Re: [PATCH v2 07/14] mm: khugepaged: collapse_pte_mapped_thp() use pte_offset_map_rw_nolock()

[Qi Zheng <zhengqi.arch@bytedance.com>]

On 2024/9/5 14:32, Muchun Song wrote:
> 
> 
>> On Aug 30, 2024, at 14:54, Qi Zheng <zhengqi.arch@bytedance.com> wrote:
>>
>>
>>
>> On 2024/8/29 16:10, Muchun Song wrote:
>>> On 2024/8/22 15:13, Qi Zheng wrote:
>>>> In collapse_pte_mapped_thp(), we may modify the pte and pmd entry after
>>>> acquring the ptl, so convert it to using pte_offset_map_rw_nolock(). At
>>>> this time, the write lock of mmap_lock is not held, and the pte_same()
>>>> check is not performed after the PTL held. So we should get pgt_pmd and do
>>>> pmd_same() check after the ptl held.
>>>>
>>>> For the case where the ptl is released first and then the pml is acquired,
>>>> the PTE page may have been freed, so we must do pmd_same() check before
>>>> reacquiring the ptl.
>>>>
>>>> Signed-off-by: Qi Zheng <zhengqi.arch@bytedance.com>
>>>> ---
>>>>   mm/khugepaged.c | 16 +++++++++++++++-
>>>>   1 file changed, 15 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
>>>> index 53bfa7f4b7f82..15d3f7f3c65f2 100644
>>>> --- a/mm/khugepaged.c
>>>> +++ b/mm/khugepaged.c
>>>> @@ -1604,7 +1604,7 @@ int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
>>>>       if (userfaultfd_armed(vma) && !(vma->vm_flags & VM_SHARED))
>>>>           pml = pmd_lock(mm, pmd);
>>>> -    start_pte = pte_offset_map_nolock(mm, pmd, haddr, &ptl);
>>>> +    start_pte = pte_offset_map_rw_nolock(mm, pmd, haddr, &pgt_pmd, &ptl);
>>>>       if (!start_pte)        /* mmap_lock + page lock should prevent this */
>>>>           goto abort;
>>>>       if (!pml)
>>>> @@ -1612,6 +1612,9 @@ int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
>>>>       else if (ptl != pml)
>>>>           spin_lock_nested(ptl, SINGLE_DEPTH_NESTING);
>>>> +    if (unlikely(!pmd_same(pgt_pmd, pmdp_get_lockless(pmd))))
>>>> +        goto abort;
>>>> +
>>>>       /* step 2: clear page table and adjust rmap */
>>>>       for (i = 0, addr = haddr, pte = start_pte;
>>>>            i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
>>>> @@ -1657,6 +1660,16 @@ int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
>>>>       /* step 4: remove empty page table */
>>>>       if (!pml) {
>>>>           pml = pmd_lock(mm, pmd);
>>>> +        /*
>>>> +         * We called pte_unmap() and release the ptl before acquiring
>>>> +         * the pml, which means we left the RCU critical section, so the
>>>> +         * PTE page may have been freed, so we must do pmd_same() check
>>>> +         * before reacquiring the ptl.
>>>> +         */
>>>> +        if (unlikely(!pmd_same(pgt_pmd, pmdp_get_lockless(pmd)))) {
>>>> +            spin_unlock(pml);
>>>> +            goto pmd_change;
>>> Seems we forget to flush TLB since we've cleared some pte entry?
>>
>> See comment above the ptep_clear():
>>
>> /*
>> * Must clear entry, or a racing truncate may re-remove it.
>> * TLB flush can be left until pmdp_collapse_flush() does it.
>> * PTE dirty? Shmem page is already dirty; file is read-only.
>> */
>>
>> The TLB flush was handed over to pmdp_collapse_flush(). If a
> 
> But you skipped pmdp_collapse_flush().

I skip it only in !pmd_same() case, at which time it must be cleared
by other thread, which will be responsible for flushing TLB:

CPU 0				CPU 1
				pmd_clear
				spin_unlock
				flushing tlb
spin_lock
if (!pmd_same)	
	goto pmd_change;
pmdp_collapse_flush

Did I miss something?

> 
>> concurrent thread free the PTE page at this time, the TLB will
>> also be flushed after pmd_clear().
>>
>>>> +        }
>>>>           if (ptl != pml)
>>>>               spin_lock_nested(ptl, SINGLE_DEPTH_NESTING);
>>>>       }
>>>> @@ -1688,6 +1701,7 @@ int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
>>>>           pte_unmap_unlock(start_pte, ptl);
>>>>       if (pml && pml != ptl)
>>>>           spin_unlock(pml);
>>>> +pmd_change:
>>>>       if (notified)
>>>>           mmu_notifier_invalidate_range_end(&range);
>>>>   drop_folio:
>