Re: [PATCH v3 0/4] RPi: SD/WiFi ACPI updates

["Ard Biesheuvel" <ardb@kernel.org>]

On Wed, 17 Feb 2021 at 08:59, Andrei Warkentin <awarkentin@vmware.com> wrote:
>
> I'm not sure I truly follow the conversation you're having (sorry), but I'll raise my hand as a _DMA user. I'm not that married to it, but I haven't added support for the alternate IORT-based DMA range clamping. If long term we want to bail on _DMA and adopt IORT exclusively, I'd love to know that so I can plan accordingly...
>

SBSA requires that all memory is DMA addressable by all non-secure
masters, but Linux/arm64 can deal with systems that violate this
requirement, as long as the DMA limit is 4 GB or higher. The reason is
that the memory region from which bounce buffers are allocated is
created early during boot, and is allocated in the first 4 GB of
memory.

For devices such as RPi4, we have relaxed this requirement somewhat
when booting Linux/arm64 in ACPI mode, by adding an early IORT scan
that finds the smallest DMA limit, and using that limit for the bounce
buffer region if it is smaller than 4 GB. (At this point during the
boot, the AML interpreter is not up and running yet)

On Linux, the _DMA method supersedes the IORT limit, and so the IORT
limit is only used for sizing the bounce buffer region, and we get the
full 3GB range for XHCI.

In this particular case, I'd assume there is no ambiguity if the IORT
and the _DMA method both produce the same 1 GB DMA limit.