[etnaviv-next v13 7/7] drm/etnaviv: Add support for vivante GPU cores attached via PCI(e)

[Sui Jingfeng]

Hi,

On 2024/2/13 22:38, Maxime Ripard wrote:
> On Sat, Feb 10, 2024 at 12:25:33AM +0800, Sui Jingfeng wrote:
>> On 2024/2/9 23:15, Maxime Ripard wrote:
>>> On Fri, Feb 09, 2024 at 12:02:48PM +0100, Daniel Vetter wrote:
>>>> On Thu, Feb 08, 2024 at 04:27:02PM +0100, Maxime Ripard wrote:
>>>>> On Wed, Feb 07, 2024 at 10:35:49AM +0100, Daniel Vetter wrote:
>>>>>> On Wed, Feb 07, 2024 at 01:27:59AM +0800, Sui Jingfeng wrote:
>>>>>>> The component helper functions are the glue, which is used to bind multiple
>>>>>>> GPU cores to a virtual master platform device. Which is fine and works well
>>>>>>> for the SoCs who contains multiple GPU cores.
>>>>>>>
>>>>>>> The problem is that usperspace programs (such as X server and Mesa) will
>>>>>>> search the PCIe device to use if it is exist. In other words, usperspace
>>>>>>> programs open the PCIe device with higher priority. Creating a virtual
>>>>>>> master platform device for PCI(e) GPUs is unnecessary, as the PCI device
>>>>>>> has been created by the time drm/etnaviv is loaded.
>>>>>>>
>>>>>>> we create virtual platform devices as a representation for the vivante GPU
>>>>>>> ip core. As all of subcomponent are attached via the PCIe master device,
>>>>>>> we reflect this hardware layout by binding all of the virtual child to the
>>>>>>> the real master.
>>>>>>>
>>>>>>> Signed-off-by: Sui Jingfeng <sui.jingfeng at linux.dev>
>>>>>> Uh so my understanding is that drivers really shouldn't create platform
>>>>>> devices of their own. For this case here I think the aux-bus framework is
>>>>>> the right thing to use. Alternatively would be some infrastructure where
>>>>>> you feed a DT tree to driver core or pci subsystem and it instantiates it
>>>>>> all for you correctly, and especially with hotunplug all done right since
>>>>>> this is pci now, not actually part of the soc that cannot be hotunplugged.
>>>>> I don't think we need intermediate platform devices at all. We just need
>>>>> to register our GPU against the PCI device and that's it. We don't need
>>>>> a platform device, we don't need the component framework.
>>>> Afaik that's what this series does. The component stuff is for the
>>>> internal structure of the gpu ip, so that the same modular approach that
>>>> works for arm-soc also works for pci chips.
>>> But there should be a single PCI device, while we have multiple "DT"
>>> devices, right? Or is there several PCI devices too on that PCI card?
>>
>> There is only a single PCI(e) device on that PCI(e) card, this single
>> PCI(e) device is selected as the component master. All other Hardware IP
>> blocks are shipped by the single PCI(e) master. It may includes Display
>> controllers, GPUs, video decoders, HDMI display bridges hardware unit etc.
>>
>> But all of those Hardware IP share the same MMIO registers PCI BAR, this
>> PCI BAR is a kind of PCI(e) MEM resource. It is a relative *big* chunk,
>> as large as 32MB in address ranges for the JingJia Macro dGPU. Therefore,
>> I break the whole registers memory(MMIO) resource into smaller pieces by
>> creating platform device manually, manually created platform device is
>> called as virtual child in this series.
>>
>> In short, we cut the whole into smaller piece, each smaller piece is a
>> single hardware IP block, thus deserve a single device driver. We will
>> have multiple platform devices if the dGPU contains multiple hardware
>> IP block. On the driver side, we bind all of the scattered driver module
>> with component.
> That's kind of my point then. If there's a single device, there's no
> need to create intermediate devices and use the component framework to
> tie them all together. You can have a simpler approach where you create
> a function that takes the memory area it operates on (and whatever
> additional resource it needs: interrupt, clocks, etc.) and call that
> directly from the PCIe device probe, and the MMIO device bind.


Yes, you are right. I have implemented the method just as you told me at
V12 of this series (see 0004 patch at [1]). But at V13, I suddenly realized
that the component code path plus(+) non-component code path yield a
"side-by-side" implement. The old non-component approach can not support
binding multiple sub-core, it can only support one Vivante GPU IP core case.
But there are dGPU which shipping two identical core.

While, the component-based approach implemented at this version is the most
universal and the most flexible and extensible. We could bind a virtual
display driver and/or a real display driver to the real master (refer to the
PCI(e) device).  The PCI(e) device is responsible for present the DRM service
to userspace, like a leader or agent. All other sub hardware and software are
hiding behind of the master.


Besides, Lucas asked me implement the driver like this way at V10 (see [2])

Lucas said:

"My favorite option would be to just always use the component path
even when the GPU is on a PCI device to keep both paths mostly aligned.
One could easily image both a 3D and a 2D core being made available
though the same PCI device."

Lucas, are you watching? How about this version? Can you help to review please?


[1] https://patchwork.freedesktop.org/series/127084/

[2] 
https://lore.kernel.org/dri-devel/0f1095ef333da7ea103486a1121ca9038815e57c.camel@pengutronix.de/


> Maxime