Re: [PATCH v2 2/5] x86/virt/tdx: Pull kexec cache flush logic into arch/x86

[Sean Christopherson <seanjc@google.com>]

On Wed, Apr 01, 2026, H. Peter Anvin wrote:
> On April 1, 2026 8:03:02 AM PDT, Dave Hansen <dave.hansen@intel.com> wrote:
> >On 3/31/26 16:04, Sean Christopherson wrote:
> >> But unless the WBINVD is actually costly, why bother getting fancy?
> >
> >WBINVD might be the most expensive single instruction in the whole ISA.
> >
> >That said, I'd much rather have a potentially unnecessary WBINVD than
> >miss one. The thing I'd be worried about would be something wonky like:
> >
> >	1. CPU offline does WBINVD
> >	2. Some other TDX call gets made, dirties caches again
> >	3. tdx_offline_cpu() skips WBINVD
> >
> >So, let's just do both for now: Do WBINVD in tdx_offline_cpu() and
> >comment that it might be redundant with other things in the CPU offline
> >procedure.
> >
> >This really needs to be solved with infrastructure and keeping data
> >about the reasons for needing WBINVD, not relying on code ordering or
> >fragile semantics.
> 
> It is, *by far*, the most expensive *uninterruptible* instruction in the ISA.
> REP string instructions can of course be arbitrarily long, but are
> interruptible and so don't really count.
> 
> Some MSRs used during very early (pre-OS) initialization might be even slower
> on some implementations, but that's not visible to Linux and no workload of
> any kind is running.

Sorry, "costly" wasn't the right word.  I know WBINVD super expensive, but unless
someone cares deeply about the latency of offlining a CPU after its down TDX stuff,
the "cost" is effectively zero.