lkml.org 
[lkml]   [2022]   [Apr]   [6]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [PATCH RFC v2 03/11] iommu/sva: Add iommu_domain type for SVA
From
On 2022-04-06 06:58, Tian, Kevin wrote:
>> From: Jason Gunthorpe <jgg@nvidia.com>
>> Sent: Wednesday, April 6, 2022 9:24 AM
>>
>> On Wed, Apr 06, 2022 at 01:00:13AM +0000, Tian, Kevin wrote:
>>
>>>> Because domains wrap more than just the IOPTE format, they have
>>>> additional data related to the IOMMU HW block itself. Imagine a SOC
>>>> with two IOMMU HW blocks that can both process the CPU IOPTE format,
>>>> but have different configuration.
>>>
>>> Curious. Is it hypothesis or real? If real can you help give a concrete
>>> example?
>>
>> Look at arm_smmu_attach_dev() - the domain has exactly one smmu
>> pointer which contains the base address for the SMMU IP block. If the
>> domain doesn't match the smmu pointer from the struct device it won't
>> allow attaching.
>>
>> I know of ARM SOCs with many copies of the SMMU IP block.
>>
>> So at least with current drivers ARM seems to have this limitation.
>>
>
> I saw that code, but before this series it is used only for stage-2 instead
> of SVA. and I didn't see similar check in the old sva related paths (though
> it doesn't use domain):
>
> arm_smmu_master_sva_enable_iopf()
> arm_smmu_master_enable_sva{}
> __arm_smmu_sva_bind()
>
> If I didn't overlook some trick hiding in the call chain of those functions,
> is there a bug in the existing SMMU sva logic or is it conceptually correct
> to not have such check for SVA?

The current SVA APIs are all device-based, so implicitly reflect
whichever SMMU instance serves the given device. Once domains come into
the picture, callers are going to have to be more aware that a domain
may be specific to a particular IOMMU instance, and potentially allocate
separate domains for separate devices to represent the same address
space, much like vfio_iommu_type1_attach_group() does.

It's not really worth IOMMU drivers trying to support a domain spanning
potentially-heterogeneous instances internally, since they can't
reasonably know what matters in any particular situation. That's
primarily why we've never tried to do it in the SMMU drivers. It's a lot
easier for relevant callers to look at what they get and figure out
whether any mismatch in capabilities is tolerable or not.

Robin.

> If the former then yes we have to take SMMU IP block into consideration
> thus could have multiple domains per CPU page table. If the latter then
> this is not a valid example for that configuration.
>
> Which one is correct?
>
> Thanks
> Kevin

\
 
 \ /
  Last update: 2022-04-06 17:25    [W:3.859 / U:0.008 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site