[PVE-User] Regression on virtio-net ? (Was: Re: Proxmox VE 4.2 released!)

[Wolfgang Bumiller]

On Tue, May 03, 2016 at 12:40:08PM +0200, Gilou wrote:
> Le 30/04/2016 07:20, Alexandre DERUMIER a écrit :
> > ok, 
> > 
> > can you check with perf command ?
> > 
> > check both qemu and vhost process
> > 
> > 
> > "perf top -p pidofqemu"
> > 
> > "perf top -p pidofvhost"
> 
> 
> OK, I did it, and hit the bug again, on a different hardware.
> Context is:
> Ubuntu 16.04 VM, single core, host CPU, using virtio-net to blast at 1
> Gbps through the physical NIC. I still have 100% of a core used by the
> kvm process, and the perf top results are here:
> 
> https://gist.github.com/gilou/15b620a7a067fd1d58a7616942e025b4#file-perf_virtionet_4-4-txt
> 
> This was done again on a 4.1 updated to a 4.2. I will probably try on a
> fresh 4.2 install to check it out, but I doubt it will change much.
> If I fall back to the 4.2.6 kernel, bug isn't there.

I'll try to bisect this next week. I was able to produce a similar perf
output (read_tsc() at >25% with 4.4, and far below with 4.2).

> I can grant you access to the host should you need that, but I'm pretty
> sure anybody can reproduce that issue. Which brings quite some worries.
> 
> Regards,
> 
> Gilles Pietri