kvm suffering from dramatic performance regression

[James Hunt]

The recent upgrade to kvm is slowing down kvm access considerably. It gives comparable performance to running kvm with no kernel modules, however I do have kvm and kvm_intel loaded.

I'm in the kvm group and am running kvm like this:

$ kvm -hda ubuntu_13.04.kvm -name "ubuntu_13.04.kvm" -m 512 -net user -net nic -redir tcp:11601::22 -serial file:/tmp/kvm.log

ProblemType: Bug
DistroRelease: Ubuntu 13.04
Package: qemu-kvm 1.3.0+dfsg-1~exp3ubuntu4
ProcVersionSignature: Ubuntu 3.8.0-1.5-generic 3.8.0-rc4
Uname: Linux 3.8.0-1-generic i686
NonfreeKernelModules: nvidia
ApportVersion: 2.8-0ubuntu2
Architecture: i386
Date: Mon Jan 21 17:01:25 2013
InstallationDate: Installed on 2010-10-21 (823 days ago)
InstallationMedia: Ubuntu 10.10 "Maverick Meerkat" - Release i386 (20101007)
MarkForUpload: True
ProcEnviron:
 TERM=xterm
 PATH=(custom, no user)
 XDG_RUNTIME_DIR=<set>
 LANG=fr_CA.UTF8
 SHELL=/bin/bash
SourcePackage: qemu
UpgradeStatus: Upgraded to raring on 2012-12-20 (32 days ago)

[James Hunt (jamesodhunt)]

Manually removing the kernel modules and re-loading shows the following in dmesg:

[32444.994884] kvm: VM_EXIT_LOAD_IA32_PERF_GLOBAL_CTRL does not work properly. Using workaround

[James Hunt (jamesodhunt)]

Looks like a behaviour change. Version 1.3.0+dfsg-1~exp3ubuntu4 seems to need '-enable-kvm' to take advantage of the kernel modules. So, my command-line now becomes:

$ kvm -enable-kvm -hda ubuntu_13.04.kvm -name "ubuntu_13.04.kvm" -m 512 -net user -net nic -redir tcp:11601::22 -serial file:/tmp/kvm.log

[Soren Hansen (soren)]

When invoked as kvm, it's a perfectly reasonable expectation that it uses kvm. If this is indeed not the case, that's a bug.

[Serge Hallyn (serge-hallyn)]

The debian qemu tree has added a kvm wrapper which adds -enable-kvm and -cpu kvm64 to the arguments. I'm not sure whether the unconditional addition of -cpu is correct for us.

(I'm also not sure the best answer isn't simply to remove the 'kvm' symlink)