Re: [PATCH RFC 1/1] hugetlbfs: introduce truncation/fault mutex to avoid races

[Mike Kravetz <mike.kravetz@oracle.com>]

On 10/8/18 1:03 AM, Kirill A. Shutemov wrote:
> On Sun, Oct 07, 2018 at 04:38:48PM -0700, Mike Kravetz wrote:
>> The following hugetlbfs truncate/page fault race can be recreated
>> with programs doing something like the following.
>>
>> A huegtlbfs file is mmap(MAP_SHARED) with a size of 4 pages.  At
>> mmap time, 4 huge pages are reserved for the file/mapping.  So,
>> the global reserve count is 4.  In addition, since this is a shared
>> mapping an entry for 4 pages is added to the file's reserve map.
>> The first 3 of the 4 pages are faulted into the file.  As a result,
>> the global reserve count is now 1.
>>
>> Task A starts to fault in the last page (routines hugetlb_fault,
>> hugetlb_no_page).  It allocates a huge page (alloc_huge_page).
>> The reserve map indicates there is a reserved page, so this is
>> used and the global reserve count goes to 0.
>>
>> Now, task B truncates the file to size 0.  It starts by setting
>> inode size to 0(hugetlb_vmtruncate).  It then unmaps all mapping
>> of the file (hugetlb_vmdelete_list).  Since task A's page table
>> lock is not held at the time, truncation is not blocked.  Truncation
>> removes the 3 pages from the file (remove_inode_hugepages).  When
>> cleaning up the reserved pages (hugetlb_unreserve_pages), it notices
>> the reserve map was for 4 pages.  However, it has only freed 3 pages.
>> So it assumes there is still (4 - 3) 1 reserved pages.  It then
>> decrements the global reserve count by 1 and it goes negative.
>>
>> Task A then continues the page fault process and adds it's newly
>> acquired page to the page cache.  Note that the index of this page
>> is beyond the size of the truncated file (0).  The page fault process
>> then notices the file has been truncated and exits.  However, the
>> page is left in the cache associated with the file.
>>
>> Now, if the file is immediately deleted the truncate code runs again.
>> It will find and free the one page associated with the file.  When
>> cleaning up reserves, it notices the reserve map is empty.  Yet, one
>> page freed.  So, the global reserve count is decremented by (0 - 1) -1.
>> This returns the global count to 0 as it should be.  But, it is
>> possible for someone else to mmap this file/range before it is deleted.
>> If this happens, a reserve map entry for the allocated page is created
>> and the reserved page is forever leaked.
>>
>> To avoid all these conditions, let's simply prevent faults to a file
>> while it is being truncated.  Add a new truncation specific rw mutex
>> to hugetlbfs inode extensions.  faults take the mutex in read mode,
>> truncation takes in write mode.
> 
> Hm. Don't we have already a lock for this? I mean i_mmap_lock.
> 

Thanks Kirill,

Yes, we could use use i_mmap_rwsem for this synchronization.  I don't
see anyone else using the mutex in this manner.  hugetlb code only
explicitly takes this mutex in write mode today.  I suspect that is not
optimal and could be improved.  Certainly, the use within
hugetlb_fault->huge_pte_alloc->huge_pmd_share would need to be changed
if we always wanted to take the mutex in read mode during faults.

I'll work on the changes to use i_mmap_rwsem.

However, right now our DB team informs me that the truncate/fault race
is not the cause of their huge page reserve count going negative issue.
So, I am searching for more bugs in this area.  Found another where an
allocation for migration could race with a fault in a VM_NORESERVE vma.
But, there were no migrations noted on the system, so there must be another
bug.  Sigh!
-- 
Mike Kravetz