Re: Management applications and CPU feature flags

[Markus Armbruster <armbru@redhat.com>]

Daniel P. Berrangé <berrange@redhat.com> writes:

> On Fri, Apr 11, 2025 at 12:40:46PM +0200, Markus Armbruster wrote:
>> Daniel P. Berrangé <berrange@redhat.com> writes:
>> 
>> > On Wed, Apr 09, 2025 at 09:58:13AM +0200, Peter Krempa via Devel wrote:
>> >> On Wed, Apr 09, 2025 at 09:39:02 +0200, Markus Armbruster via Devel wrote:
>> >> > Hi Steve, I apologize for the slow response.
>> >> > 
>> >> > Steve Sistare <steven.sistare@oracle.com> writes:
>> >> > 
>> >> > > Using qom-list and qom-get to get all the nodes and property values in a
>> >> > > QOM tree can take multiple seconds because it requires 1000's of individual
>> >> > > QOM requests.  Some managers fetch the entire tree or a large subset
>> >> > > of it when starting a new VM, and this cost is a substantial fraction of
>> >> > > start up time.
>> >> > 
>> >> > "Some managers"... could you name one?
>> >> 
>> >> libvirt is at ~500 qom-get calls during an average startup ...
>> >> 
>> >> > > To reduce this cost, consider QAPI calls that fetch more information in
>> >> > > each call:
>> >> > >   * qom-list-get: given a path, return a list of properties and values.
>> >> > >   * qom-list-getv: given a list of paths, return a list of properties and
>> >> > >     values for each path.
>> >> > >   * qom-tree-get: given a path, return all descendant nodes rooted at that
>> >> > >     path, with properties and values for each.
>> >> > 
>> >> > Libvirt developers, would you be interested in any of these?
>> >> 
>> >> YES!!!
>> >
>> > Not neccessarily, see below... !!!! 
>> >
>> >> 
>> >> The getter with value could SO MUCH optimize the startup sequence of a
>> >> VM where libvirt needs to probe CPU flags:
>> >> 
>> >> (note the 'id' field in libvirt's monitor is sequential)
>> >> 
>> >> buf={"execute":"qom-get","arguments":{"path":"/machine/unattached/device[0]","property":"realized"},"id":"libvirt-8"}
>> >> buf={"execute":"qom-get","arguments":{"path":"/machine/unattached/device[0]","property":"hotplugged"},"id":"libvirt-9"}
>> >> buf={"execute":"qom-get","arguments":{"path":"/machine/unattached/device[0]","property":"hotpluggable"},"id":"libvirt-10"}
>> >> 
>> >> [...]
>> >> 
>> >> buf={"execute":"qom-get","arguments":{"path":"/machine/unattached/device[0]","property":"hv-apicv"},"id":"libvirt-470"}
>> >> buf={"execute":"qom-get","arguments":{"path":"/machine/unattached/device[0]","property":"xd"},"id":"libvirt-471"}
>> >> buf={"execute":"qom-get","arguments":{"path":"/machine/unattached/device[0]","property":"sse4_1"},"id":"libvirt-472"}
>> >> buf={"execute":"qom-get","arguments":{"path":"/machine/unattached/device[0]","property":"unavailable-features"},"id":"libvirt-473"}
>> >> 
>> >> First and last line's timestamps:
>> >> 
>> >> 2025-04-08 14:44:28.882+0000: 1481190: info : qemuMonitorIOWrite:340 : QEMU_MONITOR_IO_WRITE: mon=0x7f4678048360 buf={"execute":"qom-get","arguments":{"path":"/machine/unattached/device[0]","property":"realized"},"id":"libvirt-8"}
>> >> 
>> >> 2025-04-08 14:44:29.149+0000: 1481190: info : qemuMonitorIOWrite:340 : QEMU_MONITOR_IO_WRITE: mon=0x7f4678048360 buf={"execute":"qom-get","arguments":{"path":"/machine/unattached/device[0]","property":"unavailable-features"},"id":"libvirt-473"}
>> >> 
>> >> Libvirt spent ~170 ms probing cpu flags.
>> >
>> > One thing I would point out is that qom-get can be considered an
>> > "escape hatch" to get information when no better QMP command exists.
>> > In this case, libvirt has made the assumption that every CPU feature
>> > is a QOM property.
>> >
>> > Adding qom-list-get doesn't appreciably change that, just makes the
>> > usage more efficient.
>> >
>> > Considering the bigger picture QMP design, when libvirt is trying to
>> > understand QEMU's CPU feature flag expansion, I would ask why we don't
>> > have something like a "query-cpu" command to tell us the current CPU
>> > expansion, avoiding the need for poking at QOM properties directly.
>> 
>> How do the existing query-cpu-FOO fall short of what management
>> applications such as libvirt needs?
>
> It has been along while since I looked at them, but IIRC they were
> returning static info about CPU models, whereas libvirt wanted info
> on the currently requested '-cpu ARGS'

Libvirt developers, please work with us on design of new commands or
improvements to existing ones to better meet libvirt's needs in this
area.