[I moved your reply to avoid top-posting]

On 12/28/20 6:03 PM, Brian Geffon wrote:

On Mon, Oct 12, 2020 at 6:34 PM Dmitry Safonov [off-list ref] wrote:

quoted

As kernel expect to see only one of such mappings, any further
operations on the VMA-copy may be unexpected by the kernel.
Maybe it's being on the safe side, but there doesn't seem to be any
expected use-case for this, so restrict it now.

Fixes: commit e346b3813067 ("mm/mremap: add MREMAP_DONTUNMAP to mremap()")
Signed-off-by: Dmitry Safonov <redacted>

I don't think this situation can ever happen MREMAP_DONTUNMAP is
already restricted to anonymous mappings (defined as not having
vm_ops) and vma_to_resize checks that the mapping is anonymous before
move_vma is called.

I've looked again now, I think it is possible. One can call
MREMAP_DONTUNMAP without MREMAP_FIXED and without resizing. So that the
old VMA is copied at some free address.

The calltrace would be: mremap()=>move_vma()
[under if (flags & MREMAP_MAYMOVE)].

On the other side I agree with you that the fix could have been better
if I realized the semantics that MREMAP_DONTUNMAP should only work with
anonymous mappings.

Probably, a better fix would be to move
:       if (flags & MREMAP_DONTUNMAP && (!vma_is_anonymous(vma) ||
:                       vma->vm_flags & VM_SHARED))
:               return ERR_PTR(-EINVAL);

from vma_to_resize() into the mremap() syscall directly.
What do you think?

-- 
          Dima