Re: [Qemu-devel] [PATCH v4 1/4] vhost-user: add new vhost user messages to support virtio config space

[Liu, Changpeng]

> -----Original Message-----
> From: Michael S. Tsirkin [mailto:address@hidden
> Sent: Friday, October 20, 2017 10:12 AM
> To: Liu, Changpeng <address@hidden>
> Cc: Paolo Bonzini <address@hidden>; address@hidden;
> address@hidden; address@hidden; address@hidden; Harris,
> James R <address@hidden>
> Subject: Re: [PATCH v4 1/4] vhost-user: add new vhost user messages to support
> virtio config space
> 
> On Fri, Oct 20, 2017 at 01:55:20AM +0000, Liu, Changpeng wrote:
> >
> >
> > > -----Original Message-----
> > > From: Michael S. Tsirkin [mailto:address@hidden
> > > Sent: Friday, October 20, 2017 8:28 AM
> > > To: Paolo Bonzini <address@hidden>
> > > Cc: Liu, Changpeng <address@hidden>; address@hidden;
> > > address@hidden; address@hidden; address@hidden;
> Harris,
> > > James R <address@hidden>
> > > Subject: Re: [PATCH v4 1/4] vhost-user: add new vhost user messages to 
> > > support
> > > virtio config space
> > >
> > > On Thu, Oct 19, 2017 at 11:04:48PM +0200, Paolo Bonzini wrote:
> > > > On 19/10/2017 19:43, Michael S. Tsirkin wrote:
> > > > > On Thu, Oct 19, 2017 at 05:43:18PM +0200, Paolo Bonzini wrote:
> > > > >> On 19/10/2017 17:39, Michael S. Tsirkin wrote:
> > > > >>>> Add VHOST_USER_GET_CONFIG/VHOST_USER_SET_CONFIG messages
> > > which can be
> > > > >>>> used for live migration of vhost user devices, also vhost user 
> > > > >>>> devices
> > > > >>>> can benefit from the messages to get/set virtio config space 
> > > > >>>> from/to the
> > > > >>>> I/O target. For the purpose to support virtio config space change,
> > > > >>>> VHOST_USER_SET_CONFIG_FD message is added as the event notifier
> > > > >>>> in case virtio config space change in the I/O target.
> > > > >>>>
> > > > >>>> Signed-off-by: Changpeng Liu <address@hidden>
> > > > >>> I don't much like it that config is getting passed through.
> > > > >>>
> > > > >>> IMO this makes managing things harder not easier.
> > > > >>>
> > > > >>> How about specific messages about specific parts of
> > > > >>> config space that you want to get from the backend?
> > > > >>
> > > > >> In the case of virtio-blk that would be all of it.  Do you have a 
> > > > >> case
> > > > >> in mind where some part of the configuration space is owned by QEMU?
> > > > >>
> > > > >> Paolo
> > > > >
> > > > > Yes. seg_max
> > > >
> > > > The seg_max limit is established by whoever reads buffers from the vring
> > > > and passes them down to the lower layer.  For vhost-blk that's the
> > > > device server, not QEMU.
> > > >
> > > > Paolo
> > >
> > > Good point. How about num_queues though?
> > num_queues  is part of virtio_blk config, vhost-user slave can set it,
> > and Qemu driver can rewrite it if user want less IO queues.
> 
> Fundamentally QEMU needs to support this # of queues for this
> device.
> 
> So whenever QEMU doesn't always expose config space as-is,
> need to document the exact semantics.
Agreed, Qemu vhost block driver should always has a default value, so I also 
added the
value as one of the parameters for vhost block driver.
> 
> Also, does backend need to know?
vhost-user slave does know how many queues are used, because vhost-user messages
such as SET_VRING_CALL/KICK are related with queues. Here the idea is 
vhost-user slave
provides the maximum io queues supported, and Qemu users can specify lower io 
queues.
> 
> 
> > >
> > > Also why is there SET_CONFIG? Does not look like blk uses it.
> > Only one possible usage when disable write cache to vhost-user slave device.
> 
> Again need to add documentation what can be written.
Agreed.
> 
> 
> > >
> > > And I wonder how do we do it for other devices.
> > >
> > > E.g. for net there's a bit in the middle of the
> > > config field that deals with migration.
> > Well, I'm okay to make those messages only valid for virtio block device, 
> > because
> it's enough
> > for virtio block to be started with vhost-user slave target.
> 
> OK but I'd rather make them at least somewhat generic so we can reuse
> them down the road.  It looks like adding offset/size pair would solve
> most of the issues. Thoughts?
Do you mean SET_CONFIG message followed with offset/size to let vhost-user slave
Know which field the master want to change?  Yes, sounds good to me.
> 
> > >
> > >
> > > --
> > > MST