Bug 1314111

Summary: vxlan processing occurs on one cpu only
Product: Red Hat Enterprise Linux 7 Reporter: Hannes Frederic Sowa <hsowa>
Component: kernelAssignee: Paolo Abeni <pabeni>
kernel sub component: Networking QA Contact: Network QE <network-qe>
Status: CLOSED NOTABUG Docs Contact:
Severity: high    
Priority: high CC: aloughla, jbenc, jeder, jpirko, jsitnick, kzhang, pabeni, ptalbert, rkhan, tgraf
Version: 7.3   
Target Milestone: rc   
Target Release: ---   
Hardware: Unspecified   
OS: Unspecified   
Whiteboard:
Fixed In Version: Doc Type: Bug Fix
Doc Text:
Story Points: ---
Clone Of: Environment:
Last Closed: 2016-03-30 15:19:48 UTC Type: Bug
Regression: --- Mount Type: ---
Documentation: --- CRM:
Verified Versions: Category: ---
oVirt Team: --- RHEL 7.3 requirements from Atomic Host:
Cloudforms Team: --- Target Upstream Version:
Embargoed:
Bug Depends On:    
Bug Blocks: 1323132    

Description Hannes Frederic Sowa 2016-03-03 00:04:28 UTC 
Investigate if vxlan only processes packets on one CPU. Different flows should actually use different flows and get send out via different tx queues. If this is not the case we need to fix this.
Comment 3 Paolo Abeni 2016-03-30 15:19:48 UTC 
Short summary: with vxlan offload, vxlan processing works correctly, without h/w offload the single cpu issue is unsolvable.

On h/w with vxlan offload, vxlan processing is correctly spread across the available cpus.

If the nic lacks vxlan offload, than the rx-hash by default ignores the udp ports and all vxlan flows collide on the same CPU (the IPs in the external header are the same in all flows).

To avoid such collision, the following setting could be used:

ethtool -N <interface> rx-flow-hash udp4 sdfn

but the above will introduce reordering if the external udp datagram is fragmented, and that may break existing applications.

Moreover without vxlan offloading we lose:

* LRO/GRO because we don't get CHECKSUM_PARTIAL frames or the depth we 
require the checksumming logic to look into the packet is not deep 
enough (outer frames don't have checksum by default)

* without CHECKSUM_PARTIAL no use of LCO (local checksum offload)

* sending checksum offload also not possible, because of the lack of 
CHECKSUM_PARTIAL in most hardware

We certainly should only consider either new networking cards with 
CHECKSUM_PARTIAL or we depend on vxlan (or later, geneve) offloading.