Re: hugetlbfs: WARNING: bad unlock balance detected during MADV_REMOVE

[Miaohe Lin <linmiaohe@huawei.com>]

On 2024/1/30 12:08, Liam R. Howlett wrote:
> * Miaohe Lin <linmiaohe@huawei.com> [240129 21:14]:
>> On 2024/1/30 0:17, Liam R. Howlett wrote:
>>> * Miaohe Lin <linmiaohe@huawei.com> [240129 07:56]:
>>>> On 2024/1/27 18:13, Miaohe Lin wrote:
>>>>> On 2024/1/26 15:50, Muchun Song wrote:
>>>>>>
>>>>>>
>>>>>>> On Jan 26, 2024, at 04:28, Thorvald Natvig <thorvald@google.com> wrote:
>>>>>>>
>>>>>>> We've found what appears to be a lock issue that results in a blocked
>>>>>>> process somewhere in hugetlbfs for shared maps; seemingly from an
>>>>>>> interaction between hugetlb_vm_op_open and hugetlb_vmdelete_list.
>>>>>>>
>>>>>>> Based on some added pr_warn, we believe the following is happening:
>>>>>>> When hugetlb_vmdelete_list is entered from the child process,
>>>>>>> vma->vm_private_data is NULL, and hence hugetlb_vma_trylock_write does
>>>>>>> not lock, since neither __vma_shareable_lock nor __vma_private_lock
>>>>>>> are true.
>>>>>>>
>>>>>>> While hugetlb_vmdelete_list is executing, the parent process does
>>>>>>> fork(), which ends up in hugetlb_vm_op_open, which in turn allocates a
>>>>>>> lock for the same vma.
>>>>>>>
>>>>>>> Thus, when the hugetlb_vmdelete_list in the child reaches the end of
>>>>>>> the function, vma->vm_private_data is now populated, and hence
>>>>>>> hugetlb_vma_unlock_write tries to unlock the vma_lock, which it does
>>>>>>> not hold.
>>>>>>
>>>>>> Thanks for your report. ->vm_private_data was introduced since the
>>>>>> series [1]. So I suspect it was caused by this. But I haven't reviewed
>>>>>> that at that time (actually, it is a little complex in pmd sharing
>>>>>> case). I saw Miaohe had reviewed many of those.
>>>>>>
>>>>>> CC Miaohe, maybe he has some ideas on this.
>>>>>>
>>>>>> [1] https://lore.kernel.org/all/20220914221810.95771-7-mike.kravetz@oracle.com/T/#m2141e4bc30401a8ce490b1965b9bad74e7f791ff
>>>>>>
>>>>>> Thanks.
>>>>>>
>>>>>>>
>>>>>>> dmesg:
>>>>>>> WARNING: bad unlock balance detected!
>>>>>>> 6.8.0-rc1+ #24 Not tainted
>>>>>>> -------------------------------------
>>>>>>> lock/2613 is trying to release lock (&vma_lock->rw_sema) at:
>>>>>>> [<ffffffffa94c6128>] hugetlb_vma_unlock_write+0x48/0x60
>>>>>>> but there are no more locks to release!
>>>>>
>>>>> Thanks for your report. It seems there's a race:
>>>>>
>>>>>  CPU 1											CPU 2
>>>>>  fork											hugetlbfs_fallocate
>>>>>   dup_mmap										 hugetlbfs_punch_hole
>>>>>    i_mmap_lock_write(mapping);								
>>>>>    vma_interval_tree_insert_after -- Child vma is visible through i_mmap tree.
>>>>>    i_mmap_unlock_write(mapping);
>>>>>    hugetlb_dup_vma_private -- Clear vma_lock outside i_mmap_rwsem!			 i_mmap_lock_write(mapping);
>>>>>    											 hugetlb_vmdelete_list
>>>>> 											  vma_interval_tree_foreach
>>>>> 											   hugetlb_vma_trylock_write -- Vma_lock is cleared.
>>>>>    tmp->vm_ops->open -- Alloc new vma_lock outside i_mmap_rwsem!
>>>>> 											   hugetlb_vma_unlock_write -- Vma_lock is assigned!!!
>>>>> 											 i_mmap_unlock_write(mapping);
>>>>>
>>>>> hugetlb_dup_vma_private and hugetlb_vm_op_open are called outside i_mmap_rwsem lock. So there will be another bugs behind it.
>>>>> But I'm not really sure. I will take a more closed look at next week.
>>>>
>>>>
>>>> This can be fixed by deferring vma_interval_tree_insert_after() until vma is fully initialized.
>>>> But I'm not sure whether there're side effects with this patch.
>>>>
>>>> linux-UJMmTI:/home/linmiaohe/mm # git diff
>>>> diff --git a/kernel/fork.c b/kernel/fork.c
>>>> index 47ff3b35352e..2ef2711452e0 100644
>>>> --- a/kernel/fork.c
>>>> +++ b/kernel/fork.c
>>>> @@ -712,21 +712,6 @@ static __latent_entropy int dup_mmap(struct mm_struct *mm,
>>>>                 } else if (anon_vma_fork(tmp, mpnt))
>>>>                         goto fail_nomem_anon_vma_fork;
>>>>                 vm_flags_clear(tmp, VM_LOCKED_MASK);
>>>> -               file = tmp->vm_file;
>>>> -               if (file) {
>>>> -                       struct address_space *mapping = file->f_mapping;
>>>> -
>>>> -                       get_file(file);
>>>> -                       i_mmap_lock_write(mapping);
>>>> -                       if (vma_is_shared_maywrite(tmp))
>>>> -                               mapping_allow_writable(mapping);
>>>> -                       flush_dcache_mmap_lock(mapping);
>>>> -                       /* insert tmp into the share list, just after mpnt */
>>>> -                       vma_interval_tree_insert_after(tmp, mpnt,
>>>> -                                       &mapping->i_mmap);
>>>> -                       flush_dcache_mmap_unlock(mapping);
>>>> -                       i_mmap_unlock_write(mapping);
>>>> -               }
>>>>
>>>>                 /*
>>>>                  * Copy/update hugetlb private vma information.
>>>> @@ -747,6 +732,22 @@ static __latent_entropy int dup_mmap(struct mm_struct *mm,
>>>>                 if (tmp->vm_ops && tmp->vm_ops->open)
>>>>                         tmp->vm_ops->open(tmp);
>>>>
>>>> +               file = tmp->vm_file;
>>>> +               if (file) {
>>>> +                       struct address_space *mapping = file->f_mapping;
>>>> +
>>>> +                       get_file(file);
>>>> +                       i_mmap_lock_write(mapping);
>>>> +                       if (vma_is_shared_maywrite(tmp))
>>>> +                               mapping_allow_writable(mapping);
>>>> +                       flush_dcache_mmap_lock(mapping);
>>>> +                       /* insert tmp into the share list, just after mpnt. */
>>>> +                       vma_interval_tree_insert_after(tmp, mpnt,
>>>> +                                       &mapping->i_mmap);
>>>> +                       flush_dcache_mmap_unlock(mapping);
>>>> +                       i_mmap_unlock_write(mapping);
>>>> +               }
>>>> +
>>>>                 if (retval) {
>>>>                         mpnt = vma_next(&vmi);
>>>>                         goto loop_out;
>>>>
>>>>
>>>
>>> How is this possible?  I thought, as specified in mm/rmap.c, that the
>>> hugetlbfs path would be holding the mmap lock (which is also held in the
>>> fork path)?
>>
>> The fork path holds the mmap lock from parent A and other childs(except first child B) while hugetlbfs path
>> holds the mmap lock from first child B. So the mmap lock won't help here because it comes from different mm.
>> Or am I miss something?
> 
> You are correct.  It is also in mm/rmap.c:
>  * hugetlbfs PageHuge() take locks in this order:
>  *   hugetlb_fault_mutex (hugetlbfs specific page fault mutex)                                                          
>  *     vma_lock (hugetlb specific lock for pmd_sharing)
>  *       mapping->i_mmap_rwsem (also used for hugetlb pmd sharing)                                                      
>  *         page->flags PG_locked (lock_page)
> 
> Does it make sense for hugetlb_dup_vma_private()  to assert
> mapping->i_mmap_rwsem is locked?  When is that necessary?

I'm afraid not. AFAICS, vma_lock(vma->vm_private_data) is only modified at the time of
vma creating or destroy. Vma_lock is not supposed to be used at that time.

> 
> I also think it might be safer to move the hugetlb_dup_vma_private()
> call up instead of the insert into the interval tree down?
> See the following comment from mmap.c:
> 
>                         /*                                                                                              
>                          * Put into interval tree now, so instantiated pages                                            
>                          * are visible to arm/parisc __flush_dcache_page
>                          * throughout; but we cannot insert into address                                                
>                          * space until vma start or end is updated.                                                     
>                          */
> 
> So there may be arch dependent reasons for this order.

Yes, it should be safer to move hugetlb_dup_vma_private() call up. But we also need to move tmp->vm_ops->open(tmp) call up.
Or the race still exists:

 CPU 1											CPU 2
 fork											hugetlbfs_fallocate
  dup_mmap										 hugetlbfs_punch_hole
   hugetlb_dup_vma_private -- Clear vma_lock.	<-- it is moved up.
   i_mmap_lock_write(mapping);								
   vma_interval_tree_insert_after -- Child vma is visible through i_mmap tree.
   i_mmap_unlock_write(mapping);
   		 									 i_mmap_lock_write(mapping);
   											 hugetlb_vmdelete_list
											  vma_interval_tree_foreach
											   hugetlb_vma_trylock_write -- Vma_lock is already cleared.
   tmp->vm_ops->open -- Alloc new vma_lock outside i_mmap_rwsem!
											   hugetlb_vma_unlock_write -- Vma_lock is assigned!!!
											 i_mmap_unlock_write(mapping);


My patch should not be a complete solution. It's used to prove and fix the race quickly. It's very great if you or
someone else can provide a better and safer solution.

Thanks.

> 
> Thanks,
> Liam
> 
> .
>