Re: [PATCH v2 1/3] xen/arm: pci: introduce PCI_PASSTHROUGH Kconfig option

[Stewart Hildebrand <stewart.hildebrand@xxxxxxx>]

On 7/20/23 05:20, Julien Grall wrote:
> Hi,
> 
> On 18/07/2023 18:35, Stewart Hildebrand wrote:
>> On 7/13/23 14:40, Julien Grall wrote:
>>> Hi Stewart,
>>>
>>> On 07/07/2023 02:47, Stewart Hildebrand wrote:
>>>> From: Rahul Singh <rahul.singh@xxxxxxx>
>>>>
>>>> Setting CONFIG_PCI_PASSTHROUGH=y will enable PCI passthrough on ARM, even 
>>>> though
>>>> the feature is not yet complete in the current upstream codebase. The 
>>>> purpose of
>>>> this is to make it easier to enable the necessary configs (HAS_PCI, 
>>>> HAS_VPCI) for
>>>> testing and development of PCI passthrough on ARM.
>>>>
>>>> Since PCI passthrough on ARM is still work in progress at this time, make 
>>>> it
>>>> depend on EXPERT.
>>>
>>> While preparing the patch for committing, I noticed that HAS_PASSTHROUGH
>>> will now allow the user to select one of the IOMMU quarantine options.
>>>
>>> There are three of them right now:
>>>    1. none
>>>    2. basic (i.e. faulting)
>>>    3. scratch page
>>>
>>> The latter is unlikely to work on Arm because we don't setup the scratch
>>> page. AFAIU, for that, we would need to implement the callback
>>> quarantine_init().
>>>
>>> I would expect 1 and 2 work. That said, I think 1. would behave like 2.
>>> because on Arm the device should not be automatically re-assigned to
>>> dom0. I know this is correct for platform device, but will it be valid
>>> for PCI as well?
>>
>> In a system with dom0 where the guest is created from the xl toolstack, we 
>> rely on "xl pci-assignable-add". Upon domain destruction, the device 
>> automatically gets assigned to domIO.
> 
> Ok. To clarify, does this mean any DMA will fault, the same as for
> platform device?

Yes, when the PCI device is assigned to domIO, any DMA from the device will 
produce a SMMU fault. The value of the quarantine= option doesn't change this 
behavior.

>> However, there's nothing preventing a user from attempting to invoke "xl 
>> pci-assignable-remove", which should assign the device back to dom0, but it 
>> is not automatic.
> 
> I don't think we want to fully prevent a user to re-assign a device to
> dom0. But we at least want to avoid re-assigning the device to dom0 by
> default. After that a user can reset the device before it gets
> re-assigned to dom0.
> 
> Cheers,
> 
> -- 
> Julien Grall