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

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <20180928213557.GD18045@cisco.lan>
Date:   Fri, 28 Sep 2018 15:35:57 -0600
From:   Tycho Andersen <tycho@...ho.ws>
To:     Jann Horn <jannh@...gle.com>
Cc:     Kees Cook <keescook@...omium.org>,
        kernel list <linux-kernel@...r.kernel.org>,
        Andy Lutomirski <luto@...capital.net>
Subject: Re: [PATCH 3/3] seccomp: introduce read protection for struct seccomp

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@...ho.ws> 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@...ho.ws> 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);
 

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

Tycho