Re: [Bug #12465] KVM guests stalling on 2.6.28 (bisected) [Bug12465]

[Kevin Shanahan]

On Sun, 2009-02-15 at 11:04 +0100, Ingo Molnar wrote:
> It would be nice to enhance this single-CPU trace some more - to more
> surgically see what is going on. Firstly, absolute timestamps would be
> nice:
> 
>   echo funcgraph-abstime  > trace_options
>   echo funcgraph-proc     > trace_options
> 
> as it's a bit hard to see the global timescale of events.

Okay, here's some more trace data. I grabbed a few samples at different
times during the ping test. I think the data in files trace6.txt and
trace8.txt coincided with some of the biggest delays.

  http://disenchant.net/tmp/bug-12465/trace-2/

This is captured on 2.6.29-rc5-tip-02057-gaad11ad. The kvm guest being
pinged is process 11211:

  flexo:~# pstree -p 11211
  qemu-system-x86(11211)âââ{qemu-system-x86}(11212)
                         ââ{qemu-system-x86}(11213)
                         ââ{qemu-system-x86}(11609)

Cheers,
Kevin.

> Secondly, not all events are included - in particular i dont really see
> the points when packets are passed. Would it be possible to add a tracing
> hypercall so that the guest kernel can inject trace events that can be seen
> on the native-side trace? Regarding ping latencies really just two things
> matter: the loopback network device's rx and tx path. We should trace the
> outgoing sequence number and the incoming sequence number of IP packets,
> and inject that to the host side. This way we can correlate the delays
> precisely.
> 
> 	Ingo


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/