[syzbot] KASAN: use-after-free Read in rxrpc_lookup_local

[David Howells]

Hillf Danton <hdanton at sina.com> wrote:

> > Hmmm...  That can't be the whole explanation.  As you say, the hlist_del is
> > done under the mutex in rxrpc_destroy_local() - the same as the
> > hlist_add/hlist_replace and the search in rxrpc_lookup_local().
> 
> The uaf is simple and due to local ep freed without being deleted from
> rxnet->local_endpoints while the list walker in rxrpc_lookup_local() stumbles
> on it. rxrpc_destroy_local() is irrelevant as per the syzbot report.

No, that's not the explanation.  We can't get that far unless the I/O thread
got started - kthread_run() has to wait for the task_struct pointer to become
available - and the I/O thread should remove the link before it exits and
before the rxrpc_local object is deallocated.

I've tracked the problem down.  It's due to kthread(), specifically this bit:

	preempt_disable();
	complete(done);
	schedule_preempt_disabled();
	preempt_enable();

	ret = -EINTR;
	if (!test_bit(KTHREAD_SHOULD_STOP, &self->flags)) {
		cgroup_kthread_ready();
		__kthread_parkme(self);
		ret = threadfn(data);
	}

So the complete() is done before we've decided if we're going to call
threadfn() or return an error.  This permits kthread_run() to resume before
we've checked KTHREAD_SHOULD_STOP - thus if kthread_stop() is called quickly
enough by the rxrpc socket being released, kthread() skips calling threadfn(),
but kthread_run() returns success.

The fact that the thread didn't start doesn't get seen until kthread_stop() is
called.

I think the above code needs rearranging slightly.  KTHREAD_SHOULD_STOP should
be tested and, if needed EINTR set, *before* complete(done).  That way
kthread_run() will return -EINTR in this case.

David