Re: [visorchipset] invalid opcode: 0000 [#1] PREEMPT SMP

[Fengguang Wu <fengguang.wu@intel.com>]

CC the KVM people: it looks like a KVM problem that can be triggered by

        qemu-system-x86_64 -cpu Haswell,+smep,+smap

On Thu, Apr 10, 2014 at 01:58:18AM +0800, Jet Chen wrote:
> On 04/09/2014 10:44 PM, Romer, Benjamin M wrote:
> > On Wed, 2014-04-09 at 02:38 +0800, Jet Chen wrote:
> > 
> >> Hi Ben,
> >>
> >> I checked my <Intel 64 and IA-32 Architectures Software Developer's Manual> which published in Feb 2014.
> >> Volume 2: Instruction Set Reference, A-Z: CPUID--CPU Identification
> >>
> > 
> > I agree completely, which is why I'm confused about KVM's behavior. If
> > bit 31 was off, the code in our drivers that uses the vmcall instruction
> > would not have been run, the kernel would not have tried to perform a
> > vmcall, and not crashed with invalid op.
> > 
> > If you look in the definition for the VMCALL instruction (Intel 64 and
> > IA32 Architectures Software Developer's Manual, volume 3C pg.30-9)
> > You'll see that a processor in VMX non-root operation should perform a
> > vmexit.
> > 
> >> Why this document not match what you said ? I am not experienced with VM, please correct me if I went for wrong document
> >>
> > 
> > According to VMWare's documentation (there is a page at
> > http://kb.vmware.com./selfservice/microsites/search.do?cmd=displayKC&externalId=1009458 ) , as well as Microsoft's hypervisor spec (at http://www.microsoft.com/en-us/download/details.aspx?id=39289 ), this bit is used to indicate the CPU is running under virtualization. KVM is also setting this bit to indicate virtualization. I believe Xen uses it as well.
> > 
> > 
> > My contention is, if KVM is going to set the ISVM bit, it needs to do a
> > vmexit, and if it's not going to set the bit, then doing an invalid op
> > is okay, but the current behavior is inconsistent.
> > 
> > -- Ben
> > 
> 
> Ben,
> 
> Really thanks for your explanation.
> Let me summary it up, please correct me where i am wrong. If it is really a KVM bug, we report it to KVM guys.
> On a real CPU, ECX 31bit always be 0 as Intel documentation filed.
> However, KVM, as a hypervisor, should emulate this bit of the virtual ECX register to 1 for guest OS to indicate it is running in a virtualization environment.
> Problem is, KVM does set this bit to 1, but does an invalid op instead of emit a VMCALL. As a result, we get this dmesg error messages.
> 
> Thanks,
> -Jet