Re: [ARM] SMMU and PCI passthrough Was: Re: Xen 4.5-rc1 update (RC1 is out 2014-Oct-24th)

[Julien Grall <julien.grall@linaro.org>]

On 03/11/2014 05:08, manish jaggi wrote:
> On 2 November 2014 15:47, Julien Grall <julien.grall@linaro.org> wrote:
>> (Renaming the subject of the thread).
>>
>> On 02/11/2014 06:03, manish jaggi wrote:
>>>
>>> On 1 November 2014 23:13, Julien Grall <julien.grall@linaro.org> wrote:
>>>>
>>>> Hi Konrad,
>>>>
>>>>
>>>> On 31/10/2014 21:01, Konrad Rzeszutek Wilk wrote:
>>>>>
>>>>>
>>>>> On Fri, Oct 31, 2014 at 02:31:41PM +0000, Julien Grall wrote:
>>>>>>
>>>>>>
>>>>>> On 10/31/2014 02:24 PM, Konrad Rzeszutek Wilk wrote:
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> *  PVH - PCI passthrough for DomU.
>>>>>>>>
>>>>>>>>
>>>>>>>> I am working on Cavium Thunder (ARM64) on this feature.
>>>>>>>> [Xen SMMU driver changes + PCI passthrough changes in Xen and Linux]
>>>>>>
>>>>>>
>>>>>>
>>>>>> FYI, I'm currently reworking the SMMU drivers to resync with Linux.
>>>>>> With
>>>>>> thoses changes, you should not need to modify the SMMU code.
>>>>>
>>>>>
>>>>>
>>>>> Thank you for the update. Put your name behind that for 4.6.
>>>>>>
>>>>>>
>>>>>>
>>>>>>> Ok, replaced Julien's name with yours. Please make sure
>>>>>>> that for the Linux patches you CC xen-devel and the
>>>>>>> maintainers (David, Stefano, Boris and me).
>>>>>>
>>>>>>
>>>>>>
>>>>>> There is 2 distinct passthrough: platform (i.e non-PCI) and PCI one.
>>>>>>
>>>>>> While Manish is working on PCI passthrough, I'm still working the
>>>>>> non-PCI one. Please don't drop my name.
>>>>>
>>>>>
>>>>>
>>>>> I thought that Arianna's patches had taken care of that (the MMIO
>>>>> part?). Or does each platform need a different implementation of
>>>>> that?
>>>>
>>>>
>>>>
>>>> To passthrough a platform device you need to be able to assign the device
>>>> to
>>>> the guest via the IOMMU and map MMIOs (done by Arianna's series) and
>>>> interrupts.
>>>>
>>> For a PCI passthrough SMMU ops are to be added. The way the smmu for a
>>> pci device is found needs to be updated in the smmu.c, so there are
>>> some substantial changes to smmu.c for pci passthrough.
>>
>>
>> The SMMU drivers in Linux already supports PCI. As I'm currently resync our
>> driver with this version PCI assignment in the SMMU should come freely.
>>
>> I expect the only plumbing for the Xen callback and few bugs fixes will be
>> necessary.
>>
> we can discuss more on design level. There are changes

What kind of changes? Do you have a tree with them?

>>> Also MMIO mapping code the same pci device to be added.
>>
>>
>> Hmmm? What do you mean? MMIO mapping code is definitely not part of the SMMU
>> drivers.
>>
>> IIRC, this should be done by either the toolstack or PCI back in Linux.
>>
>>> So in short there changes, and as they are in the same files and
>>> features are also similar,  is it possible that we work together may
>>> be julien can provide a design document (simple txt file would do).
>>
>>
>> There is no need of design document for the SMMU drivers. Everything for DT
>> passthrough is already there.
>>
> It would be helpful if you can provide a basic flow.

http://lists.xen.org/archives/html/xen-devel/2014-07/msg04090.html

-- 
Julien Grall