On Wed, Nov 04, 2026 at 07:40:05AM +0100, Tjerk Kusters via B4 Relay wrote:
> From: Tjerk Kusters <[email protected]>
>
> igb_clean_rx_irq_zc() handles every descriptor as a complete frame and
> strips the inline Rx timestamp header from each of them. If a frame
> spans several descriptors, the AF_XDP socket gets its later buffers as
> separate packets without headers. If such a buffer is shorter than the
> header, the size wraps and the XDP program sees data past data_end. On
> XDP_PASS, igb_construct_skb_zc() then fails to allocate an skb, and the
> queue can stall on that descriptor.
>
> The previous patches make sure that frames fit in one buffer while XDP
> is attached. As ixgbe does, also drop any frame that still spans several
> descriptors instead of passing its buffers on: free the buffers without
> EOP, and the EOP buffer that ends such a frame. The state is kept in a
> ring flag, as a frame can be split over two NAPI polls.
>
> This depends on the previous patch. Without it, every full-size frame
> on 2 KiB chunks spans two buffers and would be dropped.
>
> Not reproduced on hardware, as with the previous patches the I350 no
> longer splits frames in zero-copy mode.
I am missing the point of this patch then. Are you adding a dead code to
hot path?
ixgbe was probably problematic to address all supported hw (i think 82599
has no per-q registers and would need different treatment on config side?)
that's why it was chosen to carry the discard logic.
But here if you specifically are addressing the config side of things on
igb I am having a hard time to understand the purpose of this patch.
>
> Fixes: 2c6196013f84 ("igb: Add AF_XDP zero-copy Rx support")
> Cc: [email protected]
> Signed-off-by: Tjerk Kusters <[email protected]>
> ---
> drivers/net/ethernet/intel/igb/igb.h | 1 +
> drivers/net/ethernet/intel/igb/igb_xsk.c | 23 +++++++++++++++++++++++
> 2 files changed, 24 insertions(+)
>
> diff --git a/drivers/net/ethernet/intel/igb/igb.h
> b/drivers/net/ethernet/intel/igb/igb.h
> index 0495c02544d6..6613ea0eedb8 100644
> --- a/drivers/net/ethernet/intel/igb/igb.h
> +++ b/drivers/net/ethernet/intel/igb/igb.h
> @@ -393,6 +393,7 @@ enum e1000_ring_flags_t {
> IGB_RING_FLAG_TX_DETECT_HANG,
> IGB_RING_FLAG_TX_DISABLED,
> IGB_RING_FLAG_RX_ALLOC_FAILED,
> + IGB_RING_FLAG_RX_XSK_DISCARD,
> };
>
> #define ring_uses_large_buffer(ring) \
> diff --git a/drivers/net/ethernet/intel/igb/igb_xsk.c
> b/drivers/net/ethernet/intel/igb/igb_xsk.c
> index fb04df82272a..7b098069fff2 100644
> --- a/drivers/net/ethernet/intel/igb/igb_xsk.c
> +++ b/drivers/net/ethernet/intel/igb/igb_xsk.c
> @@ -269,6 +269,8 @@ void igb_clean_rx_ring_zc(struct igb_ring *rx_ring)
> u16 ntc = rx_ring->next_to_clean;
> u16 ntu = rx_ring->next_to_use;
>
> + clear_bit(IGB_RING_FLAG_RX_XSK_DISCARD, &rx_ring->flags);
> +
> while (ntc != ntu) {
> struct xdp_buff *xdp = rx_ring->rx_buffer_info_zc[ntc];
>
> @@ -390,6 +392,27 @@ int igb_clean_rx_irq_zc(struct igb_q_vector *q_vector,
> dma_rmb();
>
> xdp = rx_ring->rx_buffer_info_zc[ntc];
> +
> + /* A frame should never span several buffers here, see
> + * igb_xsk_frame_fits(). If one does, drop all its buffers.
> + */
> + if (unlikely(!igb_test_staterr(rx_desc, E1000_RXD_STAT_EOP) ||
> + test_bit(IGB_RING_FLAG_RX_XSK_DISCARD,
> + &rx_ring->flags))) {
> + if (igb_test_staterr(rx_desc, E1000_RXD_STAT_EOP))
> + clear_bit(IGB_RING_FLAG_RX_XSK_DISCARD,
> + &rx_ring->flags);
> + else
> + set_bit(IGB_RING_FLAG_RX_XSK_DISCARD,
> + &rx_ring->flags);
> +
> + xsk_buff_free(xdp);
> + ntc++;
> + if (ntc == rx_ring->count)
> + ntc = 0;
> + continue;
> + }
> +
> xsk_buff_set_size(xdp, size);
> xsk_buff_dma_sync_for_cpu(xdp);
>
>
> --
> 2.27.0
>
>
