Re: [PATCH 4/4] iommu/vt-d: Remove lazy allocation of domains
From: James Sewart
Date: Sat Mar 09 2019 - 06:57:06 EST
> On 9 Mar 2019, at 01:53, Lu Baolu <baolu.lu@xxxxxxxxxxxxxxx> wrote:
> Hi James,
> On 3/9/19 12:57 AM, James Sewart wrote:
>> Hey Lu,
>>> On 8 Mar 2019, at 03:09, Lu Baolu<baolu.lu@xxxxxxxxxxxxxxx> wrote:
>>>>> Do you mind if I work on top of your patches for further cleanups and
>>>>> sign off a v2 together with you?
>>>> Sure, sounds good. Iâll fixup patch 3 and have a go at integrating
>>>> iommu_prepare_isa into get_resv_regions. This should make the initial
>>>> domain logic here quite concise.
>>> Here attached three extra patches which I think should be added before
>>> PATCH 3/4, and some further cleanup changes which you can merge with
>>> PATCH 4/4.
>>> These two patches aim to add a generic method for vendor specific iommu
>>> drivers to specify the type of the default domain for a device. Intel
>>> iommu driver will register an ops for this since it already has its own
>>> identity map adjudicator for a long time.
>> This seems like a good idea, but as domain alloc is only called for the
>> default domain on first device attached to a group, we may miss checking
>> whether a device added later should have an identity domain. Should there
>> be paths to downgrade a groups domain if one of the devices requires one?
> Good catch!
> This is supposed to be handled in iommu_no_mapping(). But, obviously
> current code sticks to lazy domain allocation. I'm not sure whether
> there are any real such cases, but we should handle it in a clean way.
> My idea is that we could downgrade to multiple domains per group (just
> like what we have now) in this case and print a kernel message for this.
I think if a device requires an identity domain, then it should ignore
attempts to attach to something else. A print to warn a user about this
would be a good idea.
I figure during attach: if iommu_no_mapping() then attach to si_domain and
print, else continue with the given domain.
> Or any other thoughts?
> Best regards,
> Lu Baolu