Re: [PATCH v10 0/8] hmp,qmp: Add commands to introspect virtio devices

["Michael S. Tsirkin" <mst@redhat.com>]

On Mon, Dec 06, 2021 at 07:13:49PM +0100, Christian Schoenebeck wrote:
> On Montag, 6. Dezember 2021 17:09:45 CET Jonah Palmer wrote:
> > On 12/6/21 08:50, Christian Schoenebeck wrote:
> > > On Montag, 6. Dezember 2021 13:43:18 CET Jonah Palmer wrote:
> > >> This series introduces new QMP/HMP commands to dump the status of a
> > >> virtio device at different levels.
> > >> 
> > >> [Jonah: Rebasing previous patchset from Nov. 10 (v9). Original patches
> > >> 
> > >>   are by Laurent Vivier from May 2020.
> > >>   
> > >>   Rebase from v9 to v10 includes reformatting virtio.json examples and
> > >>   command structures for better consistency. Also removed all enums from
> > >>   virtio.json and replaced their purpose with string literals.
> > >>   
> > >>   Removed @ndescs from VirtioQueueElement, as the number of descriptors
> > >>   can be inferred from the length of the @descs chain.
> > >>   
> > >>   Lastly, removed the examples in hmp-commands-info.hx to fix
> > >>   'inconsistent
> > >>   literal block quoting' warning from Sphinx.]
> > > 
> > > I have not followed the entire discussion. AFAICS this is intended to
> > > monitor status information on virtio level only, like virtqueue fill
> > > status, etc.
> > > 
> > > One thing that I am looking for is monitoring device specific information
> > > above virtio level, e.g. certain performance numbers or statistics that
> > > only make sense for the specific device. That would not fit into any of
> > > these commands, right?
> > > 
> > > Best regards,
> > > Christian Schoenebeck
> > 
> > Correct. These are just one-shot commands that dump information on virtio
> > devices (including vhosts), their virtqueues, and virtqueue elements as they
> > are at the time of the command.
> > 
> > Jonah
> 
> What I would find useful though on this virtio level: also being able to query 
> the maximum and average fill state of the two ring buffers of each virtqueue. 
> That would allow to identify performance bottlenecks.
> 
> Best regards,
> Christian Schoenebeck

Adding this to vhost would need some interface work though.
Also, collecting these stats isn't free or trivial,
so I imagine we would need commands to enable/disable data collection.

-- 
MST