Re: Xen/OS layer debugging help

[Anil Madhavapeddy <anil@xxxxxxxxxx>]

On Thu, Jul 19, 2012 at 06:22:26PM +0100, Balraj Singh wrote:
> A request for some debugging help: the remaining TCP problem that I
> see can be attributed to significant pkt loss somewhere in the Xen/OS
> layer.  I see packets if I sniff the virtual interface, but a lot of
> the packets never make it to Ethif.input.  I've tried to trace the
> code from netif down to the grant tables myself, but I think I need
> help.  If you can take a look, can you see a reason why pkts would get
> lost at busy times, between the vif and Ethif.input.
> 
> This may just be my current test, but the particular packets that get
> lost are the pure ACKs (no payload) coming back to my Mirage data
> source from a linux data sink running on the same host in another VM.

This could point to a bug in Netfront not updating the pointers, but it
would be unusual to see this only happening for ACKs.  Of course, if it's
only on the transmit ring, then we might only ever be sending ACKs.

Could you share details of your test case so that Dave/I can reproduce
this?  A ping-flood to my Mirage VM resulted in no packet loss over 5
minutes, so I think this requires looking at your specific test.

-- 
Anil Madhavapeddy                                 http://anil.recoil.org