Hi, Axel,
Sorry to read this late.
On Fri, Jun 03, 2022 at 01:57:41PM -0700, Axel Rasmussen wrote:
> When fallocate() is used on a shmem file, the pages we allocate can end
> up with !PageUptodate.
>
> Since UFFDIO_CONTINUE tries to find the existing page the user wants to
> map with SGP_READ, we would fail to find such a page, since
> shmem_getpage_gfp returns with a "NULL" pagep for SGP_READ if it
> discovers !PageUptodate. As a result, UFFDIO_CONTINUE returns -EFAULT,
> as it would do if the page wasn't found in the page cache at all.
>
> This isn't the intended behavior. UFFDIO_CONTINUE is just trying to find
> if a page exists, and doesn't care whether it still needs to be cleared
> or not. So, instead of SGP_READ, pass in SGP_NOALLOC. This is the same,
> except for one critical difference: in the !PageUptodate case,
> SGP_NOALLOC will clear the page and then return it. With this change,
> UFFDIO_CONTINUE works properly (succeeds) on a shmem file which has been
> fallocated, but otherwise not modified.
>
> Fixes: 153132571f02 ("userfaultfd/shmem: support UFFDIO_CONTINUE for shmem")
> Cc: stable@xxxxxxxxxxxxxxx
> Signed-off-by: Axel Rasmussen <axelrasmussen@xxxxxxxxxx>
> ---
> mm/userfaultfd.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/mm/userfaultfd.c b/mm/userfaultfd.c
> index 4f4892a5f767..c156f7f5b854 100644
> --- a/mm/userfaultfd.c
> +++ b/mm/userfaultfd.c
> @@ -246,7 +246,7 @@ static int mcontinue_atomic_pte(struct mm_struct *dst_mm,
> struct page *page;
> int ret;
>
> - ret = shmem_getpage(inode, pgoff, &page, SGP_READ);
> + ret = shmem_getpage(inode, pgoff, &page, SGP_NOALLOC);
> if (ret)
> goto out;
> if (!page) {
It all looks sane if the page is !uptodate as you described. Though I've a
question on what'll happen if the page is actually missing rather than just
!PageUptodate().
My reading is previously it'll keep returning 0 on shmem_getpage_gfp() for
both cases, but now for the missing page shmem_getpage_gfp() will return
-ENOENT instead.
This reminded me on whether this will errornously let __mcopy_atomic() go
into the special path to copy the page without mmap lock, please see this
commit:
b6ebaedb4cb1 ("userfaultfd: avoid mmap_sem read recursion in mcopy_atomic", 2015-09-04)
Would that be a problem? Or could I read it wrong?
This also reminded me that whether we'd better need some protection in the
-ENOENT handling in __mcopy_atomic() to be always safe.
Thanks,
--
Peter Xu
