Re: [Xen-devel] [PATCH 1/3] x86/boot: Remove cached CPUID data from the trampoline

[Andrew Cooper <andrew.cooper3@xxxxxxxxxx>]

On 04/11/2019 15:31, Jan Beulich wrote:
> On 04.11.2019 16:22, Andrew Cooper wrote:
>> On 04/11/2019 15:03, Jan Beulich wrote:
>>> On 04.11.2019 15:59, Andrew Cooper wrote:
>>>> On 04/11/2019 13:25, Jan Beulich wrote:
>>>>> On 01.11.2019 21:25, Andrew Cooper wrote:
>>>>>> --- a/xen/arch/x86/cpu/intel.c
>>>>>> +++ b/xen/arch/x86/cpu/intel.c
>>>>>> @@ -270,6 +270,7 @@ static void early_init_intel(struct cpuinfo_x86 *c)
>>>>>>          if (disable) {
>>>>>>                  wrmsrl(MSR_IA32_MISC_ENABLE, misc_enable & ~disable);
>>>>>>                  bootsym(trampoline_misc_enable_off) |= disable;
>>>>>> +                bootsym(trampoline_efer) |= EFER_NX;
>>>>>>          }
>>>>> I'm fine with all other changes here, just this one concerns me:
>>>>> Before your change we latch a value into trampoline_misc_enable_off
>>>>> just to be used for subsequent IA32_MISC_ENABLE writes we do. This
>>>>> means that, on a hypervisor (like Xen itself) simply discarding
>>>>> guest writes to the MSR (which is "fine" with the described usage
>>>>> of ours up to now), with your change we'd now end up trying to set
>>>>> EFER.NX when the feature may not actually be enabled in
>>>>> IA32_MISC_ENABLE. Wouldn't such an EFER write be liable to #GP(0)?
>>>>> I.e. don't we need to read back the MSR value here, and verify
>>>>> we actually managed to clear the bit before actually OR-ing in
>>>>> EFER_NX?
>>>> If this is a problem in practice, execution won't continue beyond the
>>>> next if() condition just out of context, which set EFER.NX on the BSP
>>>> with an unguarded WRMSR.
>>> And how is this any good? This kind of crash is exactly what I'm
>>> asking to avoid.
>> What is the point of working around a theoretical edge case of broken
>> nesting under Xen which demonstrably doesn't exist in practice?
> Well, so far nothing was said about this not being an actual problem.

Its not an actual problem.  If it were, we would have had crash reports.

> I simply don't know whether hardware would refuse such an EFER write.

I've just experimented - writing EFER.NX takes a #GP fault when
MISC_ENABLE.XD is set.

> If it does, it would be appropriate for hypervisors to also refuse
> it. I.e. if we don't do so right now, correcting the behavior would
> trip the code here.

MISC_ENABLES.XD is architectural on any Intel system which enumerates
NX, and if the bit is set, it can be cleared.  (Although the semantics
described in the SDM are broken.  It is only available if NX is
enumerated, which is obfuscated by setting XD).

However, no hypervisor is going to bother virtualising this
functionality.  Either configure the VM with NX or without.  (KVM for
example doesn't virtualise MISC_ENABLES at all.)

There is one corner case on out-of-support versions of Xen (which don't
clear XD themselves) where XD would leak through and be ignored, after
which Xen will take a #GP fault trying to set EFER.NX, but I am still
firmly of the opinion that it is not worth putting in a workaround for
an obsolete issue which doesn't exist in practice.

~Andrew

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/xen-devel