On 05/03/2026 16:52, Andrew Davis wrote:
>>>> The code is not correct logically, either, because functions like
>>>> ti_sci_get_handle() and ti_sci_put_handle() are meant to modify the
>>>> handle reference counting, thus they must modify the handle.
>>>
>>> The reference counting is handled outside of the ti_sci_handle struct,
>>> the contents of the handle are never modified after it is created.
>>>
>>> The const is only added by functions return a handle to consumers.
>>> We cannot return non-const to consumer drivers or then they would
>>> be able to modify the content without a compiler warning, which would
>>> be a real problem.
>>
>> This is the same argument as making pointer to const the pointer freed
>> via kfree() (or free() in userspace). kfree() does not modify the
>> contents of the pointer, right? The same as getting putting handle does
>> not modify the handle...
>>
>
> In that argument, if we wanted the consumer of the pointer to not free()
> it we would return a const pointer, free()'ing that would result in the
> warning we want (discards const qualifier).
>
> If you could somehow malloc() from a const area in memory then free()
> doesn't modify the pointed to values, only the non-const record keeping
> which would be stored outside of the const memory. So even in this analogy
> there isn't a problem.
I am not saying about malloc. I am saying about free() which does not
modify the freed memory.
>
>> The point is that storing the reference counter outside of handle does
>> not make the argument correct. Logically when you get a reference, you
>> increase the counter, so it is not a pointer to const. And the code
>> agrees, because you must drop the const.
>>
>
> The record keeping memory is not const and can be modified.
>
> And where do we drop the const? The outer "struct ti_sci_info" was never
> const to begin with, so no dropped const.
We discuss about different points. I did not say the outer memory is
const. I said that you drop the const - EXPLICITLY - from the pointer to
handle.
And that API which gets a handle (increases reference count) via pointer
to const is completely illogical, because increasing refcnt is already
modifying it. Just because you store the refcnt outside, does not change
the fact that API is simply confusing.
>
> If the issue is that the handle is not const inside that outer struct
> we could fix that,
>
> struct ti_sci_info {
> ...
> - struct ti_sci_handle handle;
> + const struct ti_sci_handle handle;
> ...
> };
>
> And with that change even your original commit message example issue
> goes away,
>
> struct ti_sci_info *info = handle_to_ti_sci_info(handle);
> info->handle.version.abi_major = 0;
>
> would now fail to work to compile.
But you cannot do that for other reasons in your code because you DO
modify the handle in all the APIs which you call "pointer to const".
Best regards,
Krzysztof
