[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <cd137540-ae01-46a1-93d2-062bc21b827c@bytedance.com>
Date: Fri, 30 Aug 2024 14:54:16 +0800
From: Qi Zheng <zhengqi.arch@...edance.com>
To: Muchun Song <muchun.song@...ux.dev>
Cc: david@...hat.com, hughd@...gle.com, willy@...radead.org,
vbabka@...nel.org, akpm@...ux-foundation.org, rppt@...nel.org,
vishal.moola@...il.com, peterx@...hat.com, ryan.roberts@....com,
christophe.leroy2@...soprasteria.com, linux-kernel@...r.kernel.org,
linux-mm@...ck.org, linux-arm-kernel@...ts.infradead.org,
linuxppc-dev@...ts.ozlabs.org
Subject: Re: [PATCH v2 07/14] mm: khugepaged: collapse_pte_mapped_thp() use
pte_offset_map_rw_nolock()
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@...edance.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
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:
>
Powered by blists - more mailing lists