Re: [PATCH 8/8] hugetlb: use new vma_lock for pmd sharing synchronization

[Miaohe Lin <linmiaohe@huawei.com>]

On 2022/9/14 8:50, Mike Kravetz wrote:
> On 09/13/22 10:14, Miaohe Lin wrote:
>> On 2022/9/13 7:02, Mike Kravetz wrote:
>>> On 09/05/22 11:08, Miaohe Lin wrote:
>>>> On 2022/9/3 7:07, Mike Kravetz wrote:
>>>>> On 08/30/22 10:02, Miaohe Lin wrote:
>>>>>> On 2022/8/25 1:57, Mike Kravetz wrote:
>>>>>>> The new hugetlb vma lock (rw semaphore) is used to address this race:
>>>>>>>
>>>>>>> Faulting thread                                 Unsharing thread
>>>>>>> ...                                                  ...
>>>>>>> ptep = huge_pte_offset()
>>>>>>>       or
>>>>>>> ptep = huge_pte_alloc()
>>>>>>> ...
>>>>>>>                                                 i_mmap_lock_write
>>>>>>>                                                 lock page table
>>>>>>> ptep invalid   <------------------------        huge_pmd_unshare()
>>>>>>> Could be in a previously                        unlock_page_table
>>>>>>> sharing process or worse                        i_mmap_unlock_write
>>>>>>> ...
>>>>>>>
>>>>>>> The vma_lock is used as follows:
>>>>>>> - During fault processing. the lock is acquired in read mode before
>>>>>>>   doing a page table lock and allocation (huge_pte_alloc).  The lock is
>>>>>>>   held until code is finished with the page table entry (ptep).
>>>>>>> - The lock must be held in write mode whenever huge_pmd_unshare is
>>>>>>>   called.
>>>>>>>
>>>>>>> Lock ordering issues come into play when unmapping a page from all
>>>>>>> vmas mapping the page.  The i_mmap_rwsem must be held to search for the
>>>>>>> vmas, and the vma lock must be held before calling unmap which will
>>>>>>> call huge_pmd_unshare.  This is done today in:
>>>>>>> - try_to_migrate_one and try_to_unmap_ for page migration and memory
>>>>>>>   error handling.  In these routines we 'try' to obtain the vma lock and
>>>>>>>   fail to unmap if unsuccessful.  Calling routines already deal with the
>>>>>>>   failure of unmapping.
>>>>>>> - hugetlb_vmdelete_list for truncation and hole punch.  This routine
>>>>>>>   also tries to acquire the vma lock.  If it fails, it skips the
>>>>>>>   unmapping.  However, we can not have file truncation or hole punch
>>>>>>>   fail because of contention.  After hugetlb_vmdelete_list, truncation
>>>>>>>   and hole punch call remove_inode_hugepages.  remove_inode_hugepages
>>>>>>>   check for mapped pages and call hugetlb_unmap_file_page to unmap them.
>>>>>>>   hugetlb_unmap_file_page is designed to drop locks and reacquire in the
>>>>>>>   correct order to guarantee unmap success.
>>>>>>>
>>>>>>> Signed-off-by: Mike Kravetz <mike.kravetz@oracle.com>
>>>>>>> ---
>>>>>>>  fs/hugetlbfs/inode.c |  46 +++++++++++++++++++
>>>>>>>  mm/hugetlb.c         | 102 +++++++++++++++++++++++++++++++++++++++----
>>>>>>>  mm/memory.c          |   2 +
>>>>>>>  mm/rmap.c            | 100 +++++++++++++++++++++++++++---------------
>>>>>>>  mm/userfaultfd.c     |   9 +++-
>>>>>>>  5 files changed, 214 insertions(+), 45 deletions(-)
>>>>>>>
>>>>>>> diff --git a/fs/hugetlbfs/inode.c b/fs/hugetlbfs/inode.c
>>>>>>> index b93d131b0cb5..52d9b390389b 100644
>>>>>>> --- a/fs/hugetlbfs/inode.c
>>>>>>> +++ b/fs/hugetlbfs/inode.c
>>>>>>> @@ -434,6 +434,8 @@ static void hugetlb_unmap_file_folio(struct hstate *h,
>>>>>>>  					struct folio *folio, pgoff_t index)
>>>>>>>  {
>>>>>>>  	struct rb_root_cached *root = &mapping->i_mmap;
>>>>>>> +	unsigned long skipped_vm_start;
>>>>>>> +	struct mm_struct *skipped_mm;
>>>>>>>  	struct page *page = &folio->page;
>>>>>>>  	struct vm_area_struct *vma;
>>>>>>>  	unsigned long v_start;
>>>>>>> @@ -444,6 +446,8 @@ static void hugetlb_unmap_file_folio(struct hstate *h,
>>>>>>>  	end = ((index + 1) * pages_per_huge_page(h));
>>>>>>>  
>>>>>>>  	i_mmap_lock_write(mapping);
>>>>>>> +retry:
>>>>>>> +	skipped_mm = NULL;
>>>>>>>  
>>>>>>>  	vma_interval_tree_foreach(vma, root, start, end - 1) {
>>>>>>>  		v_start = vma_offset_start(vma, start);
>>>>>>> @@ -452,11 +456,49 @@ static void hugetlb_unmap_file_folio(struct hstate *h,
>>>>>>>  		if (!hugetlb_vma_maps_page(vma, vma->vm_start + v_start, page))
>>>>>>>  			continue;
>>>>>>>  
>>>>>>> +		if (!hugetlb_vma_trylock_write(vma)) {
>>>>>>> +			/*
>>>>>>> +			 * If we can not get vma lock, we need to drop
>>>>>>> +			 * immap_sema and take locks in order.
>>>>>>> +			 */
>>>>>>> +			skipped_vm_start = vma->vm_start;
>>>>>>> +			skipped_mm = vma->vm_mm;
>>>>>>> +			/* grab mm-struct as we will be dropping i_mmap_sema */
>>>>>>> +			mmgrab(skipped_mm);
>>>>>>> +			break;
>>>>>>> +		}
>>>>>>> +
>>>>>>>  		unmap_hugepage_range(vma, vma->vm_start + v_start, v_end,
>>>>>>>  				NULL, ZAP_FLAG_DROP_MARKER);
>>>>>>> +		hugetlb_vma_unlock_write(vma);
>>>>>>>  	}
>>>>>>>  
>>>>>>>  	i_mmap_unlock_write(mapping);
>>>>>>> +
>>>>>>> +	if (skipped_mm) {
>>>>>>> +		mmap_read_lock(skipped_mm);
>>>>>>> +		vma = find_vma(skipped_mm, skipped_vm_start);
>>>>>>> +		if (!vma || !is_vm_hugetlb_page(vma) ||
>>>>>>> +					vma->vm_file->f_mapping != mapping ||
>>>>>>> +					vma->vm_start != skipped_vm_start) {
>>>>>>
>>>>>> i_mmap_lock_write(mapping) is missing here? Retry logic will do i_mmap_unlock_write(mapping) anyway.
>>>>>>
>>>>>
>>>>> Yes, that is missing.  I will add here.
>>>>>
>>>>>>> +			mmap_read_unlock(skipped_mm);
>>>>>>> +			mmdrop(skipped_mm);
>>>>>>> +			goto retry;
>>>>>>> +		}
>>>>>>> +
>>>>>>
>>>>>> IMHO, above check is not enough. Think about the below scene:
>>>>>>
>>>>>> CPU 1					CPU 2
>>>>>> hugetlb_unmap_file_folio		exit_mmap
>>>>>>   mmap_read_lock(skipped_mm);		  mmap_read_lock(mm);
>>>>>>   check vma is wanted.
>>>>>>   					  unmap_vmas
>>>>>>   mmap_read_unlock(skipped_mm);		  mmap_read_unlock
>>>>>>   					  mmap_write_lock(mm);
>>>>>>   					  free_pgtables
>>>>>>   					  remove_vma
>>>>>> 					    hugetlb_vma_lock_free
>>>>>>   vma, hugetlb_vma_lock is still *used after free*
>>>>>>   					  mmap_write_unlock(mm);
>>>>>> So we should check mm->mm_users == 0 to fix the above issue. Or am I miss something?
>>>>>
>>>>> In the retry case, we are OK because go back and look up the vma again.  Right?
>>>>>
>>>>> After taking mmap_read_lock, vma can not go away until we mmap_read_unlock.
>>>>> Before that, we do the following:
>>>>>
>>>>>>> +		hugetlb_vma_lock_write(vma);
>>>>>>> +		i_mmap_lock_write(mapping);
>>>>>
>>>>> IIUC, vma can not go away while we hold i_mmap_lock_write.  So, after this we
>>>>
>>>> I think you're right. free_pgtables() can't complete its work as unlink_file_vma() will be
>>>> blocked on i_mmap_rwsem of mapping. Sorry for reporting such nonexistent race.
>>>>
>>>>> can.
>>>>>
>>>>>>> +		mmap_read_unlock(skipped_mm);
>>>>>>> +		mmdrop(skipped_mm);
>>>>>
>>>>> We continue to hold i_mmap_lock_write as we goto retry.
>>>>>
>>>>> I could be missing something as well.  This was how I intended to keep
>>>>> vma valid while dropping and acquiring locks.
>>>>
>>>> Thanks for your clarifying.
>>>>
>>>
>>> Well, that was all correct 'in theory' but not in practice.  I did not take
>>> into account the inode lock that is taken at the beginning of truncate (or
>>> hole punch).  In other code paths, we take inode lock after mmap_lock.  So,
>>> taking mmap_lock here is not allowed.
>>
>> Considering the Lock ordering in mm/filemap.c:
>>
>>  *  ->i_rwsem
>>  *    ->invalidate_lock		(acquired by fs in truncate path)
>>  *      ->i_mmap_rwsem		(truncate->unmap_mapping_range)
>>
>>  *  ->i_rwsem			(generic_perform_write)
>>  *    ->mmap_lock		(fault_in_readable->do_page_fault)
>>
>> It seems inode_lock is taken before the mmap_lock?
> 
> Hmmmm?  I can't find a sequence where inode_lock is taken after mmap_lock.
> lockdep was complaining about taking mmap_lock after i_rwsem in the above code.
> I assumed there was such a sequence somewhere.  Might need to go back and get
> another trace/warning.

Sorry, I'm somewhat confused. Take generic_file_write_iter() as an example:

generic_file_write_iter
  inode_lock(inode); -- *inode lock is held here*
  __generic_file_write_iter
    generic_perform_write
      fault_in_iov_iter_readable -- *may cause page fault and thus take mmap_lock*
  inode_unlock(inode);

This is the documented example in the mm/filemap.c. So we should take inode_lock before
taking mmap_lock. Or this is out-dated ? And above example needs a fix?

> 
> In any case, I think the scheme below is much cleaner.  Doing another round of
> benchmarking before sending.

That should be a good alternative. Thanks for your work. :)

Thanks,
Miaohe Lin

> 
>>> I came up with another way to make this work.  As discussed above, we need to
>>> drop the i_mmap lock before acquiring the vma_lock.  However, once we drop
>>> i_mmap, the vma could go away.  My solution is to make the 'vma_lock' be a
>>> ref counted structure that can live on after the vma is freed.  Therefore,
>>> this code can take a reference while under i_mmap then drop i_mmap and wait
>>> on the vma_lock.  Of course, once it acquires the vma_lock it needs to check
>>> and make sure the vma still exists.  It may sound complicated, but I think
>>> it is a bit simpler than the code here.  A new series will be out soon.
>>>
>