Removal of bus->msi assignment breaks MSI with stacked domains

[wangyijing0307@gmail.com (Yijing Wang)]

? 2014/11/21 18:29, Marc Zyngier ??:
> Hi Thomas,
>
> On 21/11/14 01:46, Thomas Gleixner wrote:
>> On Fri, 21 Nov 2014, Yijing Wang wrote:
>>> On 2014/11/21 0:31, Marc Zyngier wrote:
>>>> Bjorn, Yijing,
>>>>
>>>> I've just realized that patch c167caf8d174 (PCI/MSI: Remove useless
>>>> bus->msi assignment) completely breaks MSI on arm64 when using the new
>>>> MSI stacked domain:
>>> Sorry, this is my first part to refactor MSI related code, now how
>>> to get pci msi_controller depends arch
>>> functions(pcibios_msi_controller() or arch_setup_msi_irq()), we are
>>> working on generic pci_host_bridge, after that, we could eventually
>>> eliminate MSI arch functions and find pci dev 's msi controller by
>>> pci_host_bridge->get_msi_controller().
>> The main question is why you think that pci_host_bridge is the proper
>> place to store that information.
>>
>> On x86 we have DMAR units associated to a single device. Each DMAR
>> unit is a seperate MSI irq domain.
>>
>> Can you guarantee that the pci_host_bridge is the right point to
>> provide the association of the device to the irq domain?
>>
>> So the real question is:
>>
>>     What is the association level requirement to properly identify the
>>     irqdomain for a specific device on any given architecture with and
>>     without IOMMU, interrupt redirection etc.
>>
>> To be honest: I don't know.
>>
>> My gut feeling tells me that it's at the device level, but I really
>> leave that decision to the experts in that field.
> Given the above requirement (single device associated to DMAR), I can
> see two possibilities:
> - we represent DMAR as a single PCI bus: feels a bit artificial
> - we move the MSI domain to the device, as you suggested.
>
> The second one seems a lot more attractive to me. What I don't
> completely see is how the host bridge has all required the knowledge.

Hmmm, maybe I'm in the wrong direction,  I need to think more about it.

Thanks!
Yijing.


>
> Also, it is not clear to me what is the advantage of getting rid of the
> MSI controller. By doing so, we loose an important part of the topology
> information (the irq domain is another level of abstraction).
>
> Thanks,
>
> 	M.