Re: [PATCH 4/6] hugetlbfs: implement memfd sealing

[David Herrmann]

Hi

On Fri, Nov 3, 2017 at 6:12 PM, Mike Kravetz <mike.kravetz@xxxxxxxxxx> wrote:
> On 11/03/2017 10:03 AM, David Herrmann wrote:
>> Hi
>>
>> On Tue, Oct 31, 2017 at 7:40 PM, Marc-Andrà Lureau
>> <marcandre.lureau@xxxxxxxxxx> wrote:
>>> Implements memfd sealing, similar to shmem:
>>> - WRITE: deny fallocate(PUNCH_HOLE). mmap() write is denied in
>>>   memfd_add_seals(). write() doesn't exist for hugetlbfs.
>>> - SHRINK: added similar check as shmem_setattr()
>>> - GROW: added similar check as shmem_setattr() & shmem_fallocate()
>>>
>>> Except write() operation that doesn't exist with hugetlbfs, that
>>> should make sealing as close as it can be to shmem support.
>>
>> SEAL, SHRINK, and GROW look fine to me.
>>
>> Regarding WRITE
>
> The commit message may not be clear.  However, hugetlbfs does not support
> the write system call (or aio).  The only way to modify contents of a
> hugetlbfs file is via mmap or hole punch/truncate.  So, we do not really
> need to worry about those special (a)io cases for hugetlbfs.

This is not about the write(2) syscall. Please consider this scenario
about shmem:

You create a memfd via memfd_create() and map it writable. You now
call another kernel syscall that takes as input _any mapped page
range_. You pass your mapped memfd-addresses to it. Those syscalls
tend to use get_user_pages() to pin arbitrary user-mapped pages, as
such this also affects shmem. In this case, those pages might stay
mapped even if you munmap() your memfd!

One example of this is using AIO-read() on any other file that
supports it, passing your mapped memfd as buffer to _read into_. The
operations supported on the memfd are irrelevant here.
The selftests contain a FUSE-based test for this, since FUSE allows
user-space to GUP pages for an arbitrary amount of time.

The original fix for this is:

    commit 05f65b5c70909ef686f865f0a85406d74d75f70f
    Author: David Herrmann <dh.herrmann@xxxxxxxxx>
    Date:   Fri Aug 8 14:25:36 2014 -0700

        shm: wait for pins to be released when sealing

Please have a look at this. Your patches use shmem_add_seals() almost
unchanged, and as such you call into shmem_wait_for_pins() on
hugetlbfs. I would really like to see an explicit ACK that this works
on hugetlbfs.

Thanks
David