Merge branch 'bnxt-xdp-multi-buffer'

Michael Chan says:

====================
bnxt: Support XDP multi buffer

This series adds XDP multi buffer support, allowing MTU to go beyond
the page size limit.

v4: Rebase with latest net-next
v3: Simplify page mode buffer size calculation
    Check to make sure XDP program supports multipage packets
v2: Fix uninitialized variable warnings in patch 1 and 10.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/broadcom/bnxt/bnxt.h

@@ -591,10 +591,12 @@ struct nqe_cn {
 #define BNXT_RX_PAGE_SIZE (1 << BNXT_RX_PAGE_SHIFT)
 
 #define BNXT_MAX_MTU		9500
-#define BNXT_MAX_PAGE_MODE_MTU	\
+#define BNXT_PAGE_MODE_BUF_SIZE \
 	((unsigned int)PAGE_SIZE - VLAN_ETH_HLEN - NET_IP_ALIGN -	\
-	 XDP_PACKET_HEADROOM - \
-	 SKB_DATA_ALIGN((unsigned int)sizeof(struct skb_shared_info)))
+	 XDP_PACKET_HEADROOM)
+#define BNXT_MAX_PAGE_MODE_MTU	\
+	BNXT_PAGE_MODE_BUF_SIZE - \
+	SKB_DATA_ALIGN((unsigned int)sizeof(struct skb_shared_info))
 
 #define BNXT_MIN_PKT_SIZE	52
 
@@ -699,13 +701,12 @@ struct bnxt_sw_tx_bd {
 	};
 	DEFINE_DMA_UNMAP_ADDR(mapping);
 	DEFINE_DMA_UNMAP_LEN(len);
+	struct page		*page;
 	u8			is_gso;
 	u8			is_push;
 	u8			action;
-	union {
 	unsigned short		nr_frags;
 	u16			rx_prod;
-	};
 };
 
 struct bnxt_sw_rx_bd {
@@ -1817,6 +1818,7 @@ struct bnxt {
 #define BNXT_SUPPORTS_TPA(bp)	(!BNXT_CHIP_TYPE_NITRO_A0(bp) &&	\
 				 (!((bp)->flags & BNXT_FLAG_CHIP_P5) ||	\
 				  (bp)->max_tpa_v2) && !is_kdump_kernel())
+#define BNXT_RX_JUMBO_MODE(bp)	((bp)->flags & BNXT_FLAG_JUMBO)
 
 #define BNXT_CHIP_SR2(bp)			\
 	((bp)->chip_num == CHIP_NUM_58818)

drivers/net/ethernet/broadcom/bnxt/bnxt_ethtool.c

@@ -3491,7 +3491,7 @@ static int bnxt_run_loopback(struct bnxt *bp)
 		dev_kfree_skb(skb);
 		return -EIO;
 	}
-	bnxt_xmit_bd(bp, txr, map, pkt_size);
+	bnxt_xmit_bd(bp, txr, map, pkt_size, NULL);
 
 	/* Sync BD data before updating doorbell */
 	wmb();

drivers/net/ethernet/broadcom/bnxt/bnxt_xdp.c

@@ -24,36 +24,91 @@ DEFINE_STATIC_KEY_FALSE(bnxt_xdp_locking_key);
 
 struct bnxt_sw_tx_bd *bnxt_xmit_bd(struct bnxt *bp,
 				   struct bnxt_tx_ring_info *txr,
-				   dma_addr_t mapping, u32 len)
+				   dma_addr_t mapping, u32 len,
+				   struct xdp_buff *xdp)
 {
-	struct bnxt_sw_tx_bd *tx_buf;
+	struct skb_shared_info *sinfo;
+	struct bnxt_sw_tx_bd *tx_buf, *first_buf;
 	struct tx_bd *txbd;
+	int num_frags = 0;
 	u32 flags;
 	u16 prod;
+	int i;
+
+	if (xdp && xdp_buff_has_frags(xdp)) {
+		sinfo = xdp_get_shared_info_from_buff(xdp);
+		num_frags = sinfo->nr_frags;
+	}
 
+	/* fill up the first buffer */
 	prod = txr->tx_prod;
 	tx_buf = &txr->tx_buf_ring[prod];
+	first_buf = tx_buf;
+	tx_buf->nr_frags = num_frags;
+	if (xdp)
+		tx_buf->page = virt_to_head_page(xdp->data);
 
 	txbd = &txr->tx_desc_ring[TX_RING(prod)][TX_IDX(prod)];
-	flags = (len << TX_BD_LEN_SHIFT) | (1 << TX_BD_FLAGS_BD_CNT_SHIFT) |
-		TX_BD_FLAGS_PACKET_END | bnxt_lhint_arr[len >> 9];
+	flags = ((len) << TX_BD_LEN_SHIFT) | ((num_frags + 1) << TX_BD_FLAGS_BD_CNT_SHIFT);
 	txbd->tx_bd_len_flags_type = cpu_to_le32(flags);
 	txbd->tx_bd_opaque = prod;
 	txbd->tx_bd_haddr = cpu_to_le64(mapping);
 
+	/* now let us fill up the frags into the next buffers */
+	for (i = 0; i < num_frags ; i++) {
+		skb_frag_t *frag = &sinfo->frags[i];
+		struct bnxt_sw_tx_bd *frag_tx_buf;
+		struct pci_dev *pdev = bp->pdev;
+		dma_addr_t frag_mapping;
+		int frag_len;
+
 		prod = NEXT_TX(prod);
 		txr->tx_prod = prod;
-	return tx_buf;
+
+		/* first fill up the first buffer */
+		frag_tx_buf = &txr->tx_buf_ring[prod];
+		frag_tx_buf->page = skb_frag_page(frag);
+
+		txbd = &txr->tx_desc_ring[TX_RING(prod)][TX_IDX(prod)];
+
+		frag_len = skb_frag_size(frag);
+		frag_mapping = skb_frag_dma_map(&pdev->dev, frag, 0,
+						frag_len, DMA_TO_DEVICE);
+
+		if (unlikely(dma_mapping_error(&pdev->dev, frag_mapping)))
+			return NULL;
+
+		dma_unmap_addr_set(frag_tx_buf, mapping, frag_mapping);
+
+		flags = frag_len << TX_BD_LEN_SHIFT;
+		txbd->tx_bd_len_flags_type = cpu_to_le32(flags);
+		txbd->tx_bd_opaque = prod;
+		txbd->tx_bd_haddr = cpu_to_le64(frag_mapping);
+
+		len = frag_len;
+	}
+
+	flags &= ~TX_BD_LEN;
+	txbd->tx_bd_len_flags_type = cpu_to_le32(((len) << TX_BD_LEN_SHIFT) | flags |
+			TX_BD_FLAGS_PACKET_END);
+	/* Sync TX BD */
+	wmb();
+	prod = NEXT_TX(prod);
+	txr->tx_prod = prod;
+
+	return first_buf;
 }
 
 static void __bnxt_xmit_xdp(struct bnxt *bp, struct bnxt_tx_ring_info *txr,
-			    dma_addr_t mapping, u32 len, u16 rx_prod)
+			    dma_addr_t mapping, u32 len, u16 rx_prod,
+			    struct xdp_buff *xdp)
 {
 	struct bnxt_sw_tx_bd *tx_buf;
 
-	tx_buf = bnxt_xmit_bd(bp, txr, mapping, len);
+	tx_buf = bnxt_xmit_bd(bp, txr, mapping, len, xdp);
 	tx_buf->rx_prod = rx_prod;
 	tx_buf->action = XDP_TX;
+
 }
 
 static void __bnxt_xmit_xdp_redirect(struct bnxt *bp,
@@ -63,7 +118,7 @@ static void __bnxt_xmit_xdp_redirect(struct bnxt *bp,
 {
 	struct bnxt_sw_tx_bd *tx_buf;
 
-	tx_buf = bnxt_xmit_bd(bp, txr, mapping, len);
+	tx_buf = bnxt_xmit_bd(bp, txr, mapping, len, NULL);
 	tx_buf->action = XDP_REDIRECT;
 	tx_buf->xdpf = xdpf;
 	dma_unmap_addr_set(tx_buf, mapping, mapping);
@@ -78,7 +133,7 @@ void bnxt_tx_int_xdp(struct bnxt *bp, struct bnxt_napi *bnapi, int nr_pkts)
 	struct bnxt_sw_tx_bd *tx_buf;
 	u16 tx_cons = txr->tx_cons;
 	u16 last_tx_cons = tx_cons;
-	int i;
+	int i, j, frags;
 
 	for (i = 0; i < nr_pkts; i++) {
 		tx_buf = &txr->tx_buf_ring[tx_cons];
@@ -96,6 +151,13 @@ void bnxt_tx_int_xdp(struct bnxt *bp, struct bnxt_napi *bnapi, int nr_pkts)
 		} else if (tx_buf->action == XDP_TX) {
 			rx_doorbell_needed = true;
 			last_tx_cons = tx_cons;
+
+			frags = tx_buf->nr_frags;
+			for (j = 0; j < frags; j++) {
+				tx_cons = NEXT_TX(tx_cons);
+				tx_buf = &txr->tx_buf_ring[tx_cons];
+				page_pool_recycle_direct(rxr->page_pool, tx_buf->page);
+			}
 		}
 		tx_cons = NEXT_TX(tx_cons);
 	}
@@ -103,7 +165,52 @@ void bnxt_tx_int_xdp(struct bnxt *bp, struct bnxt_napi *bnapi, int nr_pkts)
 	if (rx_doorbell_needed) {
 		tx_buf = &txr->tx_buf_ring[last_tx_cons];
 		bnxt_db_write(bp, &rxr->rx_db, tx_buf->rx_prod);
+
+	}
+}
+
+bool bnxt_xdp_attached(struct bnxt *bp, struct bnxt_rx_ring_info *rxr)
+{
+	struct bpf_prog *xdp_prog = READ_ONCE(rxr->xdp_prog);
+
+	return !!xdp_prog;
+}
+
+void bnxt_xdp_buff_init(struct bnxt *bp, struct bnxt_rx_ring_info *rxr,
+			u16 cons, u8 **data_ptr, unsigned int *len,
+			struct xdp_buff *xdp)
+{
+	struct bnxt_sw_rx_bd *rx_buf;
+	struct pci_dev *pdev;
+	dma_addr_t mapping;
+	u32 offset;
+
+	pdev = bp->pdev;
+	rx_buf = &rxr->rx_buf_ring[cons];
+	offset = bp->rx_offset;
+
+	mapping = rx_buf->mapping - bp->rx_dma_offset;
+	dma_sync_single_for_cpu(&pdev->dev, mapping + offset, *len, bp->rx_dir);
+
+	xdp_init_buff(xdp, BNXT_PAGE_MODE_BUF_SIZE + offset, &rxr->xdp_rxq);
+	xdp_prepare_buff(xdp, *data_ptr - offset, offset, *len, false);
+}
+
+void bnxt_xdp_buff_frags_free(struct bnxt_rx_ring_info *rxr,
+			      struct xdp_buff *xdp)
+{
+	struct skb_shared_info *shinfo;
+	int i;
+
+	if (!xdp || !xdp_buff_has_frags(xdp))
+		return;
+	shinfo = xdp_get_shared_info_from_buff(xdp);
+	for (i = 0; i < shinfo->nr_frags; i++) {
+		struct page *page = skb_frag_page(&shinfo->frags[i]);
+
+		page_pool_recycle_direct(rxr->page_pool, page);
 	}
+	shinfo->nr_frags = 0;
 }
 
 /* returns the following:
@@ -111,14 +218,14 @@ void bnxt_tx_int_xdp(struct bnxt *bp, struct bnxt_napi *bnapi, int nr_pkts)
  * false   - packet should be passed to the stack.
  */
 bool bnxt_rx_xdp(struct bnxt *bp, struct bnxt_rx_ring_info *rxr, u16 cons,
-		 struct page *page, u8 **data_ptr, unsigned int *len, u8 *event)
+		 struct xdp_buff xdp, struct page *page, unsigned int *len, u8 *event)
 {
 	struct bpf_prog *xdp_prog = READ_ONCE(rxr->xdp_prog);
 	struct bnxt_tx_ring_info *txr;
 	struct bnxt_sw_rx_bd *rx_buf;
 	struct pci_dev *pdev;
-	struct xdp_buff xdp;
 	dma_addr_t mapping;
+	u32 tx_needed = 1;
 	void *orig_data;
 	u32 tx_avail;
 	u32 offset;
@@ -128,16 +235,10 @@ bool bnxt_rx_xdp(struct bnxt *bp, struct bnxt_rx_ring_info *rxr, u16 cons,
 		return false;
 
 	pdev = bp->pdev;
-	rx_buf = &rxr->rx_buf_ring[cons];
 	offset = bp->rx_offset;
 
-	mapping = rx_buf->mapping - bp->rx_dma_offset;
-	dma_sync_single_for_cpu(&pdev->dev, mapping + offset, *len, bp->rx_dir);
-
 	txr = rxr->bnapi->tx_ring;
 	/* BNXT_RX_PAGE_MODE(bp) when XDP enabled */
-	xdp_init_buff(&xdp, PAGE_SIZE, &rxr->xdp_rxq);
-	xdp_prepare_buff(&xdp, *data_ptr - offset, offset, *len, false);
 	orig_data = xdp.data;
 
 	act = bpf_prog_run_xdp(xdp_prog, &xdp);
@@ -150,26 +251,38 @@ bool bnxt_rx_xdp(struct bnxt *bp, struct bnxt_rx_ring_info *rxr, u16 cons,
 		*event &= ~BNXT_RX_EVENT;
 
 	*len = xdp.data_end - xdp.data;
-	if (orig_data != xdp.data) {
+	if (orig_data != xdp.data)
 		offset = xdp.data - xdp.data_hard_start;
-		*data_ptr = xdp.data_hard_start + offset;
-	}
+
 	switch (act) {
 	case XDP_PASS:
 		return false;
 
 	case XDP_TX:
-		if (tx_avail < 1) {
+		rx_buf = &rxr->rx_buf_ring[cons];
+		mapping = rx_buf->mapping - bp->rx_dma_offset;
+		*event = 0;
+
+		if (unlikely(xdp_buff_has_frags(&xdp))) {
+			struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(&xdp);
+
+			tx_needed += sinfo->nr_frags;
+			*event = BNXT_AGG_EVENT;
+		}
+
+		if (tx_avail < tx_needed) {
 			trace_xdp_exception(bp->dev, xdp_prog, act);
+			bnxt_xdp_buff_frags_free(rxr, &xdp);
 			bnxt_reuse_rx_data(rxr, cons, page);
 			return true;
 		}
 
-		*event = BNXT_TX_EVENT;
 		dma_sync_single_for_device(&pdev->dev, mapping + offset, *len,
 					   bp->rx_dir);
+
+		*event |= BNXT_TX_EVENT;
 		__bnxt_xmit_xdp(bp, txr, mapping + offset, *len,
-				NEXT_RX(rxr->rx_prod));
+				NEXT_RX(rxr->rx_prod), &xdp);
 		bnxt_reuse_rx_data(rxr, cons, page);
 		return true;
 	case XDP_REDIRECT:
@@ -177,6 +290,8 @@ bool bnxt_rx_xdp(struct bnxt *bp, struct bnxt_rx_ring_info *rxr, u16 cons,
 		 * redirect is coming from a frame received by the
 		 * bnxt_en driver.
 		 */
+		rx_buf = &rxr->rx_buf_ring[cons];
+		mapping = rx_buf->mapping - bp->rx_dma_offset;
 		dma_unmap_page_attrs(&pdev->dev, mapping,
 				     PAGE_SIZE, bp->rx_dir,
 				     DMA_ATTR_WEAK_ORDERING);
@@ -184,6 +299,7 @@ bool bnxt_rx_xdp(struct bnxt *bp, struct bnxt_rx_ring_info *rxr, u16 cons,
 		/* if we are unable to allocate a new buffer, abort and reuse */
 		if (bnxt_alloc_rx_data(bp, rxr, rxr->rx_prod, GFP_ATOMIC)) {
 			trace_xdp_exception(bp->dev, xdp_prog, act);
+			bnxt_xdp_buff_frags_free(rxr, &xdp);
 			bnxt_reuse_rx_data(rxr, cons, page);
 			return true;
 		}
@@ -203,6 +319,7 @@ bool bnxt_rx_xdp(struct bnxt *bp, struct bnxt_rx_ring_info *rxr, u16 cons,
 		trace_xdp_exception(bp->dev, xdp_prog, act);
 		fallthrough;
 	case XDP_DROP:
+		bnxt_xdp_buff_frags_free(rxr, &xdp);
 		bnxt_reuse_rx_data(rxr, cons, page);
 		break;
 	}
@@ -270,8 +387,9 @@ static int bnxt_xdp_set(struct bnxt *bp, struct bpf_prog *prog)
 	int tx_xdp = 0, rc, tc;
 	struct bpf_prog *old;
 
-	if (prog && bp->dev->mtu > BNXT_MAX_PAGE_MODE_MTU) {
-		netdev_warn(dev, "MTU %d larger than largest XDP supported MTU %d.\n",
+	if (prog && !prog->aux->xdp_has_frags &&
+	    bp->dev->mtu > BNXT_MAX_PAGE_MODE_MTU) {
+		netdev_warn(dev, "MTU %d larger than %d without XDP frag support.\n",
 			    bp->dev->mtu, BNXT_MAX_PAGE_MODE_MTU);
 		return -EOPNOTSUPP;
 	}
@@ -337,3 +455,26 @@ int bnxt_xdp(struct net_device *dev, struct netdev_bpf *xdp)
 	}
 	return rc;
 }
+
+struct sk_buff *
+bnxt_xdp_build_skb(struct bnxt *bp, struct sk_buff *skb, u8 num_frags,
+		   struct page_pool *pool, struct xdp_buff *xdp,
+		   struct rx_cmp_ext *rxcmp1)
+{
+	struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp);
+
+	if (!skb)
+		return NULL;
+	skb_checksum_none_assert(skb);
+	if (RX_CMP_L4_CS_OK(rxcmp1)) {
+		if (bp->dev->features & NETIF_F_RXCSUM) {
+			skb->ip_summed = CHECKSUM_UNNECESSARY;
+			skb->csum_level = RX_CMP_ENCAP(rxcmp1);
+		}
+	}
+	xdp_update_skb_shared_info(skb, num_frags,
+				   sinfo->xdp_frags_size,
+				   PAGE_SIZE * sinfo->nr_frags,
+				   xdp_buff_is_frag_pfmemalloc(xdp));
+	return skb;
+}

drivers/net/ethernet/broadcom/bnxt/bnxt_xdp.h

@@ -14,13 +14,25 @@ DECLARE_STATIC_KEY_FALSE(bnxt_xdp_locking_key);
 
 struct bnxt_sw_tx_bd *bnxt_xmit_bd(struct bnxt *bp,
 				   struct bnxt_tx_ring_info *txr,
-				   dma_addr_t mapping, u32 len);
+				   dma_addr_t mapping, u32 len,
+				   struct xdp_buff *xdp);
 void bnxt_tx_int_xdp(struct bnxt *bp, struct bnxt_napi *bnapi, int nr_pkts);
 bool bnxt_rx_xdp(struct bnxt *bp, struct bnxt_rx_ring_info *rxr, u16 cons,
-		 struct page *page, u8 **data_ptr, unsigned int *len,
+		 struct xdp_buff xdp, struct page *page, unsigned int *len,
 		 u8 *event);
 int bnxt_xdp(struct net_device *dev, struct netdev_bpf *xdp);
 int bnxt_xdp_xmit(struct net_device *dev, int num_frames,
 		  struct xdp_frame **frames, u32 flags);
 
+bool bnxt_xdp_attached(struct bnxt *bp, struct bnxt_rx_ring_info *rxr);
+
+void bnxt_xdp_buff_init(struct bnxt *bp, struct bnxt_rx_ring_info *rxr,
+			u16 cons, u8 **data_ptr, unsigned int *len,
+			struct xdp_buff *xdp);
+void bnxt_xdp_buff_frags_free(struct bnxt_rx_ring_info *rxr,
+			      struct xdp_buff *xdp);
+struct sk_buff *bnxt_xdp_build_skb(struct bnxt *bp, struct sk_buff *skb,
+				   u8 num_frags, struct page_pool *pool,
+				   struct xdp_buff *xdp,
+				   struct rx_cmp_ext *rxcmp1);
 #endif