by Christophe Leroy

[permalink] [raw]

Subject: Re: [PATCH v4 2/5] powerpc/mm/slice: Enhance for supporting PPC32

Le 12/02/2018 à 00:34, Nicholas Piggin a écrit :
> On Sun, 11 Feb 2018 21:04:42 +0530
> "Aneesh Kumar K.V" <[email protected]> wrote:
>
>> On 02/11/2018 07:29 PM, Nicholas Piggin wrote:
>>> On Sat, 10 Feb 2018 13:54:27 +0100 (CET)
>>> Christophe Leroy <[email protected]> wrote:
>>>
>>>> In preparation for the following patch which will fix an issue on
>>>> the 8xx by re-using the 'slices', this patch enhances the
>>>> 'slices' implementation to support 32 bits CPUs.
>>>>
>>>> On PPC32, the address space is limited to 4Gbytes, hence only the low
>>>> slices will be used.
>>>>
>>>> This patch moves "slices" functions prototypes from page64.h to slice.h
>>>>
>>>> The high slices use bitmaps. As bitmap functions are not prepared to
>>>> handling bitmaps of size 0, the bitmap_xxx() calls are wrapped into
>>>> slice_bitmap_xxx() functions which will void on PPC32
>>>
>>> On this last point, I think it would be better to put these with the
>>> existing slice bitmap functions in slice.c and just have a few #ifdefs
>>> for SLICE_NUM_HIGH == 0.
>>>
>>
>> We went back and forth with that. IMHO, we should avoid as much #ifdef
>> as possible across platforms. It helps to understand the platform
>> restrictions better as we have less and less access to these platforms.
>> The above change indicates that nohash 32 wants to use the slice code
>> and they have different restrictions. With that we now know that
>> book3s64 and nohash 32 are the two different configs using slice code.
>
> I don't think it's the right place to put it. It's not platform dependent
> so much as it just depends on whether or not you have 0 high slices as
> a workaround for bitmap API not accepting 0 length.
>
> Another platform that uses the slice code would just have to copy and
> paste either the nop or the bitmap implementation depending if it has
> high slices. So I don't think it's the right abstraction. And it
> implies a bitmap operation but it very specifically only works for
> struct slice_mask.high_slices bitmap, which is not clear. Better to
> just work with struct slice_mask.
>
> Some ifdefs inside .c code for small helper functions like this IMO isn't
> really a big deal -- it's not worse than having it in headers. You just
> want to avoid ifdef mess when looking at non-trivial logic.
>
> static inline void slice_or_mask(struct slice_mask *dst, struct slice_mask *src)
> {
> dst->low_slices |= src->low_slices;
> #if SLICE_NUM_HIGH > 0
> bitmap_or(result, dst->high_slices, src->high_slices, SLICE_NUM_HIGH);
> #endif
> }
>
> I think that's pretty fine. If you have a singular hatred for ifdef in .c,
> then if() works just as well.
>

To be honest, I tend to agree with you. Therefore, after a few words
with Michael, v5 gets rid of those obscure wrappers to come back to the
initial (v1) approach which was to use 'if (SLICE_NUM_HIGH)'.
Behind the fact that it in my mind looks cleared in the code than using
slice_bitmap_xxx() wrappers, it also has the advantage of significantly
reducing the size of the patch, which is a must to be able to get it
applied on stable.

Christophe