Thanks Jason,
On Fri, 4 Dec 2015 16:49:52 +0800 Jason Wang <jasow...@redhat.com> wrote:
> > @@ -2257,6 +2262,10 @@ static void vmxnet3_pci_realize(PCIDevice *pci_dev,
> > Error **errp)
> >
> > vmxnet3_net_init(s);
> >
> > + if (pci_is_express(pci_dev) && pci_bus_is_express(pci_dev->bus)) {
>
> Looks like pci_bus_is_express() has been checked in
> pcie_endpoint_cap_init().
Yes, but only for toggling between 'type = PCI_EXP_TYPE_ENDPOINT' vs.
'type = PCI_EXP_TYPE_RC_END'.
We would not like to expose the capability stating device is an express
endpoint (of any kind) unless it is located on a pcie bus.
> > +static void vmxnet3_realize(DeviceState *qdev, Error **errp)
> > +{
> > + VMXNET3Class *vc = VMXNET3_DEVICE_GET_CLASS(qdev);
> > + PCIDevice *pci_dev = PCI_DEVICE(qdev);
> > + VMXNET3State *s = VMXNET3(qdev);
> > +
> > + if (!(s->compat_flags & VMXNET3_COMPAT_FLAG_DISABLE_PCIE)) {
> > + pci_dev->cap_present |= QEMU_PCI_CAP_EXPRESS;
> > + }
> > +
> > + vc->parent_dc_realize(qdev, errp);
> > +}
>
> It's not clear that how the class helps here. Why not simply do
> everthing in vmxnet3_pci_realize()?
Since 'vmxnet3_pci_realize' is invoked too late. By the time it's
invoked, the config space is already allocated during 'pci_qdev_realize',
without correctly knowing whether the device is pci or pcie.
We must call 'pci_qdev_realize' (parent_dc_realize) only after
'cap_present' is properly set.
See discussion here:
https://lists.nongnu.org/archive/html/qemu-devel/2015-12/msg00043.html
Regards,
Shmulik
