Re: [PATCH] x86/virt/tdx: Make TDX and kexec mutually exclusive at runtime

["Edgecombe, Rick P" <rick.p.edgecombe@intel.com>]

On Wed, 2025-05-07 at 14:03 -0700, Rick Edgecombe wrote:
> Kai and I have been discussing this internally. Here I'll try to move the
> discussion external and continue it.
> 
> The problem with this approach turned out to be per_cpu(newbool) = 1 part is
> racy with the SEAMCALL. The task could reschedule between SEAMCALL and the per-
> cpu set. Disabling preemption around every SEAMCALL should be possible, but it
> seems a bit heavyweight for what originally appeared to be an easy way to reduce
> but not eliminate the chances of hitting the race.


Well, after more offline discussion, this time with Dave, we'll go with the per-
cpu approach. The reasons are:
 - global var is not horrible, but...
 - preempt_disable/enable() is not heavyweight. Since the SEAMCALL is not
preemptible, the non-preemptible cycles difference is only a few instructions.
so "preempt_disable is bad" is not a good reason.
 - Is the complexity still too high for what we are trying to do here (shrink a
race)? No, and it's simpler to understand. wbinvd is per-cpu, so the
needs_to_flush bool should also be per-cpu.

So back to the per-cpu approach.