On Thu, 2026-10-15 at 11:37 +0800, Jay Chang wrote: > In the Xilinx AXI Ethernet device model, U-Boot network commands > (ping, dhcp, tftp) all fail due to a packet duplication bug. > > The original implementation had a packet duplication issue where the > same > packet could exist in both the internal buffer and the incoming > queue: > > Problem Flow: > 1. eth_rx() receives a packet from the network layer. > 2. Packet is processed and stored in s->rxmem buffer. > 3. axienet_eth_rx_notify() is called to push data to DMA. > 4. If DMA cannot accept data immediately, the rxpos is 0 and wth_rx > will return 0 > 5. Because eth_rx callback returns 0, the network layer queues this > packet again in the incoming queue. > 6. Now the same packet exists in TWO places: > - Already copied in s->rxmem buffer > - Queued in the incoming packet queue > 7. This causes duplicate packets to be delivered to the guest > 8. Network protocols fail due to receiving identical packets twice > > Root Cause: > eth_rx() did not check if the DMA stream could accept data before > copying the packet into the internal buffer. This caused unnecessary > buffering and potential double-processing. > > Fix: > 1. Check stream_can_push() at the beginning of eth_rx() > - If DMA cannot accept data, set need_flush and return 0 > immediately > - This prevents packet buffering when DMA is not ready > - The packet will be redelivered by QEMU's network layer > > 2. Move need_flush check outside the while loop in > axienet_eth_rx_notify() > - Ensures flush happens even when rxsize == 0 > - Handles edge cases where the buffer is already empty but flush > is needed > > Signed-off-by: Jay Chang <[email protected]> > Reviewed-by: Frank Chang <[email protected]> > Reviewed-by: Jim Shu <[email protected]>
Acked-by: Alistair Francis <[email protected]> Alistair > --- > hw/net/xilinx_axienet.c | 12 ++++++++++++ > 1 file changed, 12 insertions(+) > > diff --git a/hw/net/xilinx_axienet.c b/hw/net/xilinx_axienet.c > index d35f4a847d..cb9aa47707 100644 > --- a/hw/net/xilinx_axienet.c > +++ b/hw/net/xilinx_axienet.c > @@ -699,6 +699,12 @@ static void axienet_eth_rx_notify(void *opaque) > } > } > } > + > + /* Check need_flush outside the loop for an empty RX buffer. */ > + if (!s->rxsize && s->need_flush) { > + s->need_flush = false; > + qemu_flush_queued_packets(qemu_get_queue(s->nic)); > + } > enet_update_irq(s); > } > > @@ -722,6 +728,12 @@ static ssize_t eth_rx(NetClientState *nc, const > uint8_t *buf, size_t size) > return 0; > } > > + /* Check if DMA can accept data before processing packet */ > + if (!stream_can_push(s->tx_data_dev, axienet_eth_rx_notify, s)) > { > + s->need_flush = true; > + return 0; > + } > + > unicast = ~buf[0] & 0x1; > broadcast = memcmp(buf, sa_bcast, 6) == 0; > multicast = !unicast && !broadcast;
