On Mon, Oct 18, 2021, Borislav Petkov wrote:
On Wed, Oct 13, 2021 at 03:22:30PM +0100, Jane Malalane wrote: Totally untested of course.
static void early_probe_null_seg_clearing_base(struct cpuinfo_x86 *c) { /* * A hypervisor may sythesize the bit, but may also hide it * for migration safety, so do not probe for model-specific * behaviour when virtualised. */ if (cpu_has(c, X86_FEATURE_HYPERVISOR))
This isn't correct. When running as a guest, the intended behavior is to fully trust the CPUID.0x80000021 bit. If bit 6 is set, yay, the hypervisor has told the kernel that it will only ever run on hardware without the bug. If bit 6 is clear and HYPERVISOR is true, then the FMS crud can't be trusted because the kernel _may_ run on affected hardware in the future even if the current underlying hardware is not affected.
return;/* Zen3 CPUs advertise Null Selector Clears Base in CPUID. */ if (c->extended_cpuid_level >= 0x80000021 && cpuid_eax(0x80000021) & BIT(6)) return;
/* Zen2 CPUs also have this behaviour, but no CPUID bit. */ if (c->x86 == 0x17 && check_null_seg_clears_base(c)) return;
/* All the remaining ones are affected */ set_cpu_bug(c, X86_BUG_NULL_SEG); }
...
@@ -1457,8 +1457,6 @@ static void generic_identify(struct cpuinfo_x86 *c) get_model_name(c); /* Default name */
- detect_null_seg_behavior(c);
- /*
- ESPFIX is a strange bug. All real CPUs have it. Paravirt
- systems that run Linux at CPL > 0 may or may not have the
So this function is called on all x86 CPUs. Are you sure others besides AMD and Hygon do not have the same issue?
IOW, I wouldn't remove that call here.
I agree. If the argument for this patch is that the kernel can be migrated to older hardware, then it stands to reason that the kernel could also be migrated to a different CPU vendor entirely. E.g. start on Intel, migrate to Zen1, kaboom.