On Thu, 2010-03-11 at 11:40 -0800, Brian Haley wrote:

I can only reproduce this on one system out of many, so it's either a
race condition or bad hardware.  The only thing I can confirm at the
moment is that it's the code at the bottom of bnx2_set_coalesce()
that's causing it, I'm trying to go through all those codepaths now.

The NETDEV WATCHDOG is caused by stopping the TX queues with
->trans_start older than dev->watchdog_timeo which is set to 5 seconds
in bnx2.  Please try this patch below to update the ->trans_start first
before stopping the TX queues:

diff --git a/drivers/net/bnx2.c b/drivers/net/bnx2.c
index d3f739a..c0f4aa7 100644
--- a/drivers/net/bnx2.c
+++ b/drivers/net/bnx2.c

@@ -656,7 +656,6 @@ bnx2_netif_stop(struct bnx2 *bp)
 		int i;
 
 		bnx2_napi_disable(bp);
-		netif_tx_disable(bp->dev);
 		/* prevent tx timeout */
 		for (i = 0; i <  bp->dev->num_tx_queues; i++) {
 			struct netdev_queue *txq;

@@ -664,6 +663,7 @@ bnx2_netif_stop(struct bnx2 *bp)
 			txq = netdev_get_tx_queue(bp->dev, i);
 			txq->trans_start = jiffies;
 		}
+		netif_tx_disable(bp->dev);
 	}
 	bnx2_disable_int_sync(bp);
 }