On 1/11/2023 22:27, Rik van Riel wrote:
On Wed, 2023-11-01 at 14:36 +0800, Edward Adam Davis wrote:
When obtaining resv_map from vma, it is necessary to simultaneously
determine
the flag HPAGE_RESV_OWNER of vm_private_data.
Only when they are met simultaneously, resv_map is valid.
Reported-and-tested-by:
syzbot+6ada951e7c0f7bc8a71e@xxxxxxxxxxxxxxxxxxxxxxxxx
Fixes: bf4916922c60 ("hugetlbfs: extend hugetlb_vma_lock to private
VMAs")
Signed-off-by: Edward Adam Davis <eadavis@xxxxxx>
---
include/linux/hugetlb.h | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)
diff --git a/include/linux/hugetlb.h b/include/linux/hugetlb.h
index 47d25a5e1933..1a3ec1aee1a3 100644
--- a/include/linux/hugetlb.h
+++ b/include/linux/hugetlb.h
@@ -1265,9 +1265,11 @@ static inline bool __vma_shareable_lock(struct
vm_area_struct *vma)
return (vma->vm_flags & VM_MAYSHARE) && vma->vm_private_data;
}
+#define HPAGE_RESV_OWNER (1UL << 0)
static inline bool __vma_private_lock(struct vm_area_struct *vma)
{
- return (!(vma->vm_flags & VM_MAYSHARE)) && vma-
vm_private_data;
+ return (!(vma->vm_flags & VM_MAYSHARE)) && vma-
vm_private_data &&
+ ((unsigned long)vma->vm_private_data &
HPAGE_RESV_OWNER);
}
I am wondering whether this line should be:
return (!(vma->vm_flags & VM_MAYSHARE)) &&
((unsigned long)vma->vm_private_data & HPAGE_RESV_OWNER);
or even:
return (unsigned long)vma->vm_private_data & HPAGE_RESV_OWNER;
because mm/hugetlb.c has:
"We know private mapping must have HPAGE_RESV_OWNER set."
This could be cleaned up a bit by moving the HPAGE_RESV_OWNER
definition (and its friends) into hugetlb.h, as well as the
is_vma_resv_set() helper function.
Then __vma_private_lock() can just call is_vma_resv_set(),
and open coding a duplicate of the same code.
Not having duplicates of the code will make it much harder
to "miss a spot" with future changes.
I am still struggling to find a place where we might leave
HPAGE_RESV_OWNER behind on a pointer that is otherwise NULL,
but if your tests show this fixes the issue, I'm all for it :)
Like you said, vma->vm_private_data is cleared during fork().
But I saw following possible code path in child process:
hugetlb_wp()
unmap_ref_private() if alloc_hugetlb_folio fails
unmap_hugepage_range()
__unmap_hugepage_range()
set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED)
vm_private_data become 0x2 which is none NULL without HPAGE_RESV_OWNER
bit.
