Re: [PATCH] PCI/MSI: don't disable AMD IOMMU MSI on Xen dom0

[Wei Wang <wei.wang2@amd.com>]

On 06/21/2012 02:45 PM, Jan Beulich wrote:
>>>> On 21.06.12 at 14:28, Wei Wang<wei.wang2@amd.com>  wrote:
>> On 06/21/2012 02:06 PM, Jan Beulich wrote:
>>>>>> On 21.06.12 at 13:21, Wei Wang<wei.wang2@amd.com>   wrote:
>>>> I also evaluated the possibility of hiding iommu device from dom0. I
>>>> think the change is no quite a lot, at least, for io based pcicfg
>>>> access. A proof-of-concept patch is attached.
>>>
>>> This completely hides the device from Dom0, but only when
>>> config space is accessed via method 1. Did you not see my
>>> earlier patch doing this for MCFG as well
>> Could you please provide a particular c/s number?... (I saw too many c/s
>> might be related to this topic). so that I could work out a patch to
>> support both i/o and mmcfg.
>
> I sent this to you yesterday, so you'd be able to test whether
> it actually fulfills its purpose before we discuss whether this is
> acceptable for 4.2. See
> http://lists.xen.org/archives/html/xen-devel/2012-06/msg01129.html

Oh, yes I found it, my email filter did not work well so I did not see 
it at the right folder. I will test right now.

>
>> (albeit only disallowing
>>> writes, so allowing the device to still be seen by Dom0)?
>> Sounds better to me...this still allows user to check iommu status from
>> lspci.
>>
>>> Whether completely hiding the device is actually okay I can't
>>> easily tell: Would IOMMUs always be either at func 0 of a single-
>>> unction device, or at a non-zero func of a multi-function one? If
>>> not, other devices may get hidden implicitly.
>>
>> AMD IOMMU is an independent pci-e endpoint, and this function will not
>> be used for other purposes other than containing an iommu. So I don't
>> see that iommu will share bdf value with other devices.
>
> The question is not regarding bdf, but regarding whether under
> the same seg:bus:dev there might be multiple functions, one of
> which is the IOMMU, and if so, whether the IOMMU would be
> guaranteed to have a non-zero function number.

In a real system (single or multiple iommu), amd iommu shares the same 
device number with north bridge but has function number 2.. (e.g 
bus:00.2) Howerver according to spec, it does not guaranteed to have 
non-zero function number. So what is the problem you see if iommu uses 
fun0 on a multi-func device?

Thanks,
Wei

> Jan
>
>