Re: [fedora-virt] kvm "tablet" device burning cpu cycles

[Eric Blake]

[redirecting to the fedora virt list]

On 12/06/2011 03:36 PM, Dennis Jacobfeuerborn wrote:
> Hi,
> I noticed that when you install a virtual system with virt-manager it 
> always installs a tablet device by default.
> The problem is that this device consumes 10% of cpu resources of the host 
> according to "top" which should probably be considered a bug.

The poor behavior of a tablet device should be raised against the qemu
team, to see if it can be made more efficient.
> What is worse is that removing the tablet also means no mouse pointer 
> showing up in a centos 6 guest which means you are pretty much forced to 
> use the tablet device.

Meanwhile, sufficiently new qemu provides USB device emulation, which is
more efficient than the PCI tablet emulation; although it is not as
apparent whether typical guests support that device driver in their
default install (one of the reasons we default to the older tablet
emulation is that it is likely to have a driver pre-installed on most
guest OSs).  The idea of swapping the default to a USB pointer should be
raised against virt-manager and/or python-virtinst, when installing a
new-enough guest OS with new enough qemu support, while continuing to
provide the compatible fallback of the tablet for older setups.

kvm "tablet" device burning cpu cycles

[Andrew Cathrow]

----- Original Message -----
> From: "Eric Blake" 
> To: "For testing and quality assurance of Fedora releases" 
> Cc: [e-mail]
> Sent: Tuesday, December 6, 2011 6:33:43 PM
> Subject: Re: [fedora-virt] kvm "tablet" device burning cpu cycles
>
> [redirecting to the fedora virt list]
>
> On 12/06/2011 03:36 PM, Dennis Jacobfeuerborn wrote:
> > Hi,
> > I noticed that when you install a virtual system with virt-manager
> > it
> > always installs a tablet device by default.
> > The problem is that this device consumes 10% of cpu resources of
> > the host
> > according to "top" which should probably be considered a bug.
>
> The poor behavior of a tablet device should be raised against the
> qemu
> team, to see if it can be made more efficient.
> > What is worse is that removing the tablet also means no mouse
> > pointer
> > showing up in a centos 6 guest which means you are pretty much
> > forced to
> > use the tablet device.
>
> Meanwhile, sufficiently new qemu provides USB device emulation, which
> is
> more efficient than the PCI tablet emulation; although it is not as
> apparent whether typical guests support that device driver in their
> default install (one of the reasons we default to the older tablet
> emulation is that it is likely to have a driver pre-installed on most
> guest OSs).  The idea of swapping the default to a USB pointer should
> be
> raised against virt-manager and/or python-virtinst, when installing a
> new-enough guest OS with new enough qemu support, while continuing to
> provide the compatible fallback of the tablet for older setups.

Also we've found Spice more efficient in our testing.
> --
> Eric Blake   [e-mail]    [phone]
> Libvirt virtualization library http://libvirt.org
>
> virt mailing list
> [e-mail]
> https://admin.fedoraproject.org/mailman/listinfo/virt

kvm "tablet" device burning cpu cycles

[Daniel P. Berrange]

On Tue, Dec 06, 2011 at 04:33:43PM -0700, Eric Blake wrote:
> [redirecting to the fedora virt list]
>
> On 12/06/2011 03:36 PM, Dennis Jacobfeuerborn wrote:
> > Hi,
> > I noticed that when you install a virtual system with virt-manager it 
> > always installs a tablet device by default.
> > The problem is that this device consumes 10% of cpu resources of the host 
> > according to "top" which should probably be considered a bug.
>
> The poor behavior of a tablet device should be raised against the qemu
> team, to see if it can be made more efficient.

This is a long standing problem, but it shouldn't be taking anywhree
near 10% of CPU.

If using SPICE with the SPICE guest agent installed there is a parvirt
mouse available, avoiding the need for the USB tablet.
> > What is worse is that removing the tablet also means no mouse pointer 
> > showing up in a centos 6 guest which means you are pretty much forced to 
> > use the tablet device.

That shouldn't be the case. If you boot the guest with no USB tablet
present, X should auto-detect the PS2 mouse and use that. You would
only have trouble if either hot-unpluging the USB tablet, or if you
have hardcoded its use in /etc/X11/xorg.conf
> Meanwhile, sufficiently new qemu provides USB device emulation, which is
> more efficient than the PCI tablet emulation;

Actually the tablet is a USB device, not a PCI device.

Daniel

kvm "tablet" device burning cpu cycles

[Dennis Jacobfeuerborn]

On 12/07/2011 10:20 AM, Daniel P. Berrange wrote:
> On Tue, Dec 06, 2011 at 04:33:43PM -0700, Eric Blake wrote:
> > [redirecting to the fedora virt list]
> >
> > On 12/06/2011 03:36 PM, Dennis Jacobfeuerborn wrote:
> > > Hi,
> > > I noticed that when you install a virtual system with virt-manager it
> > > always installs a tablet device by default.
> > > The problem is that this device consumes 10% of cpu resources of the host
> > > according to "top" which should probably be considered a bug.
> >
> > The poor behavior of a tablet device should be raised against the qemu
> > team, to see if it can be made more efficient.
>
> This is a long standing problem, but it shouldn't be taking anywhree
> near 10% of CPU.
>
> If using SPICE with the SPICE guest agent installed there is a parvirt
> mouse available, avoiding the need for the USB tablet.

I did some more testing today and also installed a Fedora 16 VM. Here I
don't see the big overhead and an idle guest is also idle on the host so it 
seems this is an issue with the driver on the guest side. While I tried
SPICE I didn't play with the guest agent yet. When using SPICE I switch the 
display device from VNC to SPICE and let virt-manager add an additional 
channel device (it asks me if it should do so). Is it sufficient to install 
spice-vdagent in the guest to make use of the client mouse mode?
> > > What is worse is that removing the tablet also means no mouse pointer
> > > showing up in a centos 6 guest which means you are pretty much forced to
> > > use the tablet device.
>
> That shouldn't be the case. If you boot the guest with no USB tablet
> present, X should auto-detect the PS2 mouse and use that. You would
> only have trouble if either hot-unpluging the USB tablet, or if you
> have hardcoded its use in /etc/X11/xorg.conf

I experimented a bit and I only see this:
1) when I use "qxl" as video model
2) when I keep the graphical console open in virt-manager after booting

When I close the window and reopen it again the mouse pointer is there and
when I use "cirrus" as model the problem doesn't happen at all.

Regards,
    Dennis