[ovs-dev] bug in flow stats for VLAN acceleration?

[Ben Pfaff]

I sent out a patch that should fix this:
        http://openvswitch.org/pipermail/dev/2014-December/049972.html

It will take some time, probably a few days, for it to get into the
kernel and the OVS tree's kernel module.  After that, I think you'll be
ready to go.  (However, the buggy behavior will still be present if a
user upgrades userspace without upgrading their kernel module.  I doubt
that that is really a big deal though.)

On Thu, Dec 11, 2014 at 12:36:51AM +0900, Motonori Shindo wrote:
> Ben and Jesse,
> 
> Thanks for picking this issue up.
> 
> Although current behavior is not consistent in VLAN and no-VLAN cases, as far as NetFlow concerns it doesn’t matter in fact because NetFlow only counts L3 packet size anyway.
> 
> Here’s the story. First, I wanted to fix an issue that OVS currently reports L2 packet size, not L3 packet size in NetFlow flow records (NetFlow should report L3 packet size). I thought the easiest way to fix this problem was to subtract the ethernet header length when reporting a NetFlow flow record. Then, I realized that I should also take VLAN header into account if it is present. I checked how NetFlow behaves in VLAN case and no-VLAN case and found out that OVS reports the same packet size in NetFlow regardless to the presence of VLAN header.
> 
> Once this packet size issue related to VLAN header gets solved by someone else (as this is probably beyond of my understanding of the relevant code), I will work on fixing the L2 vs L3 packet size issue specifically in NetFlow.
> 
> Regards,
> 
> On 2014/12/09 3:22, "Jesse Gross" <jesse at nicira.com <mailto:jesse at nicira.com>> wrote:
> 
> On Mon, Dec 8, 2014 at 9:58 AM, Ben Pfaff <blp at nicira.com <mailto:blp at nicira.com>> wrote:
> Motonori Shindo (CCed) reported to me recently that NetFlow does not
> include VLAN headers in the byte counts that it exports.  Taking a
> look at the kernel datapath code, I see that ovs_flow_stats_update()
> accounts raw skb->len to byte counters.  I suspect that it should add
> 4 if a VLAN header was present but removed by VLAN acceleration.
> 
> Presumably we should also do this for port stats as well to be consistent.
> 
> 
> 
> ---
> Motonori Shindo