Re: [RFC ixgbe 1/2] ixgbe: Implement support for ndo_xdp_xmit in skb mode

[Maciej Fijalkowski <maciej.fijalkowski@intel.com>]

On Mon, Nov 03, 2025 at 12:40:14PM -0500, Nabil S. Alramli wrote:
> On 11/3/25 11:38, Maciej Fijalkowski wrote:
> > On Thu, Oct 09, 2025 at 03:28:30PM -0400, Nabil S. Alramli wrote:
> >> This commit adds support for `ndo_xdp_xmit` in skb mode in the ixgbe
> >> ethernet driver, by allowing the call to continue to transmit the packets
> >> using `dev_direct_xmit`.
> >>
> >> Previously, the driver did not support the operation in skb mode. The
> >> handler `ixgbe_xdp_xmit` had the following condition:
> >>
> >> ```
> >> 	ring = adapter->xdp_prog ? ixgbe_determine_xdp_ring(adapter) : NULL;
> >> 	if (unlikely(!ring))
> >> 		return -ENXIO;
> >> ```
> >>
> >> That only works in native mode. In skb mode, `adapter->xdp_prog == NULL` so
> >> the call returned an error, which prevented the ability to send packets
> >> using `bpf_prog_test_run_opts` with the `BPF_F_TEST_XDP_LIVE_FRAMES` flag.
> > 
> > Hi Nabil,
> > 
> > What stops you from loading a dummy XDP program to interface? This has
> > been an approach that we follow when we want to use anything that utilizes
> > XDP resources (XDP Tx queues).
> > 
> 
> Hi Maciej,
> 
> Thank you for your response. In one use case we have multiple XDP programs
> already loaded on an interface in SKB mode using the dispatcher, and we want
> to use bpf_prog_test_run_opts to egress packets from another XDP program. We
> want to avoid having to unload the dispatcher or be forced to use it in native
> mode. Without this patch, that does not seem possible currently, correct?

Why does it have to be bpf_prog_test_run_opts?
You're trying to use an interface which was designed for native XDP from a
different layer. Generic XDP has support for redirect and tx.

> 
> >>
> >> Signed-off-by: Nabil S. Alramli <dev@nalramli.com>
> >> ---
> >>  drivers/net/ethernet/intel/ixgbe/ixgbe.h      |  8 ++++
> >>  drivers/net/ethernet/intel/ixgbe/ixgbe_main.c | 43 +++++++++++++++++--
> >>  2 files changed, 47 insertions(+), 4 deletions(-)
> >>
> >> diff --git a/drivers/net/ethernet/intel/ixgbe/ixgbe.h b/drivers/net/ethernet/intel/ixgbe/ixgbe.h
> >> index e6a380d4929b..26c378853755 100644
> >> --- a/drivers/net/ethernet/intel/ixgbe/ixgbe.h
> >> +++ b/drivers/net/ethernet/intel/ixgbe/ixgbe.h
> >> @@ -846,6 +846,14 @@ struct ixgbe_ring *ixgbe_determine_xdp_ring(struct ixgbe_adapter *adapter)
> >>  	return adapter->xdp_ring[index];
> >>  }
> >>  
> >> +static inline
> >> +struct ixgbe_ring *ixgbe_determine_tx_ring(struct ixgbe_adapter *adapter)
> >> +{
> >> +	int index = ixgbe_determine_xdp_q_idx(smp_processor_id());
> >> +
> >> +	return adapter->tx_ring[index];
> >> +}
> >> +
> >>  static inline u8 ixgbe_max_rss_indices(struct ixgbe_adapter *adapter)
> >>  {
> >>  	switch (adapter->hw.mac.type) {
> >> diff --git a/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c b/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
> >> index 467f81239e12..fed70cbdb1b2 100644
> >> --- a/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
> >> +++ b/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
> >> @@ -10748,7 +10748,8 @@ static int ixgbe_xdp_xmit(struct net_device *dev, int n,
> >>  	/* During program transitions its possible adapter->xdp_prog is assigned
> >>  	 * but ring has not been configured yet. In this case simply abort xmit.
> >>  	 */
> >> -	ring = adapter->xdp_prog ? ixgbe_determine_xdp_ring(adapter) : NULL;
> >> +	ring = adapter->xdp_prog ? ixgbe_determine_xdp_ring(adapter) :
> >> +		ixgbe_determine_tx_ring(adapter);
> >>  	if (unlikely(!ring))
> >>  		return -ENXIO;
> >>  
> >> @@ -10762,9 +10763,43 @@ static int ixgbe_xdp_xmit(struct net_device *dev, int n,
> >>  		struct xdp_frame *xdpf = frames[i];
> >>  		int err;
> >>  
> >> -		err = ixgbe_xmit_xdp_ring(ring, xdpf);
> >> -		if (err != IXGBE_XDP_TX)
> >> -			break;
> >> +		if (adapter->xdp_prog) {
> >> +			err = ixgbe_xmit_xdp_ring(ring, xdpf);
> >> +			if (err != IXGBE_XDP_TX)
> >> +				break;
> >> +		} else {
> >> +			struct xdp_buff xdp = {0};
> >> +			unsigned int metasize = 0;
> >> +			unsigned int size = 0;
> >> +			unsigned int truesize = 0;
> >> +			struct sk_buff *skb = NULL;
> >> +
> >> +			xdp_convert_frame_to_buff(xdpf, &xdp);
> >> +			size = xdp.data_end - xdp.data;
> >> +			metasize = xdp.data - xdp.data_meta;
> >> +			truesize = SKB_DATA_ALIGN(xdp.data_end - xdp.data_hard_start) +
> >> +				   SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
> >> +
> >> +			skb = napi_alloc_skb(&ring->q_vector->napi, truesize);
> >> +			if (likely(skb)) {
> >> +				skb_reserve(skb, xdp.data - xdp.data_hard_start);
> >> +				skb_put_data(skb, xdp.data, size);
> >> +				build_skb_around(skb, skb->data, truesize);
> >> +				if (metasize)
> >> +					skb_metadata_set(skb, metasize);
> >> +				skb->dev = dev;
> >> +				skb->queue_mapping = ring->queue_index;
> >> +
> >> +				err = dev_direct_xmit(skb, ring->queue_index);
> >> +				if (!dev_xmit_complete(err))
> >> +					break;
> >> +			} else {
> >> +				break;
> >> +			}
> >> +
> >> +			xdp_return_frame_rx_napi(xdpf);
> >> +		}
> >> +
> >>  		nxmit++;
> >>  	}
> >>  
> >> -- 
> >> 2.43.0
> >>
> >>
>