Re: [PATCH v2 16/19] gendwarfksyms: Add support for reserved structure fields

[Benno Lossin]

On 12.09.24 18:06, Sami Tolvanen wrote:
> On Wed, Sep 11, 2024 at 4:43 AM Petr Pavlu <petr.pavlu@xxxxxxxx> wrote:
>>
>> On 8/31/24 02:05, Sami Tolvanen wrote:
>>> On Fri, Aug 30, 2024 at 9:34 AM Miroslav Benes <mbenes@xxxxxxx> wrote:
>>>>
>>>> yes, this is one of the approaches we use in SLES. We add kabi paddings
>>>> to some structures in advance (see [1] as a random example) and then use
>>>> it later if needed.
>>>>
>>>> It is not the only approach. Much more often we do not have a padding and
>>>> use alignment holes ([5]), addition of a new member to the end of a
>>>> structure ([2] or [3]) and such "tricks" ([4] for a newly fully defined
>>>> structure).
>>>
>>> Thanks for bringing this up! Sounds like we're also going to need a
>>> way to completely exclude specific fields from the output then. I
>>> think we can use a similar union approach, but instead of instructing
>>> the tool to use another type, we can just indicate that the field
>>> should be skipped. I'll come up with a solution for v3.
>>
>> It might have been mentioned previously, not sure, but one more case to
>> consider is handling of enum declarations. New enumerators can be
>> typically added without breaking ABI, e.g. 'enum E { OLD1, OLD2, NEW }'.
>> It would be then great to have some ability to hide them from
>> gendwarfksyms.
>>
>> I think neither of the __kabi_reserved or __gendwarfksyms_declonly
>> mechanism can currently help with that.
> 
> I thought about this a bit and I wonder if we need a separate
> mechanism for that, or is it sufficient to just #define any additional
> hidden values you want to add instead of including them in the enum?
> 
>   enum e {
>       A,
>       B,
>   #define C (B + 1)
>   #define D (C + 1)
>   };
>
> 
> Do you see any issues with this approach? I think Clang would complain
> about this with -Wassign-enum, but I'm not sure if we even enable that
> in the kernel, and as long as you don't overflow the underlying type,
> which is a requirement for not breaking the ABI anyway, it should be
> fine.

Rust has problems with `#define`-style enums, because bindgen (the tool
that generates definitions for Rust to be able to call C code) isn't
able to convert them to Rust enums.

So if you can come up with an approach that allows you to continue to
use C enums instead of `#define`, we would appreciate that, since it
would make our lives a lot easier.

---
Cheers,
Benno