Re: [PATCH v9 1/5] dt-bindings: reserved-memory: Document iommu-addresses

[Robin Murphy <robin.murphy@arm.com>]

On 2022-10-07 16:22, Thierry Reding wrote:
> On Fri, Oct 07, 2022 at 03:21:46PM +0100, Robin Murphy wrote:
>> On 2022-10-07 14:54, Thierry Reding wrote:
>>> On Fri, Oct 07, 2022 at 02:45:31PM +0100, Robin Murphy wrote:
>>>> On 2022-09-23 13:35, Thierry Reding wrote:
>>>>> From: Thierry Reding <treding@nvidia.com>
>>>>>
>>>>> This adds the "iommu-addresses" property to reserved-memory nodes, which
>>>>> allow describing the interaction of memory regions with IOMMUs. Two use-
>>>>> cases are supported:
>>>>>
>>>>>      1. Static mappings can be described by pairing the "iommu-addresses"
>>>>>         property with a "reg" property. This is mostly useful for adopting
>>>>>         firmware-allocated buffers via identity mappings. One common use-
>>>>>         case where this is required is if early firmware or bootloaders
>>>>>         have set up a bootsplash framebuffer that a display controller is
>>>>>         actively scanning out from during the operating system boot
>>>>>         process.
>>>>>
>>>>>      2. If an "iommu-addresses" property exists without a "reg" property,
>>>>>         the reserved-memory node describes an IOVA reservation. Such memory
>>>>>         regions are excluded from the IOVA space available to operating
>>>>>         system drivers and can be used for regions that must not be used to
>>>>>         map arbitrary buffers.
>>>>
>>>> Bah, I've only just realised: don't we also need to change the "oneOf:
>>>> required: ..." schema to permit "iommu-addresses" without "reg" or "size"?
>>>
>>> Hm... good point. I think at least we'll want another:
>>>
>>>        - required:
>>>            - iommu-addresses
>>>
>>> in there. I wonder if we also need to avoid the combination of "size"
>>> and "iommu-addresses". When "size" is specified, is it guaranteed that
>>> those regions will be allocated before the direct mapping needs to be
>>> created?
>>
>> Well, it couldn't really be a direct mapping anyway. In general I don't
>> think that combination makes any sense, since the presence of
>> "iommu-addresses" means one of two things; either it says the IOVA range is
>> carved out for some special purpose or just unusable, in which case
>> allocating any memory to back it would surely be pointless, or it's saying
>> don't touch these addresses because the device is already accessing them,
>> thus the underlying physical memory must be allocated somewhere already.
> 
> I thought perhaps there could be cases where it is known that a
> controller needs to access memory in a certain I/O virtual region but
> doesn't actually care where that lives in physical memory and also does
> not rely on that memory have been previously set up (pre-filled, or
> whatever). Say you've got a micro-controller in a system that needs its
> firmware in a given region, but the OS can set up that region without
> any other limitations. One could use "size" and "iommu-addresses" to
> make sure the region is allocated with a specific size and located in a
> specific I/O virtual region. Not sure if that's perhaps a bit exotic,
> though.

Yeah, that was the closest case I could think of as well, but I'd really 
rather not encourage people to abuse DT that way. If a kernel driver is 
loading firmware and initialising the device from scratch then it should 
be able to sort everything out at runtime without DT involvement. Even 
if the firmware is somehow massive enough to warrant an early dynamic 
carveout rather than a regular page/CMA allocation, there's still no 
good excuse for the driver not to manage its own address space constraints.

On the other hand if the device really does need its firmware at a 
specific hard-coded address than that would need a fixed physical 
reservation anyway, since the DT can't assume that the OS is definitely 
going to use IOMMU translation.

Thanks,
Robin.