Re: [RFC PATCH v4 4/8] hugetlbfs: catch and handle truncate racing with page faults

[David Hildenbrand <david@redhat.com>]

On 28.07.22 18:45, Mike Kravetz wrote:
> On 07/28/22 10:02, Miaohe Lin wrote:
>> On 2022/7/28 3:00, Mike Kravetz wrote:
>>> On 07/27/22 17:20, Miaohe Lin wrote:
>>>> On 2022/7/7 4:23, Mike Kravetz wrote:
>>>>> Most hugetlb fault handling code checks for faults beyond i_size.
>>>>> While there are early checks in the code paths, the most difficult
>>>>> to handle are those discovered after taking the page table lock.
>>>>> At this point, we have possibly allocated a page and consumed
>>>>> associated reservations and possibly added the page to the page cache.
>>>>>
>>>>> When discovering a fault beyond i_size, be sure to:
>>>>> - Remove the page from page cache, else it will sit there until the
>>>>>   file is removed.
>>>>> - Do not restore any reservation for the page consumed.  Otherwise
>>>>>   there will be an outstanding reservation for an offset beyond the
>>>>>   end of file.
>>>>>
>>>>> The 'truncation' code in remove_inode_hugepages must deal with fault
>>>>> code potentially removing a page/folio from the cache after the page was
>>>>> returned by filemap_get_folios and before locking the page.  This can be
>>>>> discovered by a change in folio_mapping() after taking folio lock.  In
>>>>> addition, this code must deal with fault code potentially consuming
>>>>> and returning reservations.  To synchronize this, remove_inode_hugepages
>>>>> will now take the fault mutex for ALL indices in the hole or truncated
>>>>> range.  In this way, it KNOWS fault code has finished with the page/index
>>>>> OR fault code will see the updated file size.
>>>>>
>>>>> Signed-off-by: Mike Kravetz <mike.kravetz@oracle.com>
>>>>> ---
>>>>
>>>> <snip>
>>>>
>>>>> @@ -5606,8 +5610,10 @@ static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
>>>>>  
>>>>>  	ptl = huge_pte_lock(h, mm, ptep);
>>>>>  	size = i_size_read(mapping->host) >> huge_page_shift(h);
>>>>> -	if (idx >= size)
>>>>> +	if (idx >= size) {
>>>>> +		beyond_i_size = true;
>>>>
>>>> Thanks for your patch. There is one question:
>>>>
>>>> Since races between hugetlb pagefault and truncate is guarded by hugetlb_fault_mutex,
>>>> do we really need to check it again after taking the page table lock?
>>>>
>>>
>>> Well, the fault mutex can only guard a single hugetlb page.  The fault mutex
>>> is actually an array/table of mutexes hashed by mapping address and file index.
>>> So, during truncation we take take the mutex for each page as they are
>>> unmapped and removed.  So, the fault mutex only synchronizes operations
>>> on one specific page.  The idea with this patch is to coordinate the fault
>>> code and truncate code when operating on the same page.
>>>
>>> In addition, changing the file size happens early in the truncate process
>>> before taking any locks/mutexes.
>>
>> I wonder whether we can somewhat live with it to make code simpler. When changing the file size happens
>> after checking i_size but before taking the page table lock in hugetlb_fault, the truncate code would
>> remove the hugetlb page from the page cache for us after hugetlb_fault finishes if we don't roll back
>> when checking i_size again under the page table lock?
>>
>> In a word, if hugetlb_fault see a truncated inode, back out early. If not, let truncate code does its
>> work. So we don't need to complicate the already complicated error path. Or am I miss something?
>>
> 
> Thank you! I believe your observations and suggestions are correct.
> 
> We can just let the fault code proceed after the early "idx >= size",
> and let the truncation code remove the page.  This also eliminates the
> need for patch 3 (hugetlbfs: move routine remove_huge_page to hugetlb.c).

At least remaining the functions would be very welcome nonetheless :)

> 
> I will make these changes in the next version.

Just so I understand correctly, we want to let fault handling code back
out early if we find any incompatible change, and simply retry the
fault? I'm thinking about some kind of a high-level seqcount.

-- 
Thanks,

David / dhildenb