On Wed, Dec 22, 2021 at 04:15:02PM +0100, Paolo Bonzini wrote:
> On 12/22/21 22:47, Yang Zhong wrote:
> >+ /* Trigger #NM exception */
> >+ __tileloadd(tiledata);
> >+ GUEST_SYNC(10);
> >+
> >+ GUEST_DONE();
> >+}
> >+
> >+void guest_nm_handler(struct ex_regs *regs)
> >+{
> >+ /* Check if #NM is triggered by XFEATURE_MASK_XTILEDATA */
> >+ GUEST_SYNC(7);
> >+ GUEST_ASSERT(rdmsr(MSR_IA32_XFD_ERR) == XFEATURE_MASK_XTILEDATA);
> >+ GUEST_SYNC(8);
> >+ GUEST_ASSERT(rdmsr(MSR_IA32_XFD_ERR) == XFEATURE_MASK_XTILEDATA);
> >+ /* Clear xfd_err */
> >+ wrmsr(MSR_IA32_XFD_ERR, 0);
> >+ regs->rip += 3;
> >+ GUEST_SYNC(9);
> >+}
>
> I don't understand why "regs->rip += 3" is needed though.
>
> My idea was that, after GUEST_SYNC(9) and IRET, the guest would
> execute __tileloadd again; this time without generating #NM, so that
> after GUEST_SYNC(10) the host sees the loaded data in TMM0.
>
Sorry, I didn't capture it before, and in that time, I wanted to skip this
load intruction.
Let me enbale amx before GUEST_SYNC(9), and then load tiles again.
The new v3 will address this, thanks!
Yang
> Thanks,
>
> Paolo
