Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
From:   Kuniyuki Iwashima <kuniyu@amazon.com>
To:     <keescook@chromium.org>
CC:     <ayudutta@amazon.com>, <brauner@kernel.org>, <kuni1840@gmail.com>,
        <kuniyu@amazon.com>, <linux-kernel@vger.kernel.org>,
        <luto@amacapital.net>,
        <syzbot+ab17848fe269b573eb71@syzkaller.appspotmail.com>,
        <wad@chromium.org>
Subject: Re: [PATCH v1] seccomp: Release filter when copy_process() fails.
Date:   Mon, 22 Aug 2022 14:49:35 -0700
Message-ID: <20220822214935.29842-1-kuniyu@amazon.com>
In-Reply-To: <202208221414.A0E13E7@keescook>
References: <202208221414.A0E13E7@keescook>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Content-Type: text/plain
Precedence: bulk

From:   Kees Cook <keescook@chromium.org>
Date:   Mon, 22 Aug 2022 14:16:03 -0700
> On Mon, Aug 22, 2022 at 01:44:36PM -0700, Kuniyuki Iwashima wrote:
> > Our syzbot instance reported memory leaks in do_seccomp() [0], similar
> > to the report [1].  It shows that we miss freeing struct seccomp_filter
> > and some objects included in it.
> > 
> > We can reproduce the issue with the program below [2] which calls one
> > seccomp() and two clone() syscalls.
> > 
> > The first clone()d child exits earlier than its parent and sends a
> > signal to kill it during the second clone(), more precisely before the
> > fatal_signal_pending() test in copy_process().  When the parent receives
> > the signal, it has to destroy the embryonic process and return -EINTR to
> > user space.  In the failure path, we have to call seccomp_filter_release()
> > to decrement the filter's ref count.
> > 
> > Initially, we called it in free_task() called from the failure path, but
> > the commit 3a15fb6ed92c ("seccomp: release filter after task is fully
> > dead") moved it to release_task() to notify user space as early as possible
> > that the filter is no longer used.
> > 
> > To keep the change, let's call seccomp_filter_release() in copy_process()
> > and add a WARN_ON_ONCE() in free_task() for future debugging.
> 
> Thanks for tracking this down! I think I'd prefer to avoid changing the
> semantics around the existing seccomp refcount lifetime, so what about
> just moving copy_seccomp() below the last possible error path?

Actually, I also thought of it but avoid it because it means we move the
signal check relatively earlier than before, so would-be-killed processes
could consume more resouces.

What do you think about this?

> 
> 
> diff --git a/kernel/fork.c b/kernel/fork.c
> index 90c85b17bf69..e7f4e7f1e01e 100644
> --- a/kernel/fork.c
> +++ b/kernel/fork.c
> @@ -2409,12 +2409,6 @@ static __latent_entropy struct task_struct *copy_process(
>  
>  	spin_lock(&current->sighand->siglock);
>  
> -	/*
> -	 * Copy seccomp details explicitly here, in case they were changed
> -	 * before holding sighand lock.
> -	 */
> -	copy_seccomp(p);
> -
>  	rv_task_fork(p);
>  
>  	rseq_fork(p, clone_flags);
> @@ -2431,6 +2425,14 @@ static __latent_entropy struct task_struct *copy_process(
>  		goto bad_fork_cancel_cgroup;
>  	}
>  
> +	/* No more failures paths after this point. */
> +
> +	/*
> +	 * Copy seccomp details explicitly here, in case they were changed
> +	 * before holding sighand lock.
> +	 */
> +	copy_seccomp(p);
> +
>  	init_task_pid_links(p);
>  	if (likely(p->pid)) {
>  		ptrace_init_task(p, (clone_flags & CLONE_PTRACE) || trace);
> 
> 
> Totally untested, but I think it would fix this?
> 
> -Kees
> 
> -- 
> Kees Cook