On 7/13/21 8:24 AM, Joao Martins wrote:
> commit 82e5d378b0e47 ("mm/hugetlb: refactor subpage recording")
> refactored the count of subpages but missed an edge case when @vaddr is
> not aligned to PAGE_SIZE e.g. when close to vma->vm_end. It would then
> errousnly set @refs to 0 and record_subpages_vmas() wouldn't set the
> @pages array element to its value, consequently causing the reported
> null-deref by syzbot.
>
> Fix it by aligning down @vaddr by PAGE_SIZE in @refs calculation.
Thanks for finding and fixing!
>
> Reported-by: syzbot+a3fcd59df1b372066f5a@xxxxxxxxxxxxxxxxxxxxxxxxx
> Fixes: 82e5d378b0e47 ("mm/hugetlb: refactor subpage recording")
> Signed-off-by: Joao Martins <joao.m.martins@xxxxxxxxxx>
> ---
> An alternate approach is to have record_subpages_vmas() iterate while
> addr < vm_end and renaming @refs to nr_pages (which would limit how many
> pages we should store). But I felt that this approach would be slightly
> more convoluted?
I prefer the approach you have taken in this patch.
>
> Side-Note: I could add a WARN_ON_ONCE(!refs) and/or create an
> helper like vma_pages() but with a ulong addr argument e.g.
> vma_pages_from(vma, vaddr).
IIUC, the only way refs could be zero is if there was error in
caluclations within this routine. Correct?
IMO, the only reason to add a warning would be if there are any assumptions
based on things outside this routine which could cause refs to be zero.
> The syzbot repro no longer reproduces after this patch. Additionally, ran
> the libhugetlbfs tests (which were passing without this), gup_test and an
> extra gup_test extension that take an offset to exercise gup() starting
> address not being page aligned.
> ---
> mm/hugetlb.c | 5 +++--
> 1 file changed, 3 insertions(+), 2 deletions(-)
>
> diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> index 924553aa8f78..dfc940d5221d 100644
> --- a/mm/hugetlb.c
> +++ b/mm/hugetlb.c
> @@ -5440,8 +5440,9 @@ long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
> continue;
> }
>
> - refs = min3(pages_per_huge_page(h) - pfn_offset,
> - (vma->vm_end - vaddr) >> PAGE_SHIFT, remainder);
> + /* vaddr may not be aligned to PAGE_SIZE */
> + refs = min3(pages_per_huge_page(h) - pfn_offset, remainder,
> + (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT);
>
> if (pages || vmas)
> record_subpages_vmas(mem_map_offset(page, pfn_offset),
>
Reviewed-by: Mike Kravetz <mike.kravetz@xxxxxxxxxx>
--
Mike Kravetz
