On Tue, Nov 16, 2021, Paolo Bonzini wrote:
> On 11/16/21 19:55, Thomas Gleixner wrote:
> > We can do that, but I'm unhappy about this conditional in schedule(). So
> > I was asking for doing a simple KVM only solution first:
> >
> > vcpu_run()
> > kvm_load_guest_fpu()
> > wrmsrl(XFD, guest_fpstate->xfd);
> > XRSTORS
> > do {
> >
> > local_irq_disable();
> >
> > if (test_thread_flag(TIF_NEED_FPU_LOAD))
> > switch_fpu_return()
> > wrmsrl(XFD, guest_fpstate->xfd);
> >
> > do {
> > vmenter(); // Guest modifies XFD
> > } while (reenter);
> >
> > update_xfd_state(); // Restore consistency
> >
> > local_irq_enable();
> >
> > and check how bad that is for KVM in terms of overhead on AMX systems.
>
> I agree, this is how we handle SPEC_CTRL for example and it can be extended
> to XFD. We should first do that, then switch to the MSR lists. Hacking
> into schedule() should really be the last resort.
Agreed as well.
> > local_irq_enable(); <- Problem starts here
> >
> > preempt_enable(); <- Becomes wider here
>
> It doesn't become that much wider because there's always preempt notifiers.
> So if it's okay to save XFD in the XSAVES wrapper and in
> kvm_arch_vcpu_put(), that might be already remove the need to do it
> schedule().
Assuming AMX can be accessed from (soft) IRQ context, hooking the preempt notifiers
isn't sufficient. That's also why KVM waits until IRQs are disabled before
handling TIF_NEED_FPU_LOAD.
