Re: [PATCH] binfmt_elf: Take the mmap lock when walking the VMA list

["Eric W. Biederman" <ebiederm@xmission.com>]

Matthew Wilcox <willy@infradead.org> writes:

> On Mon, Jan 31, 2022 at 10:03:31AM -0600, Eric W. Biederman wrote:
>> "Matthew Wilcox (Oracle)" <willy@infradead.org> writes:
>> 
>> > I'm not sure if the VMA list can change under us, but dump_vma_snapshot()
>> > is very careful to take the mmap_lock in write mode.  We only need to
>> > take it in read mode here as we do not care if the size of the stack
>> > VMA changes underneath us.
>> >
>> > If it can be changed underneath us, this is a potential use-after-free
>> > for a multithreaded process which is dumping core.
>> 
>> The problem is not multi-threaded process so much as processes that
>> share their mm.
>
> I don't understand the difference.  I appreciate that another process can
> get read access to an mm through, eg, /proc, but how can another process
> (that isn't a thread of this process) modify the VMAs?

There are a couple of ways.

A classic way is a multi-threads process can call vfork, and the
mm_struct is shared with the child until exec is called.

A process can do this more deliberately by forking a child using
clone(CLONE_VM) and not including CLONE_THREAD.   Supporting this case
is a hold over from before CLONE_THREAD was supported in the kernel and
such processes were used to simulate threads.

The practical difference between a CLONE_THREAD thread and a
non-CLONE_THREAD process is that the signal handling is not shared.
Without sharing the signal handlers it does not make sense for a fatal
signal to kill the other process.

From the perspective of coredump generation it stops the execution of
all CLONE_THREAD threads that are going to be part of the coredump
and allows anyone else who shared the mm_struct to keep running.


It also happens that there are subsystems in the kernel that do things
like kthread_use_mm that can also be modifying the mm during a coredump.

Which is why we have dump_vma_snapshot.  Preventing the mm_struct and
the vmas from being modified during a coredump is not really practical.


>> I think rather than take a lock we should be using the snapshot captured
>> with dump_vma_snapshot.  Otherwise we have the very real chance that the
>> two get out of sync.  Which would result in a non-sense core file.
>> 
>> Probably that means that dump_vma_snapshot needs to call get_file on
>> vma->vm_file store it in core_vma_metadata.
>> 
>> Do you think you can fix it something like that?
>
> Uhh .. that seems like it needs a lot more understanding of binfmt_elf
> than I currently possess.  I'd rather spend my time working on folios
> than learning much more about binfmt_elf.  I was just trying to fix an
> assertion failure with the maple tree patches (we now assert that you're
> holding a lock when walking the list of VMAs).

Fair enough.  I will put it on my list of things to address.

Eric