Re: [Xen-devel] MCE/EDAC Status/Updating?

[Andrew Cooper <andrew.cooper3@xxxxxxxxxx>]

On 15/02/2019 18:20, Elliott Mitchell wrote:
> On Fri, Feb 15, 2019 at 03:58:49AM -0700, Jan Beulich wrote:
>>>>> On 15.02.19 at 05:23, <ehem+xen@xxxxxxx> wrote:
>>> The MCE/EDAC support code appears to be in rather poor shape.
>>>
>>> The AMD code mentions Family 10h, which might have been available 10
>>> years ago.  They've likely been findable used with difficulty more
>>> recently, but no hardware made in the past 5 years matches this profile.
>> Well, Fam10 is mentioned explicitly, but as per the use of e.g.
>> mcheck_amd_famXX newer ones are supported by this code
>> as well.
> In that case sometime between Xen 4.1 and Xen 4.4, the AMD MCE/EDAC code
> was completely broken and hasn't been fixed.

I take it you've got a usecase which now doesn't work?

>>> Given the recent trends in Xen's development I'd tend to suggest going a
>>> different direction from the existing code.  The existing code was
>>> attempting to handle MCE/EDAC errors by emulating them and passing them
>>> to the effected domain.  Instead of this approach, let Domain 0 handle
>>> talking to MCE/EDAC hardware and merely have Xen decode addresses.
>>>
>>> If errors/warnings are occuring, you need those reports centralized,
>>> which points to handling them in Domain 0.  If an uncorrectable error
>>> occurs, Domain 0 should choose whether to kill a given VM or panic the
>>> entire machine.  Either way, Domain 0 really needs to be alerted that
>>> hardware is misbehaving and may need to be replaced.
>> But the point of the virtualization is to allow guests to more or less
>> gracefully recover (at least as far as the theory of it goes), e.g. by
>> killing just a process, rather than getting blindly killed.
>>
>> As to panic-ing the entire machine - if that's necessary, Dom0 is
>> unlikely to be in the right position. There's way too high a chance for
>> further things to go wrong until the event has even just arrived in
>> Dom0, let alone it having taken a decision.
> I'll agree it does make sense to try sending a corrupted memory alert to
> the effected domain, rather than nuking the entire VM.  Alerting the
> owner of the hardware though should be higher priority as they will then
> know they need to schedule a downtime and replace the module.
>
>
>>> The other part is alerting Domain 0 is *far* more likely to get the
>>> correct type of attention.  A business owning a Domain U on a random
>>> machine, may run a kernel without MCE/EDAC support or could miss the
>>> entries in their system log, nor would they necessarily know the correct
>>> personel to contact about hardware failing.
>> Alerting Dom0 alongside the affected DomU may indeed be desirable,
>> but mainly for the purpose of logging, only as a last resort for the
>> purpose of killing a guest.
> I think alerting Dom0 should be rather higher priority than alerting
> DomUs.  A given DomU may see one correctable memory error per month,
> which might seem harmless until you find there are a hundred DomUs on
> that hardware and every one of them is seeing one error per month.
>
> The only real useful place to report correctable errors like that is to
> Dom0.  Meanwhile uncorrectable errors are likely better to send a PV
> message to the DomU.  Let QEMU turn it into something which looks like
> real hardware if needed.  Meanwhile Dom0 may have a more up to date
> driver for the hardware than Xen.

I don't think anyone can defend the current state of MCE
handling/reporting in Xen, and I would certainly like to see it improved.

However, its not a simple as "let dom0 handle everything".  Dom0 is just
a VM, like all other domains.  It cant access the MCE MSR banks, and
even if it could, it would have a pcpu vs vcpu problem when trying to
interpret the data.

Xen is the entity which needs to handle the #MC, and do first-pass
processing.  If we want to give it to dom0 for further processing, it
either has to be virtualised in an architectural manner, or passed via a
paravirt channel so dom0 definitely knows it is dealing with data in
different enumeration spaces.

I expect there are also some non-trivial ACPI interactions here, which
are also made complicated by the Xen/dom0
interface-turned-undoucmented-mess.

Another issue we should look into is if we are going to make
improvements here, how do we go about ensuring that we don't regress
behaviour again.  I have no experience in this area (other than
bugfixing the #MC handler until it appears to behave as it did before),
but surely there are some ways of testing?

~Andrew

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