Re: [PATCH 1/2] x86,kvm: move qemu/guest FPU switching out to vcpu_run

[Quan Xu]

On 2017-11-16 12:21, Rik van Riel wrote:
> On Thu, 2017-11-16 at 10:50 +0800, Quan Xu wrote:
> > On 2017-11-15 22:43, Rik van Riel wrote:
> > > Can you explain why you believe that?
> > for example, a vcpu thread is running in kvm mode under cretical
> > condition to stop. QEMU send an IPI to cause a VM-exit to happen
> > immediately, and this IPI doesn't make vcpu return to QEMU. IIUC
> > this vcpu thread will still continue to run in kvm mode when is
> > waked up at targer machine. with your patch, I don't see a chance
> > to load guest FPU or XSTATE, until return to QEMU and run kvm mode
> > again.
> >
> > then the FPU or XSTATE status is inconsistent for a small window,
> > what's
> > even
> > worse is that the vcpu is running.
> >
> > Did I misunderstand?
> At context switch time, the context switch code will save
> the guest FPU state to current->thread.fpu when the
> VCPU thread is scheduled out.
>
> When the VCPU thread is scheduled back in, the context
> switch code will restore current->thread.fpu to the FPU
> registers.


good catch!
Also as your comment, PKRU is switched out separately at
VMENTER and VMEXIT time, but with a lots of IF conditions..

the pkru may be restored with host pkru after VMEXIT. when
vcpu thread is scheduled out, the pkru value in current->thread.fpu.state
may be the host pkru value, instead of guest pkru value (of course,
this _assumes_ that the pkru is in current->thread.fpu.state as well).
in this way, the pkru may be a coner case.

VM migration again, in case,
ÂÂÂÂÂÂÂÂÂÂ source_host_pkru_value != guest_pkru_value,
ÂÂÂÂÂÂÂÂÂÂ target_host_pkru_value == guest_pkru_value..

the pkru status would be inconsistent..



Quan
Alibaba Cloud

> The VCPU thread will never run with anything else than
> the guest FPU state, while inside the KVM_RUN code.