Re: [PATCH net-next 4/6] net: bcmgenet: add XDP_TX support

[Nicolai Buchwitz <nb@tipi-net.de>]

On 13.3.2026 12:37, Subbaraya Sundeep wrote:
> Hi,

Hi Sundeep

> 
> On 2026-03-13 at 14:50:59, Nicolai Buchwitz (nb@tipi-net.de) wrote:
>> Implement XDP_TX by submitting XDP frames through the default TX ring
>> (DESC_INDEX). The frame is DMA-mapped and placed into a single TX
>> descriptor with SOP|EOP|APPEND_CRC flags.
>> 
>> The xdp_frame pointer is stored in the TX control block so that
>> bcmgenet_free_tx_cb() can call xdp_return_frame() on TX completion,
>> returning the page to the originating page_pool.
>> 
>> The page_pool DMA direction is changed from DMA_FROM_DEVICE to
>> DMA_BIDIRECTIONAL to support the TX DMA mapping of received pages.
>> 
>> Signed-off-by: Nicolai Buchwitz <nb@tipi-net.de>
>> ---
>>  .../net/ethernet/broadcom/genet/bcmgenet.c    | 73 
>> ++++++++++++++++++-
>>  .../net/ethernet/broadcom/genet/bcmgenet.h    |  1 +
>>  2 files changed, 71 insertions(+), 3 deletions(-)
>> 
>> diff --git a/drivers/net/ethernet/broadcom/genet/bcmgenet.c 
>> b/drivers/net/ethernet/broadcom/genet/bcmgenet.c
>> index d43729fc2b1b..373ba5878ca1 100644
>> --- a/drivers/net/ethernet/broadcom/genet/bcmgenet.c
>> +++ b/drivers/net/ethernet/broadcom/genet/bcmgenet.c
>> @@ -1893,6 +1893,12 @@ static struct sk_buff 
>> *bcmgenet_free_tx_cb(struct device *dev,
>>  		if (cb == GENET_CB(skb)->last_cb)
>>  			return skb;
>> 
>> +	} else if (cb->xdpf) {
>> +		dma_unmap_single(dev, dma_unmap_addr(cb, dma_addr),
>> +				 dma_unmap_len(cb, dma_len), DMA_TO_DEVICE);
>> +		dma_unmap_addr_set(cb, dma_addr, 0);
>> +		xdp_return_frame(cb->xdpf);
>> +		cb->xdpf = NULL;
>>  	} else if (dma_unmap_addr(cb, dma_addr)) {
>>  		dma_unmap_page(dev,
>>  			       dma_unmap_addr(cb, dma_addr),
>> @@ -2299,10 +2305,62 @@ static struct sk_buff 
>> *bcmgenet_xdp_build_skb(struct bcmgenet_rx_ring *ring,
>>  	return skb;
>>  }
>> 
>> +static bool bcmgenet_xdp_xmit_frame(struct bcmgenet_priv *priv,
>> +				     struct xdp_frame *xdpf)
>> +{
>> +	struct bcmgenet_tx_ring *ring = &priv->tx_rings[DESC_INDEX];
>> +	struct device *kdev = &priv->pdev->dev;
>> +	struct enet_cb *tx_cb_ptr;
>> +	dma_addr_t mapping;
>> +	u32 len_stat;
>> +
>> +	spin_lock(&ring->lock);
>> +
>> +	if (ring->free_bds < 1) {
>> +		spin_unlock(&ring->lock);
>> +		return false;
>> +	}
>> +
>> +	tx_cb_ptr = bcmgenet_get_txcb(priv, ring);
>> +
>> +	mapping = dma_map_single(kdev, xdpf->data, xdpf->len, 
>> DMA_TO_DEVICE);
> 
> AFAIU you are transmitting the frame received on a RQ which is from the 
> page pool
> and already dma mapped. Do you have to do dma_map again?
> 
> Thanks,
> Sundeep
> 

You're right. Since the page_pool is configured with DMA_BIDIRECTIONAL,
the pages are already mapped and we can reuse the existing mapping for
XDP_TX frames. The initial implementation took the simple route of
mapping everything uniformly, but that's unnecessary overhead for the
local XDP_TX case.

In v2 I'll add a bool dma_map parameter to bcmgenet_xdp_xmit_frame()
(following the mvneta/stmmac pattern): XDP_TX will reuse the page_pool
mapping via page_pool_get_dma_addr() + dma_sync_single_for_device(),
while ndo_xdp_xmit will keep dma_map_single() for foreign frames. The
cleanup path will be split accordingly.

Regards
Nicolai

>> [...]