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;
+	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_debugfs.c

@@ -42,6 +42,7 @@ 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]);
 
+		if (!r_vec->xsk_pool) {
 			frag = READ_ONCE(rx_ring->rxbufs[i].frag);
 			if (frag)
 				seq_printf(file, " frag=%p", frag);
@@ -49,6 +50,11 @@ static int nfp_rx_q_show(struct seq_file *file, void *data)
 			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.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_ */