[PATCH 0/3] Apple M1 DART IOMMU driver

[Sven Peter]

On Fri, Mar 26, 2021, at 20:59, Arnd Bergmann wrote:
> On Fri, Mar 26, 2021 at 6:51 PM Sven Peter <sven at svenpeter.dev> wrote:
> >   dart-sio: 0021c000 fbde4000 (at least their Secure Enclave/TPM co-processor)
> 
> Same here:
>     dart-sio {
>        vm-base = <0x0>;
>        vm-size = <0xfc000000>;
>        pio-vm-base = <0xfd000000>;
>       pio-vm-size = <0x2000000>;
>       pio-granularity = <0x1000000>;
>    }
> 
> There are clearly two distinct ranges that split up the 4GB space again,
> with a small hole of 16MB (==pio-granularity) at the end of each range.
> 
> The "pio" name might indicate that this is a range of addresses that
> can be programmed to point at I/O registers in another device, rather
> than pointing to RAM.
> 
>            Arnd
>

Very interesting observation!

Mark and I have discussed this a little bit further on IRC. Mark also successfully
used the PCIe DARTs with a DMA window outside of the one specified by vm-base/vm-size
in the ADT.

I believe that the (pio-)vm-base/size properties merely specify the ranges their
allocator uses and do not describe actual hardware limitations. Mark also suggested
that they might reserve memory at the beginning to find bugs similar to how one
might not allow to map memory at 0x0.

I have also done a few more experiments and figured out that if I put the IOMMU
into bypass mode (which doesn't seem to work for all IOMMUs/master combinations
which is why I'll leave it out of this series for now until I figure out more
details) I *can* use the full address space. I think the limitation
is therefore imposed by the translation hardware inside the IOMMU and not by
the bus/the interconnect.

If that's correct I think the right place to enforce this is to just limit
the aperture inside the DART driver to a 32bit address space whenever address
translation is enabled.


Thanks,

Sven