Re: [PATCH v10 3/4] mm, thp: introduce FOLL_SPLIT_PMD

[Song Liu <songliubraving@fb.com>]

> On Jul 30, 2019, at 9:11 AM, Oleg Nesterov <oleg@redhat.com> wrote:
> 
> I don't understand this code, so I can't review, but.
> 
> On 07/29, Song Liu wrote:
>> 
>> This patches introduces a new foll_flag: FOLL_SPLIT_PMD. As the name says
>> FOLL_SPLIT_PMD splits huge pmd for given mm_struct, the underlining huge
>> page stays as-is.
>> 
>> FOLL_SPLIT_PMD is useful for cases where we need to use regular pages,
>> but would switch back to huge page and huge pmd on. One of such example
>> is uprobe. The following patches use FOLL_SPLIT_PMD in uprobe.
> 
> So after the next patch we have a single user of FOLL_SPLIT_PMD (uprobes)
> and a single user of FOLL_SPLIT: arch/s390/mm/gmap.c:thp_split_mm().
> 
> Hmm.

I think this is what we want. :) 

FOLL_SPLIT is the fallback solution for users who cannot handle THP. With
more THP aware code, there will be fewer users of FOLL_SPLIT. 

> 
>> @@ -399,7 +399,7 @@ static struct page *follow_pmd_mask(struct vm_area_struct *vma,
>> 		spin_unlock(ptl);
>> 		return follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
>> 	}
>> -	if (flags & FOLL_SPLIT) {
>> +	if (flags & (FOLL_SPLIT | FOLL_SPLIT_PMD)) {
>> 		int ret;
>> 		page = pmd_page(*pmd);
>> 		if (is_huge_zero_page(page)) {
>> @@ -408,7 +408,7 @@ static struct page *follow_pmd_mask(struct vm_area_struct *vma,
>> 			split_huge_pmd(vma, pmd, address);
>> 			if (pmd_trans_unstable(pmd))
>> 				ret = -EBUSY;
>> -		} else {
>> +		} else if (flags & FOLL_SPLIT) {
>> 			if (unlikely(!try_get_page(page))) {
>> 				spin_unlock(ptl);
>> 				return ERR_PTR(-ENOMEM);
>> @@ -420,6 +420,10 @@ static struct page *follow_pmd_mask(struct vm_area_struct *vma,
>> 			put_page(page);
>> 			if (pmd_none(*pmd))
>> 				return no_page_table(vma, flags);
>> +		} else {  /* flags & FOLL_SPLIT_PMD */
>> +			spin_unlock(ptl);
>> +			split_huge_pmd(vma, pmd, address);
>> +			ret = pte_alloc(mm, pmd);
> 
> I fail to understand why this differs from the is_huge_zero_page() case above.

split_huge_pmd() handles is_huge_zero_page() differently. In this case, we 
cannot use the pmd_trans_unstable() check. 

> 
> Anyway, ret = pte_alloc(mm, pmd) can't be correct. If __pte_alloc() fails pte_alloc()
> will return 1. This will fool the IS_ERR(page) check in __get_user_pages().

Great catch! Let me fix it.

Thanks,
Song