Re: [PATCH] mm: simplify size2index conversion of __kmalloc_index

[Vlastimil Babka <vbabka@suse.cz>]

On 8/30/22 07:51, Christophe Leroy wrote:
> 
> 
> Le 29/08/2022 à 16:21, Vlastimil Babka a écrit :
>> On 8/29/22 05:36, Hyeonggon Yoo wrote:
>>> On Mon, Aug 29, 2022 at 04:11:04AM +0100, Matthew Wilcox wrote:
>>>> On Sun, Aug 28, 2022 at 11:14:48PM +0800, Dawei Li wrote:
>>>>> Current size2index is implemented by one to one hardcode mapping,
>>>>> which can be improved by order_base_2().
>>>>> Must be careful to not violate compile-time optimization rule.
>>>>
>>>> This patch has been NACKed before (when submitted by other people).
>>>
>>>
>>> Hmm right.
>>> https://lkml.iu.edu/hypermail/linux/kernel/1606.2/05402.html
>>>
>>> Christoph Lameter wrote:
>>>> On Wed, 22 Jun 2016, Yury Norov wrote:
>>>>> There will be no fls() for constant at runtime because ilog2() calculates
>>>>> constant values at compile-time as well. From this point of view,
>>>>> this patch removes code duplication, as we already have compile-time
>>>>> log() calculation in kernel, and should re-use it whenever possible.\
>>>
>>>> The reason not to use ilog there was that the constant folding did not
>>>> work correctly with one or the other architectures/compilers. If you want
>>>> to do this then please verify that all arches reliably do produce a
>>>> constant there.
>>>
>>> Can we re-evaluate this?
>>
>> Is there a way to turn inability of compile-time calculation to a
>> compile-time error? (when size_is_constant=true etc). Then we could try and
>> see if anything breaks in -next.
>>
>>
> 
> The following will generate a build error if the function
> constant_check() is not called with a buildtime constant argument.
> 
> static void __always_inline constant_check(unsigned long val)
> {
> 	BUILD_BUG_ON(!__builtin_constant_p(val));
> }
> 
> Is that what you are looking for ?
Maybe, if we can rely on these two being equivalent:
- __kmalloc_index(x) is evaluated compile-time
- __builtin_constant_p(__kmalloc_index(x)) is true

Logically such equivalency should be expected, and a quick attempt 
locally with recent gcc seems to work fine, but I guess we'll have to 
try in -next for a bit and see if anything comes out.