Re: [PATCH 1/2] seccomp: notify user trap about unused filter

From: Christian Brauner <christian.brauner@ubuntu.com>
To: Kees Cook <keescook@chromium.org>
Cc: linux-kernel@vger.kernel.org, Andy Lutomirski <luto@kernel.org>,
	Tycho Andersen <tycho@tycho.ws>,
	Matt Denton <mpdenton@google.com>,
	Sargun Dhillon <sargun@sargun.me>, Jann Horn <jannh@google.com>,
	Chris Palmer <palmer@google.com>,
	Aleksa Sarai <cyphar@cyphar.com>,
	Robert Sesek <rsesek@google.com>,
	Jeffrey Vander Stoep <jeffv@google.com>,
	Linux Containers <containers@lists.linux-foundation.org>
Subject: Re: [PATCH 1/2] seccomp: notify user trap about unused filter
Date: Thu, 28 May 2020 01:16:46 +0200	[thread overview]
Message-ID: <20200527231646.4v743erjpzh6qe5f@wittgenstein> (raw)
In-Reply-To: <20200527224501.jddwcmvtvjtjsmsx@wittgenstein>

On Thu, May 28, 2020 at 12:45:02AM +0200, Christian Brauner wrote:
> On Wed, May 27, 2020 at 03:37:58PM -0700, Kees Cook wrote:
> > On Thu, May 28, 2020 at 12:05:32AM +0200, Christian Brauner wrote:
> > > The main question also is, is there precedence where the kernel just
> > > closes the file descriptor for userspace behind it's back? I'm not sure
> > > I've heard of this before. That's not how that works afaict; it's also
> > > not how we do pidfds. We don't just close the fd when the task
> > > associated with it goes away, we notify and then userspace can close.
> > 
> > But there's a mapping between pidfd and task struct that is separate
> > from task struct itself, yes? I.e. keeping a pidfd open doesn't pin
> > struct task in memory forever, right?
> 
> No, but that's an implementation detail and we discussed that. It pins
> struct pid instead of task_struct. Once the process is fully gone you
> just get ESRCH.
> For example, fds to /proc/<pid>/<tid>/ fds aren't just closed once the
> task has gone away, userspace will just get ESRCH when it tries to open
> files under there but the fd remains valid until close() is called.
> 
> In addition, of all the anon inode fds, none of them have the "close the
> file behind userspace back" behavior: io_uring, signalfd, timerfd, btf,
> perf_event, bpf-prog, bpf-link, bpf-map, pidfd, userfaultfd, fanotify,
> inotify, eventpoll, fscontext, eventfd. These are just core kernel ones.
> I'm pretty sure that it'd be very odd behavior if we did that. I'd
> rather just notify userspace and leave the close to them. But maybe I'm
> missing something.

I'm also starting to think this isn't even possible or currently doable
safely.
The fdtable in the kernel would end up with a dangling pointer, I would
think. Unless you backtrack all fds that still have a reference into the
fdtable and refer to that file and close them all in the kernel which I
don't think is possible and also sounds very dodgy. This also really
seems like we would be breaking a major contract, namely that fds stay
valid until userspace calls close, execve(), or exits.

Christian