SG off on domU middle node drops packets

SG off on domU middle node drops packets

[Lennart Schulte]

Hi,
I use a Xen domU with vanilla Linux 3.19 kernel as a middle node. SG is 
switched off via ethtool. I get random packet drops for about 7% of 
small packets (pings, TCP ACKs/SYNs/little payload).
 From the code in Linux 4.0 it seems that the behavior should be the same.

I posted on Xen-users before with more information:
http://lists.xen.org/archives/html/xen-users/2015-05/msg00170.html

I narrowed down the issue, and the below patch seems to work. Any 
comments are highly welcome.

Thanks,
Lennart


diff --git a/drivers/net/xen-netfront.c b/drivers/net/xen-netfront.c
index d8c1076..f5e403b 100644
--- a/drivers/net/xen-netfront.c
+++ b/drivers/net/xen-netfront.c
@@ -611,11 +611,19 @@ static int xennet_start_xmit(struct sk_buff *skb, 
struct net_device *dev)
         spin_lock_irqsave(&queue->tx_lock, flags);

         if (unlikely(!netif_carrier_ok(dev) ||
-                    (slots > 1 && !xennet_can_sg(dev)) ||
                      netif_needs_gso(dev, skb, netif_skb_features(skb)))) {
spin_unlock_irqrestore(&queue->tx_lock, flags);
                 goto drop;
         }
+       if (unlikely(slots > 1 && !xennet_can_sg(dev))) {
+               if (skb_linearize(skb)) {
+ spin_unlock_irqrestore(&queue->tx_lock, flags);
+                       goto drop;
+               }
+               data = skb->data;
+               offset = offset_in_page(data);
+               len = skb_headlen(skb);
+       }

         i = queue->tx.req_prod_pvt;

Re: SG off on domU middle node drops packets

[David Vrabel]

On 26/05/15 08:55, Lennart Schulte wrote:
> Hi,
> I use a Xen domU with vanilla Linux 3.19 kernel as a middle node. SG is
> switched off via ethtool. I get random packet drops for about 7% of
> small packets (pings, TCP ACKs/SYNs/little payload).
> From the code in Linux 4.0 it seems that the behavior should be the same.

Why would you turn this option off?

> I posted on Xen-users before with more information:
> http://lists.xen.org/archives/html/xen-users/2015-05/msg00170.html
> 
> I narrowed down the issue, and the below patch seems to work. Any
> comments are highly welcome.
> 
> Thanks,
> Lennart
> 
> 
> diff --git a/drivers/net/xen-netfront.c b/drivers/net/xen-netfront.c
> index d8c1076..f5e403b 100644
> --- a/drivers/net/xen-netfront.c
> +++ b/drivers/net/xen-netfront.c
> @@ -611,11 +611,19 @@ static int xennet_start_xmit(struct sk_buff *skb,
> struct net_device *dev)
>         spin_lock_irqsave(&queue->tx_lock, flags);
> 
>         if (unlikely(!netif_carrier_ok(dev) ||
> -                    (slots > 1 && !xennet_can_sg(dev)) ||
>                      netif_needs_gso(dev, skb, netif_skb_features(skb)))) {
> spin_unlock_irqrestore(&queue->tx_lock, flags);
>                 goto drop;
>         }
> +       if (unlikely(slots > 1 && !xennet_can_sg(dev))) {
> +               if (skb_linearize(skb)) {
> + spin_unlock_irqrestore(&queue->tx_lock, flags);
> +                       goto drop;
> +               }
> +               data = skb->data;
> +               offset = offset_in_page(data);
> +               len = skb_headlen(skb);
> +       }

Looks ok, can you send a proper patch to devel and xen-devel, please?

David