Re: [PATCH 1/3] net: lpc_eth: Replace WARN() trace with simple pr_warn()

On Mon, 2012-06-11 at 10:36 +0200, Roland Stigge wrote:

I encountered cases where this happened for me on a custom board under
heavy load.

I discussed this with Kevin Wells, the original driver author. We
identified the case of xmit()'s TX request (from .ndo_start_xmit) with
full TX driver buffers as valid when ethernet is busy.

But maybe this is wrong. Can you please give me a hint how the net
subsystem makes sure that this doesn't happen under normal circumstances?

When TX ring is about to be filler, driver lpc_eth_hard_start_xmit()
calls netif_stop_queue(ndev);

So network stack should not call again lpc_eth_hard_start_xmit().

I would say the bug(s) come from __lpc_handle_xmit(), since it does :

if (netif_queue_stopped(ndev))
	netif_wake_queue(ndev);

without making sure some room is available in TX ring.

cumulative patch :

diff --git a/drivers/net/ethernet/nxp/lpc_eth.c b/drivers/net/ethernet/nxp/lpc_eth.c
index 8d2666f..59b37c8 100644
--- a/drivers/net/ethernet/nxp/lpc_eth.c
+++ b/drivers/net/ethernet/nxp/lpc_eth.c

@@ -946,16 +946,16 @@ static void __lpc_handle_xmit(struct net_device *ndev)
 			/* Update stats */
 			ndev->stats.tx_packets++;
 			ndev->stats.tx_bytes += skb->len;
-
-			/* Free buffer */
-			dev_kfree_skb_irq(skb);
 		}
+		dev_kfree_skb_irq(skb);
 
 		txcidx = readl(LPC_ENET_TXCONSUMEINDEX(pldat->net_base));
 	}
 
-	if (netif_queue_stopped(ndev))
-		netif_wake_queue(ndev);
+	if (pldat->num_used_tx_buffs <= ENET_TX_DESC/2) { 
+		if (netif_queue_stopped(ndev))
+			netif_wake_queue(ndev);
+	}
 }
 
 static int __lpc_handle_recv(struct net_device *ndev, int budget)