On Sun, Nov 27, 2011 at 1:31 AM, Andrea Arcangeli <aarcange@xxxxxxxxxx> wrote:
> On Sat, Nov 26, 2011 at 10:43:15PM +0800, Hillf Danton wrote:
>> In VM core, window is opened for MIPS to use THP.
>>
>> And two simple helper functions are added to easy MIPS a bit.
>>
>> Signed-off-by: Hillf Danton <dhillf@xxxxxxxxx>
>> ---
>>
>> --- a/mm/Kconfig Thu Nov 24 21:12:00 2011
>> +++ b/mm/Kconfig Sat Nov 26 22:12:56 2011
>> @@ -307,7 +307,7 @@ config NOMMU_INITIAL_TRIM_EXCESS
>>
>> config TRANSPARENT_HUGEPAGE
>> bool "Transparent Hugepage Support"
>> - depends on X86 && MMU
>> + depends on MMU
>> select COMPACTION
>> help
>> Transparent Hugepages allows the kernel to use huge pages and
>
> Then the build will break for all archs if they enable it, better to
> limit the option to those archs that supports it.
>
>> --- a/mm/huge_memory.c Thu Nov 24 21:12:48 2011
>> +++ b/mm/huge_memory.c Sat Nov 26 22:30:24 2011
>> @@ -17,6 +17,7 @@
>> #include <linux/khugepaged.h>
>> #include <linux/freezer.h>
>> #include <linux/mman.h>
>> +#include <linux/pagemap.h>
>> #include <asm/tlb.h>
>> #include <asm/pgalloc.h>
>> #include "internal.h"
>> @@ -135,6 +136,30 @@ static int set_recommended_min_free_kbyt
>> }
>> late_initcall(set_recommended_min_free_kbytes);
>>
>> +/* helper function for MIPS to call pmd_page() indirectly */
>> +static inline struct page *__pmd_page(pmd_t pmd)
>> +{
>> + struct page *page;
>> +
>> +#ifdef __HAVE_ARCH_THP_PMD_PAGE
>> + page = thp_pmd_page(pmd);
>> +#else
>> + page = pmd_page(pmd);
>> +#endif
>> + return page;
>> +}
>
> Why do you need this and also a branch in thp_pmd_page checking for
> pmd_trans_huge? If you fallback in pmd_page that would mean you're
> called by hugetlbfs. Doesn't make much sense to fallback in pmd_page
> if the hugepmd format for thp and hugetlbfs is different.
>
> Couldn't you set a different _PAGE_HUGE flag in the pmd in the thp
> case to avoid the above? Then you could have a pmd_page that works on
> both. Ok it'll be slower and require 1 more branch (but you already
> have a branch for something that doesn't seem needed).
>
> pmd_page is only called by hugetlbfs/thp, rest uses pte_offset* so I
> don't think a branch would be a big deal and you could hide the fact
> he format of the pmd between hugetlbfs and thp is different with a
> bitflag on the pmd (if any reserved is available to use to software).
>
>> +
>> +/* helper function for MIPS to call update_mmu_cache() indirectly */
>> +static inline void __update_mmu_cache(struct vm_area_struct *vma,
>> + unsigned long addr, pmd_t *pmdp)
>> +{
>> +#ifdef __HAVE_ARCH_UPDATE_MMU_THP
>> + update_mmu_thp(vma, addr, pmdp);
>> +#else
>> + update_mmu_cache(vma, addr, pmdp);
>> +#endif
>> +}
>
> Maybe here same, check pmd_trans_huge (and make it succeed only in the
> thp case and not the hugetlbfs case) and avoid it the __ and the ifdefs.
>
Got and thanks.
Hillf
