Re: [PATCH] x86: always park offline CPUs

[George Dunlap <gwd@xxxxxxxxxxxxxx>]

On Wed, Aug 26, 2026 at 1:36 PM Jan Beulich <jbeulich@xxxxxxxx> wrote:
>
> While on AMD (or Hygon) CPUs the situation isn't as bad wrt broadcasting
> of #MC, some "multicast" can still happen. Therefore the reasoning to park
> CPUs rather than fully offlining them applies everywhere.
>
> Don't retain the dependency on the "mce=" cmdline option either: That
> option may best be dropped as well, as not enabling MCE will result in a
> shutdown when #MC would otherwise be raised.
>
> Drop the global variable, using a #define (just like common code does)
> instead. Outside of common code, simplify expressions / code accordingly.
> (In common code we still have to cater for x86 wanting it different from
> everyone else.)
>
> Suggested-by: Andrew Cooper <andrew.cooper3@xxxxxxxxxx>
> Signed-off-by: Jan Beulich <jbeulich@xxxxxxxx>
> ---
> As it was never actually used after its introduction, we may want to
> further consider dropping CPU_REMOVE again.
>
> I was almost certain that we would have at least one place (presumably a
> CPU notifier handler) were we assumed no parking for AMD/Hygon. Yet I
> couldn't find anything; did I overlook the crucial bits?

FWIW, the per-CPU stack mapping bit of the ASI series I have would
almost certainly have tripped over such an instance if it existed, but
didn't.

I've tested this patch on an Intel box, whose behavior in theory
shouldn't change.  Not sure if that warrants a Tested-by, given that
the main change should happen on an AMD box.

 -George