Re: [RFC PATCH 2/2] kernel/sched.c: VLA in middle of struct

[Ingo Molnar <mingo@elte.hu>]

* Jeff Garzik <jeff@garzik.org> wrote:

> Ingo Molnar wrote:
>>>>>> That cpumask[] should probably be cpumask[0], to document the
>>>>>> aliasing to ->span and ->cpus properly.
>>>>> If the comment wasn't sufficient documentation, I don't think 
>>>>> that  would help :(
>>>> It's a visual helper: it matches up with how we do these 'zero size 
>>>>  array means dynamic structure continuation' tricks generally.
>>>>
>>>> I first mis-parsed the code for a second when seeing cpumask[].   
>>>> cpumask[0] stands out like a sore thumb. And we dont read comments  
>>>> anyway ;-)
>>>>
>>>> Jeff, i suspect you found this because you are working on something 
>>>>  rather interesting? :) If yes, would it help your project if we 
>>>> did  the cpumask[0] cleanup and pushed it upstream immediately?
>>> I think cpumask[0] would be more clear and consistent with the rest 
>>> of the kernel.
>>>
>>> But unfortunately for the twin projects of (a) static analysis and  
>>> checking with 'sparse', and (b) compiling under another compiler,  
>>> VLA-in-middle-of-struct is a killer in either case.
>>
>> even if at the end of the struct?
>
> Putting the VLA at the end of the struct would be a huge help, 
> yes.
>
> For example, struct sched_group and struct sched_domain are OK 
> as-is (though "[0]" would be preferred).
>
> It is the definition of struct static_sched_group and struct 
> static_sched_domain that creates the problem, because with the 
> bitmap following cpumask[] and span[], the VLA is no longer at the 
> end of the struct.
>
> VLA-in-the-middle raises the complexity required of the compiler 
> quite a bit.  As a result, VLA-in-middle is not implemented in 
> sparse or clang (LLVM's C front-end and static analyzer).

feel free to send patches for this - i dont have those build modes 
to test that it's sufficient. I'd suggest to go the simplest path: 
remove all the vla aliasing tricks: just make struct sched_domain 
use a plain struct cpumask and eliminate static_sched_domain 
altogether.

The memory overhead is marginal as most of our sched domains are 
static allocated (and full size) anyway.

	Ingo