On Thu, 2021-10-21 at 22:08 +0200, Paolo Bonzini wrote:
> On 18/10/21 19:14, butt3rflyh4ck wrote:
> > {
> > struct kvm_vcpu *vcpu = kvm_get_running_vcpu(); //-------> invoke
> > kvm_get_running_vcpu() to get a vcpu.
> >
> > WARN_ON_ONCE(vcpu->kvm != kvm); [1]
> >
> > return &vcpu->dirty_ring;
> > }
> > ```
> > but we had not called KVM_CREATE_VCPU ioctl to create a kvm_vcpu so
> > vcpu is NULL.
>
> It's not just because there was no call to KVM_CREATE_VCPU; in general
> kvm->dirty_ring_size only works if all writes are associated to a
> specific vCPU, which is not the case for the one of
> kvm_xen_shared_info_init.
>
> David, what do you think? Making dirty-page ring buffer incompatible
> with Xen is ugly and I'd rather avoid it; taking the mutex for vcpu 0 is
> not an option because, as the reporter said, you might not have even
> created a vCPU yet when you call KVM_XEN_HVM_SET_ATTR. The remaining
> option would be just "do not mark the page as dirty if the ring buffer
> is active". This is feasible because userspace itself has passed the
> shared info gfn; but again, it's ugly...
I think I am coming to quite like that 'remaining option' as long as we
rephrase it as follows:
KVM does not mark the shared_info page as dirty, and userspace is
expected to *assume* that it is dirty at all times. It's used for
delivering event channel interrupts and the overhead of marking it
dirty each time is just pointless.
I've merged the patch I sent yesterday into my pfncache series because
I realised we needed the same there, but I'll look at making the Xen
code *not* mark the shinfo page dirty when it writes the clock data
there.
Amazon Development Centre (London) Ltd. Registered in England and Wales with registration number 04543232 with its registered office at 1 Principal Place, Worship Street, London EC2A 2FA, United Kingdom.
