TLB flushing in x86's map_pages_to_xen()

TLB flushing in x86's map_pages_to_xen()

[Jan Beulich]

While I realize that the local-flush-only approach attempts to minimize overhead,
I have some difficulty understanding how this can properly deal with the effects
of speculation on remote processors. I believe this will get even worse when
considering cacheability changes, in which case it is not only an immediate TLB
flush that is needed, but also an immediate cache flush.

Thanks for any clarification, Jan

Re: TLB flushing in x86's map_pages_to_xen()

[Keir Fraser]

On 15/5/07 08:20, "Jan Beulich" <jbeulich@novell.com> wrote:

> While I realize that the local-flush-only approach attempts to minimize
> overhead,
> I have some difficulty understanding how this can properly deal with the
> effects
> of speculation on remote processors. I believe this will get even worse when
> considering cacheability changes, in which case it is not only an immediate
> TLB
> flush that is needed, but also an immediate cache flush.

Map_pages_to_xen() is only used at start-of-day, except for memory-guarding
where TLB flushing isn't actually required for correctness. If it were used
during normal Xen execution for more sophisticated purposes then yes, more
consideration would need to be given for flushing processor state.

 -- Keir