Re: [RESEND][PATCH v3 06/17] static_call: Add basic static call infrastructure

From: Nadav Amit
Date: Thu Mar 26 2020 - 15:02:39 EST


> On Mar 26, 2020, at 11:28 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> On Thu, Mar 26, 2020 at 06:09:07PM +0000, Nadav Amit wrote:
>
>> I think that the kernel underutilizes the pure attribute in general.
>> Building it with "-Wsuggest-attribute=pureâ results in many warnings.
>> Function pointers such kvm_x86_ops.get_XXX() could have been candidates to
>> use the âpureâ attribute.
>>
>> The syntax is what you would expect:
>>
>> static void __attribute__((pure))(*ptr)(void);
>
> Well, I didn't in fact expect that, because an attribute is not a
> type qualifier.

Just a small correction for my stupid example - pure function should always
return a value.

>> However, you have a point, gcc does not appear to respect âpureâ for
>> function pointers and emits a warning it is ignored. GCC apparently only
>> respects âconstâ. In contrast clang appears to respect the pure attribute
>> for function pointers.
>
> Still, we can probably make it happen for static_call(), since it is a
> direct call to the trampoline, all we need to do is make sure the
> trampoline is declared pure.
>
> It does however mean that static_call() inherits all the dangers and
> pit-falls of function pointers with some extra on top. It will be
> impossible to validate this stuff.
>
> That is, you can static_call_update() with a pointer to a !pure function
> and you get to keep the pieces.

I understand. Well, perhaps it can be added later, as anyhow GCC does not
support it.

On another note - it may be beneficial to see if the infrastructure that you
built can accommodate notifier-chains. It is not the most painful point, but
it would be nice to deal with those as well. Since many of those are changed
asynchronously, I am not sure it is the easiest thing to do.