Re: [BUG] mm/mempolicy: possible double-unlock in migrate_to_node() on 6.6

["David Hildenbrand (Arm)" <david@kernel.org>]

On 3/3/26 05:15, Matthew Wilcox wrote:
> On Tue, Mar 03, 2026 at 10:52:51AM +0800, liying3@sungrowpower.com wrote:
>> Hi Andrew,
>>
>> I found that a fix is required for this submission in branch 6.6.
>> https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/commit/?id=a13b2b9b0b0b04612c7d81e3b3dfb485c5f7abc3 
> 
> Agreed, although your email should more properly have been directed to
> David and possibly Greg.  Are you going to submit a patch to fix this,
> or do you prefer that somebody else does it?
> 
>> Regarding the difference between branches 6.6 and 6.12, in branch 6.6, mmap_read_lock() and mmap_read_unlock() are called outside the migrate_to_node() function. 
>> Therefore, mmap_read_unlock(mm) should not be invoked inside migrate_to_node(); otherwise, it would result in a double unlock of the mm.
>> diff --git a/mm/mempolicy.c b/mm/mempolicy.c
>> index 54f1b78d1b2c0f..94c74c594d102a 100644
>> --- a/mm/mempolicy.c
>> +++ b/mm/mempolicy.c
>> @@ -1071,6 +1071,10 @@ static long migrate_to_node(struct mm_struct *mm, int source, int dest,
>> VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
>> vma = find_vma(mm, 0);
>> + if (unlikely(!vma)) {
>> + mmap_read_unlock(mm); // this line should be deleted
>> + return 0;
>> + }

Right, that unlock must go. Let me know if I should prepare a fix.

-- 
Cheers,

David