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

[Nadav Amit <nadav.amit@gmail.com>]

Hello Mike,

Regarding your (old) patch:

> On May 23, 2018, at 12:42 AM, Mike Rapoport <rppt@linux.vnet.ibm.com> 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.

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.

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.

Thanks,
Nadav