Re: [PATCH 03/11] dt-bindings: Add RISC-V IOMMU bindings

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

On 2023/7/24 21:23, Zong Li wrote:
>>>>> In RISC-V IOMMU, certain devices can be set to bypass mode when the
>>>>> IOMMU is in translation mode. To identify the devices that require
>>>>> bypass mode by default, does it be sensible to add a property to
>>>>> indicate this behavior?
>>>> Bypass mode for a device is a property of that device (similar to dma-coherent)
>>>> and not of the IOMMU. Other architectures (ARM and x86) never added such
>>>> a device property for bypass mode so I guess it is NOT ADVISABLE to do it.
>>>>
>>>> If this is REALLY required then we can do something similar to the QCOM
>>>> SMMU driver where they have a whitelist of devices which are allowed to
>>>> be in bypass mode (i.e. IOMMU_DOMAIN_IDENTITY) based their device
>>>> compatible string and any device outside this whitelist is blocked by default.
>>>>
>>> I have considered that adding the property of bypass mode to that
>>> device would be more appropriate. However, if we want to define this
>>> property for the device, it might need to go through the generic IOMMU
>>> dt-bindings, but I'm not sure if other IOMMU devices need this. I am
>>> bringing up this topic here because I would like to explore if there
>>> are any solutions on the IOMMU side, such as a property that indicates
>>> the phandle of devices wishing to set bypass mode, somewhat similar to
>>> the whitelist you mentioned earlier. Do you think we should address
>>> this? After all, this is a case of RISC-V IOMMU supported.
>> Bypass mode is a common feature across IOMMUs. Other IOMMUs don't
>> have a special property for bypass mode at device-level or at IOMMU level,
>> which clearly indicates that defining a RISC-V specific property is not the
>> right way to go.
>>
>> The real question is how do we set IOMMU_DOMAIN_IDENTITY (i.e.
>> bypass/identity domain) as the default domain for certain devices ?
>>
>> One possible option is to implement def_domain_type() IOMMU operation
>> for RISC-V IOMMU which will return IOMMU_DOMAIN_IDENTITY for
>> certain devices based on compatible string matching (i.e. whitelist of
>> devices). As an example, refer qcom_smmu_def_domain_type()
>> of drivers/iommu/arm/arm-smmu/arm-smmu-qcom.c
>>
> That is indeed one way to approach it, and we can modify the
> compatible string when we want to change the mode. However, it would
> be preferable to explore a more flexible approach to achieve this
> goal. By doing so, we can avoid hard coding anything in the driver or
> having to rebuild the kernel  whenever we want to change the mode for
> certain devices. While I have considered extending a cell in the
> 'iommus' property to indicate a device's desire to set bypass mode, it
> doesn't comply with the iommu documentation and could lead to
> ambiguous definitions.

Hard coding the matching strings in the iommu driver is definitely not a
preferable way. A feasible solution from current code's point of view is
that platform opt-in the device's special requirements through DT or
ACPI. And in the def_domain_type callback, let the iommu core know that,
hence it can allocate a right type of domain for the device.

Thoughts?

Best regards,
baolu