Re: KVM, AMD & Interrupts

[Anthony Liguori <aliguori-NZpS4cJIG2HvQtjrzfazuQ@public.gmane.org>]

Matthew Hall wrote:
> Hi Avi,
>
> I spoke to you a while ago on #kvm on freenode - I had a problem with 
> ide ops on an amd svm host with kvm. The issues regarded the inability 
> to access ida devices (hda/hdc) under qemu & kvm on my host - if this 
> triggers your memory...
>
> The last point I got to was that it was kvm and amd specific. Id est, 
> the same image run under an intel vmx host with the *exact* same image, 
> kvm version and command options (svn 'trunk' from the $date).
> The issue seemed to be AMD and kvm specific (running with '-no-kvm' made 
> everything work ok, but slower, as you'd expect).
>
> I got word of a possible workaround, from one of your colleagues whilst 
> you've been away, to use the boot options 'noapic nolapic' on the host 
> o/s. Using these options solved my hd* op problems and I can boot ok now.
>   

I've seen a very similar problem on an IBM x460.  The main symptom I see 
is lost timer interrupts in the host.

I haven't spent much time looking into this but I suspect it's almost 
certainly related to the fact that KVM is using a different interrupt 
delivery method than Xen.  Xen doesn't seem to have an issue on this 
system.  I suspect that the interrupts being received that's causing a 
VMEXIT is not actually getting propagated to the host (the timer being 
the most frequent interrupt to cause a VMEXIT).  If we switch to not to 
acknowledging the interrupt before the VMEXIT and then regenerating it 
either with a software interrupt or delivering it directly in the host, 
it's possible this problem may go away.

This is all just speculation though.

regards,

Anthony Liguori

> It seems disabling the local apic was the fundamental part (iirc), as 
> without host support the guest would always start the dummy apic; so 
> running the guest with any 'special' options (ie. '-no-apic') should not 
> be required.
>
> I'm think this problem is specific to my particular cpu/motherboard 
> combination aswell.
> I also have box with a recent intel cpu (with the vmx extention) which 
> doesn't display this behaviour; and from talking with you it's not 
> purely amd specific (I believe your tests on amd cpu worked).
> Is it possible a slowdown in the interrupt controller between the guest 
> and host is a contributing factor?
> Iirc, with a 'dummy' apic, I experience a ~5% slowdown on disk and 
> network operations, which isn't so bad. I've yet to think about smp usage.
>
> Thought you might appreciate this info should you feel it deserve a wiki 
> entry. Point me in the direction of the appropriate page should you feel 
> I can put this info in myself.
>
>
> Best regards,
> Matt
>
> -------------------------------------------------------------------------
> Take Surveys. Earn Cash. Influence the Future of IT
> Join SourceForge.net's Techsay panel and you'll get the chance to share your
> opinions on IT & business topics through brief surveys - and earn cash
> http://www.techsay.com/default.php?page=join.php&p=sourceforge&CID=DEVDEV
> _______________________________________________
> kvm-devel mailing list
> kvm-devel-5NWGOfrQmneRv+LV9MX5uipxlwaOVQ5f@public.gmane.org
> https://lists.sourceforge.net/lists/listinfo/kvm-devel
>   


-------------------------------------------------------------------------
Take Surveys. Earn Cash. Influence the Future of IT
Join SourceForge.net's Techsay panel and you'll get the chance to share your
opinions on IT & business topics through brief surveys - and earn cash
http://www.techsay.com/default.php?page=join.php&p=sourceforge&CID=DEVDEV