Re: [PATCH v2 1/1] mm/hugetlb: fix possible deadlocks in hugetlb VMA unmap paths

[Hillf Danton <hdanton@sina.com>]

On Tue, 11 Nov 2025 00:39:29 +0800 Lance Yang wrote:
> On 2025/11/10 20:17, Harry Yoo wrote:
> > On Mon, Nov 10, 2025 at 07:15:53PM +0800, Lance Yang wrote:
> >> From: Lance Yang <lance.yang@linux.dev>
> >>
> >> The hugetlb VMA unmap path contains several potential deadlocks, as
> >> reported by syzbot. These deadlocks occur in __hugetlb_zap_begin(),
> >> move_hugetlb_page_tables(), and the retry path of
> >> hugetlb_unmap_file_folio() (affecting remove_inode_hugepages() and
> >> unmap_vmas()), where vma_lock is acquired before i_mmap_lock. This lock
> >> ordering conflicts with other paths like hugetlb_fault(), which establish
> >> the correct dependency as i_mmap_lock -> vma_lock.
> >>
> >> Possible unsafe locking scenario:
> >>
> >> CPU0                                 CPU1
> >> ----                                 ----
> >> lock(&vma_lock->rw_sema);
> >>                                       lock(&i_mmap_lock);
> >>                                       lock(&vma_lock->rw_sema);
> >> lock(&i_mmap_lock);
> >>
> >> Resolve the circular dependencies reported by syzbot across multiple call
> >> chains by reordering the locks in all conflicting paths to consistently
> >> follow the established i_mmap_lock -> vma_lock order.
> > 
> > But mm/rmap.c says:
> >> * 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)
> >> *         folio_lock
> >> */
> 
> Thanks! You are right, I was mistaken ...
> 
> > 
> > I think the commit message should explain why the locking order described
> > above is incorrect (or when it became incorrect) and fix the comment?
> 
> I think the locking order documented in mm/rmap.c (vma_lock -> i_mmap_lock)
> is indeed the correct one to follow.
> 
> This fix has it backwards then. I'll rework it to fix the actual violations.
>
Break a leg, better after taking a look at ffa1e7ada456 ("block: Make
request_queue lockdep splats show up earlier")