Re: [PATCH] userfaultfd: prevent non-cooperative events vs mcopy_atomic races

[Mike Rapoport]

Hello Nadav,

On Thu, Dec 03, 2020 at 11:57:46AM -0800, Nadav Amit wrote:
> Hello Mike,
> 
> Regarding your (old) patch:
> 
> > On May 23, 2018, at 12:42 AM, Mike Rapoport <rppt@xxxxxxxxxxxxxxxxxx> wrote:
> > 
> > If a process monitored with userfaultfd changes it's memory mappings or
> > forks() at the same time as uffd monitor fills the process memory with
> > UFFDIO_COPY, the actual creation of page table entries and copying of the
> > data in mcopy_atomic may happen either before of after the memory mapping
> > modifications and there is no way for the uffd monitor to maintain
> > consistent view of the process memory layout.
> > 
> > For instance, let's consider fork() running in parallel with
> > userfaultfd_copy():
> > 
> > process        		         |	uffd monitor
> > ---------------------------------+------------------------------
> > fork()        		         | userfaultfd_copy()
> > ...        		         | ...
> >    dup_mmap()        	         |     down_read(mmap_sem)
> >    down_write(mmap_sem)         |     /* create PTEs, copy data */
> >        dup_uffd()               |     up_read(mmap_sem)
> >        copy_page_range()        |
> >        up_write(mmap_sem)       |
> >        dup_uffd_complete()      |
> >            /* notify monitor */ |
> > 
> > If the userfaultfd_copy() takes the mmap_sem first, the new page(s) will be
> > present by the time copy_page_range() is called and they will appear in the
> > child's memory mappings. However, if the fork() is the first to take the
> > mmap_sem, the new pages won't be mapped in the child's address space.
> > 
> > Since userfaultfd monitor has no way to determine what was the order, let's
> > disallow userfaultfd_copy in parallel with the non-cooperative events. In
> > such case we return -EAGAIN and the uffd monitor can understand that
> > userfaultfd_copy() clashed with a non-cooperative event and take an
> > appropriate action.
> 
> I am struggling to understand this patch and would appreciate your
> assistance.
 
The tl;dr version is that without this commit we had failing fork tests
in CRIU [1] :)

> Specifically, I have two questions:
> 
> 1. How can memory corruption occur? If the page is already mapped and the
> handler “mistakenly" calls userfaultfd_copy(), wouldn't mcopy_atomic_pte()
> return -EEXIST once it sees the PTE already exists? In such case, I would
> presume that the handler should be able to recover gracefully by waking the
> faulting thread.
 
The issue we had was when fork() in a monitored process happened
concurrently with "background copy" of pages into the process address
space during a post-copy process migration.

The userspace has no way to tell who won the race for mmap_lock and
depending on that we can have two different cases:

* fork() took the mmap_lock, pages in the parent are still empty, so
they will be empty in the child

* userfaultfd_copy() won the lock, there is data in the parent and the
child's inherits these mappings

The uffd monotor should *know* what is the state of child's memory and
without this patch it could only guess.

> 2. How is memory ordering supposed to work here? IIUC, mmap_changing is not
> protected by any lock and there are no memory barriers that are associated
> with the assignment. Indeed, the code calls WRITE_ONCE()/READ_ONCE(), but
> AFAIK this does not guarantee ordering with non-volatile reads/writes.

There is also mmap_lock involved, so I don't see how copy can start in
parallel with fork processing. Fork sets mmap_chaning to true while
holding mmap_lock, so copy cannot start in parallel. When mmap_lock is
realeased, mmap_chaning remains true until fork event is pushed to
userspace and when this is done there is no issue with
userfaultfd_copy.

Maybe I am missing something...

[1] https://github.com/checkpoint-restore/criu/blob/criu-dev/test/zdtm/transition/fork.c

> Thanks,
> Nadav

-- 
Sincerely yours,
Mike.