Re: [patch V4 37/37] x86/smpboot/64: Implement arch_cpuhp_init_parallel_bringup() and enable it

[Peter Zijlstra <peterz@infradead.org>]

On Fri, May 12, 2023 at 11:07:56PM +0200, Thomas Gleixner wrote:
> From: Thomas Gleixner <tglx@linutronix.de>
> 
> Implement the validation function which tells the core code whether
> parallel bringup is possible.
> 
> The only condition for now is that the kernel does not run in an encrypted
> guest as these will trap the RDMSR via #VC, which cannot be handled at that
> point in early startup.
> 
> There was an earlier variant for AMD-SEV which used the GHBC protocol for
> retrieving the APIC ID via CPUID, but there is no guarantee that the
> initial APIC ID in CPUID is the same as the real APIC ID. There is no
> enforcement from the secure firmware and the hypervisor can assign APIC IDs
> as it sees fit as long as the ACPI/MADT table is consistent with that
> assignment.
> 
> Unfortunately there is no RDMSR GHCB protocol at the moment, so enabling
> AMD-SEV guests for parallel startup needs some more thought.

One option, other than adding said protocol, would be to:

 - use the APICID from CPUID -- with the expectation that it can be
   wrong.
 - (ab)use one of the high bits in cpuid_to_apicid[] as a test-and-set
   trylock. This avoids two CPUs from using the same per-cpu base, if
   CPUID is being malicious. Panic on fail.
 - validate against MSR the moment we can and panic if not matching

The trylock ensures the stacks/percpu state is not used by multiple
CPUs, and should guarantee a coherent state to get far enough along to
be able to do the #VE inducing RDMSR.