Re: [PATCH 1/3] hugetlb: fix vma lock handling during split vma and range unmapping

From: Mike Kravetz
Date: Mon Oct 17 2022 - 22:58:14 EST


On 10/15/22 09:25, Miaohe Lin wrote:
> Sorry for late respond. It's a really busy week. :)
>
> On 2022/10/5 9:17, Mike Kravetz wrote:
> > The hugetlb vma lock hangs off the vm_private_data field and is specific
> > to the vma. When vm_area_dup() is called as part of vma splitting, the
>
> Oh, I checked vm_area_dup() from callsite of copy_vma and dup_mmap but split_vma
> is missed... And yes, vma splitting can occur but vma merging won't for hugetlb
> vma. Thanks for catching this, Mike.
>
> > vma lock pointer is copied to the new vma. This will result in issues
> > such as double freeing of the structure. Update the hugetlb open vm_ops
> > to allocate a new vma lock for the new vma.
> >
> > The routine __unmap_hugepage_range_final unconditionally unset
> > VM_MAYSHARE to prevent subsequent pmd sharing. hugetlb_vma_lock_free
> > attempted to anticipate this by checking both VM_MAYSHARE and VM_SHARED.
> > However, if only VM_MAYSHARE was set we would miss the free. With the
> > introduction of the vma lock, a vma can not participate in pmd sharing
> > if vm_private_data is NULL. Instead of clearing VM_MAYSHARE in
> > __unmap_hugepage_range_final, free the vma lock to prevent sharing. Also,
> > update the sharing code to make sure vma lock is indeed a condition for
> > pmd sharing. hugetlb_vma_lock_free can then key off VM_MAYSHARE and not
> > miss any vmas.
> >
> > Fixes: "hugetlb: add vma based lock for pmd sharing"
> > Signed-off-by: Mike Kravetz <mike.kravetz@xxxxxxxxxx>
> > ---
> > mm/hugetlb.c | 43 +++++++++++++++++++++++++++----------------
> > mm/memory.c | 4 ----
> > 2 files changed, 27 insertions(+), 20 deletions(-)
> >
> > diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> > index 4443e87e814b..0129d371800c 100644
> > --- a/mm/hugetlb.c
> > +++ b/mm/hugetlb.c
> > @@ -4612,7 +4612,14 @@ static void hugetlb_vm_op_open(struct vm_area_struct *vma)
> > kref_get(&resv->refs);
> > }
> >
> > - hugetlb_vma_lock_alloc(vma);
> > + /*
> > + * vma_lock structure for sharable mappings is vma specific.
> > + * Clear old pointer (if copied via vm_area_dup) and create new.
> > + */
> > + if (vma->vm_flags & VM_MAYSHARE) {
> > + vma->vm_private_data = NULL;
> > + hugetlb_vma_lock_alloc(vma);
> > + }
>
> IMHO this would lead to memoryleak. Think about the below move_vma() flow:
> move_vma
> copy_vma
> new_vma = vm_area_dup(vma);
> new_vma->vm_ops->open(new_vma); --> new_vma has its own vma lock.
> is_vm_hugetlb_page(vma)
> clear_vma_resv_huge_pages
> hugetlb_dup_vma_private --> vma->vm_private_data is set to NULL
> without put ref. So vma lock is *leaked*?

You are right, that could lead to a leak.

I have an idea about setting vma->vm_private_data to NULL for VM_MAYSHARE
vmas in routines like hugetlb_dup_vma_private(). We can check
hugetlb_vma_lock->vma and only set to NULL if,

vma->(hugetlb_vma_lock)vma->vm_private_data->vma != vma

Got sidetracked chasing down another leak today. Will send a patch
implementing this idea soon.

Thanks for looking at this!
--
Mike Kravetz