Re: [PATCH v7 12/16] arm: mm: define clear_user_highpages()

[Ankur Arora <ankur.a.arora@oracle.com>]

David Hildenbrand <david@redhat.com> writes:

>>>> assumes one of the following:
>>>>     1. clear_user_highpages is defined by the architecture or,
>>>>     2. HIGHMEM => arch defines clear_user_highpage or clear_user_page
>>>>     3. !HIGHMEM => arch defines clear_user_pages or clear_user_page
>>>> Case 2 is fine, since ARM has clear_user_highpage().
>>>> Case 3 runs into a problem since ARM doesn't have clear_user_pages()
>>>> or clear_user_page() (it does have the second, but only with !CONFIG_MMU).
>>>
>>> I think we should look into having a generic fallback version in common code
>>> instead for that case, and not require the arch to implement such a loop around
>>> clear_user_highpage().
>> So, as you suggested, I moved clear_user_pages() to mm/utils.c and
>> conditioned it on clear_user_page() also existing.
>>    #if defined(clear_user_page) && !defined(clear_user_pages)
>>    void clear_user_pages(void *addr, unsigned long vaddr, struct page *page,
>>                          unsigned int npages) {
>>                        ...
>>    }
>>    #endif
>> That fixed this issue as well since there's no more bogus reference to
>> clear_user_page().
>
> I'll have to see the resulting code to comment on details, but if we can handle it in
> common code, all good.
>
>> Are there cases in which (TRANSPARENT_HUGEPAGE || HUGETLB) might be enabled
>> on ARM?
>
> Arm has
>
> arch/arm/Kconfig:       select HAVE_ARCH_TRANSPARENT_HUGEPAGE if ARM_LPAE
>
> and supports hugetlb. So yes on both.

I tried to figure out a way forward for arm with THP/HUGETLB and AFAICS
the cleanest approach would be to have some version of this patch.

Just to reiterate the problem with arch/arm: it defines
clear_user_highpage(), but does not define clear_user_page().

This means that common code cannot usefully define clear_user_pages().

And the common definition of clear_user_highpages() either needs to use:

  - for HIGHMEM, clear_user_highpage()
  - for !HIGHMEM, clear_user_pages()

The first works but I don't see how to make the second work without adding
some such special handling:

     static inline void clear_user_highpages(struct page *page, unsigned long vaddr,
     					unsigned int npages)
vaddr, page, npages);
     		return;
     	}
     #endif

     	do {
     		clear_user_highpage(page, vaddr);
     		vaddr += PAGE_SIZE;
     		page++;
     	} while (--npages);
     }

(Even this is a bit contorted, as common code shouldn't really need to
have an associated #define for clear_user_pages().)

Thanks

--
ankur