[MODERATED] Re: Interrupts policy for SMT

[Andi Kleen <ak@linux.intel.com>]

>  Of course it is impossible to handle userspace, but the same handwavy
> justifications used for (3) apply even more there, since userspace
> should hardly run on a properly configured guest.

What do you mean?  user space on the host would be handled by
the scheduler in this case (e.g. exclusive cpuset)

This proposal was merely about interrupts.

> Since we are at it, we could add a sticky bit "I am in an interrupt"
> sticky (bit 0 = am I in an interrupt, bit 1 = have I been in an
> interrupt since the flag was cleared).  Then:
> 
> - the interrupt handler writes 3 on entry and 2 on exit;
> 
> - KVM writes 0 just before setting IF=1 and checks the sticky bit just
> before vmentry to check whether to do a L1D flush.

Not sure what the point of this is, but if we wanted something
like that the easiest would be to maintain a interrupt count
per CPU.

-Andi