On Mon, Aug 23, 2010 at 02:47:23PM +0300, Plamen Petrov wrote:

quoted

???? 21.8.2010 ??. 11:07, Jarek Poplawski ????????????:

quoted

On Sat, Aug 21, 2010 at 09:50:58AM +0200, Eric Dumazet wrote:

quoted

Le samedi 21 ao??t 2010 ?? 09:47 +0200, Jarek Poplawski a écrit :

quoted

On Fri, Aug 20, 2010 at 09:38:35PM +0200, Jarek Poplawski wrote:

quoted

Plamen Petrov wrote, On 20.08.2010 12:53:

quoted

So, I guess its David and Herbert's turn?...

If you're bored in the meantime I'd suggest to do check the realtek
driver eg:
- for locking with the patch below,
- to turn off with ethtool its tx-checksumming and/or scatter-gather,

...

quoted

Yeah, 3 days and counting, right until I decided to try the freshly
announced 2.6.36-rc2.

So I upgraded the kernel, but left the scripts that turn GRO off for
the tg3 card still run at system startup. This way the system ran for
2 and a half hours, when I decided its time to try turning GRO on.

I first tried to turn GRO on for the tg3 nic, and the system oopsed
immediately (if the panic screen is necessary - please, ask for it).

After the system came back, I tried turning GRO on for the 2 RealTek
8139 nics, too, but ethtool only accepted turning GRO off.

And unfortunately, I can't test if other nics will fail the same way
as the motherboard integrated tg3 I have does, so for now, this is
only a tg3 + GRO on problem; I don't have any other hardware to test
with available.

A little misunderstanding: I was intersted with turning off some
features on realteks to change the packet path from tg3 with gro
to realtek without gro and without tx-checksumming etc.

But maybe you could try the patch below instead (so the patched
kernel, tg3 with gro on, and realteks without any change).

Thanks,
Jarek P.

--- (for debugging only)

diff --git a/net/core/dev.c b/net/core/dev.c
index 3721fbb..51823cd 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c

@@ -1935,6 +1935,23 @@ static inline int skb_needs_linearize(struct sk_buff *skb,
 					      illegal_highdma(dev, skb))));
 }
 
+static int skb_csum_start_bug(struct sk_buff *skb)
+{
+
+	if (skb->ip_summed == CHECKSUM_PARTIAL) {
+		long csstart;
+
+		csstart = skb->csum_start - skb_headroom(skb);
+		if (WARN_ON(csstart > skb_headlen(skb))) {
+			pr_warning("csum_start %d, headroom %d, headlen %d\n",
+				   skb->csum_start, skb_headroom(skb),
+				   skb_headlen(skb));

+			return 1;
+		}
+	}
+	return 0;
+}
+
 int dev_hard_start_xmit(struct sk_buff *skb, struct net_device *dev,
 			struct netdev_queue *txq)
 {

@@ -1955,11 +1972,13 @@ int dev_hard_start_xmit(struct sk_buff *skb, struct net_device *dev,
 		skb_orphan_try(skb);
 
 		if (netif_needs_gso(dev, skb)) {
+			skb_csum_start_bug(skb);
 			if (unlikely(dev_gso_segment(skb)))
 				goto out_kfree_skb;
 			if (skb->next)
 				goto gso;
 		} else {
+			skb_csum_start_bug(skb);
 			if (skb_needs_linearize(skb, dev) &&
 			    __skb_linearize(skb))
 				goto out_kfree_skb;

@@ -1997,7 +2016,12 @@ gso:
 		if (dev->priv_flags & IFF_XMIT_DST_RELEASE)
 			skb_dst_drop(nskb);
 
-		rc = ops->ndo_start_xmit(nskb, dev);
+		if (skb_csum_start_bug(skb)) {
+			kfree_skb(skb);
+			rc = NETDEV_TX_OK;
+		} else
+			rc = ops->ndo_start_xmit(nskb, dev);
+
 		if (unlikely(rc != NETDEV_TX_OK)) {
 			if (rc & ~NETDEV_TX_MASK)
 				goto out_kfree_gso_skb;