Re: [PATCH RFC] riscv: dts: spacemit: Add DMA translation buses for K1

From: Vivian Wang
Date: Thu Jun 19 2025 - 11:44:05 EST

Hi Alex,

Thank you for your comments on this.

On 6/19/25 23:11, Alex Elder wrote:
On 6/17/25 12:21 AM, Vivian Wang wrote:
The SpacemiT K1 has various static translations of DMA accesses. Add
these as simple-bus nodes. Devices actually using these translation will
be added in later patches.

The bus names are assigned according to consensus with SpacemiT [1].

[1] https://lore.kernel.org/all/CAH1PCMaC+imcMZCFYtRdmH6ge=dPgnANn_GqVfsGRS=+YhyJCw@xxxxxxxxxxxxxx/

So what you include here very closely matches what Guodong
said in the message above.  Yours differs from his proposal
and that makes it hard to compare them.  I have a few comments
on that below.

Signed-off-by: Vivian Wang <wangruikang@xxxxxxxxxxx>
---
This is my concrete proposal for representing DMA translations for
SpacemiT K1.

It's worth acknowledging that this is derived from what Guodong
proposed (it's not "your" proposal in that respect).  That said,
yours is a more complete and "formal" RFP than what he wrote.

I had thought that since the addresses were already there in vendor's DT [2], and the names were provided by SpacemiT, anything other than the names was "well-known information". In retrospect, I should have made the chain of information of this clearer and make it explicit that this was based on Guodong's note.

So, just to be clear, the information in my proposal is based on Guodong's reply [1] (link the quoted text), which I had assumed, but not explicitly confirmed, was based on already addresses in SpacemiT's DT and names provided by SpacemiT.

[2]: https://github.com/spacemit-com/linux-k1x/blob/k1/arch/riscv/boot/dts/spacemit/k1-x.dtsi

For context, memory on the SpacemiT K1 is split into two chunks:

- 0x0000_0000 to 0x8000_0000: First 2 GiB of memory
- 0x1_0000_0000 above: Rest of memory

DMA-capable devices on the K1 all have access to the lower 2G of memory
through an identity mapping. However, for the upper region of memory,
each device falls under one of six different mappings. The mappings are
provided in this patch as simple-bus nodes that device nodes should be
added to.

This patch is an RFC because it is not meant to be applied, or at least,
not certainly meant to be applied. Instead, this is an attempt to come
to a consensus on how these bus nodes should look like.

I think the above is what Krzysztof might not have seen.  Perhaps
it could have been made more clear--maybe in the "main" description
section (above the ---) or even the subject line.

Yeah, that's my mistake in organizing the paragraphs.

More specifically, I propose that the process proceeds as follows:

- Firstly, relevant parties agree on these bus nodes given here.
- After that, each time the first user of a bus appears, the series
   should include a patch to add the bus required for that driver.
- If a driver being submitted uses the same bus as another one that has
   been submitted but hasn't yet landed, it can depend on the bus patch
   from that previous series.

Getting agreement is good, but otherwise this is basically
the process Guodong was suggesting, right?

Hmm, actually re-reading the discussion now, I realized that I may have come to this late and missed out on some previous discussions, which were alluded to in Yixun's messages. (This is again thread around link [1] in quoted text.) This led me to believe that some of these were not really agreed upon.

I also realized I think one of the things I may have not yet made clear is that I would like the bus node to be a *separate* patch. I think this makes sense, because it's dealing with two different subsystems.


For conventions regarding coding style, I propose that:

- #address-cells and #size-cells are 2 for consistency
- These bus nodes are put at the end of /soc, inside /soc
- These bus nodes are sorted alphabetically, not in vendor's order
- Devices are added into *-bus nodes directly, not appended towards the
   end with a label reference

I do like that you're trying to be more complete and explicit
on what you think needs agreement on.

Being thorough was the main goal of this RFC. If there was previous agreement on how dma-ranges should be done, I'm sorry for missing them, but from my observations on the mailing list on how these ended up into patches I really haven't seen much consistency. Maybe there was misunderstanding, which I'm hoping to clear up.

(Although see my paragraph above, maybe I haven't been thorough enough.)

The K1 DMA translations are *not* interconnects, since they do not
provide any configuration capabilities.

These bus nodes names and properties are provided compliant with
"simple-bus" bindings, and should pass "make dtbs_check".

Remaining questions:

- Should storage-bus exist? Or should drivers under it simply specify
   32-bit DMA?

Explicitly saying storage devices have one-to-one mapping
seems informative, to me.

---
  arch/riscv/boot/dts/spacemit/k1.dtsi | 53 ++++++++++++++++++++++++++++++++++++
  1 file changed, 53 insertions(+)

The short summary of what differs between your proposal
and what Guodong said is:
- You sort nodes alphabetically, Guodong did not
- You dropped the unit address
- You dropped the comments he had, which indicated which
  devices "belonged" to each mapping
- You added a compatible property to each ("simple-bus")
- You added an explicit (empty) ranges property to each
- You add #address-cells and #size-cells properties, both 2
- Your dma-ranges properties are identical to Guodong's,
  for all nodes

That was a good summary. Thanks!

My main goal of organizing the bus this way is making it actually pass "make dtbs_check". I'm not sure if Krzysztof still objects to my reading of simple-bus.yaml though.

By the way, I don't think I will be making an RFC v2 of this. I think we should get everything sorted under this one thread.

Thanks again for taking a look.

Vivian "dramforever" Wang