Re: [PATCH 3/3] seccomp: introduce read protection for struct seccomp

[Jann Horn]

On Fri, Sep 28, 2018 at 11:36 PM Tycho Andersen <tycho@xxxxxxxx> wrote:
> On Fri, Sep 28, 2018 at 11:10:48PM +0200, Jann Horn wrote:
> > On Fri, Sep 28, 2018 at 10:56 PM Tycho Andersen <tycho@xxxxxxxx> wrote:
> > >
> > > On Fri, Sep 28, 2018 at 10:33:34PM +0200, Jann Horn wrote:
> > > > On Fri, Sep 28, 2018 at 5:47 PM Tycho Andersen <tycho@xxxxxxxx> wrote:
> > > > > As Jann pointed out, there is a race between SECCOMP_FILTER_FLAG_TSYNC and
> > > > > the ptrace code that can inspect a filter of another process. Let's
> > > > > introduce read locking into the two ptrace accesses so that we don't race.
> > > >
> > > > Hmm. Is that true? The ptrace code uses get_nth_filter(), which holds
> > > > the siglock while grabbing the seccomp filter and bumping its
> > > > refcount. And TSYNC happens from seccomp_set_mode_filter(), which
> > > > takes the siglock. So this looks okay to me?
> > >
> > > Oh, yes, you're right. So I guess we should just change the comment to
> > > say we're using siglock to represent the read lock.
> >
> > Hmm... actually, looking at this closer, I think you only need the
> > siglock for writing. As far as I can tell, any read (no matter if
> > current or non-current) can just use READ_ONCE(), because once a
> > seccomp filter is in a task's seccomp filter chain, it can't be freed
> > until the task reaches free_task() and calls put_seccomp_filter() from
> > there. And if the task whose seccomp filter you're trying to read can
> > reach free_task(), you have bigger problems.
>
> Ok; looks like get_nth_filter() took the siglock anyway. Since we get
> the filters in these two functions in get_nth_filter(), I think it's
> enough just to just,
>
> diff --git a/kernel/seccomp.c b/kernel/seccomp.c
> index f65d47650ac1..79d833ed4c34 100644
> --- a/kernel/seccomp.c
> +++ b/kernel/seccomp.c
> @@ -1001,7 +1001,7 @@ static struct seccomp_filter *get_nth_filter(struct task_struct *task,
>                 return ERR_PTR(-EINVAL);
>         }
>
> -       orig = task->seccomp.filter;
> +       orig = READ_ONCE(task->seccomp.filter);
>         __get_seccomp_filter(orig);
>         spin_unlock_irq(&task->sighand->siglock);

Huh? Now you're holding the siglock *and* you're using READ_ONCE()?
I'm not sure what you're trying to accomplish here.

> since once it's returned from get_nth_filter() we don't need to worry
> about multiple accesses?
>
> Tycho