On Wed, 2014-09-10 at 10:42 +0300, Or Gerlitz wrote:

Hi Chuck, thanks for bisecting this out. Indeed, as of this kernel 3.2 
commit 936d7de "IPoIB: Stop lying about hard_header_len and use skb->cb 
to stash LL addresses" we are using the skb->cb field to enable proper 
work under GRO and avoid another historical quirk we had there... so I 
think we can definetly consider commit e0f31d849 to introduce a severe 
regression... Govindarajulu, Dave - what's your thinking here? any quick 
idea on how to fix?

I mentioned the IB stuff when patch was posted, and David replied :

"I think this is fine, IPOIB's control block will need still just 44
bytes after these changes, so there will still be 4 bytes to spare."

http://patchwork.ozlabs.org/patch/357584/

My suggestion was to reduce sch_choke usage and not store the whole
flow_keys.