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;
MIME-Version: 1.0
References: <20220902034135.2853973-1-ovt@google.com> <20220905083914.msdgd575tblq4syj@wittgenstein>
In-Reply-To: <20220905083914.msdgd575tblq4syj@wittgenstein>
From:   Oleksandr Tymoshenko <ovt@google.com>
Date:   Mon, 5 Sep 2022 15:53:54 -0700
Message-ID: <CACGj0CgTVVAhBrL7DsP_fL7R9EhZUwu+MkQ1_GpPEtHQW4ojOQ@mail.gmail.com>
Subject: Re: [PATCH] seccomp: fix refcounter leak if fork/clone is terminated
To:     Christian Brauner <brauner@kernel.org>
Cc:     linux-kernel@vger.kernel.org, Kees Cook <keescook@chromium.org>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

Hi Christian,

The patch in the seccomp tree, adapted to 5.10 branch, fixed the
memory leak in my reproducer.
Thanks for working on this, I should have checked the seccomp tree first :)
Please disregard the patch in my submission.

Thank you

On Mon, Sep 5, 2022 at 1:39 AM Christian Brauner <brauner@kernel.org> wrote:
>
> On Fri, Sep 02, 2022 at 03:41:35AM +0000, Oleksandr Tymoshenko wrote:
> > release_task, where the seccomp's filter refcounter is released, is not
> > called for the case when the fork/clone is terminated midway by a
> > signal. This leaves an extra reference that prevents filter from being
> > destroyed even after all processes using it exit leading to a BPF JIT
> > memory leak. Dereference the refcounter in the failure path of the
> > copy_process function.
> >
> > Fixes: 3a15fb6ed92c ("seccomp: release filter after task is fully dead")
> > Cc: Christian Brauner <brauner@kernel.org>
> > Cc: stable@vger.kernel.org
> > Signed-off-by: Oleksandr Tymoshenko <ovt@google.com>
> > ---
>
> Hey Oleksandr,
>
> Thanks for the patch! I'm really puzzled as to why we never noticed this
> and I'm trying to re-architect how this happend. But in any case,
> there's a patch in the seccomp tree that fixes this:
>
> https://git.kernel.org/pub/scm/linux/kernel/git/kees/linux.git/commit/?id=6d17452707ca
>
> which is slighly different from your approach in that it moves
> copy_seccomp() after the point of no return. Let us know if you see any
> issues with this!
>
> Christian