Re: [PATCH v3] mm: mincore: use pte_batch_bint() to batch process large folios

[Baolin Wang <baolin.wang@linux.alibaba.com>]

On 2025/5/9 15:51, David Hildenbrand wrote:
> On 09.05.25 02:45, Baolin Wang wrote:
>> When I tested the mincore() syscall, I observed that it takes longer with
>> 64K mTHP enabled on my Arm64 server. The reason is the 
>> mincore_pte_range()
>> still checks each PTE individually, even when the PTEs are contiguous,
>> which is not efficient.
>>
>> Thus we can use pte_batch_hint() to get the batch number of the present
>> contiguous PTEs, which can improve the performance. I tested the 
>> mincore()
>> syscall with 1G anonymous memory populated with 64K mTHP, and observed an
>> obvious performance improvement:
>>
>> w/o patch        w/ patch        changes
>> 6022us            549us            +91%
>>
>> Moreover, I also tested mincore() with disabling mTHP/THP, and did not
>> see any obvious regression for base pages.
>>
>> Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
>> ---
>> Changes from v2:
>> - Re-calculate the max_nr, per Barry.
>> Changes from v1:
>> - Change to use pte_batch_hint() to get the batch number, per Ryan.
>>
>> Note: I observed the min_t() can introduce a slight performance 
>> regression
>> for base pages, so I change to add a batch size check for base pages,
>> which can resolve the performance regression issue.
>> ---
>>   mm/mincore.c | 22 +++++++++++++++++-----
>>   1 file changed, 17 insertions(+), 5 deletions(-)
>>
>> diff --git a/mm/mincore.c b/mm/mincore.c
>> index 832f29f46767..42d6c9c8da86 100644
>> --- a/mm/mincore.c
>> +++ b/mm/mincore.c
>> @@ -21,6 +21,7 @@
>>   #include <linux/uaccess.h>
>>   #include "swap.h"
>> +#include "internal.h"
>>   static int mincore_hugetlb(pte_t *pte, unsigned long hmask, unsigned 
>> long addr,
>>               unsigned long end, struct mm_walk *walk)
>> @@ -105,6 +106,7 @@ static int mincore_pte_range(pmd_t *pmd, unsigned 
>> long addr, unsigned long end,
>>       pte_t *ptep;
>>       unsigned char *vec = walk->private;
>>       int nr = (end - addr) >> PAGE_SHIFT;
>> +    int step, i;
>>       ptl = pmd_trans_huge_lock(pmd, vma);
>>       if (ptl) {
>> @@ -118,16 +120,26 @@ static int mincore_pte_range(pmd_t *pmd, 
>> unsigned long addr, unsigned long end,
>>           walk->action = ACTION_AGAIN;
>>           return 0;
>>       }
>> -    for (; addr != end; ptep++, addr += PAGE_SIZE) {
>> +    for (; addr != end; ptep += step, addr += step * PAGE_SIZE) {
>>           pte_t pte = ptep_get(ptep);
>> +        step = 1;
>>           /* We need to do cache lookup too for pte markers */
>>           if (pte_none_mostly(pte))
>>               __mincore_unmapped_range(addr, addr + PAGE_SIZE,
>>                            vma, vec);
>> -        else if (pte_present(pte))
>> -            *vec = 1;
>> -        else { /* pte is a swap entry */
>> +        else if (pte_present(pte)) {
>> +            unsigned int batch = pte_batch_hint(ptep, pte);
>> +
>> +            if (batch > 1) {
>> +                unsigned int max_nr = (end - addr) >> PAGE_SHIFT;
> 
> Nit: probably would have called this max_step to match step.

OK. If need respin the patch, I'll rename it.

>> +
>> +                step = min_t(unsigned int, batch, max_nr);
>> +            }
>> +
>> +            for (i = 0; i < step; i++)
>> +                vec[i] = 1;
> 
> I'm surprised this micro-optimization matters that much. Probably the 

Me too.

> compiler
> defers the calculation of max_nr. I am not convinced we need that level of
> micro-optimization in this code ...
> 
> 
> But if we're already micro-optimizing, you could have optimized out the 
> loop as
> well for order-0:
> 
>      unsigned int batch = pte_batch_hint(ptep, pte);
> 
>      if (batch > 1) {
>          unsigned int max_nr = (end - addr) >> PAGE_SHIFT;
> 
>          step = min_t(unsigned int, batch, max_nr);
>          for (i = 0; i < step; i++)
>              vec[i] = 1;
>      } else {
>          *vec = 1;
>      }

I tried this method, and it had no impact on performance.

> In any case
> 
> Acked-by: David Hildenbrand <david@redhat.com>

Thanks.