On Mon, Jan 19, 2026 at 12:07:55PM +0100, Lorenzo Bianconi wrote:

quoted

quoted

Airoha folks reported the NPU hw can't provide the PCIe Vendor/Device ID info
of the connected WiFi chip.
I guess we have the following options here:
- Rely on the firmware-name property as proposed in v1
- Access the PCIe bus from the NPU driver during probe in order to enumerate
  the PCIe devices and verify WiFi chip PCIe Vendor/Device ID
- During mt76 probe trigger the NPU fw reload if required. This approach would
  require adding a new callback in airoha_npu ops struct (please note I have
  not tested this approach and I not sure this is really doable).

What i'm wondering about is if the PCIe slots are hard coded in the
firmware.  If somebody builds a board using different slots, they
would then have different firmware? Or if they used the same slots,
but swapped around the Ethernet and the WiFi, would it need different
firmware?

As pointed out by Benjamin, the NPU is a generic Risc-V cpu cluster and it is
used to move packets from/to ethernet DMA rings to/from WiFi DMA rings without
involving the host cpu (similar to what we have for MTK with WED module).
I think the PCIe slot info is not necessary for the NPU to work since it is
configured by ethernet (airoha-eth) and wireless drivers (mt76) with DMA ring
addresses to use via the airoha npu ops APIs, NPU just moves data between the
DMA rings according to my understanding.

quoted

So is the firmware name a property of the board?

We need to run different binaries on the NPU based on the MT76 WiFi chip
available on the board since the MT76 DMA rings layout changes between MT76 SoC
revisions (e.g. Egle MT7996 vs Kite MT7992). In this sense, I agree, the
firmware name is a board property.

quoted

If the PCIe slots are actually hard coded in the NPU silicon, cannot
be changed, then we might have a different solution, the firmware name
might be placed into a .dtsi file, or even hard coded in the driver?

IIUC what you mean here, it seems the solution I proposed in v1 (using
firmware-name property), right?
In this case we can't hard code the firmware name in the NPU driver since
we can't understand the MT76 WiFi chip revision running on the board at
the moment (MT76 would need to provide this info during MT76 probe,
please take a look to the option 3 in my previous email).

quoted

quoted

What do you think? Which one do you prefer?

I prefer to try to extract more information for the Airoha folks. What
actually defines the firmware? Does the slots used matter? Does it
matter what device goes in what slots? Is it all hard coded in
silicon? Is there only one true hardware design and if you do anything
else your board design is FUBAR, never to be supported?

I think the firmware is defined by the board hw configuration (e.g. MT76
SoC revision) and not by the specific PCIe slot used.
I do not think we have these info hardcoded in the silicon since NPU is a
generic RiscV cpu (this has been confirmed by airoha folks).

Just to make sure everything is clear and talking on this in very
simple words, there isn't anything ""hardcoded"" or strange.

For """""""reasons""""""" (I assume space constraints or NPU CPU
limitation) it's not possible to have a single NPU firmware to support
both WiFi card.

The NPU do simple task like configuring WED registers and handling DMA
descriptor/some WiFi offload. Such configuration is specific to the WiFi
card and it's not the same between MT7996 and MT7992.

This is why specific firmware is needed. The specific NPU firmware have
support for only ONE of the 2 WiFi card and doesn't support configuring
and handling stuff for the other. (the code is not built in the
firmware)

From the kernel side (in the MT76 code) we just instruct the NPU to
start offloading stuff (if present) and all the SoC feature for WiFi
offload are used. (WED, special DMA path, ...)

The possible combination that NPU can be used currently are the
following:
- Ethernet offload (all NPU firmware)
- Ethernet offload + WiFi MT7996 (NPU firmware with MT7996 support)
- Ethernet offload + WiFi MT7992 (NPU firmware with MT7992 support)

The NPU makes use of feature already present in the SoC and makes use of
reserved space in RAM for DMA handling so it really don't care of where
the WiFi card is present (this is what I mean with nothing is hardcoded)

I hope we are not getting annoying with insisting on the firmware-names
solution.

My personal taste on this is that hardcoding the name in the driver
seems a bit wrong and creating a way to dynamically select the firmware
based on what is present in the hardware would be great but would
introduce LOTS of COMPLEXITY for WiFi router that ship with a single
WiFi card and would have their own dedicated .dts

To make this generic enough an idea might be to have simple .dtsi with
prefilled firmware names.

- en7581-npu-mt7992.dtsi
- en7581-npu-mt7996.dtsi

But they would only contain a single node with a single string.

Hope this more practical explaination clears any doubt of the
implementation.

-- 
	Ansuel