Re: Nested KVM

[Alexander Graf <agraf@suse.de>]

On 23.12.2008, at 18:05, Avi Kivity wrote:

> Alexander Graf wrote:
>>
>>> I have successfully built the latest kvm, kvm-userspace from git.
>>>
>>> I loaded the kvm_amd module with the nested=1 option.
>>>
>>> The kvm guest that I am working with has an older version
>>> of kvm (Ubuntu 8.10 + kvm package) and during boot I get:
>>>
>>> [   23.673107] Pid: 4278, comm: modprobe Tainted: G S         
>>> 2.6.27-7-generic #1
>>> [   23.673107] RIP: 0010:[<ffffffff8022d35a>]  [<ffffffff8022d35a>]
>>> native_write_msr_safe+0xa/0x10
>>> [   23.673107] RSP: 0018:ffff880018dc3dd8  EFLAGS: 00010002
>>> [   23.673107] RAX: 0000000000001d01 RBX: ffff880018dc3e04 RCX:  
>>> 00000000c0000080
>>>
>>> Maybe it is just that a newer version of KVM is needed?
>>>
>>> I will also try to build a newer version of KVM and test.
>>>
>>> Thanks for any comments.
>>
>> Your guest is writing to an MSR that is unknown to the kvm msr  
>> emulation. What does dmesg on the host say?
>>
>
> You can actually see that it's writing to EFER (rcx = 0xc00000080).   
> But that should always be implemented on AMD.  It's probably  
> complaining about the svme bit, but that can only happen if nested=0?

Oh, right. But then again dmesg spits out a lot of these:

[51056.178705] kvm: 14970: cpu0 unhandled wrmsr: 0xc0010117 data 0

So writing HSAVE fails too, which should only happen on older KVM  
versions.

But nevertheless, setting EFER only works if nested=1 is given as  
module option and -enable-nesting is used on qemu (read: the SVM CPUID  
capability is set).

Alex