Ways to deal with broken machine types

[Dr. David Alan Gilbert]

* Igor Mammedov (imammedo at redhat.com) wrote:
> On Tue, 23 Mar 2021 17:40:36 +0000
> Daniel P. Berrangé <berrange at redhat.com> wrote:
> 
> > On Tue, Mar 23, 2021 at 05:54:47PM +0100, Igor Mammedov wrote:
> > > Let me hijack this thread for beyond this case scope.
> > > 
> > > I agree that for this particular bug we've done all we could, but
> > > there is broader issue to discuss here.
> > > 
> > > We have machine versions to deal with hw compatibility issues and that covers most of the cases,
> > > but occasionally we notice problem well after release(s),
> > > so users may be stuck with broken VM and need to manually fix configuration (and/or VM).
> > > Figuring out what's wrong and how to fix it is far from trivial. So lets discuss if we
> > > can help to ease this pain, yes it will be late for first victims but it's still
> > > better than never.  
> > 
> > To summarize the problem situation
> > 
> >  - We rely on a machine type version to encode a precise guest ABI.
> >  - Due a bug, we are in a situation where the same machine type
> >    encodes two distinct guest ABIs due to a mistake introduced
> >    betwen QEMU N-2 and N-1
> >  - We want to fix the bug in QEMU N
> >  - For incoming migration there is no way to distinguish between
> >    the ABIs used in N-2 and N-1, to pick the right one
> > 
> > So we're left with an unwinnable problem:
> > 
> >   - Not fixing the bug =>
> > 
> >        a) user migrating N-2 to N-1 have ABI change
> >        b) user migrating N-2 to N have ABI change
> >        c) user migrating N-1 to N are fine
> > 
> >     No mitigation for (a) or (b)
> > 
> >   - Fixing the bug =>
> > 
> >        a) user migrating N-2 to N-1 have ABI change.
> >        b) user migrating N-2 to N are fine
> >        c) user migrating N-1 to N have ABI change
> > 
> >     Bad situations (a) and (c) are mitigated by
> >     backporting fix to N-1-stable too.
> > 
> > Generally we have preferred to fix the bug, because we have
> > usually identified them fairly quickly after release, and
> > backporting the fix to stable has been sufficient mitigation
> > against ill effects. Basically the people left broken are a
> > relatively small set out of the total userbase.
> > 
> > The real challenge arises when we are slow to identify the
> > problem, such that we have a large number of people impacted.
> > 
> > 
> > > I'll try to sum up idea Michael suggested (here comes my unorganized brain-dump),
> > > 
> > > 1. We can keep in VM's config QEMU version it was created on
> > >    and as minimum warn user with a pointer to known issues if version in
> > >    config mismatches version of actually used QEMU, with a knob to silence
> > >    it for particular mismatch.
> > > 
> > > When an issue becomes know and resolved we know for sure how and what
> > > changed and embed instructions on what options to use for fixing up VM's
> > > config to preserve old HW config depending on QEMU version VM was installed on.  
> > 
> > > some more ideas:
> > >    2. let mgmt layer to keep fixup list and apply them to config if available
> > >        (user would need to upgrade mgmt or update fixup list somehow)
> > >    3. let mgmt layer to pass VM's QEMU version to currently used QEMU, so
> > >       that QEMU could maintain and apply fixups based on QEMU version + machine type.
> > >       The user will have to upgrade to newer QEMU to get/use new fixups.  
> > 
> > The nice thing about machine type versioning is that we are treating the
> > versions as opaque strings which represent a specific ABI, regardless of
> > the QEMU version. This means that even if distros backport fixes for bugs
> > or even new features, the machine type compatibility check remains a
> > simple equality comparsion.
> > 
> > As soon as you introduce the QEMU version though, we have created a
> > large matrix for compatibility. This matrix is expanded if a distro
> > chooses to backport fixes for any of the machine type bugs to their
> > stable streams. This can get particularly expensive when there are
> > multiple streams a distro is maintaining.
> > 
> > *IF* the original N-1 qemu has a property that could be queried by
> > the mgmt app to identify a machine type bug, then we could potentially
> > apply a fixup automatically.
> > 
> > eg query-machines command in QEMU version N could report against
> > "pc-i440fx-5.0", that there was a regression fix that has to be
> > applied if property "foo" had value "bar".
> > 
> > Now, the mgmt app wants to migrate from QEMU N-2 or N-1 to QEMU N.
> > It can query the value of "foo" on the source QEMU with qom-get.
> > It now knows whether it has to override this property "foo" when
> > spawning QEMU N on the target host.
> > 
> > Of course this doesn't help us if neither N-1 or N-2 QEMU had a
> > property that can be queried to identify the bug - ie if the
> > property in question was newly introduced in QEMU N to fix the
> > bug.
> > 
> > > In my opinion both would lead to explosion of 'possibly needed' properties for each
> > > change we introduce in hw/firmware(read ACPI) and very possibly a lot of conditional
> > > branches in QEMU code. And I'm afraid it will become hard to maintain QEMU =>
> > > more bugs in future.
> > > Also it will lead to explosion of test matrix for downstreams who care about testing.
> > > 
> > > If we proactively gate changes on properties, we can just update fixup lists in mgmt,
> > > without need to update QEMU (aka Insite rules) at a cost of complexity on QMEU side.
> > > 
> > > Alternatively we can be conservative in spawning new properties, that means creating
> > > them only when issue is fixed and require users to update QEMU, so that fixups could
> > > be applied to VM.
> > > 
> > > Feel free to shoot the messenger down or suggest ways how we can deal with the problem.  
> > 
> > The best solution is of course to not have introduced the ABI change in
> > the first place. We have lots of testing, but upstream at least, I don't
> > think we have anything that is explicitly recording the ABI associated
> > with each machine type and validating that it hasn't changed. We rely on
> > the developers to follow the coding practices wrt setting machine type
> > defaults for back compat, and while we're good, we inevitably screw up
> > every now & then.
> > 
> > Downstreams do have some of this ABI testing - several problems like the
> > one we have there, have been identified when RHEL downstream QE did
> > migration tests and found a change in RHEL machine types, which then
> > was traced back to upstream.
> > 
> > I feel like we need some standard tool which can be run inside a VM
> > that dumps all the possible ABI relevant information about the virtual
> > machine in a nice data format.
> > 
> > We would have to run this for each machine type, and save the
> > results to git immediately after release. Then for every change to
> > master, we would have to run the test again for every historic
> > machine type version and compare to the recorded ABI record.
> 
> Like Michael said we don't know that something is broken until it's
> too late and this particular case it's not even broken (strictly speaking
> change is correct) and is not even a part of ABI (it's ACPI code, i.e. firmware).
> 
> Problem is in the way virtio drivers enumerate devices, which makes the same
> device appear as a new one. We can work around issue on hypervisor side so user
> won't loose network connectivity or would be able to boot guest after QEMU upgrade.
> 
> We can suggest user re-installing their Windows (method that fixes almost all Win issues)
> or to try to make it pain-less for user in these rare cases, by upgrading to
> new QEMU (or fixed stable) which has workaround, so only the first few has to suffer.
> 
> (I think downstreams would even more benefit from this, there were similar problems
> there before).
> 
> Yes, It surely will expand test matrix, but it should be limited to specific cases
> we implemented fixups for.

My suggestion from a long while ago (which no one liked) was to
include the source qemu version and then have a quirks list of things to
fix up.

Dave

> > 
> > Regards,
> > Daniel
> 
> 
-- 
Dr. David Alan Gilbert / dgilbert at redhat.com / Manchester, UK