On 2022/3/7 15:35, Alistair Popple wrote:
> Miaohe Lin <linmiaohe@xxxxxxxxxx> writes:
>
>> __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.
>
> [...]
>
>> diff --git a/mm/filemap.c b/mm/filemap.c
>> index 8f7e6088ee2a..18c353d52aae 100644
>> --- a/mm/filemap.c
>> +++ b/mm/filemap.c
>> @@ -1389,6 +1389,7 @@ static inline int folio_wait_bit_common(struct folio *folio, int bit_nr,
>> * @ptep: mapped pte pointer. Will return with the ptep unmapped. Only required
>> * for pte entries, pass NULL for pmd entries.
>> * @ptl: already locked ptl. This function will drop the lock.
>> + * @unmap: indicating whether ptep need to be unmapped.
>> *
>> * Wait for a migration entry referencing the given page to be removed. This is
>> * equivalent to put_and_wait_on_page_locked(page, TASK_UNINTERRUPTIBLE) except
>> @@ -1402,7 +1403,7 @@ static inline int folio_wait_bit_common(struct folio *folio, int bit_nr,
>> * there.
>> */
>> void migration_entry_wait_on_locked(swp_entry_t entry, pte_t *ptep,
>> - spinlock_t *ptl)
>> + spinlock_t *ptl, bool unmap)
>
> It's a pity we have to pass unmap all the way down to
> migration_entry_wait_on_locked().
>
>> {
>> struct wait_page_queue wait_page;
>> wait_queue_entry_t *wait = &wait_page.wait;
>> @@ -1439,10 +1440,9 @@ void migration_entry_wait_on_locked(swp_entry_t entry, pte_t *ptep,
>> * a valid reference to the page, and it must take the ptl to remove the
>> * migration entry. So the page is valid until the ptl is dropped.
>> */
>> - if (ptep)
>> - pte_unmap_unlock(ptep, ptl);
>> - else
>> - spin_unlock(ptl);
>> + spin_unlock(ptl);
>> + if (unmap && ptep)
>> + pte_unmap(ptep);
>
> However we might not have to - afaict this is the only usage of ptep so callers
> could do the pte_unmap() prior to calling migration_entry_wait_on_locked(). We
> could then remove both the `ptep` and `unmap` parameters. Ie:
>
> migration_entry_wait_on_locked(swp_entry_t entry, spinlock_t *ptl)
>
> This does assume it's ok to change the order of pte unmap/ptl unlock operations.
> I'm not terribly familiar with CONFIG_HIGHPTE systems, but it seems like that
> should be ok.
>
Looks like a good idea. We can leave the pte_unmap to the caller as only page table
spin_lock is needed to make sure the validity of the page.
Will try to do this in V2. Many thanks.
> - Alistair
>
>>
>> for (;;) {
>> unsigned int flags;
>> diff --git a/mm/hugetlb.c b/mm/hugetlb.c
>> index 07668781c246..8088128c25db 100644
>> --- a/mm/hugetlb.c
>> +++ b/mm/hugetlb.c
>> @@ -6713,7 +6713,7 @@ 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(mm, (pte_t *)pmd, ptl, false);
>> goto retry;
>> }
>> /*
>> diff --git a/mm/migrate.c b/mm/migrate.c
>> index 98a968e6f465..5519261f54fe 100644
>> --- a/mm/migrate.c
>> +++ b/mm/migrate.c
>> @@ -281,7 +281,7 @@ void remove_migration_ptes(struct folio *src, struct folio *dst, bool locked)
>> * When we return from this function the fault will be retried.
>> */
>> void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
>> - spinlock_t *ptl)
>> + spinlock_t *ptl, bool unmap)
>> {
>> pte_t pte;
>> swp_entry_t entry;
>> @@ -295,10 +295,12 @@ void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
>> if (!is_migration_entry(entry))
>> goto out;
>>
>> - migration_entry_wait_on_locked(entry, ptep, ptl);
>> + migration_entry_wait_on_locked(entry, ptep, ptl, unmap);
>> return;
>> out:
>> - pte_unmap_unlock(ptep, ptl);
>> + spin_unlock(ptl);
>> + if (unmap)
>> + pte_unmap(ptep);
>> }
>>
>> void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
>> @@ -306,14 +308,14 @@ void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
>> {
>> spinlock_t *ptl = pte_lockptr(mm, pmd);
>> pte_t *ptep = pte_offset_map(pmd, address);
>> - __migration_entry_wait(mm, ptep, ptl);
>> + __migration_entry_wait(mm, ptep, ptl, true);
>> }
>>
>> 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);
>> + __migration_entry_wait(mm, pte, ptl, false);
>> }
>>
>> #ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
>> @@ -324,7 +326,7 @@ void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
>> ptl = pmd_lock(mm, pmd);
>> if (!is_pmd_migration_entry(*pmd))
>> goto unlock;
>> - migration_entry_wait_on_locked(pmd_to_swp_entry(*pmd), NULL, ptl);
>> + migration_entry_wait_on_locked(pmd_to_swp_entry(*pmd), NULL, ptl, false);
>> return;
>> unlock:
>> spin_unlock(ptl);
