(struct am65_cpsw_common), GFP_KERNEL);
 	if (!common)
 		return -ENOMEM;
+
 	common->dev = dev;
+	common->xdp_zc_queues = devm_bitmap_zalloc(dev, AM65_CPSW_MAX_QUEUES,
+						   GFP_KERNEL);
+	if (!common->xdp_zc_queues)
+		return -ENOMEM;
 
 	of_id = of_match_device(am65_cpsw_nuss_of_mtable, dev);
 	if (!of_id)
diff --git a/drivers/net/ethernet/ti/am65-cpsw-nuss.h b/drivers/net/ethernet/ti/am65-cpsw-nuss.h
index 917c37e4e89b..e80e74a74d71 100644
--- a/drivers/net/ethernet/ti/am65-cpsw-nuss.h
+++ b/drivers/net/ethernet/ti/am65-cpsw-nuss.h
@@ -23,8 +23,14 @@ struct am65_cpts;
 
 #define AM65_CPSW_MAX_QUEUES	8	/* both TX & RX */
 
+#define AM65_CPSW_MIN_PACKET_SIZE	VLAN_ETH_ZLEN
+#define AM65_CPSW_MAX_PACKET_SIZE	2024
+
 #define AM65_CPSW_PORT_VLAN_REG_OFFSET	0x014
 
+#define AM65_CPSW_RX_DMA_ATTR	(DMA_ATTR_SKIP_CPU_SYNC |\
+				 DMA_ATTR_WEAK_ORDERING)
+
 struct am65_cpsw_slave_data {
 	bool				mac_only;
 	struct cpsw_sl			*mac_sl;
@@ -190,6 +196,9 @@ struct am65_cpsw_common {
 	unsigned char switch_id[MAX_PHYS_ITEM_ID_LEN];
 	/* only for suspend/resume context restore */
 	u32			*ale_context;
+	/* XDP Zero Copy */
+	unsigned long		*xdp_zc_queues;
+	int			xsk_port_id[AM65_CPSW_MAX_QUEUES];
 };
 
 struct am65_cpsw_ndev_priv {
@@ -228,4 +237,15 @@ int am65_cpsw_nuss_update_tx_rx_chns(struct am65_cpsw_common *common,
 
 bool am65_cpsw_port_dev_check(const struct net_device *dev);
 
+int am65_cpsw_create_rxq(struct am65_cpsw_common *common, int id);
+void am65_cpsw_destroy_rxq(struct am65_cpsw_common *common, int id);
+int am65_cpsw_create_txq(struct am65_cpsw_common *common, int id);
+void am65_cpsw_destroy_txq(struct am65_cpsw_common *common, int id);
+int am65_cpsw_xsk_setup_pool(struct net_device *ndev,
+			     struct xsk_buff_pool *pool, u16 qid);
+int am65_cpsw_xsk_wakeup(struct net_device *ndev, u32 qid, u32 flags);
+static inline bool am65_cpsw_xdp_is_enabled(struct am65_cpsw_port *port)
+{
+	return !!READ_ONCE(port->xdp_prog);
+}
 #endif /* AM65_CPSW_NUSS_H_ */
diff --git a/drivers/net/ethernet/ti/am65-cpsw-xdp.c b/drivers/net/ethernet/ti/am65-cpsw-xdp.c
new file mode 100644
index 000000000000..e1ab81cb4548
--- /dev/null
+++ b/drivers/net/ethernet/ti/am65-cpsw-xdp.c
@@ -0,0 +1,122 @@
+// SPDX-License-Identifier: GPL-2.0
+/* Texas Instruments K3 AM65 Ethernet Switch SubSystem Driver
+ *
+ * Copyright (C) 2025 Texas Instruments Incorporated - http://www.ti.com/
+ *
+ */
+
+#include <net/xsk_buff_pool.h>
+#include <net/xdp_sock_drv.h>
+#include "am65-cpsw-nuss.h"
+
+static int am65_cpsw_xsk_pool_enable(struct am65_cpsw_port *port,
+				     struct xsk_buff_pool *pool, u16 qid)
+{
+	struct am65_cpsw_common *common = port->common;
+	struct am65_cpsw_rx_chn *rx_chn;
+	bool need_update;
+	u32 frame_size;
+	int ret;
+
+	/*
+	 * As queues are shared between ports we can no longer
+	 * support the case where zero copy (XSK Pool) is enabled
+	 * for the queue on one port but not for other ports.
+	 *
+	 * Current solution is to drop the packet if Zero copy
+	 * is not enabled for that port + queue but enabled for
+	 * some other port + same queue.
+	 */
+	if (test_bit(qid, common->xdp_zc_queues))
+		return -EINVAL;
+
+	rx_chn = &common->rx_chns;
+	if (qid >= common->rx_ch_num_flows || qid >= common->tx_ch_num)
+		return -EINVAL;
+
+	frame_size = xsk_pool_get_rx_frame_size(pool);
+	if (frame_size < AM65_CPSW_MAX_PACKET_SIZE)
+		return -EOPNOTSUPP;
+
+	ret = xsk_pool_dma_map(pool, rx_chn->dma_dev, AM65_CPSW_RX_DMA_ATTR);
+	if (ret) {
+		netdev_err(port->ndev, "Failed to map xsk pool\n");
+		return ret;
+	}
+
+	need_update = common->usage_count &&
+		      am65_cpsw_xdp_is_enabled(port);
+	if (need_update) {
+		am65_cpsw_destroy_rxq(common, qid);
+		am65_cpsw_destroy_txq(common, qid);
+	}
+
+	set_bit(qid, common->xdp_zc_queues);
+	common->xsk_port_id[qid] = port->port_id;
+	if (need_update) {
+		am65_cpsw_create_rxq(common, qid);
+		am65_cpsw_create_txq(common, qid);
+	}
+
+	return 0;
+}
+
+static int am65_cpsw_xsk_pool_disable(struct am65_cpsw_port *port,
+				      struct xsk_buff_pool *pool, u16 qid)
+{
+	struct am65_cpsw_common *common = port->common;
+	bool need_update;
+
+	if (qid >= common->rx_ch_num_flows || qid >= common->tx_ch_num)
+		return -EINVAL;
+
+	if (!test_bit(qid, common->xdp_zc_queues))
+		return -EINVAL;
+
+	pool = xsk_get_pool_from_qid(port->ndev, qid);
+	if (!pool)
+		return -EINVAL;
+
+	need_update = common->usage_count && am65_cpsw_xdp_is_enabled(port);
+	if (need_update) {
+		am65_cpsw_destroy_rxq(common, qid);
+		am65_cpsw_destroy_txq(common, qid);
+		synchronize_rcu();
+	}
+
+	xsk_pool_dma_unmap(pool, AM65_CPSW_RX_DMA_ATTR);
+	clear_bit(qid, common->xdp_zc_queues);
+	common->xsk_port_id[qid] = -EINVAL;
+	if (need_update) {
+		am65_cpsw_create_rxq(common, qid);
+		am65_cpsw_create_txq(common, qid);
+	}
+
+	return 0;
+}
+
+int am65_cpsw_xsk_setup_pool(struct net_device *ndev,
+			     struct xsk_buff_pool *pool, u16 qid)
+{
+	struct am65_cpsw_port *port = am65_ndev_to_port(ndev);
+
+	return pool ? am65_cpsw_xsk_pool_enable(port, pool, qid) :
+		      am65_cpsw_xsk_pool_disable(port, pool, qid);
+}
+
+int am65_cpsw_xsk_wakeup(struct net_device *ndev, u32 qid, u32 flags)
+{
+	struct am65_cpsw_common *common = am65_ndev_to_common(ndev);
+	struct am65_cpsw_port *port = am65_ndev_to_port(ndev);
+
+	if (!netif_running(ndev) || !netif_carrier_ok(ndev))
+		return -ENETDOWN;
+
+	if (!am65_cpsw_xdp_is_enabled(port))
+		return -EINVAL;
+
+	if (qid >= common->rx_ch_num_flows || qid >= common->tx_ch_num)
+		return -EINVAL;
+
+	return 0;
+}

-- 
2.34.1

[PATCH RFC net-next 3/5] net: ethernet: ti: am65-cpsw: Add AF_XDP zero copy for RX

[Roger Quadros]

Add zero copy support to RX path.

Introduce xsk_pool and xsk_port_id to struct am65_cpsw_rx_flow.
This way we can quickly check if the flow is setup as XSK pool
and for which port.

If the RX flow is setup as XSK pool then register it as
MEM_TYPE_XSK_BUFF_POOL. At queue creation get free frames from
the XSK pool and push it to the RX ring.

Signed-off-by: Roger Quadros <rogerq@kernel.org>
---
 drivers/net/ethernet/ti/am65-cpsw-nuss.c | 307 +++++++++++++++++++++++++++----
 drivers/net/ethernet/ti/am65-cpsw-nuss.h |  12 +-
 drivers/net/ethernet/ti/am65-cpsw-xdp.c  |  24 +++
 3 files changed, 309 insertions(+), 34 deletions(-)

diff --git a/drivers/net/ethernet/ti/am65-cpsw-nuss.c b/drivers/net/ethernet/ti/am65-cpsw-nuss.c
index a946bcd770c4..5fbf04880722 100644
--- a/drivers/net/ethernet/ti/am65-cpsw-nuss.c
+++ b/drivers/net/ethernet/ti/am65-cpsw-nuss.c
@@ -429,6 +429,55 @@ static void am65_cpsw_nuss_ndo_host_tx_timeout(struct net_device *ndev,
 	}
 }
 
+static int am65_cpsw_nuss_rx_push_zc(struct am65_cpsw_rx_flow *flow,
+				     struct xdp_buff *xdp)
+{
+	struct am65_cpsw_rx_chn *rx_chn = &flow->common->rx_chns;
+	struct cppi5_host_desc_t *desc_rx;
+	struct am65_cpsw_swdata *swdata;
+	u32 flow_id = flow->id;
+	dma_addr_t desc_dma;
+	dma_addr_t buf_dma;
+	int buf_len;
+
+	desc_rx = k3_cppi_desc_pool_alloc(rx_chn->desc_pool);
+	if (!desc_rx)
+		return -ENOMEM;
+
+	desc_dma = k3_cppi_desc_pool_virt2dma(rx_chn->desc_pool, desc_rx);
+	buf_dma = xsk_buff_xdp_get_dma(xdp);
+	cppi5_hdesc_init(desc_rx, CPPI5_INFO0_HDESC_EPIB_PRESENT,
+			 AM65_CPSW_NAV_PS_DATA_SIZE);
+	k3_udma_glue_rx_dma_to_cppi5_addr(rx_chn->rx_chn, &buf_dma);
+	buf_len = xsk_pool_get_rx_frame_size(flow->xsk_pool);
+	cppi5_hdesc_attach_buf(desc_rx, buf_dma, buf_len, buf_dma, buf_len);
+	swdata = cppi5_hdesc_get_swdata(desc_rx);
+	swdata->xdp = xdp;
+	swdata->flow_id = flow_id;
+
+	return k3_udma_glue_push_rx_chn(rx_chn->rx_chn, flow_id,
+					desc_rx, desc_dma);
+}
+
+static int am65_cpsw_nuss_rx_alloc_zc(struct am65_cpsw_rx_flow *flow,
+				      int budget)
+{
+	struct xdp_buff *xdp;
+	int i, ret;
+
+	for (i = 0; i < budget; i++) {
+		xdp = xsk_buff_alloc(flow->xsk_pool);
+		if (!xdp)
+			break;
+
+		ret = am65_cpsw_nuss_rx_push_zc(flow, xdp);
+		if (ret < 0)
+			break;
+	}
+
+	return i;
+}
+
 static int am65_cpsw_nuss_rx_push(struct am65_cpsw_common *common,
 				  struct page *page, u32 flow_idx)
 {
@@ -529,6 +578,9 @@ void am65_cpsw_destroy_rxq(struct am65_cpsw_common *common, int id)
 		page_pool_destroy(flow->page_pool);
 		flow->page_pool = NULL;
 	}
+
+	flow->xsk_pool = NULL;
+	flow->xsk_port_id = -EINVAL;
 }
 
 static void am65_cpsw_destroy_rxqs(struct am65_cpsw_common *common)
@@ -568,6 +620,7 @@ int am65_cpsw_create_rxq(struct am65_cpsw_common *common, int id)
 	struct page_pool *pool;
 	struct page *page;
 	int port, ret, i;
+	int port_id;
 
 	flow = &rx_chn->flows[id];
 	pp_params.napi = &flow->napi_rx;
@@ -582,9 +635,24 @@ int am65_cpsw_create_rxq(struct am65_cpsw_common *common, int id)
 	/* using same page pool is allowed as no running rx handlers
 	 * simultaneously for both ndevs
 	 */
+
+	/* get first port with XSK pool & XDP program set */
 	for (port = 0; port < common->port_num; port++) {
-		if (!common->ports[port].ndev)
-		/* FIXME should we BUG here? */
+		flow->xsk_pool = am65_cpsw_xsk_get_pool(&common->ports[port],
+							id);
+		if (flow->xsk_pool)
+			break;
+	}
+
+	port_id = common->ports[port].port_id;
+	flow->xsk_port_id = flow->xsk_pool ? port_id : -EINVAL;
+	for (port = 0; port < common->port_num; port++) {
+		port_id = common->ports[port].port_id;
+
+		/* NOTE: if queue is XSK then only register it
+		 * for the relevant port it was assigned to
+		 */
+		if (flow->xsk_pool && port_id != flow->xsk_port_id)
 			continue;
 
 		rxq = &common->ports[port].xdp_rxq[id];
@@ -593,29 +661,44 @@ int am65_cpsw_create_rxq(struct am65_cpsw_common *common, int id)
 		if (ret)
 			goto err;
 
-		ret = xdp_rxq_info_reg_mem_model(rxq,
-						 MEM_TYPE_PAGE_POOL,
-						 pool);
+		if (flow->xsk_pool) {
+			ret = xdp_rxq_info_reg_mem_model(rxq,
+							 MEM_TYPE_XSK_BUFF_POOL,
+							 NULL);
+			xsk_pool_set_rxq_info(flow->xsk_pool, rxq);
+		} else {
+			ret = xdp_rxq_info_reg_mem_model(rxq,
+							 MEM_TYPE_PAGE_POOL,
+							 pool);
+		}
+
 		if (ret)
 			goto err;
 	}
 
-	for (i = 0; i < AM65_CPSW_MAX_RX_DESC; i++) {
-		page = page_pool_dev_alloc_pages(flow->page_pool);
-		if (!page) {
-			dev_err(common->dev, "cannot allocate page in flow %d\n",
-				id);
-			ret = -ENOMEM;
-			goto err;
-		}
+	if (flow->xsk_pool) {
+		/* get pages from xsk_pool and push to RX ring
+		 * queue as much as possible
+		 */
+		am65_cpsw_nuss_rx_alloc_zc(flow, AM65_CPSW_MAX_RX_DESC);
+	} else {
+		for (i = 0; i < AM65_CPSW_MAX_RX_DESC; i++) {
+			page = page_pool_dev_alloc_pages(flow->page_pool);
+			if (!page) {
+				dev_err(common->dev, "cannot allocate page in flow %d\n",
+					id);
+				ret = -ENOMEM;
+				goto err;
+			}
 
-		ret = am65_cpsw_nuss_rx_push(common, page, id);
-		if (ret < 0) {
-			dev_err(common->dev,
-				"cannot submit page to rx channel flow %d, error %d\n",
-				id, ret);
-			am65_cpsw_put_page(flow, page, false);
-			goto err;
+			ret = am65_cpsw_nuss_rx_push(common, page, id);
+			if (ret < 0) {
+				dev_err(common->dev,
+					"cannot submit page to rx channel flow %d, error %d\n",
+					id, ret);
+				am65_cpsw_put_page(flow, page, false);
+				goto err;
+			}
 		}
 	}
 
@@ -772,6 +855,8 @@ static void am65_cpsw_nuss_rx_cleanup(void *data, dma_addr_t desc_dma)
 	struct am65_cpsw_rx_chn *rx_chn = data;
 	struct cppi5_host_desc_t *desc_rx;
 	struct am65_cpsw_swdata *swdata;
+	struct am65_cpsw_rx_flow *flow;
+	struct xdp_buff *xdp;
 	dma_addr_t buf_dma;
 	struct page *page;
 	u32 buf_dma_len;
@@ -779,13 +864,20 @@ static void am65_cpsw_nuss_rx_cleanup(void *data, dma_addr_t desc_dma)
 
 	desc_rx = k3_cppi_desc_pool_dma2virt(rx_chn->desc_pool, desc_dma);
 	swdata = cppi5_hdesc_get_swdata(desc_rx);
-	page = swdata->page;
 	flow_id = swdata->flow_id;
 	cppi5_hdesc_get_obuf(desc_rx, &buf_dma, &buf_dma_len);
 	k3_udma_glue_rx_cppi5_to_dma_addr(rx_chn->rx_chn, &buf_dma);
-	dma_unmap_single(rx_chn->dma_dev, buf_dma, buf_dma_len, DMA_FROM_DEVICE);
 	k3_cppi_desc_pool_free(rx_chn->desc_pool, desc_rx);
-	am65_cpsw_put_page(&rx_chn->flows[flow_id], page, false);
+	flow = &rx_chn->flows[flow_id];
+	if (flow->xsk_pool) {
+		xdp = swdata->xdp;
+		xsk_buff_free(xdp);
+	} else {
+		page = swdata->page;
+		dma_unmap_single(rx_chn->dma_dev, buf_dma, buf_dma_len,
+				 DMA_FROM_DEVICE);
+		am65_cpsw_put_page(flow, page, false);
+	}
 }
 
 static void am65_cpsw_nuss_xmit_free(struct am65_cpsw_tx_chn *tx_chn,
@@ -1264,6 +1356,151 @@ static void am65_cpsw_nuss_rx_csum(struct sk_buff *skb, u32 csum_info)
 	}
 }
 
+static struct sk_buff *am65_cpsw_create_skb_zc(struct am65_cpsw_rx_flow *flow,
+					       struct xdp_buff *xdp)
+{
+	unsigned int metasize = xdp->data - xdp->data_meta;
+	unsigned int datasize = xdp->data_end - xdp->data;
+	struct sk_buff *skb;
+
+	skb = napi_alloc_skb(&flow->napi_rx,
+			     xdp->data_end - xdp->data_hard_start);
+	if (unlikely(!skb))
+		return NULL;
+
+	skb_reserve(skb, xdp->data - xdp->data_hard_start);
+	memcpy(__skb_put(skb, datasize), xdp->data, datasize);
+	if (metasize)
+		skb_metadata_set(skb, metasize);
+
+	return skb;
+}
+
+static void am65_cpsw_dispatch_skb_zc(struct am65_cpsw_rx_flow *flow,
+				      struct am65_cpsw_port *port,
+				      struct xdp_buff *xdp, u32 csum_info)
+{
+	struct am65_cpsw_common *common = flow->common;
+	unsigned int len = xdp->data_end - xdp->data;
+	struct am65_cpsw_ndev_priv *ndev_priv;
+	struct net_device *ndev = port->ndev;
+	struct sk_buff *skb;
+
+	skb = am65_cpsw_create_skb_zc(flow, xdp);
+	if (!skb) {
+		ndev->stats.rx_dropped++;
+		return;
+	}
+
+	ndev_priv = netdev_priv(ndev);
+	am65_cpsw_nuss_set_offload_fwd_mark(skb, ndev_priv->offload_fwd_mark);
+	if (port->rx_ts_enabled)
+		am65_cpts_rx_timestamp(common->cpts, skb);
+
+	skb_mark_for_recycle(skb);
+	skb->protocol = eth_type_trans(skb, ndev);
+	am65_cpsw_nuss_rx_csum(skb, csum_info);
+	napi_gro_receive(&flow->napi_rx, skb);
+	dev_sw_netstats_rx_add(ndev, len);
+}
+
+static int am65_cpsw_nuss_rx_zc(struct am65_cpsw_rx_flow *flow, int budget)
+{
+	struct am65_cpsw_rx_chn *rx_chn = &flow->common->rx_chns;
+	u32 buf_dma_len, pkt_len, port_id = 0, csum_info;
+	struct am65_cpsw_common *common = flow->common;
+	struct cppi5_host_desc_t *desc_rx;
+	struct device *dev = common->dev;
+	struct am65_cpsw_swdata *swdata;
+	dma_addr_t desc_dma, buf_dma;
+	struct am65_cpsw_port *port;
+	struct net_device *ndev;
+	u32 flow_idx = flow->id;
+	struct xdp_buff	*xdp;
+	int count = 0;
+	int xdp_status = 0;
+	u32 *psdata;
+	int ret;
+
+	while (count < budget) {
+		ret = k3_udma_glue_pop_rx_chn(rx_chn->rx_chn, flow_idx,
+					      &desc_dma);
+		if (ret) {
+			if (ret != -ENODATA)
+				dev_err(dev, "RX: pop chn fail %d\n",
+					ret);
+			break;
+		}
+
+		if (cppi5_desc_is_tdcm(desc_dma)) {
+			dev_dbg(dev, "%s RX tdown flow: %u\n",
+				__func__, flow_idx);
+			if (common->pdata.quirks & AM64_CPSW_QUIRK_DMA_RX_TDOWN_IRQ)
+				complete(&common->tdown_complete);
+			continue;
+		}
+
+		desc_rx = k3_cppi_desc_pool_dma2virt(rx_chn->desc_pool,
+						     desc_dma);
+		dev_dbg(dev, "%s flow_idx: %u desc %pad\n",
+			__func__, flow_idx, &desc_dma);
+
+		swdata = cppi5_hdesc_get_swdata(desc_rx);
+		xdp = swdata->xdp;
+		cppi5_hdesc_get_obuf(desc_rx, &buf_dma, &buf_dma_len);
+		k3_udma_glue_rx_cppi5_to_dma_addr(rx_chn->rx_chn, &buf_dma);
+		pkt_len = cppi5_hdesc_get_pktlen(desc_rx);
+		cppi5_desc_get_tags_ids(&desc_rx->hdr, &port_id, NULL);
+		dev_dbg(dev, "%s rx port_id:%d\n", __func__, port_id);
+		port = am65_common_get_port(common, port_id);
+		ndev = port->ndev;
+		psdata = cppi5_hdesc_get_psdata(desc_rx);
+		csum_info = psdata[2];
+		dev_dbg(dev, "%s rx csum_info:%#x\n", __func__, csum_info);
+		k3_cppi_desc_pool_free(rx_chn->desc_pool, desc_rx);
+		count++;
+		xsk_buff_set_size(xdp, pkt_len);
+		xsk_buff_dma_sync_for_cpu(xdp);
+		/* check if this port has XSK enabled. else drop packet */
+		if (port_id != flow->xsk_port_id) {
+			dev_dbg(dev, "discarding non xsk port data\n");
+			xsk_buff_free(xdp);
+			ndev->stats.rx_dropped++;
+			continue;
+		}
+
+		ret = am65_cpsw_run_xdp(flow, port, xdp, &pkt_len);
+		switch (ret) {
+		case AM65_CPSW_XDP_PASS:
+			am65_cpsw_dispatch_skb_zc(flow, port, xdp, csum_info);
+			xsk_buff_free(xdp);
+			break;
+		case AM65_CPSW_XDP_CONSUMED:
+			xsk_buff_free(xdp);
+			break;
+		case AM65_CPSW_XDP_TX:
+		case AM65_CPSW_XDP_REDIRECT:
+			xdp_status |= ret;
+			break;
+		}
+	}
+
+	if (xdp_status & AM65_CPSW_XDP_REDIRECT)
+		xdp_do_flush();
+
+	ret = am65_cpsw_nuss_rx_alloc_zc(flow, count);
+
+	if (xsk_uses_need_wakeup(flow->xsk_pool)) {
+		/* We set wakeup if we are exhausted of new requests */
+		if (ret < count)
+			xsk_set_rx_need_wakeup(flow->xsk_pool);
+		else
+			xsk_clear_rx_need_wakeup(flow->xsk_pool);
+	}
+
+	return count;
+}
+
 static int am65_cpsw_nuss_rx_packets(struct am65_cpsw_rx_flow *flow,
 				     int *xdp_state)
 {
@@ -1403,17 +1640,21 @@ static int am65_cpsw_nuss_rx_poll(struct napi_struct *napi_rx, int budget)
 	int num_rx = 0;
 
 	/* process only this flow */
-	cur_budget = budget;
-	while (cur_budget--) {
-		ret = am65_cpsw_nuss_rx_packets(flow, &xdp_state);
-		xdp_state_or |= xdp_state;
-		if (ret)
-			break;
-		num_rx++;
-	}
+	if (flow->xsk_pool) {
+		num_rx = am65_cpsw_nuss_rx_zc(flow, budget);
+	} else {
+		cur_budget = budget;
+		while (cur_budget--) {
+			ret = am65_cpsw_nuss_rx_packets(flow, &xdp_state);
+			xdp_state_or |= xdp_state;
+			if (ret)
+				break;
+			num_rx++;
+		}
 
-	if (xdp_state_or & AM65_CPSW_XDP_REDIRECT)
-		xdp_do_flush();
+		if (xdp_state_or & AM65_CPSW_XDP_REDIRECT)
+			xdp_do_flush();
+	}
 
 	dev_dbg(common->dev, "%s num_rx:%d %d\n", __func__, num_rx, budget);
 
diff --git a/drivers/net/ethernet/ti/am65-cpsw-nuss.h b/drivers/net/ethernet/ti/am65-cpsw-nuss.h
index e80e74a74d71..0e44d8a6cd68 100644
--- a/drivers/net/ethernet/ti/am65-cpsw-nuss.h
+++ b/drivers/net/ethernet/ti/am65-cpsw-nuss.h
@@ -15,6 +15,7 @@
 #include <linux/soc/ti/k3-ringacc.h>
 #include <net/devlink.h>
 #include <net/xdp.h>
+#include <net/xdp_sock_drv.h>
 #include "am65-cpsw-qos.h"
 
 struct am65_cpts;
@@ -107,6 +108,8 @@ struct am65_cpsw_rx_flow {
 	struct hrtimer rx_hrtimer;
 	unsigned long rx_pace_timeout;
 	struct page_pool *page_pool;
+	struct xsk_buff_pool *xsk_pool;
+	int xsk_port_id;
 	char name[32];
 };
 
@@ -120,7 +123,10 @@ struct am65_cpsw_tx_swdata {
 
 struct am65_cpsw_swdata {
 	u32 flow_id;
-	struct page *page;
+	union {
+		struct page *page;
+		struct xdp_buff *xdp;
+	};
 };
 
 struct am65_cpsw_rx_chn {
@@ -248,4 +254,8 @@ static inline bool am65_cpsw_xdp_is_enabled(struct am65_cpsw_port *port)
 {
 	return !!READ_ONCE(port->xdp_prog);
 }
+
+struct xsk_buff_pool *am65_cpsw_xsk_get_pool(struct am65_cpsw_port *port,
+					     u32 qid);
+
 #endif /* AM65_CPSW_NUSS_H_ */
diff --git a/drivers/net/ethernet/ti/am65-cpsw-xdp.c b/drivers/net/ethernet/ti/am65-cpsw-xdp.c
index e1ab81cb4548..e71ff38f851f 100644
--- a/drivers/net/ethernet/ti/am65-cpsw-xdp.c
+++ b/drivers/net/ethernet/ti/am65-cpsw-xdp.c
@@ -108,6 +108,9 @@ int am65_cpsw_xsk_wakeup(struct net_device *ndev, u32 qid, u32 flags)
 {
 	struct am65_cpsw_common *common = am65_ndev_to_common(ndev);
 	struct am65_cpsw_port *port = am65_ndev_to_port(ndev);
+	struct am65_cpsw_rx_flow *rx_flow;
+
+	rx_flow = &common->rx_chns.flows[qid];
 
 	if (!netif_running(ndev) || !netif_carrier_ok(ndev))
 		return -ENETDOWN;
@@ -118,5 +121,26 @@ int am65_cpsw_xsk_wakeup(struct net_device *ndev, u32 qid, u32 flags)
 	if (qid >= common->rx_ch_num_flows || qid >= common->tx_ch_num)
 		return -EINVAL;
 
+	if (!rx_flow->xsk_pool)
+		return -EINVAL;
+
+	if (flags & XDP_WAKEUP_RX) {
+		if (!napi_if_scheduled_mark_missed(&rx_flow->napi_rx)) {
+			if (likely(napi_schedule_prep(&rx_flow->napi_rx)))
+				__napi_schedule(&rx_flow->napi_rx);
+		}
+	}
+
 	return 0;
 }
+
+struct xsk_buff_pool *am65_cpsw_xsk_get_pool(struct am65_cpsw_port *port,
+					     u32 qid)
+{
+	if (!am65_cpsw_xdp_is_enabled(port) ||
+	    !test_bit(qid, port->common->xdp_zc_queues) ||
+	    port->common->xsk_port_id[qid] != port->port_id)
+		return NULL;
+
+	return xsk_get_pool_from_qid(port->ndev, qid);
+}

-- 
2.34.1

[PATCH RFC net-next 4/5] net: ethernet: ti: am65-cpsw: Add AF_XDP zero copy for TX

[Roger Quadros]

Add zero copy support to TX path.

Introduce xsk_pool and xsk_port_id to struct am65_cpsw_tx_chn.
This way we can quickly check if the flow is setup as XSK pool
and for which port.

If the TX channel is setup as XSK pool then get the frames from
the pool and send it to the TX channel.

Signed-off-by: Roger Quadros <rogerq@kernel.org>
---
 drivers/net/ethernet/ti/am65-cpsw-nuss.c | 170 ++++++++++++++++++++++++++++---
 drivers/net/ethernet/ti/am65-cpsw-nuss.h |   5 +
 drivers/net/ethernet/ti/am65-cpsw-xdp.c  |  11 +-
 3 files changed, 170 insertions(+), 16 deletions(-)

diff --git a/drivers/net/ethernet/ti/am65-cpsw-nuss.c b/drivers/net/ethernet/ti/am65-cpsw-nuss.c
index 5fbf04880722..e89b3cefcb05 100644
--- a/drivers/net/ethernet/ti/am65-cpsw-nuss.c
+++ b/drivers/net/ethernet/ti/am65-cpsw-nuss.c
@@ -747,6 +747,8 @@ void am65_cpsw_destroy_txq(struct am65_cpsw_common *common, int id)
 	k3_udma_glue_reset_tx_chn(tx_chn->tx_chn, tx_chn,
 				  am65_cpsw_nuss_tx_cleanup);
 	k3_udma_glue_disable_tx_chn(tx_chn->tx_chn);
+	tx_chn->xsk_pool = NULL;
+	tx_chn->xsk_port_id = -EINVAL;
 }
 
 static void am65_cpsw_destroy_txqs(struct am65_cpsw_common *common)
@@ -775,12 +777,22 @@ static void am65_cpsw_destroy_txqs(struct am65_cpsw_common *common)
 int am65_cpsw_create_txq(struct am65_cpsw_common *common, int id)
 {
 	struct am65_cpsw_tx_chn *tx_chn = &common->tx_chns[id];
-	int ret;
+	int port, ret;
 
 	ret = k3_udma_glue_enable_tx_chn(tx_chn->tx_chn);
 	if (ret)
 		return ret;
 
+	/* get first port with XSK pool & XDP program set */
+	for (port = 0; port < common->port_num; port++) {
+		tx_chn->xsk_pool = am65_cpsw_xsk_get_pool(&common->ports[port],
+							  id);
+		if (tx_chn->xsk_pool)
+			break;
+	}
+
+	tx_chn->xsk_port_id = tx_chn->xsk_pool ?
+			      common->ports[port].port_id : -EINVAL;
 	napi_enable(&tx_chn->napi_tx);
 
 	return 0;
@@ -881,15 +893,18 @@ static void am65_cpsw_nuss_rx_cleanup(void *data, dma_addr_t desc_dma)
 }
 
 static void am65_cpsw_nuss_xmit_free(struct am65_cpsw_tx_chn *tx_chn,
-				     struct cppi5_host_desc_t *desc)
+				     struct cppi5_host_desc_t *desc,
+				     enum am65_cpsw_tx_buf_type buf_type)
 {
 	struct cppi5_host_desc_t *first_desc, *next_desc;
 	dma_addr_t buf_dma, next_desc_dma;
 	u32 buf_dma_len;
 
 	first_desc = desc;
-	next_desc = first_desc;
+	if (buf_type == AM65_CPSW_TX_BUF_TYPE_XSK_TX)
+		goto free_pool;
 
+	next_desc = first_desc;
 	cppi5_hdesc_get_obuf(first_desc, &buf_dma, &buf_dma_len);
 	k3_udma_glue_tx_cppi5_to_dma_addr(tx_chn->tx_chn, &buf_dma);
 
@@ -912,6 +927,7 @@ static void am65_cpsw_nuss_xmit_free(struct am65_cpsw_tx_chn *tx_chn,
 		k3_cppi_desc_pool_free(tx_chn->desc_pool, next_desc);
 	}
 
+free_pool:
 	k3_cppi_desc_pool_free(tx_chn->desc_pool, first_desc);
 }
 
@@ -921,21 +937,32 @@ static void am65_cpsw_nuss_tx_cleanup(void *data, dma_addr_t desc_dma)
 	enum am65_cpsw_tx_buf_type buf_type;
 	struct am65_cpsw_tx_swdata *swdata;
 	struct cppi5_host_desc_t *desc_tx;
+	struct xsk_buff_pool *xsk_pool;
 	struct xdp_frame *xdpf;
 	struct sk_buff *skb;
 
 	desc_tx = k3_cppi_desc_pool_dma2virt(tx_chn->desc_pool, desc_dma);
 	swdata = cppi5_hdesc_get_swdata(desc_tx);
 	buf_type = am65_cpsw_nuss_buf_type(tx_chn, desc_dma);
-	if (buf_type == AM65_CPSW_TX_BUF_TYPE_SKB) {
+	switch (buf_type) {
+	case AM65_CPSW_TX_BUF_TYPE_SKB:
 		skb = swdata->skb;
 		dev_kfree_skb_any(skb);
-	} else {
+		break;
+	case AM65_CPSW_TX_BUF_TYPE_XDP_TX:
+	case AM65_CPSW_TX_BUF_TYPE_XDP_NDO:
 		xdpf = swdata->xdpf;
 		xdp_return_frame(xdpf);
+		break;
+	case AM65_CPSW_TX_BUF_TYPE_XSK_TX:
+		xsk_pool = swdata->xsk_pool;
+		xsk_tx_completed(xsk_pool, 1);
+		break;
+	default:
+		break;
 	}
 
-	am65_cpsw_nuss_xmit_free(tx_chn, desc_tx);
+	am65_cpsw_nuss_xmit_free(tx_chn, desc_tx, buf_type);
 }
 
 static struct sk_buff *am65_cpsw_build_skb(void *page_addr,
@@ -1180,6 +1207,82 @@ static int am65_cpsw_nuss_ndo_slave_open(struct net_device *ndev)
 	return ret;
 }
 
+static int am65_cpsw_xsk_xmit_zc(struct net_device *ndev,
+				 struct am65_cpsw_tx_chn *tx_chn)
+{
+	struct am65_cpsw_common *common = tx_chn->common;
+	struct xsk_buff_pool *pool = tx_chn->xsk_pool;
+	struct xdp_desc *xdp_descs = pool->tx_descs;
+	struct cppi5_host_desc_t *host_desc;
+	struct am65_cpsw_tx_swdata *swdata;
+	dma_addr_t dma_desc, dma_buf;
+	int num_tx = 0, pkt_len;
+	int descs_avail, ret;
+	int i;
+
+	descs_avail = k3_cppi_desc_pool_avail(tx_chn->desc_pool);
+	/* ensure that TX ring is not filled up by XDP, always MAX_SKB_FRAGS
+	 * will be available for normal TX path and queue is stopped there if
+	 * necessary
+	 */
+	if (descs_avail <= MAX_SKB_FRAGS)
+		return 0;
+
+	descs_avail -= MAX_SKB_FRAGS;
+	descs_avail = xsk_tx_peek_release_desc_batch(pool, descs_avail);
+
+	for (i = 0; i < descs_avail; i++) {
+		host_desc = k3_cppi_desc_pool_alloc(tx_chn->desc_pool);
+		if (unlikely(!host_desc))
+			break;
+
+		am65_cpsw_nuss_set_buf_type(tx_chn, host_desc,
+					    AM65_CPSW_TX_BUF_TYPE_XSK_TX);
+		dma_buf = xsk_buff_raw_get_dma(pool, xdp_descs[i].addr);
+		pkt_len = xdp_descs[i].len;
+		xsk_buff_raw_dma_sync_for_device(pool, dma_buf, pkt_len);
+
+		cppi5_hdesc_init(host_desc, CPPI5_INFO0_HDESC_EPIB_PRESENT,
+				 AM65_CPSW_NAV_PS_DATA_SIZE);
+		cppi5_hdesc_set_pkttype(host_desc, AM65_CPSW_CPPI_TX_PKT_TYPE);
+		cppi5_hdesc_set_pktlen(host_desc, pkt_len);
+		cppi5_desc_set_pktids(&host_desc->hdr, 0,
+				      AM65_CPSW_CPPI_TX_FLOW_ID);
+		cppi5_desc_set_tags_ids(&host_desc->hdr, 0,
+					tx_chn->xsk_port_id);
+
+		k3_udma_glue_tx_dma_to_cppi5_addr(tx_chn->tx_chn, &dma_buf);
+		cppi5_hdesc_attach_buf(host_desc, dma_buf, pkt_len, dma_buf,
+				       pkt_len);
+
+		swdata = cppi5_hdesc_get_swdata(host_desc);
+		swdata->ndev = ndev;
+		swdata->xsk_pool = pool;
+
+		dma_desc = k3_cppi_desc_pool_virt2dma(tx_chn->desc_pool,
+						      host_desc);
+		if (AM65_CPSW_IS_CPSW2G(common)) {
+			ret = k3_udma_glue_push_tx_chn(tx_chn->tx_chn,
+						       host_desc, dma_desc);
+		} else {
+			spin_lock_bh(&tx_chn->lock);
+			ret = k3_udma_glue_push_tx_chn(tx_chn->tx_chn,
+						       host_desc, dma_desc);
+			spin_unlock_bh(&tx_chn->lock);
+		}
+
+		if (ret) {
+			ndev->stats.tx_errors++;
+			k3_cppi_desc_pool_free(tx_chn->desc_pool, host_desc);
+			break;
+		}
+
+		num_tx++;
+	}
+
+	return num_tx;
+}
+
 static int am65_cpsw_xdp_tx_frame(struct net_device *ndev,
 				  struct am65_cpsw_tx_chn *tx_chn,
 				  struct xdp_frame *xdpf,
@@ -1703,15 +1806,19 @@ static int am65_cpsw_nuss_tx_compl_packets(struct am65_cpsw_common *common,
 	struct netdev_queue *netif_txq;
 	unsigned int total_bytes = 0;
 	struct net_device *ndev;
+	int xsk_frames_done = 0;
 	struct xdp_frame *xdpf;
 	unsigned int pkt_len;
 	struct sk_buff *skb;
 	dma_addr_t desc_dma;
 	int res, num_tx = 0;
+	int xsk_tx = 0;
 
 	tx_chn = &common->tx_chns[chn];
 
 	while (true) {
+		pkt_len = 0;
+
 		if (!single_port)
 			spin_lock(&tx_chn->lock);
 		res = k3_udma_glue_pop_tx_chn(tx_chn->tx_chn, &desc_dma);
@@ -1733,25 +1840,36 @@ static int am65_cpsw_nuss_tx_compl_packets(struct am65_cpsw_common *common,
 		swdata = cppi5_hdesc_get_swdata(desc_tx);
 		ndev = swdata->ndev;
 		buf_type = am65_cpsw_nuss_buf_type(tx_chn, desc_dma);
-		if (buf_type == AM65_CPSW_TX_BUF_TYPE_SKB) {
+		switch (buf_type) {
+		case AM65_CPSW_TX_BUF_TYPE_SKB:
 			skb = swdata->skb;
 			am65_cpts_tx_timestamp(tx_chn->common->cpts, skb);
 			pkt_len = skb->len;
 			napi_consume_skb(skb, budget);
-		} else {
+			total_bytes += pkt_len;
+			break;
+		case AM65_CPSW_TX_BUF_TYPE_XDP_TX:
+		case AM65_CPSW_TX_BUF_TYPE_XDP_NDO:
 			xdpf = swdata->xdpf;
 			pkt_len = xdpf->len;
+			total_bytes += pkt_len;
 			if (buf_type == AM65_CPSW_TX_BUF_TYPE_XDP_TX)
 				xdp_return_frame_rx_napi(xdpf);
 			else
 				xdp_return_frame(xdpf);
+			break;
+		case AM65_CPSW_TX_BUF_TYPE_XSK_TX:
+			pkt_len = cppi5_hdesc_get_pktlen(desc_tx);
+			xsk_frames_done++;
+			break;
+		default:
+			break;
 		}
 
-		total_bytes += pkt_len;
 		num_tx++;
-		am65_cpsw_nuss_xmit_free(tx_chn, desc_tx);
+		am65_cpsw_nuss_xmit_free(tx_chn, desc_tx, buf_type);
 		dev_sw_netstats_tx_add(ndev, 1, pkt_len);
-		if (!single_port) {
+		if (!single_port && buf_type != AM65_CPSW_TX_BUF_TYPE_XSK_TX) {
 			/* as packets from multi ports can be interleaved
 			 * on the same channel, we have to figure out the
 			 * port/queue at every packet and report it/wake queue.
@@ -1768,6 +1886,19 @@ static int am65_cpsw_nuss_tx_compl_packets(struct am65_cpsw_common *common,
 		am65_cpsw_nuss_tx_wake(tx_chn, ndev, netif_txq);
 	}
 
+	if (tx_chn->xsk_pool) {
+		if (xsk_frames_done)
+			xsk_tx_completed(tx_chn->xsk_pool, xsk_frames_done);
+
+		if (xsk_uses_need_wakeup(tx_chn->xsk_pool))
+			xsk_set_tx_need_wakeup(tx_chn->xsk_pool);
+
+		ndev = common->ports[tx_chn->xsk_port_id].ndev;
+		netif_txq = netdev_get_tx_queue(ndev, chn);
+		txq_trans_cond_update(netif_txq);
+		xsk_tx = am65_cpsw_xsk_xmit_zc(ndev, tx_chn);
+	}
+
 	dev_dbg(dev, "%s:%u pkt:%d\n", __func__, chn, num_tx);
 
 	return num_tx;
@@ -1778,7 +1909,11 @@ static enum hrtimer_restart am65_cpsw_nuss_tx_timer_callback(struct hrtimer *tim
 	struct am65_cpsw_tx_chn *tx_chns =
 			container_of(timer, struct am65_cpsw_tx_chn, tx_hrtimer);
 
-	enable_irq(tx_chns->irq);
+	if (tx_chns->irq_disabled) {
+		tx_chns->irq_disabled = false;
+		enable_irq(tx_chns->irq);
+	}
+
 	return HRTIMER_NORESTART;
 }
 
@@ -1799,7 +1934,10 @@ static int am65_cpsw_nuss_tx_poll(struct napi_struct *napi_tx, int budget)
 				      ns_to_ktime(tx_chn->tx_pace_timeout),
 				      HRTIMER_MODE_REL_PINNED);
 		} else {
-			enable_irq(tx_chn->irq);
+			if (tx_chn->irq_disabled) {
+				tx_chn->irq_disabled = false;
+				enable_irq(tx_chn->irq);
+			}
 		}
 	}
 
@@ -1821,6 +1959,7 @@ static irqreturn_t am65_cpsw_nuss_tx_irq(int irq, void *dev_id)
 {
 	struct am65_cpsw_tx_chn *tx_chn = dev_id;
 
+	tx_chn->irq_disabled = true;
 	disable_irq_nosync(irq);
 	napi_schedule(&tx_chn->napi_tx);
 
@@ -1985,14 +2124,14 @@ static netdev_tx_t am65_cpsw_nuss_ndo_slave_xmit(struct sk_buff *skb,
 	return NETDEV_TX_OK;
 
 err_free_descs:
-	am65_cpsw_nuss_xmit_free(tx_chn, first_desc);
+	am65_cpsw_nuss_xmit_free(tx_chn, first_desc, AM65_CPSW_TX_BUF_TYPE_SKB);
 err_free_skb:
 	ndev->stats.tx_dropped++;
 	dev_kfree_skb_any(skb);
 	return NETDEV_TX_OK;
 
 busy_free_descs:
-	am65_cpsw_nuss_xmit_free(tx_chn, first_desc);
+	am65_cpsw_nuss_xmit_free(tx_chn, first_desc, AM65_CPSW_TX_BUF_TYPE_SKB);
 busy_stop_q:
 	netif_tx_stop_queue(netif_txq);
 	return NETDEV_TX_BUSY;
@@ -2246,6 +2385,7 @@ static const struct net_device_ops am65_cpsw_nuss_netdev_ops = {
 	.ndo_set_tx_maxrate	= am65_cpsw_qos_ndo_tx_p0_set_maxrate,
 	.ndo_bpf		= am65_cpsw_ndo_bpf,
 	.ndo_xdp_xmit		= am65_cpsw_ndo_xdp_xmit,
+	.ndo_xsk_wakeup		= am65_cpsw_xsk_wakeup,
 };
 
 static void am65_cpsw_disable_phy(struct phy *phy)
diff --git a/drivers/net/ethernet/ti/am65-cpsw-nuss.h b/drivers/net/ethernet/ti/am65-cpsw-nuss.h
index 0e44d8a6cd68..0152767e8436 100644
--- a/drivers/net/ethernet/ti/am65-cpsw-nuss.h
+++ b/drivers/net/ethernet/ti/am65-cpsw-nuss.h
@@ -72,6 +72,7 @@ enum am65_cpsw_tx_buf_type {
 	AM65_CPSW_TX_BUF_TYPE_SKB,
 	AM65_CPSW_TX_BUF_TYPE_XDP_TX,
 	AM65_CPSW_TX_BUF_TYPE_XDP_NDO,
+	AM65_CPSW_TX_BUF_TYPE_XSK_TX,
 };
 
 struct am65_cpsw_host {
@@ -97,6 +98,9 @@ struct am65_cpsw_tx_chn {
 	unsigned char dsize_log2;
 	char tx_chn_name[128];
 	u32 rate_mbps;
+	struct xsk_buff_pool *xsk_pool;
+	int xsk_port_id;
+	bool irq_disabled;
 };
 
 struct am65_cpsw_rx_flow {
@@ -118,6 +122,7 @@ struct am65_cpsw_tx_swdata {
 	union {
 		struct sk_buff *skb;
 		struct xdp_frame *xdpf;
+		struct xsk_buff_pool *xsk_pool;
 	};
 };
 
diff --git a/drivers/net/ethernet/ti/am65-cpsw-xdp.c b/drivers/net/ethernet/ti/am65-cpsw-xdp.c
index e71ff38f851f..b8b35ce702b1 100644
--- a/drivers/net/ethernet/ti/am65-cpsw-xdp.c
+++ b/drivers/net/ethernet/ti/am65-cpsw-xdp.c
@@ -109,8 +109,10 @@ int am65_cpsw_xsk_wakeup(struct net_device *ndev, u32 qid, u32 flags)
 	struct am65_cpsw_common *common = am65_ndev_to_common(ndev);
 	struct am65_cpsw_port *port = am65_ndev_to_port(ndev);
 	struct am65_cpsw_rx_flow *rx_flow;
+	struct am65_cpsw_tx_chn *tx_ch;
 
 	rx_flow = &common->rx_chns.flows[qid];
+	tx_ch = &common->tx_chns[qid];
 
 	if (!netif_running(ndev) || !netif_carrier_ok(ndev))
 		return -ENETDOWN;
@@ -121,9 +123,16 @@ int am65_cpsw_xsk_wakeup(struct net_device *ndev, u32 qid, u32 flags)
 	if (qid >= common->rx_ch_num_flows || qid >= common->tx_ch_num)
 		return -EINVAL;
 
-	if (!rx_flow->xsk_pool)
+	if (!rx_flow->xsk_pool && !tx_ch->xsk_pool)
 		return -EINVAL;
 
+	if (flags & XDP_WAKEUP_TX) {
+		if (!napi_if_scheduled_mark_missed(&tx_ch->napi_tx)) {
+			if (likely(napi_schedule_prep(&tx_ch->napi_tx)))
+				__napi_schedule(&tx_ch->napi_tx);
+		}
+	}
+
 	if (flags & XDP_WAKEUP_RX) {
 		if (!napi_if_scheduled_mark_missed(&rx_flow->napi_rx)) {
 			if (likely(napi_schedule_prep(&rx_flow->napi_rx)))

-- 
2.34.1

[PATCH RFC net-next 5/5] net: ethernet: ti: am65-cpsw: enable zero copy in XDP features

[Roger Quadros]

Now that we have the plumbing in for XDP zero copy RX and TX, enable
the zero copy feature flag.

Signed-off-by: Roger Quadros <rogerq@kernel.org>
---
 drivers/net/ethernet/ti/am65-cpsw-nuss.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/drivers/net/ethernet/ti/am65-cpsw-nuss.c b/drivers/net/ethernet/ti/am65-cpsw-nuss.c
index e89b3cefcb05..894a0bd2a810 100644
--- a/drivers/net/ethernet/ti/am65-cpsw-nuss.c
+++ b/drivers/net/ethernet/ti/am65-cpsw-nuss.c
@@ -3173,7 +3173,8 @@ am65_cpsw_nuss_init_port_ndev(struct am65_cpsw_common *common, u32 port_idx)
 			       NETIF_F_HW_VLAN_CTAG_FILTER;
 	port->ndev->xdp_features = NETDEV_XDP_ACT_BASIC |
 				   NETDEV_XDP_ACT_REDIRECT |
-				   NETDEV_XDP_ACT_NDO_XMIT;
+				   NETDEV_XDP_ACT_NDO_XMIT |
+				   NETDEV_XDP_ACT_XSK_ZEROCOPY;
 	port->ndev->vlan_features |=  NETIF_F_SG;
 	port->ndev->netdev_ops = &am65_cpsw_nuss_netdev_ops;
 	port->ndev->ethtool_ops = &am65_cpsw_ethtool_ops_slave;

-- 
2.34.1

Re: [PATCH RFC net-next 0/5] net: ethernet: ti: am65-cpsw: add AF_XDP zero copy support

[Malladi, Meghana]

Hi Roger,

On 5/20/2025 3:53 PM, Roger Quadros wrote:
> This series adds AF_XDP zero coppy support to am65-cpsw driver.
> 
> Tests were performed on AM62x-sk with xdpsock application [1].
> 
> A clear improvement is seen in 64 byte packets on Transmit (txonly)
> and receive (rxdrop).
> 1500 byte test seems to be limited by line rate (1G link) so no
> improvement seen there in packet rate. A test on higher speed link
> (or PHY-less setup) might be worthwile.
> 
> There is some issue during l2fwd with 64 byte packets and benchmark
> results show 0. I'm still investigating this issue.
> 
> AF_XDP performance using 64 byte packets in Kpps.
> Benchmark:	XDP-SKB		XDP-Native	XDP-Native(ZeroCopy)
> rxdrop		317		504		824
> txonly		400		405		757
> l2fwd 		207		264		0
> 
> AF_XDP performance using 1500 byte packets in Kpps.
> Benchmark:	XDP-SKB		XDP-Native	XDP-Native(ZeroCopy)
> rxdrop		82		82		82
> txonly		82		82		82
> l2fwd 		82		82		82
> 
> [1]: https://github.com/xdp-project/bpf-examples/tree/master/AF_XDP-example
> 
> To:
> 
> Signed-off-by: Roger Quadros <rogerq@kernel.org>

This series crashes Linux on am64xx-hsevm, when I tried nfs boot using 
AM65-CPSW-NUSS driver:
logs: 
https://gist.github.com/MeghanaMalladiTI/d655a1c8ca88113ee7f5f57d6ab0ec4c

Seems like you have reverted the fix for the same bug which was reported 
by Siddharth and fixed by Julien: 
https://lore.kernel.org/all/7f7fb71a-6d15-46f1-b63c-b569a2e230b7@baylibre.com/

reverted lines:
		if (!common->ports[port].ndev)
		/* FIXME should we BUG here? */
			continue;

Can you please take a look at it.

> ---
> Roger Quadros (5):
>        net: ethernet: ti: am65-cpsw: fix BPF Program change on multi-port CPSW
>        net: ethernet: ti: am65-cpsw: add XSK pool helpers
>        net: ethernet: ti: am65-cpsw: Add AF_XDP zero copy for RX
>        net: ethernet: ti: am65-cpsw: Add AF_XDP zero copy for TX
>        net: ethernet: ti: am65-cpsw: enable zero copy in XDP features
> 
>   drivers/net/ethernet/ti/Makefile         |   2 +-
>   drivers/net/ethernet/ti/am65-cpsw-nuss.c | 526 +++++++++++++++++++++++++++----
>   drivers/net/ethernet/ti/am65-cpsw-nuss.h |  37 ++-
>   drivers/net/ethernet/ti/am65-cpsw-xdp.c  | 155 +++++++++
>   4 files changed, 656 insertions(+), 64 deletions(-)
> ---
> base-commit: 9f607dc39b6658ba8ea647bd99725e68c66071b7
> change-id: 20250225-am65-cpsw-xdp-zc-2af9e4be1356
> 
> Best regards,

-- 
Thanks,
Meghana Malladi

Re: [PATCH RFC net-next 0/5] net: ethernet: ti: am65-cpsw: add AF_XDP zero copy support

[Jakub Kicinski]

On Mon, 14 Jul 2025 14:50:05 +0530 Malladi, Meghana wrote:
> > AF_XDP performance using 64 byte packets in Kpps.
> > Benchmark:	XDP-SKB		XDP-Native	XDP-Native(ZeroCopy)
> > rxdrop		317		504		824
> > txonly		400		405		757
> > l2fwd 		207		264		0
> > 
> > AF_XDP performance using 1500 byte packets in Kpps.
> > Benchmark:	XDP-SKB		XDP-Native	XDP-Native(ZeroCopy)
> > rxdrop		82		82		82
> > txonly		82		82		82
> > l2fwd 		82		82		82
> > 
> > [1]: https://github.com/xdp-project/bpf-examples/tree/master/AF_XDP-example
> > 
> > To:
> > 
> > Signed-off-by: Roger Quadros <rogerq@kernel.org>  
> 
> This series crashes Linux on am64xx-hsevm, when I tried nfs boot using 
> AM65-CPSW-NUSS driver:
> logs: 
> https://gist.github.com/MeghanaMalladiTI/d655a1c8ca88113ee7f5f57d6ab0ec4c
> 
> Seems like you have reverted the fix for the same bug which was reported 
> by Siddharth and fixed by Julien: 
> https://lore.kernel.org/all/7f7fb71a-6d15-46f1-b63c-b569a2e230b7@baylibre.com/
> 
> reverted lines:
> 		if (!common->ports[port].ndev)
> 		/* FIXME should we BUG here? */
> 			continue;
> 
> Can you please take a look at it.

Just to be clear -- you're reporting this problem to Roger so that its
fixed before the series is reposted? I don't see this in the tree, I
wanted to make sure it's not something I need to track as a regression.

Re: [EXTERNAL] Re: [PATCH RFC net-next 0/5] net: ethernet: ti: am65-cpsw: add AF_XDP zero copy support

[Malladi, Meghana]

Hi Jakub,

Sorry for the duplicate mail.

On 7/14/2025 8:36 PM, Jakub Kicinski wrote:
> On Mon, 14 Jul 2025 14: 50: 05 +0530 Malladi, Meghana wrote: > > AF_XDP 
> performance using 64 byte packets in Kpps. > > Benchmark: XDP-SKB XDP- 
> Native XDP-Native(ZeroCopy) > > rxdrop 317 504 824 > > txonly 400 405 757 >
> ZjQcmQRYFpfptBannerStart
> This message was sent from outside of Texas Instruments.
> Do not click links or open attachments unless you recognize the source 
> of this email and know the content is safe.
> Report Suspicious
> <https://us-phishalarm-ewt.proofpoint.com/EWT/v1/G3vK! 
> uDdqXRfP1m37CoZlPNNDnQgOintsvKy- 
> cENuCwB1b5Qxa66rT1SFJDmyny6jsjalW7Wur6ukCSGrdQ$>
> ZjQcmQRYFpfptBannerEnd
> 
> On Mon, 14 Jul 2025 14:50:05 +0530 Malladi, Meghana wrote:
>> > AF_XDP performance using 64 byte packets in Kpps.
>> > Benchmark:	XDP-SKB		XDP-Native	XDP-Native(ZeroCopy)
>> > rxdrop		317		504		824
>> > txonly		400		405		757
>> > l2fwd 		207		264		0
>> > 
>> > AF_XDP performance using 1500 byte packets in Kpps.
>> > Benchmark:	XDP-SKB		XDP-Native	XDP-Native(ZeroCopy)
>> > rxdrop		82		82		82
>> > txonly		82		82		82
>> > l2fwd 		82		82		82
>> > 
>> > [1]: https://urldefense.com/v3/__https://github.com/xdp-project/bpf-examples/ 
> tree/master/AF_XDP-example__;!!G3vK!Sv1p-bFPBDlzD-YMO2sjo- 
> X2gv3CW5uHD_O771StRVzMR8Vr75k7tTGQJ27MRy_fz3d9m40aZg$ <https://urldefense.com/v3/__https://github.com/xdp-project/bpf-examples/tree/master/AF_XDP-example__;!!G3vK!Sv1p-bFPBDlzD-YMO2sjo-X2gv3CW5uHD_O771StRVzMR8Vr75k7tTGQJ27MRy_fz3d9m40aZg$>
>> > 
>> > To:
>> > 
>> > Signed-off-by: Roger Quadros <rogerq@kernel.org>  
>> 
>> This series crashes Linux on am64xx-hsevm, when I tried nfs boot using 
>> AM65-CPSW-NUSS driver:
>> logs: 
>> https://urldefense.com/v3/__https://gist.github.com/MeghanaMalladiTI/ 
> d655a1c8ca88113ee7f5f57d6ab0ec4c__;!!G3vK!Sv1p-bFPBDlzD-YMO2sjo- 
> X2gv3CW5uHD_O771StRVzMR8Vr75k7tTGQJ27MRy_fz3ecuWN_dw$ <https://urldefense.com/v3/__https://gist.github.com/MeghanaMalladiTI/d655a1c8ca88113ee7f5f57d6ab0ec4c__;!!G3vK!Sv1p-bFPBDlzD-YMO2sjo-X2gv3CW5uHD_O771StRVzMR8Vr75k7tTGQJ27MRy_fz3ecuWN_dw$>
>> 
>> Seems like you have reverted the fix for the same bug which was reported 
>> by Siddharth and fixed by Julien: 
>> https://urldefense.com/v3/__https://lore.kernel.org/ 
> all/7f7fb71a-6d15-46f1-b63c-b569a2e230b7@baylibre.com/__;!!G3vK!Sv1p- 
> bFPBDlzD-YMO2sjo-X2gv3CW5uHD_O771StRVzMR8Vr75k7tTGQJ27MRy_fz3exh7VnCw$ <https://urldefense.com/v3/__https://lore.kernel.org/all/7f7fb71a-6d15-46f1-b63c-b569a2e230b7@baylibre.com/__;!!G3vK!Sv1p-bFPBDlzD-YMO2sjo-X2gv3CW5uHD_O771StRVzMR8Vr75k7tTGQJ27MRy_fz3exh7VnCw$>
>> 
>> reverted lines:
>> 		if (!common->ports[port].ndev)
>> 		/* FIXME should we BUG here? */
>> 			continue;
>> 
>> Can you please take a look at it.
> 
> Just to be clear -- you're reporting this problem to Roger so that its
> fixed before the series is reposted? I don't see this in the tree, I
> wanted to make sure it's not something I need to track as a regression.
> 

Yes you are right. This isn't a regression, I reported this as part of 
my testing for this RFC patch series.

-- 
Thanks,
Meghana Malladi