Merge branch 'nfp-AF_XDP-zero-copy'

Simon Horman says:

====================
Add AF_XDP zero-copy support for NFP

Niklas Söderlund says:

This series adds AF_XDP zero-copy support for the NFP driver. The series
is based on previous work done by Jakub Kicinski.

Patch 1/5 and 2/5 prepares the driver for AF_XDP support by refactoring
functions that will act differently once AF_XDP is active or not making
the driver easier to read and by preparing some functions to be reused
outside the local file scope. Patch 3/5 and 4/5 prepares the driver for
dealing the UMEM while finally patch 5/5 adds AF_XDP support.

Based on work by Jakub Kicinski.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/netronome/nfp/Makefile

@@ -31,6 +31,7 @@ nfp-objs := \
 	    nfp_net_main.o \
 	    nfp_net_repr.o \
 	    nfp_net_sriov.o \
+	    nfp_net_xsk.o \
 	    nfp_netvf_main.o \
 	    nfp_port.o \
 	    nfp_shared_buf.o \

drivers/net/ethernet/netronome/nfp/nfp_net.h

@@ -109,6 +109,7 @@ struct nfp_eth_table_port;
 struct nfp_net;
 struct nfp_net_r_vector;
 struct nfp_port;
+struct xsk_buff_pool;
 
 /* Convenience macro for wrapping descriptor index on ring size */
 #define D_IDX(ring, idx)	((idx) & ((ring)->cnt - 1))
@@ -170,11 +171,14 @@ struct nfp_net_tx_desc {
  * struct nfp_net_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.
  */
@@ -182,10 +186,18 @@ struct nfp_net_tx_buf {
 	union {
 		struct sk_buff *skb;
 		void *frag;
+		struct xdp_buff *xdp;
 	};
 	dma_addr_t dma_addr;
-	short int fidx;
-	u16 pkt_cnt;
+	union {
+		struct {
+			short int fidx;
+			u16 pkt_cnt;
+		};
+		struct {
+			bool is_xsk_tx;
+		};
+	};
 	u32 real_len;
 };
 
@@ -314,6 +326,16 @@ struct nfp_net_rx_buf {
 	dma_addr_t dma_addr;
 };
 
+/**
+ * struct nfp_net_xsk_rx_buf - software RX XSK buffer descriptor
+ * @dma_addr:	DMA mapping address of the buffer
+ * @xdp:	XSK buffer pool handle (for AF_XDP)
+ */
+struct nfp_net_xsk_rx_buf {
+	dma_addr_t dma_addr;
+	struct xdp_buff *xdp;
+};
+
 /**
  * struct nfp_net_rx_ring - RX ring structure
  * @r_vec:      Back pointer to ring vector structure
@@ -324,6 +346,7 @@ struct nfp_net_rx_buf {
  * @fl_qcidx:   Queue Controller Peripheral (QCP) queue index for the freelist
  * @qcp_fl:     Pointer to base of the QCP freelist queue
  * @rxbufs:     Array of transmitted FL/RX buffers
+ * @xsk_rxbufs: Array of transmitted FL/RX buffers (for AF_XDP)
  * @rxds:       Virtual address of FL/RX ring in host memory
  * @xdp_rxq:    RX-ring info avail for XDP
  * @dma:        DMA address of the FL/RX ring
@@ -342,6 +365,7 @@ struct nfp_net_rx_ring {
 	u8 __iomem *qcp_fl;
 
 	struct nfp_net_rx_buf *rxbufs;
+	struct nfp_net_xsk_rx_buf *xsk_rxbufs;
 	struct nfp_net_rx_desc *rxds;
 
 	struct xdp_rxq_info xdp_rxq;
@@ -360,6 +384,7 @@ struct nfp_net_rx_ring {
  * @tx_ring:        Pointer to TX ring
  * @rx_ring:        Pointer to RX ring
  * @xdp_ring:	    Pointer to an extra TX ring for XDP
+ * @xsk_pool:	    XSK buffer pool active on vector queue pair (for AF_XDP)
  * @irq_entry:      MSI-X table entry (use for talking to the device)
  * @event_ctr:	    Number of interrupt
  * @rx_dim:	    Dynamic interrupt moderation structure for RX
@@ -431,6 +456,7 @@ struct nfp_net_r_vector {
 	u64 rx_replace_buf_alloc_fail;
 
 	struct nfp_net_tx_ring *xdp_ring;
+	struct xsk_buff_pool *xsk_pool;
 
 	struct u64_stats_sync tx_sync;
 	u64 tx_pkts;
@@ -501,6 +527,7 @@ struct nfp_stat_pair {
  * @num_stack_tx_rings:	Number of TX rings used by the stack (not XDP)
  * @num_rx_rings:	Currently configured number of RX rings
  * @mtu:		Device MTU
+ * @xsk_pools:		XSK buffer pools, @max_r_vecs in size (for AF_XDP).
  */
 struct nfp_net_dp {
 	struct device *dev;
@@ -537,6 +564,8 @@ struct nfp_net_dp {
 	unsigned int num_rx_rings;
 
 	unsigned int mtu;
+
+	struct xsk_buff_pool **xsk_pools;
 };
 
 /**
@@ -965,6 +994,7 @@ int nfp_net_mbox_reconfig_and_unlock(struct nfp_net *nn, u32 mbox_cmd);
 void nfp_net_mbox_reconfig_post(struct nfp_net *nn, u32 update);
 int nfp_net_mbox_reconfig_wait_posted(struct nfp_net *nn);
 
+void nfp_net_irq_unmask(struct nfp_net *nn, unsigned int entry_nr);
 unsigned int
 nfp_net_irqs_alloc(struct pci_dev *pdev, struct msix_entry *irq_entries,
 		   unsigned int min_irqs, unsigned int want_irqs);
@@ -973,6 +1003,19 @@ void
 nfp_net_irqs_assign(struct nfp_net *nn, struct msix_entry *irq_entries,
 		    unsigned int n);
 
+void nfp_net_tx_xmit_more_flush(struct nfp_net_tx_ring *tx_ring);
+void nfp_net_tx_complete(struct nfp_net_tx_ring *tx_ring, int budget);
+
+bool
+nfp_net_parse_meta(struct net_device *netdev, struct nfp_meta_parsed *meta,
+		   void *data, void *pkt, unsigned int pkt_len, int meta_len);
+
+void nfp_net_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);
+
 struct nfp_net_dp *nfp_net_clone_dp(struct nfp_net *nn);
 int nfp_net_ring_reconfig(struct nfp_net *nn, struct nfp_net_dp *new,
 			  struct netlink_ext_ack *extack);

drivers/net/ethernet/netronome/nfp/nfp_net_common.c

@@ -38,6 +38,7 @@
 
 #include <net/tls.h>
 #include <net/vxlan.h>
+#include <net/xdp_sock_drv.h>
 
 #include "nfpcore/nfp_nsp.h"
 #include "ccm.h"
@@ -45,6 +46,7 @@
 #include "nfp_net_ctrl.h"
 #include "nfp_net.h"
 #include "nfp_net_sriov.h"
+#include "nfp_net_xsk.h"
 #include "nfp_port.h"
 #include "crypto/crypto.h"
 #include "crypto/fw.h"
@@ -381,7 +383,7 @@ int nfp_net_mbox_reconfig_and_unlock(struct nfp_net *nn, u32 mbox_cmd)
  *
  * Clear the ICR for the IRQ entry.
  */
-static void nfp_net_irq_unmask(struct nfp_net *nn, unsigned int entry_nr)
+void nfp_net_irq_unmask(struct nfp_net *nn, unsigned int entry_nr)
 {
 	nn_writeb(nn, NFP_NET_CFG_ICR(entry_nr), NFP_NET_CFG_ICR_UNMASKED);
 	nn_pci_flush(nn);
@@ -923,7 +925,7 @@ static void nfp_net_tls_tx_undo(struct sk_buff *skb, u64 tls_handle)
 #endif
 }
 
-static void nfp_net_tx_xmit_more_flush(struct nfp_net_tx_ring *tx_ring)
+void nfp_net_tx_xmit_more_flush(struct nfp_net_tx_ring *tx_ring)
 {
 	wmb();
 	nfp_qcp_wr_ptr_add(tx_ring->qcp_q, tx_ring->wr_ptr_add);
@@ -1142,7 +1144,7 @@ static netdev_tx_t nfp_net_tx(struct sk_buff *skb, struct net_device *netdev)
  * @tx_ring:	TX ring structure
  * @budget:	NAPI budget (only used as bool to determine if in NAPI context)
  */
-static void nfp_net_tx_complete(struct nfp_net_tx_ring *tx_ring, int budget)
+void nfp_net_tx_complete(struct nfp_net_tx_ring *tx_ring, int budget)
 {
 	struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
 	struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
@@ -1315,6 +1317,9 @@ nfp_net_tx_ring_reset(struct nfp_net_dp *dp, struct nfp_net_tx_ring *tx_ring)
 		tx_ring->rd_p++;
 	}
 
+	if (tx_ring->is_xdp)
+		nfp_net_xsk_tx_bufs_free(tx_ring);
+
 	memset(tx_ring->txds, 0, tx_ring->size);
 	tx_ring->wr_p = 0;
 	tx_ring->rd_p = 0;
@@ -1338,24 +1343,43 @@ static void nfp_net_tx_timeout(struct net_device *netdev, unsigned int txqueue)
 /* Receive processing
  */
 static unsigned int
-nfp_net_calc_fl_bufsz(struct nfp_net_dp *dp)
+nfp_net_calc_fl_bufsz_data(struct nfp_net_dp *dp)
 {
-	unsigned int fl_bufsz;
+	unsigned int fl_bufsz = 0;
 
-	fl_bufsz = NFP_NET_RX_BUF_HEADROOM;
-	fl_bufsz += dp->rx_dma_off;
 	if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC)
 		fl_bufsz += NFP_NET_MAX_PREPEND;
 	else
 		fl_bufsz += dp->rx_offset;
 	fl_bufsz += ETH_HLEN + VLAN_HLEN * 2 + dp->mtu;
 
+	return fl_bufsz;
+}
+
+static unsigned int nfp_net_calc_fl_bufsz(struct nfp_net_dp *dp)
+{
+	unsigned int fl_bufsz;
+
+	fl_bufsz = NFP_NET_RX_BUF_HEADROOM;
+	fl_bufsz += dp->rx_dma_off;
+	fl_bufsz += nfp_net_calc_fl_bufsz_data(dp);
+
 	fl_bufsz = SKB_DATA_ALIGN(fl_bufsz);
 	fl_bufsz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 
 	return fl_bufsz;
 }
 
+static unsigned int nfp_net_calc_fl_bufsz_xsk(struct nfp_net_dp *dp)
+{
+	unsigned int fl_bufsz;
+
+	fl_bufsz = XDP_PACKET_HEADROOM;
+	fl_bufsz += nfp_net_calc_fl_bufsz_data(dp);
+
+	return fl_bufsz;
+}
+
 static void
 nfp_net_free_frag(void *frag, bool xdp)
 {
@@ -1484,10 +1508,14 @@ static void nfp_net_rx_ring_reset(struct nfp_net_rx_ring *rx_ring)
 	/* Move the empty entry to the end of the list */
 	wr_idx = D_IDX(rx_ring, rx_ring->wr_p);
 	last_idx = rx_ring->cnt - 1;
-	rx_ring->rxbufs[wr_idx].dma_addr = rx_ring->rxbufs[last_idx].dma_addr;
-	rx_ring->rxbufs[wr_idx].frag = rx_ring->rxbufs[last_idx].frag;
-	rx_ring->rxbufs[last_idx].dma_addr = 0;
-	rx_ring->rxbufs[last_idx].frag = NULL;
+	if (rx_ring->r_vec->xsk_pool) {
+		rx_ring->xsk_rxbufs[wr_idx] = rx_ring->xsk_rxbufs[last_idx];
+		memset(&rx_ring->xsk_rxbufs[last_idx], 0,
+		       sizeof(*rx_ring->xsk_rxbufs));
+	} else {
+		rx_ring->rxbufs[wr_idx] = rx_ring->rxbufs[last_idx];
+		memset(&rx_ring->rxbufs[last_idx], 0, sizeof(*rx_ring->rxbufs));
+	}
 
 	memset(rx_ring->rxds, 0, rx_ring->size);
 	rx_ring->wr_p = 0;
@@ -1509,6 +1537,9 @@ nfp_net_rx_ring_bufs_free(struct nfp_net_dp *dp,
 {
 	unsigned int i;
 
+	if (nfp_net_has_xsk_pool_slow(dp, rx_ring->idx))
+		return;
+
 	for (i = 0; i < rx_ring->cnt - 1; i++) {
 		/* NULL skb can only happen when initial filling of the ring
 		 * fails to allocate enough buffers and calls here to free
@@ -1536,6 +1567,9 @@ nfp_net_rx_ring_bufs_alloc(struct nfp_net_dp *dp,
 	struct nfp_net_rx_buf *rxbufs;
 	unsigned int i;
 
+	if (nfp_net_has_xsk_pool_slow(dp, rx_ring->idx))
+		return 0;
+
 	rxbufs = rx_ring->rxbufs;
 
 	for (i = 0; i < rx_ring->cnt - 1; i++) {
@@ -1560,6 +1594,9 @@ nfp_net_rx_ring_fill_freelist(struct nfp_net_dp *dp,
 {
 	unsigned int i;
 
+	if (nfp_net_has_xsk_pool_slow(dp, rx_ring->idx))
+		return nfp_net_xsk_rx_ring_fill_freelist(rx_ring);
+
 	for (i = 0; i < rx_ring->cnt - 1; i++)
 		nfp_net_rx_give_one(dp, rx_ring, rx_ring->rxbufs[i].frag,
 				    rx_ring->rxbufs[i].dma_addr);
@@ -1587,10 +1624,10 @@ static int nfp_net_rx_csum_has_errors(u16 flags)
  * @meta: Parsed metadata prepend
  * @skb: Pointer to SKB
  */
-static void nfp_net_rx_csum(struct nfp_net_dp *dp,
-			    struct nfp_net_r_vector *r_vec,
-			    struct nfp_net_rx_desc *rxd,
-			    struct nfp_meta_parsed *meta, struct sk_buff *skb)
+void nfp_net_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)
 {
 	skb_checksum_none_assert(skb);
 
@@ -1668,7 +1705,7 @@ nfp_net_set_hash_desc(struct net_device *netdev, struct nfp_meta_parsed *meta,
 			 &rx_hash->hash);
 }
 
-static bool
+bool
 nfp_net_parse_meta(struct net_device *netdev, struct nfp_meta_parsed *meta,
 		   void *data, void *pkt, unsigned int pkt_len, int meta_len)
 {
@@ -2540,7 +2577,11 @@ static void nfp_net_rx_ring_free(struct nfp_net_rx_ring *rx_ring)
 
 	if (dp->netdev)
 		xdp_rxq_info_unreg(&rx_ring->xdp_rxq);
-	kvfree(rx_ring->rxbufs);
+
+	if (nfp_net_has_xsk_pool_slow(dp, rx_ring->idx))
+		kvfree(rx_ring->xsk_rxbufs);
+	else
+		kvfree(rx_ring->rxbufs);
 
 	if (rx_ring->rxds)
 		dma_free_coherent(dp->dev, rx_ring->size,
@@ -2548,6 +2589,7 @@ static void nfp_net_rx_ring_free(struct nfp_net_rx_ring *rx_ring)
 
 	rx_ring->cnt = 0;
 	rx_ring->rxbufs = NULL;
+	rx_ring->xsk_rxbufs = NULL;
 	rx_ring->rxds = NULL;
 	rx_ring->dma = 0;
 	rx_ring->size = 0;
@@ -2563,8 +2605,18 @@ static void nfp_net_rx_ring_free(struct nfp_net_rx_ring *rx_ring)
 static int
 nfp_net_rx_ring_alloc(struct nfp_net_dp *dp, struct nfp_net_rx_ring *rx_ring)
 {
+	enum xdp_mem_type mem_type;
+	size_t rxbuf_sw_desc_sz;
 	int err;
 
+	if (nfp_net_has_xsk_pool_slow(dp, rx_ring->idx)) {
+		mem_type = MEM_TYPE_XSK_BUFF_POOL;
+		rxbuf_sw_desc_sz = sizeof(*rx_ring->xsk_rxbufs);
+	} else {
+		mem_type = MEM_TYPE_PAGE_ORDER0;
+		rxbuf_sw_desc_sz = sizeof(*rx_ring->rxbufs);
+	}
+
 	if (dp->netdev) {
 		err = xdp_rxq_info_reg(&rx_ring->xdp_rxq, dp->netdev,
 				       rx_ring->idx, rx_ring->r_vec->napi.napi_id);
@@ -2572,6 +2624,10 @@ nfp_net_rx_ring_alloc(struct nfp_net_dp *dp, struct nfp_net_rx_ring *rx_ring)
 			return err;
 	}
 
+	err = xdp_rxq_info_reg_mem_model(&rx_ring->xdp_rxq, mem_type, NULL);
+	if (err)
+		goto err_alloc;
+
 	rx_ring->cnt = dp->rxd_cnt;
 	rx_ring->size = array_size(rx_ring->cnt, sizeof(*rx_ring->rxds));
 	rx_ring->rxds = dma_alloc_coherent(dp->dev, rx_ring->size,
@@ -2583,10 +2639,17 @@ nfp_net_rx_ring_alloc(struct nfp_net_dp *dp, struct nfp_net_rx_ring *rx_ring)
 		goto err_alloc;
 	}
 
-	rx_ring->rxbufs = kvcalloc(rx_ring->cnt, sizeof(*rx_ring->rxbufs),
-				   GFP_KERNEL);
-	if (!rx_ring->rxbufs)
-		goto err_alloc;
+	if (nfp_net_has_xsk_pool_slow(dp, rx_ring->idx)) {
+		rx_ring->xsk_rxbufs = kvcalloc(rx_ring->cnt, rxbuf_sw_desc_sz,
+					       GFP_KERNEL);
+		if (!rx_ring->xsk_rxbufs)
+			goto err_alloc;
+	} else {
+		rx_ring->rxbufs = kvcalloc(rx_ring->cnt, rxbuf_sw_desc_sz,
+					   GFP_KERNEL);
+		if (!rx_ring->rxbufs)
+			goto err_alloc;
+	}
 
 	return 0;
 
@@ -2638,6 +2701,27 @@ static void nfp_net_rx_rings_free(struct nfp_net_dp *dp)
 	kfree(dp->rx_rings);
 }
 
+static void
+nfp_net_napi_add(struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec, int idx)
+{
+	if (dp->netdev)
+		netif_napi_add(dp->netdev, &r_vec->napi,
+			       nfp_net_has_xsk_pool_slow(dp, idx) ?
+			       nfp_net_xsk_poll : nfp_net_poll,
+			       NAPI_POLL_WEIGHT);
+	else
+		tasklet_enable(&r_vec->tasklet);
+}
+
+static void
+nfp_net_napi_del(struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec)
+{
+	if (dp->netdev)
+		netif_napi_del(&r_vec->napi);
+	else
+		tasklet_disable(&r_vec->tasklet);
+}
+
 static void
 nfp_net_vector_assign_rings(struct nfp_net_dp *dp,
 			    struct nfp_net_r_vector *r_vec, int idx)
@@ -2648,6 +2732,17 @@ nfp_net_vector_assign_rings(struct nfp_net_dp *dp,
 
 	r_vec->xdp_ring = idx < dp->num_tx_rings - dp->num_stack_tx_rings ?
 		&dp->tx_rings[dp->num_stack_tx_rings + idx] : NULL;
+
+	if (nfp_net_has_xsk_pool_slow(dp, idx) || r_vec->xsk_pool) {
+		r_vec->xsk_pool = dp->xdp_prog ? dp->xsk_pools[idx] : NULL;
+
+		if (r_vec->xsk_pool)
+			xsk_pool_set_rxq_info(r_vec->xsk_pool,
+					      &r_vec->rx_ring->xdp_rxq);
+
+		nfp_net_napi_del(dp, r_vec);
+		nfp_net_napi_add(dp, r_vec, idx);
+	}
 }
 
 static int
@@ -2656,23 +2751,14 @@ nfp_net_prepare_vector(struct nfp_net *nn, struct nfp_net_r_vector *r_vec,
 {
 	int err;
 
-	/* Setup NAPI */
-	if (nn->dp.netdev)
-		netif_napi_add(nn->dp.netdev, &r_vec->napi,
-			       nfp_net_poll, NAPI_POLL_WEIGHT);
-	else
-		tasklet_enable(&r_vec->tasklet);
+	nfp_net_napi_add(&nn->dp, r_vec, idx);
 
 	snprintf(r_vec->name, sizeof(r_vec->name),
 		 "%s-rxtx-%d", nfp_net_name(nn), idx);
 	err = request_irq(r_vec->irq_vector, r_vec->handler, 0, r_vec->name,
 			  r_vec);
 	if (err) {
-		if (nn->dp.netdev)
-			netif_napi_del(&r_vec->napi);
-		else
-			tasklet_disable(&r_vec->tasklet);
-
+		nfp_net_napi_del(&nn->dp, r_vec);
 		nn_err(nn, "Error requesting IRQ %d\n", r_vec->irq_vector);
 		return err;
 	}
@@ -2690,11 +2776,7 @@ static void
 nfp_net_cleanup_vector(struct nfp_net *nn, struct nfp_net_r_vector *r_vec)
 {
 	irq_set_affinity_hint(r_vec->irq_vector, NULL);
-	if (nn->dp.netdev)
-		netif_napi_del(&r_vec->napi);
-	else
-		tasklet_disable(&r_vec->tasklet);
-
+	nfp_net_napi_del(&nn->dp, r_vec);
 	free_irq(r_vec->irq_vector, r_vec);
 }
 
@@ -2808,8 +2890,11 @@ static void nfp_net_clear_config_and_disable(struct nfp_net *nn)
 	if (err)
 		nn_err(nn, "Could not disable device: %d\n", err);
 
-	for (r = 0; r < nn->dp.num_rx_rings; r++)
+	for (r = 0; r < nn->dp.num_rx_rings; r++) {
 		nfp_net_rx_ring_reset(&nn->dp.rx_rings[r]);
+		if (nfp_net_has_xsk_pool_slow(&nn->dp, nn->dp.rx_rings[r].idx))
+			nfp_net_xsk_rx_bufs_free(&nn->dp.rx_rings[r]);
+	}
 	for (r = 0; r < nn->dp.num_tx_rings; r++)
 		nfp_net_tx_ring_reset(&nn->dp, &nn->dp.tx_rings[r]);
 	for (r = 0; r < nn->dp.num_r_vecs; r++)
@@ -3296,6 +3381,15 @@ struct nfp_net_dp *nfp_net_clone_dp(struct nfp_net *nn)
 
 	*new = nn->dp;
 
+	new->xsk_pools = kmemdup(new->xsk_pools,
+				 array_size(nn->max_r_vecs,
+					    sizeof(new->xsk_pools)),
+				 GFP_KERNEL);
+	if (!new->xsk_pools) {
+		kfree(new);
+		return NULL;
+	}
+
 	/* Clear things which need to be recomputed */
 	new->fl_bufsz = 0;
 	new->tx_rings = NULL;
@@ -3306,10 +3400,18 @@ struct nfp_net_dp *nfp_net_clone_dp(struct nfp_net *nn)
 	return new;
 }
 
+static void nfp_net_free_dp(struct nfp_net_dp *dp)
+{
+	kfree(dp->xsk_pools);
+	kfree(dp);
+}
+
 static int
 nfp_net_check_config(struct nfp_net *nn, struct nfp_net_dp *dp,
 		     struct netlink_ext_ack *extack)
 {
+	unsigned int r, xsk_min_fl_bufsz;
+
 	/* XDP-enabled tests */
 	if (!dp->xdp_prog)
 		return 0;
@@ -3322,6 +3424,18 @@ nfp_net_check_config(struct nfp_net *nn, struct nfp_net_dp *dp,
 		return -EINVAL;
 	}
 
+	xsk_min_fl_bufsz = nfp_net_calc_fl_bufsz_xsk(dp);
+	for (r = 0; r < nn->max_r_vecs; r++) {
+		if (!dp->xsk_pools[r])
+			continue;
+
+		if (xsk_pool_get_rx_frame_size(dp->xsk_pools[r]) < xsk_min_fl_bufsz) {
+			NL_SET_ERR_MSG_MOD(extack,
+					   "XSK buffer pool chunk size too small\n");
+			return -EINVAL;
+		}
+	}
+
 	return 0;
 }
 
@@ -3389,7 +3503,7 @@ int nfp_net_ring_reconfig(struct nfp_net *nn, struct nfp_net_dp *dp,
 
 	nfp_net_open_stack(nn);
 exit_free_dp:
-	kfree(dp);
+	nfp_net_free_dp(dp);
 
 	return err;
 
@@ -3398,7 +3512,7 @@ int nfp_net_ring_reconfig(struct nfp_net *nn, struct nfp_net_dp *dp,
 err_cleanup_vecs:
 	for (r = dp->num_r_vecs - 1; r >= nn->dp.num_r_vecs; r--)
 		nfp_net_cleanup_vector(nn, &nn->r_vecs[r]);
-	kfree(dp);
+	nfp_net_free_dp(dp);
 	return err;
 }
 
@@ -3716,6 +3830,9 @@ static int nfp_net_xdp(struct net_device *netdev, struct netdev_bpf *xdp)
 		return nfp_net_xdp_setup_drv(nn, xdp);
 	case XDP_SETUP_PROG_HW:
 		return nfp_net_xdp_setup_hw(nn, xdp);
+	case XDP_SETUP_XSK_POOL:
+		return nfp_net_xsk_setup_pool(netdev, xdp->xsk.pool,
+					      xdp->xsk.queue_id);
 	default:
 		return nfp_app_bpf(nn->app, nn, xdp);
 	}
@@ -3766,6 +3883,7 @@ const struct net_device_ops nfp_net_netdev_ops = {
 	.ndo_features_check	= nfp_net_features_check,
 	.ndo_get_phys_port_name	= nfp_net_get_phys_port_name,
 	.ndo_bpf		= nfp_net_xdp,
+	.ndo_xsk_wakeup		= nfp_net_xsk_wakeup,
 	.ndo_get_devlink_port	= nfp_devlink_get_devlink_port,
 };
 
@@ -3893,6 +4011,14 @@ nfp_net_alloc(struct pci_dev *pdev, void __iomem *ctrl_bar, bool needs_netdev,
 	nn->dp.num_r_vecs = max(nn->dp.num_tx_rings, nn->dp.num_rx_rings);
 	nn->dp.num_r_vecs = min_t(unsigned int,
 				  nn->dp.num_r_vecs, num_online_cpus());
+	nn->max_r_vecs = nn->dp.num_r_vecs;
+
+	nn->dp.xsk_pools = kcalloc(nn->max_r_vecs, sizeof(nn->dp.xsk_pools),
+				   GFP_KERNEL);
+	if (!nn->dp.xsk_pools) {
+		err = -ENOMEM;
+		goto err_free_nn;
+	}
 
 	nn->dp.txd_cnt = NFP_NET_TX_DESCS_DEFAULT;
 	nn->dp.rxd_cnt = NFP_NET_RX_DESCS_DEFAULT;
@@ -3932,6 +4058,7 @@ void nfp_net_free(struct nfp_net *nn)
 	WARN_ON(timer_pending(&nn->reconfig_timer) || nn->reconfig_posted);
 	nfp_ccm_mbox_free(nn);
 
+	kfree(nn->dp.xsk_pools);
 	if (nn->dp.netdev)
 		free_netdev(nn->dp.netdev);
 	else

drivers/net/ethernet/netronome/nfp/nfp_net_debugfs.c

@@ -42,13 +42,19 @@ static int nfp_rx_q_show(struct seq_file *file, void *data)
 		seq_printf(file, "%04d: 0x%08x 0x%08x", i,
 			   rxd->vals[0], rxd->vals[1]);
 
-		frag = READ_ONCE(rx_ring->rxbufs[i].frag);
-		if (frag)
-			seq_printf(file, " frag=%p", frag);
-
-		if (rx_ring->rxbufs[i].dma_addr)
-			seq_printf(file, " dma_addr=%pad",
-				   &rx_ring->rxbufs[i].dma_addr);
+		if (!r_vec->xsk_pool) {
+			frag = READ_ONCE(rx_ring->rxbufs[i].frag);
+			if (frag)
+				seq_printf(file, " frag=%p", frag);
+
+			if (rx_ring->rxbufs[i].dma_addr)
+				seq_printf(file, " dma_addr=%pad",
+					   &rx_ring->rxbufs[i].dma_addr);
+		} else {
+			if (rx_ring->xsk_rxbufs[i].dma_addr)
+				seq_printf(file, " dma_addr=%pad",
+					   &rx_ring->xsk_rxbufs[i].dma_addr);
+		}
 
 		if (i == rx_ring->rd_p % rxd_cnt)
 			seq_puts(file, " H_RD ");
@@ -103,20 +109,23 @@ static int nfp_tx_q_show(struct seq_file *file, void *data)
 		   tx_ring->rd_p, tx_ring->wr_p, d_rd_p, d_wr_p);
 
 	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 == r_vec->tx_ring) {
-			struct sk_buff *skb = READ_ONCE(tx_ring->txbufs[i].skb);
-
+		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 {
-			seq_printf(file, " frag=%p",
-				   READ_ONCE(tx_ring->txbufs[i].frag));
+			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)

drivers/net/ethernet/netronome/nfp/nfp_net_xsk.c

+// SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause)
+/* Copyright (C) 2018 Netronome Systems, Inc */
+/* Copyright (C) 2021 Corigine, Inc */
+
+#include <linux/dma-direction.h>
+#include <linux/dma-mapping.h>
+#include <linux/slab.h>
+#include <net/xdp_sock_drv.h>
+#include <trace/events/xdp.h>
+
+#include "nfp_app.h"
+#include "nfp_net.h"
+#include "nfp_net_xsk.h"
+
+static int nfp_net_tx_space(struct nfp_net_tx_ring *tx_ring)
+{
+	return tx_ring->cnt - tx_ring->wr_p + tx_ring->rd_p - 1;
+}
+
+static void nfp_net_xsk_tx_free(struct nfp_net_tx_buf *txbuf)
+{
+	xsk_buff_free(txbuf->xdp);
+
+	txbuf->dma_addr = 0;
+	txbuf->xdp = NULL;
+}
+
+void nfp_net_xsk_tx_bufs_free(struct nfp_net_tx_ring *tx_ring)
+{
+	struct nfp_net_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_net_xsk_tx_free(txbuf);
+
+			xsk_tx_completed(tx_ring->r_vec->xsk_pool, 1);
+		}
+	}
+}
+
+static bool nfp_net_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_net_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_net_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_net_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_net_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(txd,
+					      xsk_buff_raw_get_dma(xsk_pool,
+								   desc[i].addr
+								   ));
+			txd->offset_eop = PCIE_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);
+}
+
+static bool
+nfp_net_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_net_tx_buf *txbuf;
+	struct nfp_net_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 = PCIE_DESC_TX_EOP;
+	txd->dma_len = cpu_to_le16(pkt_len);
+	nfp_desc_set_dma_addr(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 int nfp_net_rx_space(struct nfp_net_rx_ring *rx_ring)
+{
+	return rx_ring->cnt - rx_ring->wr_p + rx_ring->rd_p - 1;
+}
+
+static void
+nfp_net_xsk_rx_bufs_stash(struct nfp_net_rx_ring *rx_ring, unsigned int idx,
+			  struct xdp_buff *xdp)
+{
+	unsigned int headroom;
+
+	headroom = xsk_pool_get_headroom(rx_ring->r_vec->xsk_pool);
+
+	rx_ring->rxds[idx].fld.reserved = 0;
+	rx_ring->rxds[idx].fld.meta_len_dd = 0;
+
+	rx_ring->xsk_rxbufs[idx].xdp = xdp;
+	rx_ring->xsk_rxbufs[idx].dma_addr =
+		xsk_buff_xdp_get_frame_dma(xdp) + headroom;
+}
+
+static void nfp_net_xsk_rx_unstash(struct nfp_net_xsk_rx_buf *rxbuf)
+{
+	rxbuf->dma_addr = 0;
+	rxbuf->xdp = NULL;
+}
+
+static void nfp_net_xsk_rx_free(struct nfp_net_xsk_rx_buf *rxbuf)
+{
+	if (rxbuf->xdp)
+		xsk_buff_free(rxbuf->xdp);
+
+	nfp_net_xsk_rx_unstash(rxbuf);
+}
+
+void nfp_net_xsk_rx_bufs_free(struct nfp_net_rx_ring *rx_ring)
+{
+	unsigned int i;
+
+	if (!rx_ring->cnt)
+		return;
+
+	for (i = 0; i < rx_ring->cnt - 1; i++)
+		nfp_net_xsk_rx_free(&rx_ring->xsk_rxbufs[i]);
+}
+
+void nfp_net_xsk_rx_ring_fill_freelist(struct nfp_net_rx_ring *rx_ring)
+{
+	struct nfp_net_r_vector *r_vec = rx_ring->r_vec;
+	struct xsk_buff_pool *pool = r_vec->xsk_pool;
+	unsigned int wr_idx, wr_ptr_add = 0;
+	struct xdp_buff *xdp;
+
+	while (nfp_net_rx_space(rx_ring)) {
+		wr_idx = D_IDX(rx_ring, rx_ring->wr_p);
+
+		xdp = xsk_buff_alloc(pool);
+		if (!xdp)
+			break;
+
+		nfp_net_xsk_rx_bufs_stash(rx_ring, wr_idx, xdp);
+
+		nfp_desc_set_dma_addr(&rx_ring->rxds[wr_idx].fld,
+				      rx_ring->xsk_rxbufs[wr_idx].dma_addr);
+
+		rx_ring->wr_p++;
+		wr_ptr_add++;
+	}
+
+	/* Ensure all records are visible before incrementing write counter. */
+	wmb();
+	nfp_qcp_wr_ptr_add(rx_ring->qcp_fl, wr_ptr_add);
+}
+
+static void nfp_net_xsk_rx_drop(struct nfp_net_r_vector *r_vec,
+				struct nfp_net_xsk_rx_buf *xrxbuf)
+{
+	u64_stats_update_begin(&r_vec->rx_sync);
+	r_vec->rx_drops++;
+	u64_stats_update_end(&r_vec->rx_sync);
+
+	nfp_net_xsk_rx_free(xrxbuf);
+}
+
+static void nfp_net_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;
+	}
+	memcpy(skb_put(skb, pkt_len), 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_net_rx_csum(dp, r_vec, rxd, meta, skb);
+
+	if (rxd->rxd.flags & PCIE_DESC_RX_VLAN)
+		__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q),
+				       le16_to_cpu(rxd->rxd.vlan));
+	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_net_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, r_vec->xsk_pool);
+		net_prefetch(xrxbuf->xdp->data);
+
+		if (meta_len) {
+			if (unlikely(nfp_net_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_net_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_net_xsk_rx_skb(rx_ring, rxd, xrxbuf, &meta, pkt_len,
+					   true, skbs_polled);
+			break;
+		case XDP_TX:
+			if (!nfp_net_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_map();
+
+	if (tx_ring->wr_ptr_add)
+		nfp_net_tx_xmit_more_flush(tx_ring);
+
+	return pkts_polled;
+}
+
+static void nfp_net_xsk_pool_unmap(struct device *dev,
+				   struct xsk_buff_pool *pool)
+{
+	return xsk_pool_dma_unmap(pool, 0);
+}
+
+static int nfp_net_xsk_pool_map(struct device *dev, struct xsk_buff_pool *pool)
+{
+	return xsk_pool_dma_map(pool, dev, 0);
+}
+
+int nfp_net_xsk_setup_pool(struct net_device *netdev,
+			   struct xsk_buff_pool *pool, u16 queue_id)
+{
+	struct nfp_net *nn = netdev_priv(netdev);
+
+	struct xsk_buff_pool *prev_pool;
+	struct nfp_net_dp *dp;
+	int err;
+
+	/* Reject on old FWs so we can drop some checks on datapath. */
+	if (nn->dp.rx_offset != NFP_NET_CFG_RX_OFFSET_DYNAMIC)
+		return -EOPNOTSUPP;
+	if (!nn->dp.chained_metadata_format)
+		return -EOPNOTSUPP;
+
+	/* Install */
+	if (pool) {
+		err = nfp_net_xsk_pool_map(nn->dp.dev, pool);
+		if (err)
+			return err;
+	}
+
+	/* Reconfig/swap */
+	dp = nfp_net_clone_dp(nn);
+	if (!dp) {
+		err = -ENOMEM;
+		goto err_unmap;
+	}
+
+	prev_pool = dp->xsk_pools[queue_id];
+	dp->xsk_pools[queue_id] = pool;
+
+	err = nfp_net_ring_reconfig(nn, dp, NULL);
+	if (err)
+		goto err_unmap;
+
+	/* Uninstall */
+	if (prev_pool)
+		nfp_net_xsk_pool_unmap(nn->dp.dev, prev_pool);
+
+	return 0;
+err_unmap:
+	if (pool)
+		nfp_net_xsk_pool_unmap(nn->dp.dev, pool);
+
+	return err;
+}
+
+int nfp_net_xsk_wakeup(struct net_device *netdev, u32 queue_id, u32 flags)
+{
+	struct nfp_net *nn = netdev_priv(netdev);
+
+	/* queue_id comes from a zero-copy socket, installed with XDP_SETUP_XSK_POOL,
+	 * so it must be within our vector range.  Moreover, our napi structs
+	 * are statically allocated, so we can always kick them without worrying
+	 * if reconfig is in progress or interface down.
+	 */
+	napi_schedule(&nn->r_vecs[queue_id].napi);
+
+	return 0;
+}
+
+int nfp_net_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_net_xsk_rx(r_vec->rx_ring, budget, &skbs);
+
+	if (pkts_polled < budget) {
+		if (r_vec->tx_ring)
+			nfp_net_tx_complete(r_vec->tx_ring, budget);
+
+		if (!nfp_net_xsk_complete(r_vec->xdp_ring))
+			pkts_polled = budget;
+
+		nfp_net_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;
+}

drivers/net/ethernet/netronome/nfp/nfp_net_xsk.h

+/* SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause) */
+/* Copyright (C) 2018 Netronome Systems, Inc */
+/* Copyright (C) 2021 Corigine, Inc */
+
+#ifndef _NFP_XSK_H_
+#define _NFP_XSK_H_
+
+#include <net/xdp_sock_drv.h>
+
+#define NFP_NET_XSK_TX_BATCH 16		/* XSK TX transmission batch size. */
+
+static inline bool nfp_net_has_xsk_pool_slow(struct nfp_net_dp *dp,
+					     unsigned int qid)
+{
+	return dp->xdp_prog && dp->xsk_pools[qid];
+}
+
+int nfp_net_xsk_setup_pool(struct net_device *netdev, struct xsk_buff_pool *pool,
+			   u16 queue_id);
+
+void nfp_net_xsk_tx_bufs_free(struct nfp_net_tx_ring *tx_ring);
+void nfp_net_xsk_rx_bufs_free(struct nfp_net_rx_ring *rx_ring);
+
+void nfp_net_xsk_rx_ring_fill_freelist(struct nfp_net_rx_ring *rx_ring);
+
+int nfp_net_xsk_wakeup(struct net_device *netdev, u32 queue_id, u32 flags);
+int nfp_net_xsk_poll(struct napi_struct *napi, int budget);
+
+#endif /* _NFP_XSK_H_ */