On 2022/4/11 19:41, David Hildenbrand wrote:
> On 09.04.22 09:38, Miaohe Lin wrote:
>> __migration_entry_wait and migration_entry_wait_on_locked assume pte is
>> always mapped from caller. But this is not the case when it's called from
>> migration_entry_wait_huge and follow_huge_pmd. And a parameter unmap to
>> indicate whether pte needs to be unmapped to fix this issue.
>
> Hm.
>
>
> migration_entry_wait_on_locked documents
>
> "@ptep: mapped pte pointer. Will return with the ptep unmapped. Only
> required for pte entries, pass NULL for pmd entries."
>
> Setting ptep implies that we have a *mapped pte* pointer that requires unmap.
> If some code sets that although that's not guaranteed, that calling code
> is wrong and needs to be fixed to not pass a ptep.
>
>
> hugetlbfs never requires a map/unmap. I really don't see we there is need to
> adjust migration_entry_wait_on_locked(): just don't pass a ptep as documented.
>
> What's really nasty here is that hugetlbfs actually mostly works on PMD/PUD,
> but we call it PTEs. One corner case might be CONT PTEs, but they are also
> accessed without a map+unmap.
>
> Regarding __migration_entry_wait(), I think we should just stop using it for
> hugetlbfs and have a proper hugetlbfs variant that calls
> hugetlb_migration_entry_wait(ptep == NULL), and knows that although we're
> handling ptes, we're usually not actually holding ptes in our hands
> that need a map+unmap.
>
>
> Something like (including some cleanups mm parameter):
This really helps! Many thanks! Will try to do this in next version. :)
>
>
> diff --git a/include/linux/swapops.h b/include/linux/swapops.h
> index 32d517a28969..898c407ad8f7 100644
> --- a/include/linux/swapops.h
> +++ b/include/linux/swapops.h
> @@ -234,8 +234,8 @@ extern void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
> spinlock_t *ptl);
> extern void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
> unsigned long address);
> -extern void migration_entry_wait_huge(struct vm_area_struct *vma,
> - struct mm_struct *mm, pte_t *pte);
> +extern void __migration_entry_wait_huge(pte_t *ptep, spinlock_t *ptl);
> +extern void migration_entry_wait_huge(struct vm_area_struct *vma, pte_t *pte);
> #else
> static inline swp_entry_t make_readable_migration_entry(pgoff_t offset)
> {
> @@ -261,8 +261,9 @@ static inline void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
> spinlock_t *ptl) { }
> static inline void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
> unsigned long address) { }
> +static inline void __migration_entry_wait_huge(pte_t *ptep, spinlock_t *ptl) { }
> static inline void migration_entry_wait_huge(struct vm_area_struct *vma,
> - struct mm_struct *mm, pte_t *pte) { }
> + pte_t *pte) { }
> static inline int is_writable_migration_entry(swp_entry_t entry)
> {
> return 0;
> diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> index 48740e6c3476..2b38eaaa2e60 100644
> --- a/mm/hugetlb.c
> +++ b/mm/hugetlb.c
> @@ -5622,7 +5622,7 @@ vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
> */
> entry = huge_ptep_get(ptep);
> if (unlikely(is_hugetlb_entry_migration(entry))) {
> - migration_entry_wait_huge(vma, mm, ptep);
> + migration_entry_wait_huge(vma, ptep);
> return 0;
> } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
> return VM_FAULT_HWPOISON_LARGE |
> @@ -6770,7 +6770,7 @@ follow_huge_pmd(struct mm_struct *mm, unsigned long address,
> } else {
> if (is_hugetlb_entry_migration(pte)) {
> spin_unlock(ptl);
> - __migration_entry_wait(mm, (pte_t *)pmd, ptl);
> + __migration_entry_wait_huge((pte_t *)pmd, ptl);
> goto retry;
> }
> /*
> diff --git a/mm/migrate.c b/mm/migrate.c
> index 231907e89b93..84b685a235fe 100644
> --- a/mm/migrate.c
> +++ b/mm/migrate.c
> @@ -315,11 +315,26 @@ void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
> __migration_entry_wait(mm, ptep, ptl);
> }
>
> +void __migration_entry_wait_huge(pte_t *ptep, spinlock_t *ptl)
> +{
> + swp_entry_t entry;
> + pte_t pte;
> +
> + spin_lock(ptl);
> + pte = huge_ptep_get(ptep);
> +
> + if (unlikely(!is_hugetlb_entry_migration(pte)))
> + spin_unlock(ptl);
> + else
> + migration_entry_wait_on_locked(pte_to_swp_entry(pte), NULL, ptl);
> +}
> +
> void migration_entry_wait_huge(struct vm_area_struct *vma,
> struct mm_struct *mm, pte_t *pte)
> {
> - spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), mm, pte);
> - __migration_entry_wait(mm, pte, ptl);
> + spinlock_t *ptl = huge_pte_lockptr(hstate_vma(vma), vma->mm, pte);
> +
> + __migration_entry_wait_huge(pte, ptl);
> }
>
> #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
>
