Re: [Qemu-devel] [PATCH] virtio-pci: Set the QEMU_PCI_CAP_EXPRESS capability early in its DeviceClass realize method

[Marcel Apfelbaum <marcel@redhat.com>]

On 12/02/2015 03:30 PM, Shmulik Ladkani wrote:
> Hi,
>
> On Wed, 2 Dec 2015 11:51:46 +0200, marcel@redhat.com wrote:
>> 2. We still have pci_is_express returning true, this is error prone because
>> one can use this function assuming the device is express. Maybe we should
>> call it "can_be_express" ?
>>
>> If you think this is good enough, you can simply do the same:
>>     - Instead of replacing the realize method, just advertise it with
>>       "is_express" (meaning it can be express)
>>     - Leave all the conditions as they were in prev patch.
>> As a result, the pci config space will have the right length.
>
> Oh but we can't do so, as the change of config space size is guest
> visible and breaks migration; it must depend on your x-pcie-disable
> flag :)

Indeed, we need at least to condition it on  x-pcie-disable.

>
> As I can't decide what's better, I'm following your initial suggestion
> and submit for maintainers to review.

Sure, and thanks for the patience to get to the bottom of it.

>
> However, do note that there are few more evidence that 'pci_is_express'
> is true while not necessarily placed on a pcie bus:

and this is scary ... we really should call it "pci_can_be_express"

> - pcie_endpoint_cap_init:
>    it tests for 'pci_bus_is_express' although 'dev' is guaranteed to be
>    'pci_is_express' (assertion in pcie_cap_init)
> - 058fdcf 'xhci: add endpoint cap on express bus only'


Thanks,
Marcel

>
> Thanks,
> Shmulik
>