1:30 a.m.
Eduardo Habkost <ehabkost(a)redhat.com> writes:
On Wed, May 23, 2018 at 05:53:34PM +0200, Markus Armbruster wrote:
> Eduardo Habkost <ehabkost(a)redhat.com> writes:
>
> > On Wed, May 23, 2018 at 11:17:55AM +0200, Markus Armbruster wrote:
> >> Eduardo Habkost <ehabkost(a)redhat.com> writes:
> >> > On Mon, May 21, 2018 at 04:46:36PM -0300, Daniel Henrique Barboza
wrote:
> > [...]
> >> >> Since no objection was made back then, this logic was put into
query-target
> >> >> starting
> >> >> in v2. Still, I don't have any favorites though: query-target
looks ok,
> >> >> query-machine
> >> >> looks ok and a new API looks ok too. It's all about what makes
(more) sense
> >> >> in the
> >> >> management level, I think.
> >> >
> >> > I understand the original objection from Eric: having to add a
> >> > new command for every runtime flag we want to expose to the user
> >> > looks wrong to me.
> >>
> >> Agreed.
> >>
> >> > However, extending query-machines and query-target looks wrong
> >> > too, however. query-target looks wrong because this not a
> >> > property of the target. query-machines is wrong because this is
> >> > not a static property of the machine-type, but of the running
> >> > machine instance.
> >>
> >> Of the two, query-machines looks less wrong.
> >>
> >> Arguably, -no-acpi should not exist. It's an ad hoc flag that sneakily
> >> splits a few machine types into two variants, with and without ACPI.
> >> It's silently ignored for other machine types, even APCI-capable ones.
> >>
> >> If the machine type variants with and without ACPI were separate types,
> >> wakeup-suspend-support would be a static property of the machine type.
> >>
> >> However, "separate types" probably doesn't scale: I'm
afraid we'd end up
> >> with an undesirable number of machine types. Avoiding that is exactly
> >> why we have machine types with configurable options. I suspect that's
> >> how ACPI should be configured (if at all).
> >>
> >> So, should we make -no-acpi sugar for a machine type parameter? And
> >> then deprecate -no-acpi for good measure?
> >
> > I think we should.
>
> Would you like to take care of it?
Adding to my TODO-list, I hope I will be able to do it before the
next release.
Thanks!
> >> > Can we have a new query command that could be an
obvious
> >> > container for simple machine capabilities that are not static? A
> >> > name like "query-machine" would be generic enough for that,
I
> >> > guess.
> >>
> >> Having command names differ only in a single letter is awkward, but
> >> let's focus on things other than naming now, and use
> >> query-current-machine like a working title.
> >>
> >> query-machines is wrong because wakeup-suspend-support isn't static for
> >> some machine types.
> >>
> >> query-current-machine is also kind of wrong because
> >> wakeup-suspend-support *is* static for most machine types.
> >
> > The most appropriate solution depends a lot on how/when
> > management software needs to query this.
>
> Right.
>
> > If they only need to query it at runtime for a running VM,
> > there's no reason for us to go of our way and add complexity just
> > to make it look like static data in query-machines.
> >
> > On the other hand, if they really need to query it before
> > configuring/starting a VM, it won't be useful at all to make it
> > available only at runtime.
> >
> > Daniel, when/how exactly software would need to query the new
> > flag?
> >
> >
> >> Worse, a machine type property that is static for all machine types now
> >> could conceivably become dynamic when we add a machine type
> >> configuration knob.
> >>
> >
> > This isn't the first time a machine capability that seems static
> > actually depends on other configuration arguments. We will
> > probably need to address this eventually.
>
> Then the best time to address it is now, provided we can :)
I'm not sure this is the best time. If libvirt only needs the
runtime value and don't need any info at query-machines time, I
think support for this on query-machines will be left unused and
they will only use the query-current-machine value.
Just giving libvirt the runtime data it wants
(query-current-machine) seems way better than requiring libvirt
to interpret a set of rules and independently calculate something
QEMU already knows.
I wouldn't mind adding a query-current-machine to report dynamic machine
capabilities if that helps QMP clients. query-machines could continue
to report static machine capabilities then.
However, we do need a plan on how to distribute machine capabilities
between query-machines and query-current-machine, in particular how to
handle changing staticness.
wakeup-suspend-support is static for most machine types, but dynamic for
some. Where should it go?
It needs to go into query-current-machine when its dynamic with the
current machine. It may go there just to keep things regular even if
its static with the current machine.
Does it go into query-machines, too? If not, clients lose the ability
to examine all machines efficiently. Even if this isn't an issue for
wakeup-suspend-support: are we sure this can't be an issue for any
future capabilities?
If it goes into query-machines, what should its value be for the machine
types where it's dynamic? Should it be absent, perhaps, letting clients
know they have to consult query-current-machine to find the value?
What if a capability previously thought static becomes dynamic? We can
add it to query-current-machine just fine, but removing it from
query-machines would be a compatibility break. Making it optional,
too. Should all new members of MachineInfo be optional, just in case?
These are design questions we need to ponder *now*. Picking a solution
that satisfies current needs while ignoring future needs has bitten us
in the posterior time and again. We're not going to successfully
predict *all* future needs, but not even trying should be easy to beat.
That's what I meant by "the best time to address it is now".
[...]