Re: [Qemu-devel] [PATCH v4 2/3] target-i386: add migration support for Intel LMCE

[Haozhong Zhang <haozhong.zhang@intel.com>]

On 06/17/16 14:20, Eduardo Habkost wrote:
> On Fri, Jun 17, 2016 at 10:01:05AM +0800, Haozhong Zhang wrote:
> > On 06/16/16 14:58, Eduardo Habkost wrote:
> > > On Thu, Jun 16, 2016 at 07:40:20PM +0200, Paolo Bonzini wrote:
> > > > 
> > > > 
> > > > On 16/06/2016 19:36, Eduardo Habkost wrote:
> > > > >> > 
> > > > >> > Eduardo said nice for this part in previous version [1], so we may wait
> > > > >> > for his comments?
> > > > >> > 
> > > > >> > [1] http://lists.nongnu.org/archive/html/qemu-devel/2016-06/msg01992.html
> > > > > I agree we don't need this check, but I still believe it is a
> > > > > nice thing to have.
> > > > > 
> > > > > In addition to detecting user errors, they don't hurt and are
> > > > > useful for things like "-cpu host", that don't guarantee
> > > > > live-migration compatibility but still allow migration if you
> > > > > ensure host capabilities are the same on both sides.
> > > > 
> > > > On the other hand we don't check for this on any other property, either
> > > > CPU or device, do we?  Considering "lmce=on" always breaks on an old
> > > > kernel (i.e. there's no need for an explicit ",enforce" on the -cpu
> > > > flag), I think it's unnecessary and makes things inconsistent.
> > > 
> > > We don't check that because we normally can't: we usually don't
> > > send any configuration data (or anything that could be used to
> > > detect configuration mismatches) to the destination. When we do,
> > > it's often by accident.
> > > 
> > > In this case, it looks like we never needed to send mcg_cap in
> > > the migration stream. But we already send it, so let's use it for
> > > something useful.
> > > 
> > > I believe we should have more checks like these, when possible. I
> > > have been planning for a while to send CPUID data in the
> > > migration stream, to detect migration compatibility errors
> > > (either user errors or QEMU bugs).
> > > 
> > > In theory, those checks should never be necessary. In practice I
> > > believe they would be very useful.
> > >
> > 
> > Hi Eduardo and Paolo,
> > 
> > What will be the conclusion? Do we still need this check?
> > 
> > I'm fine to remove this check if we normally didn't make such kind of
> > checks and require users to avoid configuration mismatch.
> 
> I don't know yet if Paolo is convinced that the check is still
> useful. :)
> 
> I suggest doing it as a separate patch, so we can apply the rest
> of the series now and discuss/apply the check later.
>

Yes, I'll move the check to a separate patch so that we can easily
drop it if not necessary. Thanks for the suggestion!

> > 
> > > > 
> > > > > (I was going to suggest enabling lmce automatically on "-cpu
> > > > > host" as a follow-up patch, BTW.)
> > > > 
> > > > Interesting.  Technically it comes from the host kernel, not from the
> > > > host CPU.  But it does sounds like a good idea; -cpu host pretty much
> > > > implies the same kernel (in addition to the same processor) on both
> > > > sides of the migration.
> > > 
> > > "-cpu host" already means "whatever is allowed by the host [CPU
> > > and/or kernel]", not just "host CPU". It enables x2apic on all
> > > hosts, for example.
> > >
> > 
> > Does that mean we can automatically enable LMCE for "-cpu host"?
> 
> We can automatically enable LMCE for "-cpu host" if and only if
> the host kernel supports LMCE.
>

According to our discussion for KVM Patch 3, we may have to disable it
by default by -cpu host, so that pc-2.7 will not require new kernels
unless LMCE is required explicitly by users.

Thanks,
Haozhong