Re: [libvirt] [RFC PATCH 19/19] qapi: Implement -compat deprecated-output=hide for events
Daniel P. Berrangé <berrange(a)redhat.com> writes:
On Thu, Oct 24, 2019 at 02:34:58PM +0200, Markus Armbruster wrote:
> This policy suppresses deprecated events, and thus permits "testing
> the future".
One thing that occurs to me is that this is a fairly passive impact
on libvirt. eg it may well be not at all obvious if libvirt is behaving
in a broken way due to an event not being emitted, as the code in
question simply won't be triggered.
Intented use of -compat deprecated-input=error,deprecated-output=hide is
"testing the future": make QEMU behave as if the deprecated features
were already gone. Can be useful when you want to test code that deals
with the anticipated future *now*.
It can also be used to ferret out unknown uses of deprecated interfaces:
run test suite with it, see what fails. But as you note, the
deprecated-output=hide part is somewhat problematic in that role.
With the current QMP this situation is unavoidable since QEMU
doesn't
know which events the client (libvirt) is actually using. QEMU just
unconditionally emits all events.
I've often wondered if we should have the client explicitly tell
QEMU which events it wants to receive as part of the QMP greeting
handshake.
ie, libvirt knows which events it can handle. QEMU knows which
events it can emit, and reports this via capabilities which
libvirt probes.
So on connecting libvirt can tell QEMU exactly which evnets it
wants to get back. QEMU is now able to explicitly tell libvirt
it has asked for a deprecated event, and so the logic from the
"deprecated-input" option can take effect.
QEMU already reports its events via introspection. What's missing is an
event subscription mechanism. Should be feasible.
Additional benefit: can reduce I/O.
We'd not need "deprecated-output" at that point.
If deprecated-input=error makes subscribing to a deprecated event fail,
we don't need deprecated-output=hide for events.
But events are not the only output: there's also command returns.
Consider query-cpus-fast. Returns list of CpuInfoFast. CpuInfoFast
member @arch is deprecated. deprecated-output=hide should hide it,
except it's not implemented in this series.
This is also "a fairly passive impact on libvirt", I'm afraid.
We have some 40 events, and having libvirt subscribe to the ones it
actually uses is obviously practical.
I doubt the subscription idea scales up to return values.