Re: [PATCH v2 0/7] iommu/vt-d: Fix-up device-domain relationship by refactoring to use iommu group default domain.

[Lu Baolu <baolu.lu@linux.intel.com>]

Hi James,

On 3/19/19 9:35 PM, James Sewart wrote:
> Hey Lu,
> 
>> On 15 Mar 2019, at 03:13, Lu Baolu <baolu.lu@linux.intel.com> wrote:
>>
>> Hi James,
>>
>> On 3/14/19 7:56 PM, James Sewart wrote:
>>> Patches 1 and 2 are the same as v1.
>>> v1-v2:
>>>    Refactored ISA direct mappings to be returned by iommu_get_resv_regions.
>>>    Integrated patch by Lu to defer turning on DMAR until iommu.c has mapped
>>> reserved regions.
>>>    Integrated patches by Lu to remove more unused code in cleanup.
>>> Lu: I didn't integrate your patch to set the default domain type as it
>>> isn't directly related to the aim of this patchset. Instead patch 4
>>
>> Without those patches, user experience will be affected and some devices
>> will not work on Intel platforms anymore.
>>
>> For a long time, Intel IOMMU driver has its own logic to determine
>> whether a device requires an identity domain. For example, when user
>> specifies "iommu=pt" in kernel parameter, all device will be attached
>> with the identity domain. Further more, some quirky devices require
>> an identity domain to be used before enabling DMA remapping, otherwise,
>> it will not work. This was done by adding quirk bits in Intel IOMMU
>> driver.
>>
>> So from my point of view, one way is porting all those quirks and kernel
>> parameters into IOMMU generic layer, or opening a door for vendor IOMMU
>> driver to determine the default domain type by their own. I prefer the
>> latter option since it will not impact any behaviors on other
>> architectures.
> 
> I see your point. I’m not confident that using the proposed door to set a
> groups default domain has the desired behaviour. As discussed before the
> default domain type will be set based on the desired type for only the
> first device attached to a group. I think to change the default domain
> type you would need a slightly different door that wasn’t conditioned on
> device.

I think this as another problem. Just a summarize for the ease of
discussion. We saw two problems:

1. When allocating a new group for a device, how should we determine the
type of the default domain? This is what my proposal patches trying to
address.

2. If we need to put a device into an existing group which uses a
different type of domain from what the device desires to use, we might
break the functionality of the device. For this problem I'd second your
proposal below if I get your point correctly.

> 
> For situations where individual devices require an identity domain because
> of quirks then maybe calling is_identity_map per device in
> iommu_group_get_for_dev is a better solution than the one I proposed.
> 

Do you mean if we see a quirky device requires a different domain type
other than the default domain type of the group, we will assign a new
group to it? That looks good to me as far as I can see. I suppose this
should be done in vt-d's ops callback.

Best regards,
Lu Baolu