Re: [PATCH 06/19] xen: Clean up asm-generic/device.h

[Jan Beulich <jbeulich@xxxxxxxx>]

On 05.06.2025 18:48, Alejandro Vallejo wrote:
> On Thu Jun 5, 2025 at 4:20 PM CEST, Jan Beulich wrote:
>> On 05.06.2025 16:15, Alejandro Vallejo wrote:
>>> On Mon Jun 2, 2025 at 4:24 PM CEST, Jan Beulich wrote:
>>>> On 02.06.2025 16:19, Alejandro Vallejo wrote:
>>>>> On Mon Jun 2, 2025 at 9:51 AM CEST, Jan Beulich wrote:
>>>>>> On 30.05.2025 14:02, Alejandro Vallejo wrote:
>>>>>>> --- a/xen/include/asm-generic/device.h
>>>>>>> +++ b/xen/include/asm-generic/device.h
>>>>>>> @@ -6,9 +6,7 @@
>>>>>>>  
>>>>>>>  enum device_type
>>>>>>>  {
>>>>>>> -#ifdef CONFIG_HAS_DEVICE_TREE
>>>>>>>      DEV_DT,
>>>>>>> -#endif
>>>>>>
>>>>>> Why would this enumerator need exposing on a non-DT arch? In fact I 
>>>>>> would have
>>>>>> hoped for ...
>>>>>
>>>>> A non-DT arch would not include this. x86 doesn't.
>>>>
>>>> Both here and ...
>>>>
>>>>>>>      DEV_PCI
>>>>>>
>>>>>> ... this to be hidden for arch-es not supporting PCI.
>>>>>>
>>>>>> Similar concerns elsewhere in this change.
>>>>>
>>>>> This file is exclusively used by arches supporting DT to abstract away 
>>>>> where
>>>>> the device came from. x86 does not use it at all, and while it wouldn't be
>>>>> impossible to compile-out DEV_PCI, it would needlessly pollute the 
>>>>> codebase with
>>>>> no measurable gain, because the abstractions still need to stay.
>>>>
>>>> ... here: In "xen/include/asm-generic/device.h" there's nothing at all 
>>>> saying
>>>> that this file is a DT-only one. Instead there is something in there saying
>>>> that it's suitable to use in the entirely "generic" case.
>>>>
>>>> Jan
>>>
>>> Try to use it from x86 and observe the build system catch fire. It could be 
>>> made
>>> to not go on fire, but it implies heavy refactoring in x86 (particularly 
>>> IOMMU
>>> code) for no good reason because there's no devices in a DTB to 
>>> disambiguate.
>>>
>>> How about adding this to the top of the header?
>>>
>>> ```
>>>  /*
>>>   * This header helps DTB-based architectures abstract away where a 
>>> particular
>>>   * device comes from; be it the DTB itself or enumerated on a PCI bus. 
>>>   */
>>>
>>>   [snip]
>>>
>>>  #ifndef CONFIG_HAS_DEVICE_TREE
>>>  #error "Header meant to be used exclusively by DTB-base architectures."
>>>  #endif
>>> ```
>>
>> Might be fine, together with giving the file a name somewhat referring to DT.
> 
> That would bring it out of sync with x86's asm/device.h. Both of them define
> device_t and doing so in differently named headers would just be confusing for
> everyone.

Okay, then the name can't change. In which case the #ifdef-ary needs to remain,
imo, to keep the header being "generic".

Jan