Re: [PATCH v2 01/15] xen: Clean up asm-generic/device.h

[Jan Beulich <jbeulich@xxxxxxxx>]

On 06.06.2025 11:55, Alejandro Vallejo wrote:
> On Fri Jun 6, 2025 at 8:51 AM CEST, Jan Beulich wrote:
>> On 05.06.2025 21:47, Alejandro Vallejo wrote:
>>> --- a/xen/include/asm-generic/device.h
>>> +++ b/xen/include/asm-generic/device.h
>>> @@ -1,14 +1,20 @@
>>>  /* SPDX-License-Identifier: GPL-2.0-only */
>>> +/*
>>> + * This header helps DTB-based architectures abstract away where a 
>>> particular
>>> + * device came from, be it the DTB itself or enumerated on a PCI bus.
>>> + */
>>>  #ifndef __ASM_GENERIC_DEVICE_H__
>>>  #define __ASM_GENERIC_DEVICE_H__
>>>  
>>> +#ifndef CONFIG_HAS_DEVICE_TREE
>>> +#error "Header for exclusive use of DTB-based architectures"
>>> +#endif
>>> +
>>>  #include <xen/stdbool.h>
>>>  
>>>  enum device_type
>>>  {
>>> -#ifdef CONFIG_HAS_DEVICE_TREE
>>>      DEV_DT,
>>> -#endif
>>>      DEV_PCI
>>>  };
>>
>> My objection to these changes remains; as a generic header it ought to be 
>> what
>> that attribute says - generic.
> 
> It is generic for any architecture where platform DTs exist (that is, anything
> but x86).

Here you're limiting things to what Xen presently "knows". I'm sure there are
other architectures where DT is entirely unknown.

Furthermore isn't the work here part of the hyperlaunch effort, where DT will
be introduced to x86? Hence "anything but" isn't quite right either then.

Jan

> As the commit message states, these guards are useless, provide no 
> functionality
> and create the fiction that somehow this header is still relevant on an
> architecture where only PCI is available. And that's just not true. x86 being
> the sole architecture without DTs actively overrides it, and relies on 
> device_t
> (defined as struct device here) to be a "struct pci_dev" instead in
> x86/include/asm/device.h, with dev_to_pci() and pci_to_dev() being irrelevant
> because device_t* and struct pci_dev* are identical types in x86. Removing 
> that
> override header is not just a matter of performance. All the IOMMU ops are
> referencing device_t, while the drivers are assuming pci_dev, so all IOMMU
> code breaks immediately when x86 tries to use this.
> 
> To be perfectly clear, this patch isn't strictly required to do DT 
> unflattening
> on x86. But it's a piece of arm tech debt that Xen is better off without.
> 
> Cheers,
> Alejandro