Messages in this thread | | | From | Dan Williams <> | Date | Tue, 8 Jun 2021 21:21:49 -0700 | Subject | Re: [RFC v2-fix-v4 1/1] x86/tdx: Skip WBINVD instruction for TDX guest |
| |
On Tue, Jun 8, 2021 at 9:02 PM Andy Lutomirski <luto@kernel.org> wrote: > > On 6/8/21 8:40 PM, Dan Williams wrote: > > On Tue, Jun 8, 2021 at 6:10 PM Kuppuswamy Sathyanarayanan > > <sathyanarayanan.kuppuswamy@linux.intel.com> wrote: > >> > >> Current TDX spec does not have support to emulate the WBINVD > >> instruction. If any feature that uses WBINVD is enabled/used > >> in TDX guest, it will lead to un-handled #VE exception, which > >> will be handled as #GP fault. > >> > >> ACPI drivers also uses WBINVD instruction for cache flushes in > >> reboot or shutdown code path. Since TDX guest has requirement > >> to support shutdown feature, skip WBINVD instruction usage > >> in ACPI drivers for TDX guest. > > > > This sounds awkward... > > > >> Since cache is always coherent in TDX guests, making wbinvd as > > > > This is incorrect, ACPI cache flushing is not about I/O or CPU coherency... > > > >> noop should not cause any issues in above mentioned code path. > > > > ..."should" is a famous last word... > > > >> The end-behavior is the same as KVM guest (treat as noops). > > > > ..."KVM gets away with it" is not a justification that TDX can stand > > on otherwise we would not be here fixing up ACPICA properly. > > > > How about: > > > > "TDX guests use standard ACPI mechanisms to signal sleep state entry > > (including reboot) to the host. The ACPI specification mandates WBINVD > > on any sleep state entry with the expectation that the platform is > > only responsible for maintaining the state of memory over sleep > > states, not preserving dirty data in any CPU caches. ACPI cache > > flushing requirements pre-date the advent of virtualization. Given TDX > > guest sleep state entry does not affect any host power rails it is not > > required to flush caches. The host is responsible for maintaining > > cache state over its own bare metal sleep state transitions that > > power-off the cache. If the host fails to manage caches over its sleep > > state transitions the guest..." > > > > I like this description, but shouldn't the logic be: > > if (!CPUID has hypervisor bit set) > wbinvd(); > > As far as I know, most hypervisors will turn WBINVD into a noop and, > even if they don't, it seems to be that something must be really quite > wrong for a guest to need to WBINVD for ACPI purposes.
Agree, a well behaved guest should not pretend its callouts to the virtual ACPI BIOS actually affect a host power rail.
| |