Re: [Qemu-devel] [PATCH 3/5] PXB: convert to realize()

[Marcel Apfelbaum]

On 12/20/2015 12:48 PM, Cao jin wrote:
> Hi,
>
> On 12/20/2015 06:22 PM, Marcel Apfelbaum wrote:
> > Hi,
> >
> > On 12/18/2015 01:03 PM, Cao jin wrote:
> > > Signed-off-by: Cao jin <address@hidden>
> > > ---
> > >   hw/pci-bridge/pci_expander_bridge.c | 24 ++++++++++++++----------
> > >   1 file changed, 14 insertions(+), 10 deletions(-)
> > >
> > > diff --git a/hw/pci-bridge/pci_expander_bridge.c
> > > b/hw/pci-bridge/pci_expander_bridge.c
> > > index 57f8a37..cc975f6 100644
> > > --- a/hw/pci-bridge/pci_expander_bridge.c
> > > +++ b/hw/pci-bridge/pci_expander_bridge.c
> > > @@ -145,19 +145,19 @@ static const TypeInfo pxb_host_info = {
> > >    * Returns 0 on successs, -1 if i440fx host was not
> > >    * found or the bus number is already in use.
> > >    */
> > > -static int pxb_register_bus(PCIDevice *dev, PCIBus *pxb_bus)
> > > +static int pxb_register_bus(PCIDevice *dev, PCIBus *pxb_bus, Error
> > > **errp)
> >
> > If you add an err parameter, maybe the function should return void.
>
> Ok, will modify it in V2. Actually, both style are fine with me:)
>
> > >   {
> > >       PCIBus *bus = dev->bus;
> > >       int pxb_bus_num = pci_bus_num(pxb_bus);
> > >
> > >       if (bus->parent_dev) {
> > > -        error_report("PXB devices can be attached only to root bus.");
> > > +        error_setg(errp, "PXB devices can be attached only to root
> > > bus.");
> > >           return -1;
> > >       }
> > >
> > >       QLIST_FOREACH(bus, &bus->child, sibling) {
> > >           if (pci_bus_num(bus) == pxb_bus_num) {
> > > -            error_report("Bus %d is already in use.", pxb_bus_num);
> > > +            error_setg(errp, "Bus %d is already in use.", pxb_bus_num);
> > >               return -1;
> > >           }
> > >       }
> > > @@ -193,7 +193,7 @@ static gint pxb_compare(gconstpointer a,
> > > gconstpointer b)
> > >              0;
> > >   }
> > >
> > > -static int pxb_dev_initfn(PCIDevice *dev)
> > > +static void pxb_dev_realize(PCIDevice *dev, Error **errp)
> > >   {
> > >       PXBDev *pxb = PXB_DEV(dev);
> > >       DeviceState *ds, *bds;
> > > @@ -202,8 +202,8 @@ static int pxb_dev_initfn(PCIDevice *dev)
> > >
> > >       if (pxb->numa_node != NUMA_NODE_UNASSIGNED &&
> > >           pxb->numa_node >= nb_numa_nodes) {
> > > -        error_report("Illegal numa node %d.", pxb->numa_node);
> > > -        return -EINVAL;
> > > +        error_setg(errp, "Illegal numa node %d.", pxb->numa_node);
> > > +        return;
> > >       }
> > >
> > >       if (dev->qdev.id && *dev->qdev.id) {
> > > @@ -225,8 +225,8 @@ static int pxb_dev_initfn(PCIDevice *dev)
> > >
> > >       PCI_HOST_BRIDGE(ds)->bus = bus;
> > >
> > > -    if (pxb_register_bus(dev, bus)) {
> > > -        return -EINVAL;
> > > +    if (pxb_register_bus(dev, bus, errp)) {
> > > +        goto err_register_bus;
> > >       }
> > >
> > >       qdev_init_nofail(ds);
> > > @@ -237,7 +237,11 @@ static int pxb_dev_initfn(PCIDevice *dev)
> > >       pci_config_set_class(dev->config, PCI_CLASS_BRIDGE_HOST);
> > >
> > >       pxb_dev_list = g_list_insert_sorted(pxb_dev_list, pxb,
> > > pxb_compare);
> > > -    return 0;
> > > +
> > > +err_register_bus:
> > > +    object_unref(OBJECT(ds));
> > > +    object_unref(OBJECT(bds));
> > > +    object_unref(OBJECT(bus));
> >
> >
> > The order should be in the reverse order of creation:
> >      bds, bus, ds
>
> Ok, I can do that. But it seems the order here doesn`t matter? Is there 
> dependency among these three?

Yes, there is a dependency:
At first the pxb host (ds) is created, then the bus (bus) is created as host's 
child (see pci_bus_new)
and in the end a pci bridge (bds) is attached to the bus (see qdev_create).

By the way, indeed you should call object_unparent at least for the 
pxb_host(ds), but you may want to
check the others parent relationship as well.

> > >   }
> > >
> > >   static void pxb_dev_exitfn(PCIDevice *pci_dev)
> > > @@ -259,7 +263,7 @@ static void pxb_dev_class_init(ObjectClass *klass,
> > > void *data)
> > >       DeviceClass *dc = DEVICE_CLASS(klass);
> > >       PCIDeviceClass *k = PCI_DEVICE_CLASS(klass);
> > >
> > > -    k->init = pxb_dev_initfn;
> > > +    k->realize = pxb_dev_realize;
> > >       k->exit = pxb_dev_exitfn;
> >
> > If init is converted to realize, maybe the exit should be converted to
> > unrealize?
>
> Yup, I agree with you from the point that the names should be antonym. But it 
> seems there is no PCIDeviceClass.unrealize:(

You are right. The pci_qdev_unrealize ultimately calls exit. But the same goes for 
init, pci_qdev_realize calls for pc->realize.
This is the reason I chose to use init/exit instead of the strange realize/exit.

But since the intention is to get rid of init, I am not against it.

> And I am also not aware why there is no comment for .exit while there is for 
> .init. It is appreciated if somebody could tell me the history O:-)

I'll add Markus, Andreas and Michael (the PCI maintainer), maybe they have a 
better insight to this.

On the other hand you should continue with the patch and leave the "unrealize" 
until we'll know more :)

Thanks,
Marcel

> > Thanks,
> > Marcel
> >
> > >       k->vendor_id = PCI_VENDOR_ID_REDHAT;
> > >       k->device_id = PCI_DEVICE_ID_REDHAT_PXB;
> >
> >
> >
> > .