[PATCH v2 05/14] dt-bindings: arm: mediatek: document the pcie mirror node on MT7622

[Felix Fietkau]

On 07.04.22 19:16, Rob Herring wrote:
> On Wed, Apr 06, 2022 at 01:01:06PM +0200, Felix Fietkau wrote:
>> 
>> On 06.04.22 10:20, Krzysztof Kozlowski wrote:
>> > On 05/04/2022 21:57, Felix Fietkau wrote:
>> > > From: Lorenzo Bianconi <lorenzo at kernel.org>
>> > > 
>> > > This patch adds the pcie mirror document bindings for MT7622 SoC.
>> > > The feature is used for intercepting PCIe MMIO access for the WED core
>> > > Add related info in mediatek-net bindings.
>> > > 
>> > > Signed-off-by: Lorenzo Bianconi <lorenzo at kernel.org>
>> > > Signed-off-by: Felix Fietkau <nbd at nbd.name>
>> > > ---
>> > >  .../mediatek/mediatek,mt7622-pcie-mirror.yaml | 42 +++++++++++++++++++
>> > 
>> > Eh, I wanted to ask to not put it inside arm/, but judging by your usage
>> > - you did not create drivers for both of these (WED and PCIe mirror).
>> > 
>> > You only need them to expose address spaces via syscon.
>> > 
>> > This actually looks hacky. Either WED and PCIe mirror are part of
>> > network driver, then add the address spaces via "reg". If they are not,
>> > but instead they are separate blocks, why you don't have drivers for them?
>> The code that uses the WED block is built into the Ethernet driver, but not
>> all SoCs that use this ethernet core have it. Also, there are two WED
>> blocks, and I'm not sure if future SoCs might have a different number of
>> them at some point.
>> The WED code also needs to access registers of the ethernet MAC.
>> One reason for having a separate device is this:
>> As long as WED is not in use, ethernet supports coherent DMA for increased
>> performance. When the first wireless device attaches to WED, IO coherency
>> gets disabled and the ethernet DMA rings are cleaned up and allocated again,
>> this time with the struct device of WED (which doesn't have the dma-coherent
>> property).
> 
> I'm pretty sure there are assumptions in the driver core that coherency
> is not changing on the fly. In any case, if it is, using 'dma-coherent'
> is not appropriate. You obviously have another method to determine
> whether you are coherent or not.
It's not really on the fly. Before changing coherency, all DMA memory is 
freed, and the subsequent reallocation uses the struct device of the WED 
core, which does not have the dma-coherent property.

- Felix