Re: [PATCH] mm: khugepaged: skip huge page collapse for special files

[Song Liu]

> On Oct 27, 2021, at 12:52 PM, Yang Shi <shy828301@xxxxxxxxx> wrote:
> 
> The read-only THP for filesystems would collapse THP for file opened
> readonly and mapped with VM_EXEC, the intended usecase is to avoid TLB
> miss for large text segment.  But it doesn't restrict the file types so
> THP could be collapsed for non-regular file, for example, block device,
> if it is opened readonly and mapped with EXEC permission.  This may
> cause bugs, like [1] and [2].
> 
> This is definitely not intended usecase, so just collapsing THP for regular
> file in order to close the attack surface.
> 
> [1] https://lore.kernel.org/lkml/CACkBjsYwLYLRmX8GpsDpMthagWOjWWrNxqY6ZLNQVr6yx+f5vA@xxxxxxxxxxxxxx/
> [2] https://lore.kernel.org/linux-mm/000000000000c6a82505ce284e4c@xxxxxxxxxx/
> 
> Fixes: 99cb0dbd47a1 ("mm,thp: add read-only THP support for (non-shmem) FS")
> Reported-by: Hao Sun <sunhao.th@xxxxxxxxx>
> Reported-by: syzbot+aae069be1de40fb11825@xxxxxxxxxxxxxxxxxxxxxxxxx
> Cc: Hao Sun <sunhao.th@xxxxxxxxx>
> Cc: Matthew Wilcox <willy@xxxxxxxxxxxxx>
> Cc: Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx>
> Cc: Song Liu <songliubraving@xxxxxx>
> Cc: Andrea Righi <andrea.righi@xxxxxxxxxxxxx>
> Cc: <stable@xxxxxxxxxxxxxxx>
> Signed-off-by: Hugh Dickins <hughd@xxxxxxxxxx>
> Signed-off-by: Yang Shi <shy828301@xxxxxxxxx>
> ---
> The patch is basically based off the proposal from Hugh
> (https://lore.kernel.org/linux-mm/a07564a3-b2fc-9ffe-3ace-3f276075ea5c@xxxxxxxxxx/).
> It seems Hugh is too busy to prepare the patch for formal submission (I
> didn't hear from him by pinging him a couple of times on mailing list),
> so I prepared the patch and added his SOB.
> 
> mm/khugepaged.c | 17 ++++++++++-------
> 1 file changed, 10 insertions(+), 7 deletions(-)
> 
> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
> index 045cc579f724..e91b7271275e 100644
> --- a/mm/khugepaged.c
> +++ b/mm/khugepaged.c
> @@ -445,22 +445,25 @@ static bool hugepage_vma_check(struct vm_area_struct *vma,
> 	if (!transhuge_vma_enabled(vma, vm_flags))
> 		return false;
> 
> -	/* Enabled via shmem mount options or sysfs settings. */
> -	if (shmem_file(vma->vm_file) && shmem_huge_enabled(vma)) {
> +	if (vma->vm_file)
> 		return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
> 				HPAGE_PMD_NR);

Am I misreading this? If we return here for vma->vm_file, the following 
logic (shmem_file(), etc.) would be skipped, no? 

Thanks,
Song

> -	}
> +
> +	/* Enabled via shmem mount options or sysfs settings. */
> +	if (shmem_file(vma->vm_file))
> +		return shmem_huge_enabled(vma);
> 
> 	/* THP settings require madvise. */
> 	if (!(vm_flags & VM_HUGEPAGE) && !khugepaged_always())
> 		return false;
> 
> -	/* Read-only file mappings need to be aligned for THP to work. */
> +	/* Only regular file is valid */
> 	if (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && vma->vm_file &&
> -	    !inode_is_open_for_write(vma->vm_file->f_inode) &&
> 	    (vm_flags & VM_EXEC)) {
> -		return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
> -				HPAGE_PMD_NR);
> +		struct inode *inode = vma->vm_file->f_inode;
> +
> +		return !inode_is_open_for_write(inode) &&
> +			S_ISREG(inode->i_mode);
> 	}
> 
> 	if (!vma->anon_vma || vma->vm_ops)
> -- 
> 2.26.2
>