<maarten.lankhorst@linux.intel.com>, Maxime Ripard
<mripard@kernel.org>, Thomas Zimmermann <tzimmermann@suse.de>

filp->pid is supposed to be a refcounted pointer; however, before this
patch, drm_file_update_pid() only increments the refcount of a struct
pid after storing a pointer to it in filp->pid and dropping the
dev->filelist_mutex, making the following race possible:

process A               process B
=========               =========
                        begin drm_file_update_pid
                        mutex_lock(&dev->filelist_mutex)
                        rcu_replace_pointer(filp->pid, <pid B>, 1)
                        mutex_unlock(&dev->filelist_mutex)
begin drm_file_update_pid
mutex_lock(&dev->filelist_mutex)
rcu_replace_pointer(filp->pid, <pid A>, 1)
mutex_unlock(&dev->filelist_mutex)
get_pid(<pid A>)
synchronize_rcu()
put_pid(<pid B>)   *** pid B reaches refcount 0 and is freed here ***
                        get_pid(<pid B>)   *** UAF ***
                        synchronize_rcu()
                        put_pid(<pid A>)

As far as I know, this race can only occur with CONFIG_PREEMPT_RCU=y
because it requires RCU to detect a quiescent state in code that is not
explicitly calling into the scheduler.

This race leads to use-after-free of a "struct pid".
It is probably somewhat hard to hit because process A has to pass
through a synchronize_rcu() operation while process B is between
mutex_unlock() and get_pid().

Fix it by ensuring that by the time a pointer to the current task's pid
is stored in the file, an extra reference to the pid has been taken.

This fix also removes the condition for synchronize_rcu(); I think
that optimization is unnecessary complexity, since in that case we
would usually have bailed out on the lockless check above.

Fixes: 1c7a387f ("drm: Update file owner during use")
Cc: <stable@vger.kernel.org>
Signed-off-by: Jann Horn <jannh@google.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>