> On May 30, 2019, at 4:14 AM, Kirill A. Shutemov <kirill@xxxxxxxxxxxxx> wrote:
>
> On Thu, May 30, 2019 at 02:10:15PM +0300, Kirill A. Shutemov wrote:
>> On Wed, May 29, 2019 at 02:20:46PM -0700, Song Liu wrote:
>>> @@ -2133,10 +2133,15 @@ static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
>>> VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
>>> VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
>>> && !pmd_devmap(*pmd));
>>> + /* only file backed vma need preallocate pgtable*/
>>> + VM_BUG_ON(vma_is_anonymous(vma) && prealloc_pgtable);
>>>
>>> count_vm_event(THP_SPLIT_PMD);
>>>
>>> - if (!vma_is_anonymous(vma)) {
>>> + if (prealloc_pgtable) {
>>> + pgtable_trans_huge_deposit(mm, pmd, prealloc_pgtable);
>>> + mm_inc_nr_pmds(mm);
>>> + } else if (!vma_is_anonymous(vma)) {
>>> _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
>>> /*
>>> * We are going to unmap this huge page. So
>>
>> Nope. This going to leak a page table for architectures where
>> arch_needs_pgtable_deposit() is true.
>
> And I don't there's correct handling of dirty bit.
>
> And what about DAX? Will it blow up? I think so.
>
Let me look into these cases. Thanks for the feedback!
Song
> --
> Kirill A. Shutemov
