Re: [PATCH v3 32/40] KVM: x86: Move LLDT assembly wrappers into VMX

[Yosry Ahmed]

On Mon, Jun 01, 2026 at 08:17:55AM -0700, Sean Christopherson wrote:
> On Sat, May 30, 2026, Yosry Ahmed wrote:
> > On Fri, May 29, 2026 at 03:22:15PM -0700, Sean Christopherson wrote:
> > > Move kvm_{load,read}_ldt() into vmx.c, as they are exclusively used by VMX
> > > to save/restore host state, and have no business being globally visible.
> > 
> > But they are generic helpers. I agree with mvoing them out of
> > kvm_host.h, but maybe into another header? 
> 
> No, absolutely not.  The fact that KVM has to manually save, and conditionally
> restore, LDT is a quirk of the VMX architecture.  If it weren't for the fact that
> VMX's segment_base() also needs to query the LDT, I would just open code the asm().
> 
> If it's a naming concern, I'll happily rename them to vmx_{load,store}_ldt().

I was actually going to suggest those exact names initially, so yeah
this looks better.

> 
> Hmm, arch/x86/include/asm/desc.h provides {load,store}_ldt(), but load_ldt()
> is only available for CONFIG_PARAVIRT_XXL=n builds, and both #define unnecessarily
> constrain the output to memory.  Those are both fixable, e.g. load_ldt() isn't
> used _anywhere_ AFAICT, so burying it under CONFIG_PARAVIRT_XXL=n is just some
> weird historical quirk, and I can't show_fault_oops() would care store_ldt() were
> allowed to store to a register.
> 
> So after the dust settles, it probably makes sense to improve desc.h's versions
> and then use those in KVM.  But I don't want to do that in this series, because
> it will be challenging enough to land all of this code movement without also
> having to coordinate with the tip tree.

Even better, makes sense for a separate cleanup. Thanks.