On Thu, Feb 20, 2025 at 3:00 PM Dave Hansen <dave.hansen@xxxxxxxxx> wrote:
>
> On 2/20/25 13:16, Vishal Annapurve wrote:
> > Direct HLT instruction execution causes #VEs for TDX VMs which is routed
> > to hypervisor via TDCALL. safe_halt() routines execute HLT in STI-shadow
> > so IRQs need to remain disabled until the TDCALL to ensure that pending
> > IRQs are correctly treated as wake events.
>
> This isn't quite true. There's only one paravirt safe_halt() and it
> doesn't do HLT or STI.
pv_native_safe_halt() -> native_safe_halt() -> "sti; hlt".
>
> I think it's more true to say that "safe" halts are entered with IRQs
> disabled. They logically do the halt operation and then enable
> interrupts before returning.
>
> > So "sti;hlt" sequence needs to be replaced for TDX VMs with "TDCALL;
> > *_irq_enable()" to keep interrupts disabled during TDCALL execution.
> But this isn't new. TDX already tried to avoid "sti;hlt". It just
> screwed up the implementation.
>
> > Commit bfe6ed0c6727 ("x86/tdx: Add HLT support for TDX guests")
> > prevented the idle routines from using "sti;hlt". But it missed the
> > paravirt routine which can be reached like this as an example:
> > acpi_safe_halt() =>
> > raw_safe_halt() =>
> > arch_safe_halt() =>
> > irq.safe_halt() =>
> > pv_native_safe_halt()
>
> This, on the other hand, *is* important.
>
> > Modify tdx_safe_halt() to implement the sequence "TDCALL;
> > raw_local_irq_enable()" and invoke tdx_halt() from idle routine which just
> > executes TDCALL without toggling interrupt state. Introduce dependency
> > on CONFIG_PARAVIRT and override paravirt halt()/safe_halt() routines for
> > TDX VMs.
>
> This changelog glosses over one of the key points: Why *MUST* TDX use
> paravirt? It further confuses the reasoning by alluding to the idea that
> "Direct HLT instruction execution ... is routed to hypervisor via TDCALL".
>
> It gives background and a solution, but it's not obvious what the
> problem is or how the solution _fixes_ the problem.
>
> What must TDX now depend on PARAVIRT?
>
> Why not just route the HLT to a TDXCALL via the #VE code?
>
>
Makes sense, I will update the commit message in the next version to
clearly answer these questions and address the above comments.
