Re: Configuration vs. compat hints [was Re: [Qemu-devel] [PATCHv3 03/13] qemu: add routines to manage PCI capabilities]

From: Mark McLoughlin <markmc@redhat.com>
To: Anthony Liguori <anthony@codemonkey.ws>
Cc: Jamie Lokier <jamie@shareable.org>,
	"Michael S. Tsirkin" <mst@redhat.com>,
	Carsten Otte <cotte@de.ibm.com>,
	kvm@vger.kernel.org, Glauber Costa <glommer@redhat.com>,
	Rusty Russell <rusty@rustcorp.com.au>,
	qemu-devel@nongnu.org, virtualization@lists.linux-foundation.org,
	Blue Swirl <blauwirbel@gmail.com>,
	Christian Borntraeger <borntraeger@de.ibm.com>,
	Paul Brook <paul@codesourcery.com>, Avi Kivity <avi@redhat.com>
Subject: Re: Configuration vs. compat hints [was Re: [Qemu-devel] [PATCHv3 03/13] qemu: add routines to manage PCI capabilities]
Date: Fri, 12 Jun 2009 17:48:23 +0100	[thread overview]
Message-ID: <1244825303.26769.19.camel@blaa> (raw)
In-Reply-To: <4A327E4A.7010300@codemonkey.ws>

On Fri, 2009-06-12 at 11:11 -0500, Anthony Liguori wrote:
> Mark McLoughlin wrote:
> > On Fri, 2009-06-12 at 09:51 -0500, Anthony Liguori wrote:
> >   
> >> Mark McLoughlin wrote:
> >>     
> >>> On Wed, 2009-06-10 at 20:27 +0100, Jamie Lokier wrote:
> >>>   
> >>>       
> >>>> Michael S. Tsirkin wrote:
> >>>>     
> >>>>         
> >>>>>> I think the right long term answer to all this is a way to get QEMU to
> >>>>>> dump it's current machine configuration in glorious detail as a file
> >>>>>> which can be reloaded as a machine configuration.
> >>>>>>         
> >>>>>>             
> >>>>> And then we'll have the same set of problems there.
> >>>>>       
> >>>>>           
> >>>> We will, and the solution will be the same: options to create devices
> >>>> as they were in older versions of QEMU.  It only needs to cover device
> >>>> features which matter to guests, not every bug fix.
> >>>>
> >>>> However with a machine configuration which is generated by QEMU,
> >>>> there's less worry about proliferation of obscure options, compared
> >>>> with the command line.  You don't necessarily have to document every
> >>>> backward-compatibility option in any detail, you just have to make
> >>>> sure it's written and read properly, which is much the same thing as
> >>>> the snapshot code does.
> >>>>     
> >>>>         
> >>> This is a sensible plan, but I don't think we should mix these compat
> >>> options in with the VM manager supplied configuration.
> >>>
> >>> There are two problems with that approach.
> >>>
> >>> = Problem 1 - VM manager needs to parse qemu config =
> >>>
> >>> Your proposal implies:
> >>>
> >>>   - VM manager supplies a basic configuration to qemu
> >>>
> >>>   - It then immediately asks qemu for a dump of the machine 
> >>>     configuration in all its glorious detail and retains that
> >>>     config
> >>>
> >>>   - If the VM manager wishes to add a new device it needs to parse the 
> >>>     qemu config and add it, rather than just generate an entirely new 
> >>>     config
> >>>   
> >>>       
> >> What's the problem with parsing the device config and modifying it?  Is 
> >> it just complexity?
> >>     
> >
> > Yes, complexity is the issue.
> >
> >   
> >> If we provided a mechanism to simplify manipulating a device config, 
> >> would that eliminate the concern here?
> >>     
> >
> > In libvirt's case, a lot of the complexity would come from needing to
> > figure out what to change.
> >   
> 
> Right, libvirt wants to be able to easily say "add a scsi block device 
> to this VM".  The way I see this working is that there would be a 
> default pc.dtc.  We would still have a -drive file=foo.img,if=scsi 
> option that would really just be a wrapper around first searching for an 
> existing LSI controller, if one exists, attaching the lun, if not, 
> create one, etc.
> 
> libvirt could continue to use this sort of interface.  However, as it 
> wants to do more advanced things, it may have to dive into the device 
> tree itself.
> 
> On live migration, QEMU will save a copy of the device tree somewhere 
> and libvirt needs to keep track of it.  It can treat it as opaque.  -M 
> /path/to/foo.dtc -drive file=foo.img,if=scsi should continue working as 
> expected IMHO.

So, when libvirt creates a guest for the first time, it makes a copy of
the device tree and continues to use that even if qemu is upgraded.
That's enough to ensure compat is retained for all built-in devices.

However, in order to retain compat for that SCSI device (e.g. ensuring
the PCI address doesn't change as other devices are added an removed),
we're back to the same problem ... either:

  1) Use '-drive file=foo.img,if=scsi,pci_addr=foo'; in order to figure 
     out what address to use, libvirt would need to query qemu for what 
     address was originally allocated to device or it would do all the 
     PCI address allocation itself ... or:

  2) Don't use the command line, instead get a dump of the entire 
     device tree (including the SCSI device) - if the device is to be 
     removed or modified in future, libvirt would need to modify the 
     device tree

The basic problem would be that the command line config would have very
limited ability to override the device tree config.

Cheers,
Mark.