Re: Question about vmstate_register(), dc->vmsd and instance_id

[David Gibson]

On Fri, Mar 18, 2022 at 04:51:10PM -0300, Daniel Henrique Barboza wrote:
> 
> 
> On 3/18/22 00:43, David Gibson wrote:
> > On Thu, Mar 17, 2022 at 04:29:14PM +0000, Dr. David Alan Gilbert wrote:
> > > * Peter Maydell (peter.maydell@linaro.org) wrote:
> > > > On Thu, 17 Mar 2022 at 14:03, Daniel Henrique Barboza
> > > > <danielhb413@gmail.com> wrote:
> > > > > I've been looking into converting some vmstate_register() calls to 
> > > > > use dc->vmsd,
> > > > > using as a base the docs in docs/devel/migration.rst. This doc 
> > > > > mentions that we
> > > > > can either register the vmsd by using vmstate_register() or we can 
> > > > > use dc->vmsd
> > > > > for qdev-based devices.
> > > > > 
> > > > > When trying to convert this vmstate() call for the qdev alternative 
> > > > > (hw/ppc/spapr_drc.c,
> > > > > drc_realize()) I found this:
> > > > > 
> > > > >       vmstate_register(VMSTATE_IF(drc), spapr_drc_index(drc), 
> > > > > &vmstate_spapr_drc,
> > > > >                        drc);
> > > > > 
> > > > > spapr_drc_index() is an unique identifier for these DRC devices and 
> > > > > it's being used
> > > > > as instance_id. It is not clear to me how we can keep using this same 
> > > > > instance_id when
> > > > > using the dc->vmsd alternative. By looking a bit into migration files 
> > > > > I understood
> > > > > that if dc->vmsd is being used the instance_id is always 
> > > > > autogenerated. Is that correct?
> > > > 
> > > > Not entirely. It is the intended common setup, but because changing
> > > > the ID value breaks migration compatibility there is a mechanism
> > > > for saying "my device is special and needs to set the instance ID
> > > > to something else" -- qdev_set_legacy_instance_id().
> > > 
> > > Yes, this is normally only an issue for 'system' or memory mapped
> > > devices;  for things hung off a bus that has it's own device naming,
> > > then each instance of a device has it's own device due to the bus name
> > > so instance_id's aren't used.  Where you've got a few of the
> > > same device with the same name, and no bus for them to be named by, then
> > > the instance_id is used to uniquify them.
> 
> 
> Thanks for the info. qdev_set_legacy_instance_id() was the missing piece I was
> looking for to continue with the dc->vmsd transition I'd like to do.
> 
> 
> > 
> > Thanks for the information.  I remember deciding at the time that just
> > using vmsd wouldn't work for the DRCs because we needed this fixed
> > index.  At the time either qdev_set_legacy_instance_id() didn't exist,
> > or I didn't know about it, hence the explicit vmstate_register() call
> > so that an explicit instance id could be supplied.
> > 
> 
> This is the commit that introduced DRC migration:
> 
> 
> commit a50919dddf148b0a2008db4a0593dbe69e1059c0
> Author: Daniel Henrique Barboza <danielhb@linux.vnet.ibm.com>
> Date:   Mon May 22 16:35:49 2017 -0300
> 
>     hw/ppc: migrating the DRC state of hotplugged devices
> 
> 
> I'd say you can cut yourself some slack this time. Blame that guy
> instead.

Man, not that guy again! ;-)

I think I must have done something similar with some other migration
component.

-- 
David Gibson                    | I'll have my music baroque, and my code
david AT gibson.dropbear.id.au  | minimalist, thank you.  NOT _the_ _other_
                                | _way_ _around_!
http://www.ozlabs.org/~dgibson

signature.asc
Description: PGP signature