Re: [PATCH 4/7] ppc/pnv: Simplify pnv_bmc_create()

[David Gibson]

On Thu, Jan 28, 2021 at 01:04:28PM +0100, Greg Kurz wrote:
> On Thu, 28 Jan 2021 08:46:01 +0100
> Cédric Le Goater <clg@kaod.org> wrote:
> 
> > On 1/28/21 1:46 AM, Joel Stanley wrote:
> > > On Tue, 26 Jan 2021 at 17:14, Cédric Le Goater <clg@kaod.org> wrote:
> > >>
> > >> and reuse pnv_bmc_set_pnor() to share the setting of the PNOR.
> > >>
> > >> Signed-off-by: Cédric Le Goater <clg@kaod.org>
> > >> ---
> > >>  hw/ppc/pnv_bmc.c | 7 +------
> > >>  1 file changed, 1 insertion(+), 6 deletions(-)
> > >>
> > >> diff --git a/hw/ppc/pnv_bmc.c b/hw/ppc/pnv_bmc.c
> > >> index 67ebb16c4d5f..86d16b493539 100644
> > >> --- a/hw/ppc/pnv_bmc.c
> > >> +++ b/hw/ppc/pnv_bmc.c
> > >> @@ -260,13 +260,8 @@ IPMIBmc *pnv_bmc_create(PnvPnor *pnor)
> > >>      Object *obj;
> > >>
> > >>      obj = object_new(TYPE_IPMI_BMC_SIMULATOR);
> > >> -    object_ref(OBJECT(pnor));
> > >> -    object_property_add_const_link(obj, "pnor", OBJECT(pnor));
> > > 
> > > I assume it's ok to move the link set to after the realise of the BMC 
> > > object?
> >  
> > 
> > When 2 objects need to be linked, one has to be realized first. 
> 
> Realize isn't a QOM concept in the first place...
> 
> > I suppose this is why it is allowed but I am not expert in that area. 
> > 
> 
> ... so no surprise object_property_add_const_link() doesn't care
> about it.
> 
> > Greg  ?
> > 
> 
> What is important though is that a property with a given name
> can only be added *once* to an object during its lifetime.
> Doing the contrary is a bug and QEMU aborts. So, with this
> in mind, it seems to me that adding QOM properties to a
> device object should only be done from some init path
> that is only called once.
> 
> > That was the case already when defining a "ipmi-bmc-sim" device on the 
> > command line. 
> > 
> 
> Yeah and the property is added during machine reset... which
> is typically a path that can be taken several times during
> the machine lifetime. The potential crash is avoided because
> pnv_reset() doesn't call pnv_bmc_set_pnor() if pnv->bmc is
> already set, but this is a fragile workaround...

Oof, yeah.  In general we should avoid creating or realizing objects
at machine reset time.  There might be a few exceptions, but they're
rare.

> A QOM link doesn't look like the correct way to model the BMC
> accesses to the PNOR. Since the only user is hiomap_cmd(),
> it seems you could reach the same result with a pointer to
> the PNOR object being passed to ipmi_sim_register_netfn()
> and later passed to hiomap_cmd().
> 
> > C. 
> > 
> > 
> > >>      qdev_realize(DEVICE(obj), NULL, &error_fatal);
> > >> -
> > >> -    /* Install the HIOMAP protocol handlers to access the PNOR */
> > >> -    ipmi_sim_register_netfn(IPMI_BMC_SIMULATOR(obj), IPMI_NETFN_OEM,
> > >> -                            &hiomap_netfn);
> > >> +    pnv_bmc_set_pnor(IPMI_BMC(obj), pnor);
> > >>
> > >>      return IPMI_BMC(obj);
> > >>  }
> > >>
> > >>
> > 
> 

-- 
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