On Wed, May 13, 2015 at 3:30 PM, David Miller [off-list ref] wrote:

From: Tom Herbert <redacted>
Date: Tue, 12 May 2015 08:22:58 -0700

quoted

@@ -15,6 +15,13 @@
  * All the members, except thoff, are in network byte order.
  */
 struct flow_keys {
+     u16     thoff;
+     u16     padding1;
+#define FLOW_KEYS_HASH_START_FIELD   n_proto
+     __be16  n_proto;
+     u8      ip_proto;
+     u8      padding;
+

This padding works if everyone consistently zero initializes the whole
key structure, but for whatever reason (performance, unintentional
oversight, etc.) not all paths do.

So, for example, inet_set_txhash() is going to have random crap in
keys.padding, so the hashes computed are not stable for a given flow
key tuple.

That's just the first code path I found with this issue, there are
probably several others.

memset zero is in the second patch for inet_set_txhash and
ip6_set_txhash. I can respin so those are in the first patch.

Tom