[RFC PATCH 1/1] drm/amdgpu: add initial support for pci error handler

[Luben Tuikov]

On 2020-08-14 4:10 p.m., Alex Deucher wrote:
>> I see DRM as more of a unifying layer (perhaps long term), as opposed
>> to a *library* which LLDDs call into. Then LLDDs would provide an interface
>> to the hardware. This will help us avoid many of the deadlocks and
>> synchronization issues which were outlined in a recent memory presentation.
>> It will also help us avoid code redundancy in LLDDs.
> 
> We are actually trying to move away from that model.  drm started out
> as a midlayer and it caused all kinds of pain.  E.g., some drm devices
> are usb devices, some are pci, some are platform devices, etc.  A flow
> that worked for one didn't always work well for another.  In general
> we should follow the driver model for the specific device model of
> your driver and then use drm helper functions to deal with the core
> drm functionality.  E.g., drm shouldn't be in the business of dealing
> with pci power management.  The driver should interact directly with
> the pci subsystem for that.

Ah, thanks Alex for the clarification. So a library of sorts--got it.

Regards,
Luben