[PATCH] arm64: calculate the various pages number to show

[zhongjiang@huawei.com (zhong jiang)]

On 2015/11/25 23:04, Mark Rutland wrote:
> On Wed, Nov 25, 2015 at 09:41:12PM +0800, zhongjiang wrote:
>> This patch add the interface to show the number of 4KB or 64KB page,
>> aims to statistics the number of different types of pages.
> 
> What is this useful for? Why do we want it?
> 
> What does it account for, just the swapper?
> 

The patch is wirtten when I was in backport set_memory_ro. It can be used to
detect whether there is a large page spliting and merging. large page will
significantly reduce the TLB miss, and improve the system performance.

>> Signed-off-by: zhongjiang <zhongjiang@huawei.com>
>> ---
>>  arch/arm64/include/asm/pgtable-types.h |   24 ++++++++++++++++++++++++
>>  arch/arm64/mm/mmu.c                    |   28 ++++++++++++++++++++++++++++
>>  arch/arm64/mm/pageattr.c               |   31 +++++++++++++++++++++++++++++++
>>  3 files changed, 83 insertions(+), 0 deletions(-)
>>
>> diff --git a/arch/arm64/include/asm/pgtable-types.h b/arch/arm64/include/asm/pgtable-types.h
>> index 2b1bd7e..aa52546 100644
>> --- a/arch/arm64/include/asm/pgtable-types.h
>> +++ b/arch/arm64/include/asm/pgtable-types.h
>> @@ -86,6 +86,30 @@ typedef pteval_t pgprot_t;
>>  
>>  #endif /* STRICT_MM_TYPECHECKS */
>>  
>> +struct seq_file;
>> +extern void arch_report_meminfo(struct seq_file *m);
>> +
>> +enum pg_level {
>> +	PG_LEVEL_NONE,
>> +#ifdef CONFIG_ARM64_4K_PAGES
>> +	PG_LEVEL_4K,
>> +	PG_LEVEL_2M,
>> +	PG_LEVEL_1G,
>> +#else
>> +	PG_LEVEL_64K,
>> +	PG_LEVEL_512M,
>> +#endif
>> +	PG_LEVEL_NUM
>> +};
> 
> This doesn't account for 16K pages, and it means each call site has to
> handle the various page sizes directly.
> 
> It would be better to simply count PTE/PMD/PUD/PGD, then handle the size
> conversion at the end when logging.
> 

yes, now I only consider the 4kb and 64kb. if the patch is approved ,I will
improve it.
each call site need two different varialbes to statistics, aiming to distinguish
diffent pages. I think it will no more simple.


>> @@ -85,6 +86,11 @@ void split_pmd(pmd_t *pmd, pte_t *pte)
>>  		set_pte(pte, pfn_pte(pfn, prot));
>>  		pfn++;
>>  	} while (pte++, i++, i < PTRS_PER_PTE);
>> +#ifdef CONFIG_ARM64_4K_PAGES
>> +	split_page_count(PG_LEVEL_2M);
>> +#else
>> +	split_page_count(PG_LEVEL_512M);
>> +#endif
>>  }
> 
> e.g. here you'd just count PG_LEVEL_PMD, which would work regardless of
> page size.
> 
>> diff --git a/arch/arm64/mm/pageattr.c b/arch/arm64/mm/pageattr.c
>> index 7a5ff11..c1888b9 100644
>> --- a/arch/arm64/mm/pageattr.c
>> +++ b/arch/arm64/mm/pageattr.c
>> @@ -15,12 +15,43 @@
>>  #include <linux/module.h>
>>  #include <linux/sched.h>
>>  
>> +#include <linux/seq_file.h>
>>  #include <asm/pgalloc.h>
>>  #include <asm/pgtable.h>
>>  #include <asm/tlbflush.h>
>>  
>>  #include "mm.h"
>>  
>> +static unsigned long direct_pages_count[PG_LEVEL_NUM];
> 
> This doesn't match reality by the time we start executing the kernel,
> given we created page tables in head.S.
> 
>> +
>> +void update_page_count(int level, unsigned long pages)
>> +{
>> +	direct_pages_count[level] += pages;
>> +}
>> +
>> +void split_page_count(int level)
>> +{
>> +	direct_pages_count[level]--;
>> +	direct_pages_count[level-1] += PTRS_PER_PTE;
>> +}
>> +
>> +void arch_report_meminfo(struct seq_file *m)
>> +{
>> +#ifdef CONFIG_ARM64_4K_PAGES
>> +	seq_printf(m, "DirectMap4k:     %8lu kB\n",
>> +			direct_pages_count[PG_LEVEL_4K] << 2);
>> +	seq_printf(m, "DirectMap2M:     %8lu kB\n",
>> +			direct_pages_count[PG_LEVEL_2M] << 11);
>> +	seq_printf(m, "DirectMap1G:     %8lu kB\n",
>> +			direct_pages_count[PG_LEVEL_1G] << 20);
>> +#else
>> +	seq_printf(m, "DirectMap64k:     %8lu kB\n",
>> +			direct_pages_count[PG_LEVEL_64K] << 6);
>> +	seq_printf(m, "DirectMap512M:     %8lu kB\n",
>> +			direct_pages_count[PG_LEVEL_512M] << 19);
>> +#endif
>> +}
> 
> You could dynamuically determine the sizes here for each field, and not
> have to have #ifdefs.> 

I don't understand what you mean. I think it can be more readable and operability.

Thanks
zhongjiang

> That all said, I don't see what this is useful for, and it looks very
> fragile.
> 
> Thanks,
> Mark.
> 
> .
>