Re: [RFC] vxlan: use ether header as fallback hash

On Thu, Oct 4, 2012 at 9:43 AM, Stephen Hemminger [off-list ref] wrote:

On Thu, 4 Oct 2012 09:27:14 -0700
Jesse Gross [off-list ref] wrote:

quoted

On Wed, Oct 3, 2012 at 9:39 PM, Stephen Hemminger [off-list ref] wrote:

quoted

VXLAN bases source UDP port based on flow to help the
receiver to be able to load balance based on outer header
contents.

This patches changes the algorithm to better handle packets
that can not be categorized by the rxhash() function.
It adds a fallback to use jhash on the Ether header.

It also fixes a bug where the old code could assign 0 as a port
value.


Signed-off-by: Stephen Hemminger <redacted>

---
RFC for now, compile tested only

--- a/drivers/net/vxlan.c       2012-10-03 21:25:43.747968165 -0700
+++ b/drivers/net/vxlan.c       2012-10-03 21:36:10.213805422 -0700

@@ -622,12 +622,30 @@ static inline u8 vxlan_ecn_encap(u8 tos,
        return INET_ECN_encapsulate(tos, inner);
 }

+/* Compute hash to use for source port
+ *   first choice to use L4 flow hash since it will spread
+ *     better and maybe available from hardware
+ *   secondary choice is to use jhash on the Ethernet header
+ * Always returns non-zero value
+ */
+static u16 vxlan_flow_hash(struct sk_buff *skb)
+{
+       u16 hash = skb_get_rxhash(skb);
+
+       if (!hash)
+               hash = jhash(skb->data, 3, skb->protocol);

Shouldn't this be jhash2 for words?

No. for a couple of reasons. First, the ethernet header may not be aligned.
Second we want to get source/destination and type.  The source/destination
is 12 bytes (3 words) and the ether type is already in skb->protocol.

OK, it might not be aligned but jhash takes it's length as the number
of bytes, not words.