amd64/186061: FreeBSD 10 crashes as KVM guest on GNU/Linux on AMD family 10h CPUs

[Simon Matter]

> On Wednesday, February 12, 2014 4:04:53 pm Simon Matter wrote:
>> > On Wednesday, February 12, 2014 2:40:01 am Simon Matter wrote:
>> >> The following reply was made to PR amd64/186061; it has been noted by
>> >> GNATS.
>> >>
>> >> From: "Simon Matter" <simon.matter at invoca.ch>
>> >> To: bug-followup at FreeBSD.org
>> >> Cc: simon.matter at invoca.ch
>> >> Subject: Re: amd64/186061: FreeBSD 10 crashes as KVM guest on
>> GNU/Linux
>> >> on
>> >>  AMD family 10h CPUs
>> >> Date: Wed, 12 Feb 2014 08:30:51 +0100
>> >>
>> >>  ------=_20140212083051_97180
>> >>  Content-Type: text/plain; charset="iso-8859-1"
>> >>  Content-Transfer-Encoding: 8bit
>> >>
>> >>  As noted by John Baldwin the change to mca.c is not needed. Attached
>> >> patch
>> >>  is what I'm using now with success.
>> >>
>> >>  BTW: setting vm.pmap.pg_ps_enabled="0" in loader.conf does also
>> >> mitigate
>> >>  the issue but I guess it's not the optimal solution.
>> >
>> > Talking with Alan Cox, we do think the right fix is to change the test
>> to
>> > enable the workaround.  However, we'd rather not penalize VM's on
>> other
>>
>> I'm afraid that will not work in all situations, no matter how good the
>> tests are (see below why I think so). So as a last resort, I suggest
>> that
>> it should be possible to enable the "AMD Erratum 383" workaround via
>> loader.conf.
>
> I think you misunderstand.  We would only use flags that are never set
> on an AMD 10h CPU, so they can never be set in a KVM guest that would
> ever migrate to an AMD 10h CPU.  If those flags are present, we know that
> we would _not_ need the workaround.  If none of those flags are present,
> we would enable the workaround.  Does that make sense?

OK, I think I understand now. Unfortunately I don't know enough about CPU
flags to know how well this could work. If I understand correctly, KVM can
also emulate some CPU features which are not implemented in the real CPU.
If that's true then it becomes quite complicated I guess.

What about having a sysctl flag so that the workaround can be enabled
manually? Wouldn't that make sense for cases where auto detection doesn't
work well.

Regards,
Simon