On 13/06/18 07:30, Jan Beulich wrote:
>>>> On 12.06.18 at 18:33, <andrew.coop...@citrix.com> wrote:
>> On 12/06/18 10:00, Jan Beulich wrote:
>>>>>> On 12.06.18 at 10:51, <andrew.coop...@citrix.com> wrote:
>>>> On 12/06/2018 09:15, Jan Beulich wrote:
>>>>>>>> On 08.06.18 at 20:48, <andrew.coop...@citrix.com> wrote:
>>>>>> @@ -3106,14 +3104,13 @@ static int vmx_msr_write_intercept(unsigned int
>>>>>> msr,
>>>> uint64_t msr_content)
>>>>>> for ( ; (rc == 0) && lbr->count; lbr++ )
>>>>>> for ( i = 0; (rc == 0) && (i < lbr->count); i++ )
>>>>>> if ( (rc = vmx_add_guest_msr(v, lbr->base + i)) ==
>>>>>> 0 )
>>>>>> - {
>>>>>> vmx_clear_msr_intercept(v, lbr->base + i,
>>>> VMX_MSR_RW);
>>>>>> - if ( lbr_tsx_fixup_needed )
>>>>>> - v->arch.hvm_vmx.lbr_fixup_enabled |=
>>>> FIXUP_LBR_TSX;
>>>>>> - if ( bdw_erratum_bdf14_fixup_needed )
>>>>>> - v->arch.hvm_vmx.lbr_fixup_enabled |=
>>>>>> - FIXUP_BDW_ERRATUM_BDF14;
>>>>>> - }
>>>>>> +
>>>>>> + v->arch.hvm_vmx.lbr_flags |= LBR_MSRS_INSERTED;
>>>>>> + if ( lbr_tsx_fixup_needed )
>>>>>> + v->arch.hvm_vmx.lbr_flags |= LBR_FIXUP_TSX;
>>>>>> + if ( bdw_erratum_bdf14_fixup_needed )
>>>>>> + v->arch.hvm_vmx.lbr_flags |= LBR_FIXUP_BDF14;
>>>>> Note how the setting of the flags previously depended on
>>>>> vmx_add_guest_msr() having returned success at least once.
>>>> And?
>>>>
>>>> Unless this sequence returns fully successfully, we throw #MC into the
>>>> guest without setting any kind of vMCE state. It might be the least bad
>>>> option we have available, but its also not reasonable to expect the
>>>> guest to survive.
>>>>
>>>> The two ways to fail are ENOMEM which E2BIG. The former is going to be
>>>> causing other forms of chaos, and the latter isn't going to occur in
>>>> practice because current codepaths in Xen use a maximum of ~40 or the
>>>> 256 available slots. If in the unlikely case that we fail with ENOMEM
>>>> on the first entry, all the fixup logic gets short circuited due to the
>>>> missing memory allocation (so practically 0 extra overhead), and the
>>>> guest will still malfunction.
>>>>
>>>> The error handling here is sufficiently poor that I'm not worried about
>>>> changing one minor corner case. I'm actually debating whether it would
>>>> be better to make the allocation at vmcs construction time, to avoid
>>>> runtime out-of-memory issues.
>>> With further improved MSR handling down the road, I assume we'll
>>> have some entries in the list in almost all cases, so yes, I think that
>>> would be desirable.
>> For performance reasons, we'll want to keep the size of the lists to an
>> absolute minimum.
>>
>> On a closer inspection, the only uses we currently have for the
>> load/save lists are this new EFER case (on Gen1 hardware), the Global
>> Perf Ctl (for vPMU, and we really should be using the load/save support
>> like EFER), and the LBR MSRs.
>>
>> Therefore, for on non-ancient hardware, a guest which doesn't touch
>> MSR_DEBUGCTL is not going to need the memory allocation, so perhaps an
>> up-front allocation isn't the wisest of options. I'll keep this in mind
>> during the MSR work.
> Hmm, okay. In which case, if we anyway don't expect the guest to
> survive here in case of some failure, why don't we crash it right away
> instead of injecting #MC? Would make for a much easier to recognize
> cause of the guest crash.
Yeah - I was considering that as an option. -ENOSPC is definitely a
hypervisor coding issue, and can't be remedied in place. -ENOMEM is
unfortunate if a guest happens to hit it, but cleanly crashing is better
behaviour than the current #MC with no information.
~Andrew
_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel
