Resguardo inicial

linux-7.2.3/drivers/net/ethernet/netronome/nfp/nfd3/ipsec.c

@@ -0,0 +1,39 @@
+// SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause)
+/* Copyright (C) 2018 Netronome Systems, Inc */
+/* Copyright (C) 2021 Corigine, Inc */
+
+#include <net/xfrm.h>
+
+#include "../nfp_net.h"
+#include "nfd3.h"
+
+void nfp_nfd3_ipsec_tx(struct nfp_nfd3_tx_desc *txd, struct sk_buff *skb)
+{
+	struct xfrm_state *x = xfrm_input_state(skb);
+	struct xfrm_offload *xo = xfrm_offload(skb);
+	struct iphdr *iph = ip_hdr(skb);
+	int l4_proto;
+
+	if (x->xso.dev && (x->xso.dev->features & NETIF_F_HW_ESP_TX_CSUM)) {
+		txd->flags |= NFD3_DESC_TX_CSUM;
+
+		if (iph->version == 4)
+			txd->flags |= NFD3_DESC_TX_IP4_CSUM;
+
+		if (x->props.mode == XFRM_MODE_TRANSPORT)
+			l4_proto = xo->proto;
+		else if (x->props.mode == XFRM_MODE_TUNNEL)
+			l4_proto = xo->inner_ipproto;
+		else
+			return;
+
+		switch (l4_proto) {
+		case IPPROTO_UDP:
+			txd->flags |= NFD3_DESC_TX_UDP_CSUM;
+			return;
+		case IPPROTO_TCP:
+			txd->flags |= NFD3_DESC_TX_TCP_CSUM;
+			return;
+		}
+	}
+}

linux-7.2.3/drivers/net/ethernet/netronome/nfp/nfd3/nfd3.h

@@ -0,0 +1,114 @@
+/* SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause) */
+/* Copyright (C) 2015-2019 Netronome Systems, Inc. */
+
+#ifndef _NFP_DP_NFD3_H_
+#define _NFP_DP_NFD3_H_
+
+struct sk_buff;
+struct net_device;
+
+/* TX descriptor format */
+
+#define NFD3_DESC_TX_EOP		BIT(7)
+#define NFD3_DESC_TX_OFFSET_MASK	GENMASK(6, 0)
+#define NFD3_DESC_TX_MSS_MASK		GENMASK(13, 0)
+
+/* Flags in the host TX descriptor */
+#define NFD3_DESC_TX_CSUM		BIT(7)
+#define NFD3_DESC_TX_IP4_CSUM		BIT(6)
+#define NFD3_DESC_TX_TCP_CSUM		BIT(5)
+#define NFD3_DESC_TX_UDP_CSUM		BIT(4)
+#define NFD3_DESC_TX_VLAN		BIT(3)
+#define NFD3_DESC_TX_LSO		BIT(2)
+#define NFD3_DESC_TX_ENCAP		BIT(1)
+#define NFD3_DESC_TX_O_IP4_CSUM	BIT(0)
+
+struct nfp_nfd3_tx_desc {
+	union {
+		struct {
+			u8 dma_addr_hi; /* High bits of host buf address */
+			__le16 dma_len;	/* Length to DMA for this desc */
+			u8 offset_eop;	/* Offset in buf where pkt starts +
+					 * highest bit is eop flag.
+					 */
+			__le32 dma_addr_lo; /* Low 32bit of host buf addr */
+
+			__le16 mss;	/* MSS to be used for LSO */
+			u8 lso_hdrlen;	/* LSO, TCP payload offset */
+			u8 flags;	/* TX Flags, see @NFD3_DESC_TX_* */
+			union {
+				struct {
+					u8 l3_offset; /* L3 header offset */
+					u8 l4_offset; /* L4 header offset */
+				};
+				__le16 vlan; /* VLAN tag to add if indicated */
+			};
+			__le16 data_len; /* Length of frame + meta data */
+		} __packed;
+		__le32 vals[4];
+		__le64 vals8[2];
+	};
+};
+
+/**
+ * struct nfp_nfd3_tx_buf - software TX buffer descriptor
+ * @skb:	normal ring, sk_buff associated with this buffer
+ * @frag:	XDP ring, page frag associated with this buffer
+ * @xdp:	XSK buffer pool handle (for AF_XDP)
+ * @dma_addr:	DMA mapping address of the buffer
+ * @fidx:	Fragment index (-1 for the head and [0..nr_frags-1] for frags)
+ * @pkt_cnt:	Number of packets to be produced out of the skb associated
+ *		with this buffer (valid only on the head's buffer).
+ *		Will be 1 for all non-TSO packets.
+ * @is_xsk_tx:	Flag if buffer is a RX buffer after a XDP_TX action and not a
+ *		buffer from the TX queue (for AF_XDP).
+ * @real_len:	Number of bytes which to be produced out of the skb (valid only
+ *		on the head's buffer). Equal to skb->len for non-TSO packets.
+ */
+struct nfp_nfd3_tx_buf {
+	union {
+		struct sk_buff *skb;
+		void *frag;
+		struct xdp_buff *xdp;
+	};
+	dma_addr_t dma_addr;
+	union {
+		struct {
+			short int fidx;
+			u16 pkt_cnt;
+		};
+		struct {
+			bool is_xsk_tx;
+		};
+	};
+	u32 real_len;
+};
+
+void
+nfp_nfd3_rx_csum(const struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec,
+		 const struct nfp_net_rx_desc *rxd,
+		 const struct nfp_meta_parsed *meta, struct sk_buff *skb);
+bool
+nfp_nfd3_parse_meta(struct net_device *netdev, struct nfp_meta_parsed *meta,
+		    void *data, void *pkt, unsigned int pkt_len, int meta_len);
+void nfp_nfd3_tx_complete(struct nfp_net_tx_ring *tx_ring, int budget);
+int nfp_nfd3_poll(struct napi_struct *napi, int budget);
+netdev_tx_t nfp_nfd3_tx(struct sk_buff *skb, struct net_device *netdev);
+bool
+nfp_nfd3_ctrl_tx_one(struct nfp_net *nn, struct nfp_net_r_vector *r_vec,
+		     struct sk_buff *skb, bool old);
+void nfp_nfd3_ctrl_poll(struct tasklet_struct *t);
+void nfp_nfd3_rx_ring_fill_freelist(struct nfp_net_dp *dp,
+				    struct nfp_net_rx_ring *rx_ring);
+void nfp_nfd3_xsk_tx_free(struct nfp_nfd3_tx_buf *txbuf);
+int nfp_nfd3_xsk_poll(struct napi_struct *napi, int budget);
+
+#ifndef CONFIG_NFP_NET_IPSEC
+static inline void nfp_nfd3_ipsec_tx(struct nfp_nfd3_tx_desc *txd, struct sk_buff *skb)
+{
+}
+#else
+void nfp_nfd3_ipsec_tx(struct nfp_nfd3_tx_desc *txd, struct sk_buff *skb);
+#endif
+
+#endif

linux-7.2.3/drivers/net/ethernet/netronome/nfp/nfd3/rings.c

@@ -0,0 +1,278 @@
+// SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause)
+/* Copyright (C) 2015-2019 Netronome Systems, Inc. */
+
+#include <linux/seq_file.h>
+
+#include "../nfp_net.h"
+#include "../nfp_net_dp.h"
+#include "../nfp_net_xsk.h"
+#include "nfd3.h"
+
+static void nfp_nfd3_xsk_tx_bufs_free(struct nfp_net_tx_ring *tx_ring)
+{
+	struct nfp_nfd3_tx_buf *txbuf;
+	unsigned int idx;
+
+	while (tx_ring->rd_p != tx_ring->wr_p) {
+		idx = D_IDX(tx_ring, tx_ring->rd_p);
+		txbuf = &tx_ring->txbufs[idx];
+
+		txbuf->real_len = 0;
+
+		tx_ring->qcp_rd_p++;
+		tx_ring->rd_p++;
+
+		if (tx_ring->r_vec->xsk_pool) {
+			if (txbuf->is_xsk_tx)
+				nfp_nfd3_xsk_tx_free(txbuf);
+
+			xsk_tx_completed(tx_ring->r_vec->xsk_pool, 1);
+		}
+	}
+}
+
+/**
+ * nfp_nfd3_tx_ring_reset() - Free any untransmitted buffers and reset pointers
+ * @dp:		NFP Net data path struct
+ * @tx_ring:	TX ring structure
+ *
+ * Assumes that the device is stopped, must be idempotent.
+ */
+static void
+nfp_nfd3_tx_ring_reset(struct nfp_net_dp *dp, struct nfp_net_tx_ring *tx_ring)
+{
+	struct netdev_queue *nd_q;
+	const skb_frag_t *frag;
+
+	while (!tx_ring->is_xdp && tx_ring->rd_p != tx_ring->wr_p) {
+		struct nfp_nfd3_tx_buf *tx_buf;
+		struct sk_buff *skb;
+		int idx, nr_frags;
+
+		idx = D_IDX(tx_ring, tx_ring->rd_p);
+		tx_buf = &tx_ring->txbufs[idx];
+
+		skb = tx_ring->txbufs[idx].skb;
+		nr_frags = skb_shinfo(skb)->nr_frags;
+
+		if (tx_buf->fidx == -1) {
+			/* unmap head */
+			dma_unmap_single(dp->dev, tx_buf->dma_addr,
+					 skb_headlen(skb), DMA_TO_DEVICE);
+		} else {
+			/* unmap fragment */
+			frag = &skb_shinfo(skb)->frags[tx_buf->fidx];
+			dma_unmap_page(dp->dev, tx_buf->dma_addr,
+				       skb_frag_size(frag), DMA_TO_DEVICE);
+		}
+
+		/* check for last gather fragment */
+		if (tx_buf->fidx == nr_frags - 1)
+			dev_kfree_skb_any(skb);
+
+		tx_buf->dma_addr = 0;
+		tx_buf->skb = NULL;
+		tx_buf->fidx = -2;
+
+		tx_ring->qcp_rd_p++;
+		tx_ring->rd_p++;
+	}
+
+	if (tx_ring->is_xdp)
+		nfp_nfd3_xsk_tx_bufs_free(tx_ring);
+
+	memset(tx_ring->txds, 0, tx_ring->size);
+	tx_ring->wr_p = 0;
+	tx_ring->rd_p = 0;
+	tx_ring->qcp_rd_p = 0;
+	tx_ring->wr_ptr_add = 0;
+
+	if (tx_ring->is_xdp || !dp->netdev)
+		return;
+
+	nd_q = netdev_get_tx_queue(dp->netdev, tx_ring->idx);
+	netdev_tx_reset_queue(nd_q);
+}
+
+/**
+ * nfp_nfd3_tx_ring_free() - Free resources allocated to a TX ring
+ * @tx_ring:   TX ring to free
+ */
+static void nfp_nfd3_tx_ring_free(struct nfp_net_tx_ring *tx_ring)
+{
+	struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
+	struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
+
+	kvfree(tx_ring->txbufs);
+
+	if (tx_ring->txds)
+		dma_free_coherent(dp->dev, tx_ring->size,
+				  tx_ring->txds, tx_ring->dma);
+
+	tx_ring->cnt = 0;
+	tx_ring->txbufs = NULL;
+	tx_ring->txds = NULL;
+	tx_ring->dma = 0;
+	tx_ring->size = 0;
+}
+
+/**
+ * nfp_nfd3_tx_ring_alloc() - Allocate resource for a TX ring
+ * @dp:        NFP Net data path struct
+ * @tx_ring:   TX Ring structure to allocate
+ *
+ * Return: 0 on success, negative errno otherwise.
+ */
+static int
+nfp_nfd3_tx_ring_alloc(struct nfp_net_dp *dp, struct nfp_net_tx_ring *tx_ring)
+{
+	struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
+
+	tx_ring->cnt = dp->txd_cnt;
+
+	tx_ring->size = array_size(tx_ring->cnt, sizeof(*tx_ring->txds));
+	tx_ring->txds = dma_alloc_coherent(dp->dev, tx_ring->size,
+					   &tx_ring->dma,
+					   GFP_KERNEL | __GFP_NOWARN);
+	if (!tx_ring->txds) {
+		netdev_warn(dp->netdev, "failed to allocate TX descriptor ring memory, requested descriptor count: %d, consider lowering descriptor count\n",
+			    tx_ring->cnt);
+		goto err_alloc;
+	}
+
+	tx_ring->txbufs = kvzalloc_objs(*tx_ring->txbufs, tx_ring->cnt);
+	if (!tx_ring->txbufs)
+		goto err_alloc;
+
+	if (!tx_ring->is_xdp && dp->netdev)
+		netif_set_xps_queue(dp->netdev, &r_vec->affinity_mask,
+				    tx_ring->idx);
+
+	return 0;
+
+err_alloc:
+	nfp_nfd3_tx_ring_free(tx_ring);
+	return -ENOMEM;
+}
+
+static void
+nfp_nfd3_tx_ring_bufs_free(struct nfp_net_dp *dp,
+			   struct nfp_net_tx_ring *tx_ring)
+{
+	unsigned int i;
+
+	if (!tx_ring->is_xdp)
+		return;
+
+	for (i = 0; i < tx_ring->cnt; i++) {
+		if (!tx_ring->txbufs[i].frag)
+			return;
+
+		nfp_net_dma_unmap_rx(dp, tx_ring->txbufs[i].dma_addr);
+		__free_page(virt_to_page(tx_ring->txbufs[i].frag));
+	}
+}
+
+static int
+nfp_nfd3_tx_ring_bufs_alloc(struct nfp_net_dp *dp,
+			    struct nfp_net_tx_ring *tx_ring)
+{
+	struct nfp_nfd3_tx_buf *txbufs = tx_ring->txbufs;
+	unsigned int i;
+
+	if (!tx_ring->is_xdp)
+		return 0;
+
+	for (i = 0; i < tx_ring->cnt; i++) {
+		txbufs[i].frag = nfp_net_rx_alloc_one(dp, &txbufs[i].dma_addr);
+		if (!txbufs[i].frag) {
+			nfp_nfd3_tx_ring_bufs_free(dp, tx_ring);
+			return -ENOMEM;
+		}
+	}
+
+	return 0;
+}
+
+static void
+nfp_nfd3_print_tx_descs(struct seq_file *file,
+			struct nfp_net_r_vector *r_vec,
+			struct nfp_net_tx_ring *tx_ring,
+			u32 d_rd_p, u32 d_wr_p)
+{
+	struct nfp_nfd3_tx_desc *txd;
+	u32 txd_cnt = tx_ring->cnt;
+	int i;
+
+	for (i = 0; i < txd_cnt; i++) {
+		struct xdp_buff *xdp;
+		struct sk_buff *skb;
+
+		txd = &tx_ring->txds[i];
+		seq_printf(file, "%04d: 0x%08x 0x%08x 0x%08x 0x%08x", i,
+			   txd->vals[0], txd->vals[1],
+			   txd->vals[2], txd->vals[3]);
+
+		if (!tx_ring->is_xdp) {
+			skb = READ_ONCE(tx_ring->txbufs[i].skb);
+			if (skb)
+				seq_printf(file, " skb->head=%p skb->data=%p",
+					   skb->head, skb->data);
+		} else {
+			xdp = READ_ONCE(tx_ring->txbufs[i].xdp);
+			if (xdp)
+				seq_printf(file, " xdp->data=%p", xdp->data);
+		}
+
+		if (tx_ring->txbufs[i].dma_addr)
+			seq_printf(file, " dma_addr=%pad",
+				   &tx_ring->txbufs[i].dma_addr);
+
+		if (i == tx_ring->rd_p % txd_cnt)
+			seq_puts(file, " H_RD");
+		if (i == tx_ring->wr_p % txd_cnt)
+			seq_puts(file, " H_WR");
+		if (i == d_rd_p % txd_cnt)
+			seq_puts(file, " D_RD");
+		if (i == d_wr_p % txd_cnt)
+			seq_puts(file, " D_WR");
+
+		seq_putc(file, '\n');
+	}
+}
+
+#define NFP_NFD3_CFG_CTRL_SUPPORTED					\
+	(NFP_NET_CFG_CTRL_ENABLE | NFP_NET_CFG_CTRL_PROMISC |		\
+	 NFP_NET_CFG_CTRL_L2BC | NFP_NET_CFG_CTRL_L2MC |		\
+	 NFP_NET_CFG_CTRL_RXCSUM | NFP_NET_CFG_CTRL_TXCSUM |		\
+	 NFP_NET_CFG_CTRL_RXVLAN | NFP_NET_CFG_CTRL_TXVLAN |		\
+	 NFP_NET_CFG_CTRL_RXVLAN_V2 | NFP_NET_CFG_CTRL_RXQINQ |		\
+	 NFP_NET_CFG_CTRL_TXVLAN_V2 |					\
+	 NFP_NET_CFG_CTRL_GATHER | NFP_NET_CFG_CTRL_LSO |		\
+	 NFP_NET_CFG_CTRL_CTAG_FILTER | NFP_NET_CFG_CTRL_CMSG_DATA |	\
+	 NFP_NET_CFG_CTRL_RINGCFG | NFP_NET_CFG_CTRL_RSS |		\
+	 NFP_NET_CFG_CTRL_IRQMOD | NFP_NET_CFG_CTRL_TXRWB |		\
+	 NFP_NET_CFG_CTRL_VEPA |					\
+	 NFP_NET_CFG_CTRL_VXLAN | NFP_NET_CFG_CTRL_NVGRE |		\
+	 NFP_NET_CFG_CTRL_BPF | NFP_NET_CFG_CTRL_LSO2 |			\
+	 NFP_NET_CFG_CTRL_RSS2 | NFP_NET_CFG_CTRL_CSUM_COMPLETE |	\
+	 NFP_NET_CFG_CTRL_LIVE_ADDR)
+
+const struct nfp_dp_ops nfp_nfd3_ops = {
+	.version		= NFP_NFD_VER_NFD3,
+	.tx_min_desc_per_pkt	= 1,
+	.cap_mask		= NFP_NFD3_CFG_CTRL_SUPPORTED,
+	.dma_mask		= DMA_BIT_MASK(40),
+	.poll			= nfp_nfd3_poll,
+	.xsk_poll		= nfp_nfd3_xsk_poll,
+	.ctrl_poll		= nfp_nfd3_ctrl_poll,
+	.xmit			= nfp_nfd3_tx,
+	.ctrl_tx_one		= nfp_nfd3_ctrl_tx_one,
+	.rx_ring_fill_freelist	= nfp_nfd3_rx_ring_fill_freelist,
+	.tx_ring_alloc		= nfp_nfd3_tx_ring_alloc,
+	.tx_ring_reset		= nfp_nfd3_tx_ring_reset,
+	.tx_ring_free		= nfp_nfd3_tx_ring_free,
+	.tx_ring_bufs_alloc	= nfp_nfd3_tx_ring_bufs_alloc,
+	.tx_ring_bufs_free	= nfp_nfd3_tx_ring_bufs_free,
+	.print_tx_descs		= nfp_nfd3_print_tx_descs
+};

linux-7.2.3/drivers/net/ethernet/netronome/nfp/nfd3/xsk.c

@@ -0,0 +1,409 @@
+// SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause)
+/* Copyright (C) 2018 Netronome Systems, Inc */
+/* Copyright (C) 2021 Corigine, Inc */
+
+#include <linux/bpf_trace.h>
+#include <linux/netdevice.h>
+
+#include "../nfp_app.h"
+#include "../nfp_net.h"
+#include "../nfp_net_dp.h"
+#include "../nfp_net_xsk.h"
+#include "nfd3.h"
+
+static bool
+nfp_nfd3_xsk_tx_xdp(const struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec,
+		    struct nfp_net_rx_ring *rx_ring,
+		    struct nfp_net_tx_ring *tx_ring,
+		    struct nfp_net_xsk_rx_buf *xrxbuf, unsigned int pkt_len,
+		    int pkt_off)
+{
+	struct xsk_buff_pool *pool = r_vec->xsk_pool;
+	struct nfp_nfd3_tx_buf *txbuf;
+	struct nfp_nfd3_tx_desc *txd;
+	unsigned int wr_idx;
+
+	if (nfp_net_tx_space(tx_ring) < 1)
+		return false;
+
+	xsk_buff_raw_dma_sync_for_device(pool, xrxbuf->dma_addr + pkt_off,
+					 pkt_len);
+
+	wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
+
+	txbuf = &tx_ring->txbufs[wr_idx];
+	txbuf->xdp = xrxbuf->xdp;
+	txbuf->real_len = pkt_len;
+	txbuf->is_xsk_tx = true;
+
+	/* Build TX descriptor */
+	txd = &tx_ring->txds[wr_idx];
+	txd->offset_eop = NFD3_DESC_TX_EOP;
+	txd->dma_len = cpu_to_le16(pkt_len);
+	nfp_desc_set_dma_addr_40b(txd, xrxbuf->dma_addr + pkt_off);
+	txd->data_len = cpu_to_le16(pkt_len);
+
+	txd->flags = 0;
+	txd->mss = 0;
+	txd->lso_hdrlen = 0;
+
+	tx_ring->wr_ptr_add++;
+	tx_ring->wr_p++;
+
+	return true;
+}
+
+static void nfp_nfd3_xsk_rx_skb(struct nfp_net_rx_ring *rx_ring,
+				const struct nfp_net_rx_desc *rxd,
+				struct nfp_net_xsk_rx_buf *xrxbuf,
+				const struct nfp_meta_parsed *meta,
+				unsigned int pkt_len,
+				bool meta_xdp,
+				unsigned int *skbs_polled)
+{
+	struct nfp_net_r_vector *r_vec = rx_ring->r_vec;
+	struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
+	struct net_device *netdev;
+	struct sk_buff *skb;
+
+	if (likely(!meta->portid)) {
+		netdev = dp->netdev;
+	} else {
+		struct nfp_net *nn = netdev_priv(dp->netdev);
+
+		netdev = nfp_app_dev_get(nn->app, meta->portid, NULL);
+		if (unlikely(!netdev)) {
+			nfp_net_xsk_rx_drop(r_vec, xrxbuf);
+			return;
+		}
+		nfp_repr_inc_rx_stats(netdev, pkt_len);
+	}
+
+	skb = napi_alloc_skb(&r_vec->napi, pkt_len);
+	if (!skb) {
+		nfp_net_xsk_rx_drop(r_vec, xrxbuf);
+		return;
+	}
+	skb_put_data(skb, xrxbuf->xdp->data, pkt_len);
+
+	skb->mark = meta->mark;
+	skb_set_hash(skb, meta->hash, meta->hash_type);
+
+	skb_record_rx_queue(skb, rx_ring->idx);
+	skb->protocol = eth_type_trans(skb, netdev);
+
+	nfp_nfd3_rx_csum(dp, r_vec, rxd, meta, skb);
+
+	if (unlikely(!nfp_net_vlan_strip(skb, rxd, meta))) {
+		dev_kfree_skb_any(skb);
+		nfp_net_xsk_rx_drop(r_vec, xrxbuf);
+		return;
+	}
+
+	if (meta_xdp)
+		skb_metadata_set(skb,
+				 xrxbuf->xdp->data - xrxbuf->xdp->data_meta);
+
+	napi_gro_receive(&rx_ring->r_vec->napi, skb);
+
+	nfp_net_xsk_rx_free(xrxbuf);
+
+	(*skbs_polled)++;
+}
+
+static unsigned int
+nfp_nfd3_xsk_rx(struct nfp_net_rx_ring *rx_ring, int budget,
+		unsigned int *skbs_polled)
+{
+	struct nfp_net_r_vector *r_vec = rx_ring->r_vec;
+	struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
+	struct nfp_net_tx_ring *tx_ring;
+	struct bpf_prog *xdp_prog;
+	bool xdp_redir = false;
+	int pkts_polled = 0;
+
+	xdp_prog = READ_ONCE(dp->xdp_prog);
+	tx_ring = r_vec->xdp_ring;
+
+	while (pkts_polled < budget) {
+		unsigned int meta_len, data_len, pkt_len, pkt_off;
+		struct nfp_net_xsk_rx_buf *xrxbuf;
+		struct nfp_net_rx_desc *rxd;
+		struct nfp_meta_parsed meta;
+		int idx, act;
+
+		idx = D_IDX(rx_ring, rx_ring->rd_p);
+
+		rxd = &rx_ring->rxds[idx];
+		if (!(rxd->rxd.meta_len_dd & PCIE_DESC_RX_DD))
+			break;
+
+		rx_ring->rd_p++;
+		pkts_polled++;
+
+		xrxbuf = &rx_ring->xsk_rxbufs[idx];
+
+		/* If starved of buffers "drop" it and scream. */
+		if (rx_ring->rd_p >= rx_ring->wr_p) {
+			nn_dp_warn(dp, "Starved of RX buffers\n");
+			nfp_net_xsk_rx_drop(r_vec, xrxbuf);
+			break;
+		}
+
+		/* Memory barrier to ensure that we won't do other reads
+		 * before the DD bit.
+		 */
+		dma_rmb();
+
+		memset(&meta, 0, sizeof(meta));
+
+		/* Only supporting AF_XDP with dynamic metadata so buffer layout
+		 * is always:
+		 *
+		 *  ---------------------------------------------------------
+		 * |  off | metadata  |             packet           | XXXX  |
+		 *  ---------------------------------------------------------
+		 */
+		meta_len = rxd->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK;
+		data_len = le16_to_cpu(rxd->rxd.data_len);
+		pkt_len = data_len - meta_len;
+
+		if (unlikely(meta_len > NFP_NET_MAX_PREPEND)) {
+			nn_dp_warn(dp, "Oversized RX packet metadata %u\n",
+				   meta_len);
+			nfp_net_xsk_rx_drop(r_vec, xrxbuf);
+			continue;
+		}
+
+		/* Stats update. */
+		u64_stats_update_begin(&r_vec->rx_sync);
+		r_vec->rx_pkts++;
+		r_vec->rx_bytes += pkt_len;
+		u64_stats_update_end(&r_vec->rx_sync);
+
+		xrxbuf->xdp->data += meta_len;
+		xrxbuf->xdp->data_end = xrxbuf->xdp->data + pkt_len;
+		xdp_set_data_meta_invalid(xrxbuf->xdp);
+		xsk_buff_dma_sync_for_cpu(xrxbuf->xdp);
+		net_prefetch(xrxbuf->xdp->data);
+
+		if (meta_len) {
+			if (unlikely(nfp_nfd3_parse_meta(dp->netdev, &meta,
+							 xrxbuf->xdp->data -
+							 meta_len,
+							 xrxbuf->xdp->data,
+							 pkt_len, meta_len))) {
+				nn_dp_warn(dp, "Invalid RX packet metadata\n");
+				nfp_net_xsk_rx_drop(r_vec, xrxbuf);
+				continue;
+			}
+
+			if (unlikely(meta.portid)) {
+				struct nfp_net *nn = netdev_priv(dp->netdev);
+
+				if (meta.portid != NFP_META_PORT_ID_CTRL) {
+					nfp_nfd3_xsk_rx_skb(rx_ring, rxd,
+							    xrxbuf, &meta,
+							    pkt_len, false,
+							    skbs_polled);
+					continue;
+				}
+
+				nfp_app_ctrl_rx_raw(nn->app, xrxbuf->xdp->data,
+						    pkt_len);
+				nfp_net_xsk_rx_free(xrxbuf);
+				continue;
+			}
+		}
+
+		act = bpf_prog_run_xdp(xdp_prog, xrxbuf->xdp);
+
+		pkt_len = xrxbuf->xdp->data_end - xrxbuf->xdp->data;
+		pkt_off = xrxbuf->xdp->data - xrxbuf->xdp->data_hard_start;
+
+		switch (act) {
+		case XDP_PASS:
+			nfp_nfd3_xsk_rx_skb(rx_ring, rxd, xrxbuf, &meta, pkt_len,
+					    true, skbs_polled);
+			break;
+		case XDP_TX:
+			if (!nfp_nfd3_xsk_tx_xdp(dp, r_vec, rx_ring, tx_ring,
+						 xrxbuf, pkt_len, pkt_off))
+				nfp_net_xsk_rx_drop(r_vec, xrxbuf);
+			else
+				nfp_net_xsk_rx_unstash(xrxbuf);
+			break;
+		case XDP_REDIRECT:
+			if (xdp_do_redirect(dp->netdev, xrxbuf->xdp, xdp_prog)) {
+				nfp_net_xsk_rx_drop(r_vec, xrxbuf);
+			} else {
+				nfp_net_xsk_rx_unstash(xrxbuf);
+				xdp_redir = true;
+			}
+			break;
+		default:
+			bpf_warn_invalid_xdp_action(dp->netdev, xdp_prog, act);
+			fallthrough;
+		case XDP_ABORTED:
+			trace_xdp_exception(dp->netdev, xdp_prog, act);
+			fallthrough;
+		case XDP_DROP:
+			nfp_net_xsk_rx_drop(r_vec, xrxbuf);
+			break;
+		}
+	}
+
+	nfp_net_xsk_rx_ring_fill_freelist(r_vec->rx_ring);
+
+	if (xdp_redir)
+		xdp_do_flush();
+
+	if (tx_ring->wr_ptr_add)
+		nfp_net_tx_xmit_more_flush(tx_ring);
+
+	return pkts_polled;
+}
+
+void nfp_nfd3_xsk_tx_free(struct nfp_nfd3_tx_buf *txbuf)
+{
+	xsk_buff_free(txbuf->xdp);
+
+	txbuf->dma_addr = 0;
+	txbuf->xdp = NULL;
+}
+
+static bool nfp_nfd3_xsk_complete(struct nfp_net_tx_ring *tx_ring)
+{
+	struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
+	u32 done_pkts = 0, done_bytes = 0, reused = 0;
+	bool done_all;
+	int idx, todo;
+	u32 qcp_rd_p;
+
+	if (tx_ring->wr_p == tx_ring->rd_p)
+		return true;
+
+	/* Work out how many descriptors have been transmitted. */
+	qcp_rd_p = nfp_qcp_rd_ptr_read(tx_ring->qcp_q);
+
+	if (qcp_rd_p == tx_ring->qcp_rd_p)
+		return true;
+
+	todo = D_IDX(tx_ring, qcp_rd_p - tx_ring->qcp_rd_p);
+
+	done_all = todo <= NFP_NET_XDP_MAX_COMPLETE;
+	todo = min(todo, NFP_NET_XDP_MAX_COMPLETE);
+
+	tx_ring->qcp_rd_p = D_IDX(tx_ring, tx_ring->qcp_rd_p + todo);
+
+	done_pkts = todo;
+	while (todo--) {
+		struct nfp_nfd3_tx_buf *txbuf;
+
+		idx = D_IDX(tx_ring, tx_ring->rd_p);
+		tx_ring->rd_p++;
+
+		txbuf = &tx_ring->txbufs[idx];
+		if (unlikely(!txbuf->real_len))
+			continue;
+
+		done_bytes += txbuf->real_len;
+		txbuf->real_len = 0;
+
+		if (txbuf->is_xsk_tx) {
+			nfp_nfd3_xsk_tx_free(txbuf);
+			reused++;
+		}
+	}
+
+	u64_stats_update_begin(&r_vec->tx_sync);
+	r_vec->tx_bytes += done_bytes;
+	r_vec->tx_pkts += done_pkts;
+	u64_stats_update_end(&r_vec->tx_sync);
+
+	xsk_tx_completed(r_vec->xsk_pool, done_pkts - reused);
+
+	WARN_ONCE(tx_ring->wr_p - tx_ring->rd_p > tx_ring->cnt,
+		  "XDP TX ring corruption rd_p=%u wr_p=%u cnt=%u\n",
+		  tx_ring->rd_p, tx_ring->wr_p, tx_ring->cnt);
+
+	return done_all;
+}
+
+static void nfp_nfd3_xsk_tx(struct nfp_net_tx_ring *tx_ring)
+{
+	struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
+	struct xdp_desc desc[NFP_NET_XSK_TX_BATCH];
+	struct xsk_buff_pool *xsk_pool;
+	struct nfp_nfd3_tx_desc *txd;
+	u32 pkts = 0, wr_idx;
+	u32 i, got;
+
+	xsk_pool = r_vec->xsk_pool;
+
+	while (nfp_net_tx_space(tx_ring) >= NFP_NET_XSK_TX_BATCH) {
+		for (i = 0; i < NFP_NET_XSK_TX_BATCH; i++)
+			if (!xsk_tx_peek_desc(xsk_pool, &desc[i]))
+				break;
+		got = i;
+		if (!got)
+			break;
+
+		wr_idx = D_IDX(tx_ring, tx_ring->wr_p + i);
+		prefetchw(&tx_ring->txds[wr_idx]);
+
+		for (i = 0; i < got; i++)
+			xsk_buff_raw_dma_sync_for_device(xsk_pool, desc[i].addr,
+							 desc[i].len);
+
+		for (i = 0; i < got; i++) {
+			wr_idx = D_IDX(tx_ring, tx_ring->wr_p + i);
+
+			tx_ring->txbufs[wr_idx].real_len = desc[i].len;
+			tx_ring->txbufs[wr_idx].is_xsk_tx = false;
+
+			/* Build TX descriptor. */
+			txd = &tx_ring->txds[wr_idx];
+			nfp_desc_set_dma_addr_40b(txd,
+						  xsk_buff_raw_get_dma(xsk_pool, desc[i].addr));
+			txd->offset_eop = NFD3_DESC_TX_EOP;
+			txd->dma_len = cpu_to_le16(desc[i].len);
+			txd->data_len = cpu_to_le16(desc[i].len);
+		}
+
+		tx_ring->wr_p += got;
+		pkts += got;
+	}
+
+	if (!pkts)
+		return;
+
+	xsk_tx_release(xsk_pool);
+	/* Ensure all records are visible before incrementing write counter. */
+	wmb();
+	nfp_qcp_wr_ptr_add(tx_ring->qcp_q, pkts);
+}
+
+int nfp_nfd3_xsk_poll(struct napi_struct *napi, int budget)
+{
+	struct nfp_net_r_vector *r_vec =
+		container_of(napi, struct nfp_net_r_vector, napi);
+	unsigned int pkts_polled, skbs = 0;
+
+	pkts_polled = nfp_nfd3_xsk_rx(r_vec->rx_ring, budget, &skbs);
+
+	if (pkts_polled < budget) {
+		if (r_vec->tx_ring)
+			nfp_nfd3_tx_complete(r_vec->tx_ring, budget);
+
+		if (!nfp_nfd3_xsk_complete(r_vec->xdp_ring))
+			pkts_polled = budget;
+
+		nfp_nfd3_xsk_tx(r_vec->xdp_ring);
+
+		if (pkts_polled < budget && napi_complete_done(napi, skbs))
+			nfp_net_irq_unmask(r_vec->nfp_net, r_vec->irq_entry);
+	}
+
+	return pkts_polled;
+}