On Tue, Jan 05, 2021 at 12:36:50AM +0100, Igor Mammedov wrote:
> On Mon, 4 Jan 2021 13:29:06 -0500
> Eduardo Habkost <ehabk...@redhat.com> wrote:
>
> > On Mon, Jan 04, 2021 at 01:54:32PM +0100, Vitaly Kuznetsov wrote:
> > > Igor Mammedov <imamm...@redhat.com> writes:
> > >
> > > >> >
> > > >> > + /* Hyper-V features enabled with 'hyperv=on' */
> > > >> > + x86mc->default_hyperv_features = BIT(HYPERV_FEAT_RELAXED) |
> > > >> > + BIT(HYPERV_FEAT_VAPIC) | BIT(HYPERV_FEAT_TIME) |
> > > >> > + BIT(HYPERV_FEAT_CRASH) | BIT(HYPERV_FEAT_RESET) |
> > > >> > + BIT(HYPERV_FEAT_VPINDEX) | BIT(HYPERV_FEAT_RUNTIME) |
> > > >> > + BIT(HYPERV_FEAT_SYNIC) | BIT(HYPERV_FEAT_STIMER) |
> > > >> > + BIT(HYPERV_FEAT_FREQUENCIES) |
> > > >> > BIT(HYPERV_FEAT_REENLIGHTENMENT) |
> > > >> > + BIT(HYPERV_FEAT_TLBFLUSH) | BIT(HYPERV_FEAT_EVMCS) |
> > > >> > + BIT(HYPERV_FEAT_IPI) | BIT(HYPERV_FEAT_STIMER_DIRECT);
> > > > I'd argue that feature bits do not belong to machine code at all.
> > > > If we have to involve machine at all then it should be a set
> > > > property/value pairs
> > > > that machine will set on CPU object (I'm not convinced that doing it
> > > > from machine code is good idea though).
> > > >
> > >
> > > These are 'features' and not feature bits. 'Bits' here are just our
> > > internal (to QEMU) representation of which features are enable and which
> > > are not, we could've just used booleans instead. These feature, when
> > > enabled, will result in some CPUID changes (not 1:1) but I don't see how
> > > it's different from
> > >
> > > " -machine q35,accel=kvm "
> > >
> > > which also results in CPUID changes.
> >
> > This is a good point, although having accel affect CPUID bits was
> > also a source of complexity for query-cpu-model-expansion and
> > other QMP queries.
>
> why was, it's still a headache (mutating CPU models depending on accelerator)
>
> >
> > >
> > > The main reason for putting this to x86 machine type is versioning, as
> > > we go along we will (hopefully) be implementing more and more Hyper-V
> > > features but we want to provide 'one knob to rule them all' but do it in
> > > a way that will allow migration. We already have 'hv_passthrough' for
> > > CPU.
> >
> > I agree completely that the set of bits needs to be on
> > MachineClass. We just need to agree on the external interface.
> That's where I disagree,
> let me exaggerate for demo purpose:
> - let's move all CPU models feature defaults to MachineClass and forget
> about compat properties
> since in that case we can opencode changes in machine_class_init
I don't see your point here. compat_props is also part of
MachineClass.
>
> It's rather hard code integration between device models, which we try
> to avoid and still refactoring QEMU code to get rid of it.
> (sure it works until it's not and someone else need to rewrite half of QEMU
> to accomplish it's own task because we mixed things together)
I don't see why using a X86CPU-specific API that is not based on
QOM properties is hard code integration. compat_props is not the
only allowed API for machines to communicate with devices.
>
> >
> > >
> > > >> >
> > > >> > + if (x86ms->hyperv_enabled) {
> > > >> > + feat = x86mc->default_hyperv_features;
> > > >> > + /* Enlightened VMCS is only available on Intel/VMX */
> > > >> > + if (!cpu_has_vmx(&cpu->env)) {
> > > >> > + feat &= ~BIT(HYPERV_FEAT_EVMCS);
> > > >> > + }
> > > >> > +
> > > >> > + cpu->hyperv_features |= feat;
> > > > that will ignore features user explicitly doesn't want,
> > > > ex:
> > > > -machine hyperv=on -cpu foo,hv-foo=off
> > > >
> > >
> > > Existing 'hv_passthrough' mode can also affect the result. Personally, I
> > > don't see where 'hv-foo=off' is needed outside of debugging and these
> > > use-cases can probably be covered by explicitly listing required
> > > features but I'm not against making this work, shouldn't be hard.
> >
> > I'm all for not wasting time supporting use cases that are not
> > necessary in practice. We just need to document the expected
> > behavior clearly, whatever we decide to do.
>
> documenting is good, but if it adds new semantics to how CPU features are
> handled
> users up the stack will need code it up as well and juggle with
> -machine + -cpu + -device cpu-foo
> not to mention poor developers who will have to figure out why we do
> set CPU properties in multiple different ways.
>
> however if we add it as CPU properties that behave the same way as other
> properties, all mgmt has to do is expose new property to user for usage.
I think we need to be careful here. Sometimes just exposing the
QOM properties used to implemented a feature is not the best user
interface. e.g.: even if using compat_props for implementing the
hyperv features preset, that doesn't automatically mean we want
hyperv=on to be a -cpu option.
I would even argue we shouldn't be focusing on implementation
details (like we are doing right now) until the desired external
interface is described clearly.
>
> it even more true when building machine from QMP interface would be available,
> where we would want '-device foo' more or less the same way instead of
> special casing some of them, i.e. I'd rather have one device to configure,
> instead of doing it in multiple places. It's not possible in reality
> but for new code we should try to minimize split brain issues.
Is split brain a practical problem here? If the new behavior is
implemented in x86_cpu_realizefn() or x86_cpu_pre_plug(), we know
it's going to affect all CPU objects.
>
> > >
> > > > not sure we would like to introduce such invariant,
> > > > in normal qom property handling the latest set property should have
> > > > effect
> > > > (all other invariants we have in x86 cpu property semantics are comming
> > > > from legacy handling
> > > > and I plan to deprecate them (it will affect x86 and sparc cpus) so
> > > > CPUs will behave like
> > > > any other QOM object when it come to property handling)
> > > >
> > > > anyways it's confusing a bit to have cpu flags to come from 2 different
> > > > places
> > > >
> > > > -cpu hyperv-use-preset=on,hv-foo=off
> > > >
> > > > looks less confusing and will heave expected effect
> > > >
> > >
> > > Honestly, 'hyperv-use-preset' is confusing even to me :-)
> > >
> > > What if we for a second stop thinking about Hyper-V features being CPU
> > > features only, e.g. if we want to create Dynamic Memory or PTP or any
> > > other Hyper-V specific device in a simple way? We'll have to put these
> > > under machine type.
> >
> > I agree. Hyper-V is not just a set of CPU features.
> me too,
> however in this case we are talking about a set of cpu features,
> if there is no way to implement it as cpu properties + compat properties
> and requires opencodding it within machine code it might be fine
> but I fail to see a very good reason for doing that at this momment.
The reason would be just simplicity of implementation.
I understand there are reasons to suggest using compat_props if
it makes things simpler, but I don't see why we would reject a
patch because the implementation is not based purely on
compat_props.
I will let Vitaly to decide how to proceed, based on our
feedback. I encourage him to use compat_props like you suggest,
but I don't plan to make this a requirement.
>
> >
> > Also, those two approaches are not mutually exclusive.
> > "-machine hyperv=on" can be implemented internally using
> > "hyperv-use-preset=on" if necessary. I don't think it has to,
> > however.
>
>
--
Eduardo
