[PATCH net-next] ixgbe: add xdp frags support to ndo_xdp_xmit

[PATCH net-next] ixgbe: add xdp frags support to ndo_xdp_xmit

[Lorenzo Bianconi]

Add the capability to map non-linear xdp frames in XDP_TX and ndo_xdp_xmit
callback.

Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
---
 drivers/net/ethernet/intel/ixgbe/ixgbe_main.c | 99 ++++++++++++-------
 1 file changed, 63 insertions(+), 36 deletions(-)

diff --git a/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c b/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
index c4a4954aa317..8b84c9b2eecc 100644
--- a/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
+++ b/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
@@ -2344,6 +2344,7 @@ static int ixgbe_clean_rx_irq(struct ixgbe_q_vector *q_vector,
 			hard_start = page_address(rx_buffer->page) +
 				     rx_buffer->page_offset - offset;
 			xdp_prepare_buff(&xdp, hard_start, offset, size, true);
+			xdp_buff_clear_frags_flag(&xdp);
 #if (PAGE_SIZE > 4096)
 			/* At larger PAGE_SIZE, frame_sz depend on len size */
 			xdp.frame_sz = ixgbe_rx_frame_truesize(rx_ring, size);
@@ -8571,57 +8572,83 @@ static u16 ixgbe_select_queue(struct net_device *dev, struct sk_buff *skb,
 int ixgbe_xmit_xdp_ring(struct ixgbe_ring *ring,
 			struct xdp_frame *xdpf)
 {
-	struct ixgbe_tx_buffer *tx_buffer;
-	union ixgbe_adv_tx_desc *tx_desc;
-	u32 len, cmd_type;
-	dma_addr_t dma;
-	u16 i;
-
-	len = xdpf->len;
+	struct skb_shared_info *sinfo = xdp_get_shared_info_from_frame(xdpf);
+	u8 nr_frags = unlikely(xdp_frame_has_frags(xdpf)) ? sinfo->nr_frags : 0;
+	u16 i = 0, index = ring->next_to_use;
+	struct ixgbe_tx_buffer *tx_head = &ring->tx_buffer_info[index];
+	struct ixgbe_tx_buffer *tx_buff = tx_head;
+	union ixgbe_adv_tx_desc *tx_desc = IXGBE_TX_DESC(ring, index);
+	u32 cmd_type, len = xdpf->len;
+	void *data = xdpf->data;
 
-	if (unlikely(!ixgbe_desc_unused(ring)))
+	if (unlikely(ixgbe_desc_unused(ring) < 1 + nr_frags))
 		return IXGBE_XDP_CONSUMED;
 
-	dma = dma_map_single(ring->dev, xdpf->data, len, DMA_TO_DEVICE);
-	if (dma_mapping_error(ring->dev, dma))
-		return IXGBE_XDP_CONSUMED;
+	tx_head->bytecount = xdp_get_frame_len(xdpf);
+	tx_head->gso_segs = 1;
+	tx_head->xdpf = xdpf;
 
-	/* record the location of the first descriptor for this packet */
-	tx_buffer = &ring->tx_buffer_info[ring->next_to_use];
-	tx_buffer->bytecount = len;
-	tx_buffer->gso_segs = 1;
-	tx_buffer->protocol = 0;
+	tx_desc->read.olinfo_status =
+		cpu_to_le32(tx_head->bytecount << IXGBE_ADVTXD_PAYLEN_SHIFT);
 
-	i = ring->next_to_use;
-	tx_desc = IXGBE_TX_DESC(ring, i);
+	for (;;) {
+		dma_addr_t dma;
 
-	dma_unmap_len_set(tx_buffer, len, len);
-	dma_unmap_addr_set(tx_buffer, dma, dma);
-	tx_buffer->xdpf = xdpf;
+		dma = dma_map_single(ring->dev, data, len, DMA_TO_DEVICE);
+		if (dma_mapping_error(ring->dev, dma))
+			goto unmap;
 
-	tx_desc->read.buffer_addr = cpu_to_le64(dma);
+		dma_unmap_len_set(tx_buff, len, len);
+		dma_unmap_addr_set(tx_buff, dma, dma);
+
+		cmd_type = IXGBE_ADVTXD_DTYP_DATA | IXGBE_ADVTXD_DCMD_DEXT |
+			   IXGBE_ADVTXD_DCMD_IFCS | len;
+		tx_desc->read.cmd_type_len = cpu_to_le32(cmd_type);
+		tx_desc->read.buffer_addr = cpu_to_le64(dma);
+		tx_buff->protocol = 0;
+
+		if (++index == ring->count)
+			index = 0;
+
+		if (i == nr_frags)
+			break;
+
+		tx_buff = &ring->tx_buffer_info[index];
+		tx_desc = IXGBE_TX_DESC(ring, index);
+		tx_desc->read.olinfo_status = 0;
 
+		data = skb_frag_address(&sinfo->frags[i]);
+		len = skb_frag_size(&sinfo->frags[i]);
+		i++;
+	}
 	/* put descriptor type bits */
-	cmd_type = IXGBE_ADVTXD_DTYP_DATA |
-		   IXGBE_ADVTXD_DCMD_DEXT |
-		   IXGBE_ADVTXD_DCMD_IFCS;
-	cmd_type |= len | IXGBE_TXD_CMD;
-	tx_desc->read.cmd_type_len = cpu_to_le32(cmd_type);
-	tx_desc->read.olinfo_status =
-		cpu_to_le32(len << IXGBE_ADVTXD_PAYLEN_SHIFT);
+	tx_desc->read.cmd_type_len |= cpu_to_le32(IXGBE_TXD_CMD);
 
 	/* Avoid any potential race with xdp_xmit and cleanup */
 	smp_wmb();
 
-	/* set next_to_watch value indicating a packet is present */
-	i++;
-	if (i == ring->count)
-		i = 0;
-
-	tx_buffer->next_to_watch = tx_desc;
-	ring->next_to_use = i;
+	tx_head->next_to_watch = tx_desc;
+	ring->next_to_use = index;
 
 	return IXGBE_XDP_TX;
+
+unmap:
+	for (;;) {
+		tx_buff = &ring->tx_buffer_info[index];
+		if (dma_unmap_len(tx_buff, len))
+			dma_unmap_page(ring->dev, dma_unmap_addr(tx_buff, dma),
+				       dma_unmap_len(tx_buff, len),
+				       DMA_TO_DEVICE);
+		dma_unmap_len_set(tx_buff, len, 0);
+		if (tx_buff == tx_head)
+			break;
+
+		if (!index)
+			index += ring->count;
+		index--;
+	}
+
+	return IXGBE_XDP_CONSUMED;
 }
 
 netdev_tx_t ixgbe_xmit_frame_ring(struct sk_buff *skb,
-- 
2.35.1

Re: [PATCH net-next] ixgbe: add xdp frags support to ndo_xdp_xmit

[Lorenzo Bianconi]

[-- Attachment #1: Type: text/plain, Size: 5034 bytes --]

> Add the capability to map non-linear xdp frames in XDP_TX and ndo_xdp_xmit
> callback.

Hi Tony,

do you have any feedbacks about this patch?

Regards,
Lorenzo

> 
> Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
> ---
>  drivers/net/ethernet/intel/ixgbe/ixgbe_main.c | 99 ++++++++++++-------
>  1 file changed, 63 insertions(+), 36 deletions(-)
> 
> diff --git a/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c b/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
> index c4a4954aa317..8b84c9b2eecc 100644
> --- a/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
> +++ b/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
> @@ -2344,6 +2344,7 @@ static int ixgbe_clean_rx_irq(struct ixgbe_q_vector *q_vector,
>  			hard_start = page_address(rx_buffer->page) +
>  				     rx_buffer->page_offset - offset;
>  			xdp_prepare_buff(&xdp, hard_start, offset, size, true);
> +			xdp_buff_clear_frags_flag(&xdp);
>  #if (PAGE_SIZE > 4096)
>  			/* At larger PAGE_SIZE, frame_sz depend on len size */
>  			xdp.frame_sz = ixgbe_rx_frame_truesize(rx_ring, size);
> @@ -8571,57 +8572,83 @@ static u16 ixgbe_select_queue(struct net_device *dev, struct sk_buff *skb,
>  int ixgbe_xmit_xdp_ring(struct ixgbe_ring *ring,
>  			struct xdp_frame *xdpf)
>  {
> -	struct ixgbe_tx_buffer *tx_buffer;
> -	union ixgbe_adv_tx_desc *tx_desc;
> -	u32 len, cmd_type;
> -	dma_addr_t dma;
> -	u16 i;
> -
> -	len = xdpf->len;
> +	struct skb_shared_info *sinfo = xdp_get_shared_info_from_frame(xdpf);
> +	u8 nr_frags = unlikely(xdp_frame_has_frags(xdpf)) ? sinfo->nr_frags : 0;
> +	u16 i = 0, index = ring->next_to_use;
> +	struct ixgbe_tx_buffer *tx_head = &ring->tx_buffer_info[index];
> +	struct ixgbe_tx_buffer *tx_buff = tx_head;
> +	union ixgbe_adv_tx_desc *tx_desc = IXGBE_TX_DESC(ring, index);
> +	u32 cmd_type, len = xdpf->len;
> +	void *data = xdpf->data;
>  
> -	if (unlikely(!ixgbe_desc_unused(ring)))
> +	if (unlikely(ixgbe_desc_unused(ring) < 1 + nr_frags))
>  		return IXGBE_XDP_CONSUMED;
>  
> -	dma = dma_map_single(ring->dev, xdpf->data, len, DMA_TO_DEVICE);
> -	if (dma_mapping_error(ring->dev, dma))
> -		return IXGBE_XDP_CONSUMED;
> +	tx_head->bytecount = xdp_get_frame_len(xdpf);
> +	tx_head->gso_segs = 1;
> +	tx_head->xdpf = xdpf;
>  
> -	/* record the location of the first descriptor for this packet */
> -	tx_buffer = &ring->tx_buffer_info[ring->next_to_use];
> -	tx_buffer->bytecount = len;
> -	tx_buffer->gso_segs = 1;
> -	tx_buffer->protocol = 0;
> +	tx_desc->read.olinfo_status =
> +		cpu_to_le32(tx_head->bytecount << IXGBE_ADVTXD_PAYLEN_SHIFT);
>  
> -	i = ring->next_to_use;
> -	tx_desc = IXGBE_TX_DESC(ring, i);
> +	for (;;) {
> +		dma_addr_t dma;
>  
> -	dma_unmap_len_set(tx_buffer, len, len);
> -	dma_unmap_addr_set(tx_buffer, dma, dma);
> -	tx_buffer->xdpf = xdpf;
> +		dma = dma_map_single(ring->dev, data, len, DMA_TO_DEVICE);
> +		if (dma_mapping_error(ring->dev, dma))
> +			goto unmap;
>  
> -	tx_desc->read.buffer_addr = cpu_to_le64(dma);
> +		dma_unmap_len_set(tx_buff, len, len);
> +		dma_unmap_addr_set(tx_buff, dma, dma);
> +
> +		cmd_type = IXGBE_ADVTXD_DTYP_DATA | IXGBE_ADVTXD_DCMD_DEXT |
> +			   IXGBE_ADVTXD_DCMD_IFCS | len;
> +		tx_desc->read.cmd_type_len = cpu_to_le32(cmd_type);
> +		tx_desc->read.buffer_addr = cpu_to_le64(dma);
> +		tx_buff->protocol = 0;
> +
> +		if (++index == ring->count)
> +			index = 0;
> +
> +		if (i == nr_frags)
> +			break;
> +
> +		tx_buff = &ring->tx_buffer_info[index];
> +		tx_desc = IXGBE_TX_DESC(ring, index);
> +		tx_desc->read.olinfo_status = 0;
>  
> +		data = skb_frag_address(&sinfo->frags[i]);
> +		len = skb_frag_size(&sinfo->frags[i]);
> +		i++;
> +	}
>  	/* put descriptor type bits */
> -	cmd_type = IXGBE_ADVTXD_DTYP_DATA |
> -		   IXGBE_ADVTXD_DCMD_DEXT |
> -		   IXGBE_ADVTXD_DCMD_IFCS;
> -	cmd_type |= len | IXGBE_TXD_CMD;
> -	tx_desc->read.cmd_type_len = cpu_to_le32(cmd_type);
> -	tx_desc->read.olinfo_status =
> -		cpu_to_le32(len << IXGBE_ADVTXD_PAYLEN_SHIFT);
> +	tx_desc->read.cmd_type_len |= cpu_to_le32(IXGBE_TXD_CMD);
>  
>  	/* Avoid any potential race with xdp_xmit and cleanup */
>  	smp_wmb();
>  
> -	/* set next_to_watch value indicating a packet is present */
> -	i++;
> -	if (i == ring->count)
> -		i = 0;
> -
> -	tx_buffer->next_to_watch = tx_desc;
> -	ring->next_to_use = i;
> +	tx_head->next_to_watch = tx_desc;
> +	ring->next_to_use = index;
>  
>  	return IXGBE_XDP_TX;
> +
> +unmap:
> +	for (;;) {
> +		tx_buff = &ring->tx_buffer_info[index];
> +		if (dma_unmap_len(tx_buff, len))
> +			dma_unmap_page(ring->dev, dma_unmap_addr(tx_buff, dma),
> +				       dma_unmap_len(tx_buff, len),
> +				       DMA_TO_DEVICE);
> +		dma_unmap_len_set(tx_buff, len, 0);
> +		if (tx_buff == tx_head)
> +			break;
> +
> +		if (!index)
> +			index += ring->count;
> +		index--;
> +	}
> +
> +	return IXGBE_XDP_CONSUMED;
>  }
>  
>  netdev_tx_t ixgbe_xmit_frame_ring(struct sk_buff *skb,
> -- 
> 2.35.1
> 

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 228 bytes --]