Re: [Qemu-devel] CPUID feature bits not saved with migration

[Andre Przywara]

(Sorry for the late reply, I had some mail troubles)

Jamie Lokier wrote:
> Andre Przywara wrote:
> > Jamie Lokier wrote:
> > > Anthony Liguori wrote:
> > > > It's unclear what to do about -cpu host.  If we did migrate cpuid 
> > > > values, then -cpu would effectively be ignored after an incoming 
> > > > migration.
> > > The new host might not support all the cpuid features of the old host,
> > > whether by -cpu host or explicit cpuid.  What happens then?
> > If you plan to migrate, you should think of this in advance.
>
> In my experience so far, for small sites, you don't plan migration for
> 2-3 years after starting the VM, because it's only then you realise
> your host hardware is quite old, you buy a replacement to consolidate,
> and you are still running the VM that you didn't know would still be
> mission critical years later.
That is one use-case for live migration. Another would be a migration
pool with lots of machines each running some VMs. If one host is loaded,
you can migrate to a lesser loaded one. Think of a hoster or cloud like
environment.

> At least, that's been my experience so far.  I've "cold migrated" a
> few VMs, and in some cases from non-virtual machines to VMs.  None of
> these could be planned when the guest was first installed, especially
> the ones where it wasn't realised the guest would outlive the host
> hardware.
Fortunately it seems like that newer CPUs only _add_ CPUID bits, so this
should not be a problem.

> I have to say, unfortunately hot migration has never been an option
> because the version of KVM running on the new host is invariably
> incompatible with the KVM running on the old host.
So far I have only seen problems like this if the target host KVM
version is older than the source one. Some of these issues could be
overcome by putting a translator application between source and target,
but I am not sure whether the effort is worth the results.
What kind of issues do you see? Are you migrating from newer KVMs to
older ones?

> > I have a rough version of a tool to compute the least common
> > denominator CPUID bits given either processor (code)names or host
> > names. In the latter case it will login into the box and query the
> > host's CPUID. The tool then generates a QEMU command line (like -cpu
> > qemu64,-mwait,-popcnt) with which the guest should be started.  This
> > should ensure that the guest always sees the same subset of the CPUs
> > capabilities.
>
> I wonder how that would be useful.  Don't you usually migrate only
> when you've acquired new hardware, who's specs you don't know at the
> time you'd want to compute the initial CPUID?
See above, this tool helps to enlarge the migration pool by computing
the best possible bit set of CPU features.

> > > For changing cpuid when migrating, as you might like to do with -cpu
> > > host for performance, is reboot-during-migrate useful?  It would make
> > > sure all disk state is committed to the image files asynchronously
> > > while the machine continues to run (just like normal migration), and
> > > at the last moment transfers control and the machine sees a reboot,
> > > permitting devices changes including cpuid change.
> > Is that really useful? After all the sexy part of live migration is the 
> > "live" component...
>
> The other sexy part is being easy and asynchronous: not stopping the
> guest for a long time during the migration.
>
> Easy: Now, at the moment you have to give all the right guest
> configuration on the destination command line, so I take your point.
>
> But if guest configuration is ever included in the saved state for
> migration, migration will really easy.  I hope it's just as easy to do
> "cold migration".
Agreed. We should have a savevm section transferring the guest config.
> Async: Do we save RAM state across reboots normally?  I know of OSes
> which expect at least some part of RAM to survive reboots, so killing
> the VM and restarting on another host would change the behaviour,
> compared with rebooting locally; that's not transparent migration,
> it's a subtle, unexpected behaviour change.  Unfortunately doing the
> right thing involves savev, which pauses the guests for a long time.
> The pause comes from saving and loading RAM, something which migration
> handles well.
Have you seen any real life problems with this? What are these OSes?

> There's also the small matter of migration having a totally different
> interface compared with savevm right now, with savevm requiring a
> dummy qcow2 disk while migration transfers control across a network
> with no temporary file.
You are right, that is pretty unfortunate. I worked around this 
limitation by using the exec: prefix with migrate to let a shell script 
dump the migration stream to disk, with the same trick you can reload 
the state again. That worked pretty well for me in the past.

> Guess which one is nicer for the user wanting "move my VM to host FOO
> (which doesn't support SSE4) with minimal downtime".
>
> > > CPU hotplug could be used for cpuid change in theory, but I doubt if
> > > any guests or guest apps would handle it well.
> > Hotplugging could work for secondary processors, but hotplugging the BSP 
> > is kind of tricky. And this does not solve the userspace issues, where 
> > libraries detect CPU capabilities during startup and use optimized code 
> > paths. AFAIK there is no mechanism of informing those libraries about a 
> > CPUID change.
>
> I agree, and it's pointless to spend much time discussion hotplug for this.
> Most guests wouldn't handle CPUs with mixed CPUIDs anyway.
>
> In theory, sometimes it'd be ok to push that problem to the user: they
> can stop and start specific apps under user control without bringing
> down a whole machine, and most apps don't use cpuid-dependent
> features, especially on servers.
But libraries do, even on servers.

> Btw, why can't hotplugging the main processor work?  For (real)
> high-reliability systems, all processors are hotpluggable, afaik.
Yes, but if you look into the Linux code, you will find many assumptions 
about the BSP not being hotpluggable. For instance you cannot offline 
CPU0 (take a look at /sys/devices/system/cpu/cpu0, there is no online, 
in opposite to all other cpu<n>). I haven't looked in detail, but I 
assume that there is no real showstopper, it is just the current code 
design that makes offlining CPU0 hard.
BTW, do you know of any x86 machines which really allow physical CPU 
hotplugging?

Regards,
Andre.

--
Andre Przywara
AMD-Operating System Research Center (OSRC), Dresden, Germany
Tel: +49 351 448 3567 12
----to satisfy European Law for business letters:
Advanced Micro Devices GmbH
Karl-Hammerschmidt-Str. 34, 85609 Dornach b. Muenchen
Geschaeftsfuehrer: Thomas M. McCoy; Giuliano Meroni
Sitz: Dornach, Gemeinde Aschheim, Landkreis Muenchen
Registergericht Muenchen, HRB Nr. 43632