Received-SPF: pass (google.com: domain of linux-bluetooth-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <20210712203813.Bluez.v1.1.Id7aa1152377161d17b442bf258773d9b6c624ca3@changeid>
In-Reply-To: <20210712203813.Bluez.v1.1.Id7aa1152377161d17b442bf258773d9b6c624ca3@changeid>
From:   Luiz Augusto von Dentz <luiz.dentz@gmail.com>
Date:   Mon, 12 Jul 2021 13:40:52 -0700
Message-ID: <CABBYNZJ-Oc=H7icDxBVk8iNAfcBjZDL_CgKq7r8k82iwT66pVQ@mail.gmail.com>
Subject: Re: [Bluez PATCH v1] a2dp: Fix crash in channel_free while waiting
 cmd resp
To:     Howard Chung <howardchung@google.com>
Cc:     "linux-bluetooth@vger.kernel.org" <linux-bluetooth@vger.kernel.org>,
        Yun-Hao Chung <howardchung@chromium.org>,
        Archie Pusaka <apusaka@chromium.org>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

Hi Howard,

On Mon, Jul 12, 2021 at 5:39 AM Howard Chung <howardchung@google.com> wrote:
>
> From: Yun-Hao Chung <howardchung@chromium.org>
>
> When channel_free is called and we are waiting for a command response
> from the peer, bluez NULL the setup->session but would not free its
> setup_cb. Since setup_cb holds a ref of setup, the setup wouldn't be
> freed and if service_removed is called after channel_free, a2dp_cancel
> tries to abort the ongoing avdtp commands, which accesses the
> setup->session and triggers a crash.
>
> This change finalizes all avdtp commands before assigning setup->session
> to NULL in channel_free.
>
> Crash stack trace:
> 0x000059f01943e688      (bluetoothd -avdtp.c:3690)
> avdtp_abort
> 0x000059f01943928a      (bluetoothd -a2dp.c:3069)
> a2dp_cancel
> 0x000059f0194377fa      (bluetoothd -sink.c:324)
> sink_unregister
> 0x000059f01948715a      (bluetoothd -service.c:177)
> service_remove
> 0x000059f01948d77c      (bluetoothd -device.c:5346)
> device_remove
> 0x000059f019476d14      (bluetoothd -adapter.c:7202)
> adapter_remove
> 0x000059f019476c3e      (bluetoothd -adapter.c:10827)
> adapter_cleanup
> 0x000059f01949d8d7      (bluetoothd -main.c:1114)               main
> 0x0000787b36185d74      (libc.so.6 -libc-start.c:308)
> __libc_start_main
> 0x000059f019433e39      (bluetoothd + 0x00026e39)               _start
> 0x00007fff2d2c0127
>
> Reviewed-by: Archie Pusaka <apusaka@chromium.org>
> ---
> There are two other options to fix this crash.
> 1. add a NULL check in a2dp_cancel before calling avdtp_abort.
> 2. call setup_cb_free to every setup_cb in setup->cb in channel_free.
>
> Since each setup_cb needs setup->session, I think there is no need to
> keep the setup_cb after assigning setup->session to NULL. So the first
> option is not ideal. If the second option is adopted, there would be
> some time that sink/source->connect_id/disconnect_id is not zero, but
> there is no corresponding setup_cb.
>
> Test steps:
> Reproduce the crash with the following steps. Verify the crash is
> no longer observed after this change.
> 1. ignore AVDTP_SET_CONFIGURATION resp by modifying avdtp.c
> 2. turn on a paired headset
> 3. check the bluetooth.log, while bluez is waiting for
>    AVDTP_SET_CONFIGURATION resp, stop bluetoothd immediately.
>    This will trigger:
>    session_cb (I/O error) -> connection_lost
>    -> avdtp_set_state (AVDTP_SESSION_STATE_DISCONNECTED)
>    -> avdtp_state_cb -> channel_remove -> channel_free
>    then:
>    adapter_cleanup -> adapter_remove -> device_remove -> service_remove
>    -> a2dp_sink_remove -> sink_unregister -> sink_free -> a2dp_cancel
> 4. check if bluetoothd crash
> The above steps can trigger the crash 100%.
>
>  profiles/audio/a2dp.c | 7 ++++++-
>  1 file changed, 6 insertions(+), 1 deletion(-)
>
> diff --git a/profiles/audio/a2dp.c b/profiles/audio/a2dp.c
> index d31ed845cbe7..f201b98c79d0 100644
> --- a/profiles/audio/a2dp.c
> +++ b/profiles/audio/a2dp.c
> @@ -1540,9 +1540,14 @@ static void channel_free(void *data)
>         setup = find_setup_by_session(chan->session);
>         if (setup) {
>                 setup->chan = NULL;
> +               /* Finalize pending commands before we NULL setup->session */
> +               finalize_setup_errno(setup, -ENOTCONN, finalize_discover,
> +                                                       finalize_select,
> +                                                       finalize_config,
> +                                                       finalize_resume,
> +                                                       finalize_suspend, NULL);

While the analysis seems correct I wonder if wouldn't be a better idea
to have a finalize_all so we don't have to iterate the callback list
multiple times, other than that the patch looks good and I would
prioritize it to have it included asap.

>                 avdtp_unref(setup->session);
>                 setup->session = NULL;
> -               finalize_setup_errno(setup, -ENOTCONN, NULL);
>         }
>
>         g_free(chan);
> --
> 2.32.0.93.g670b81a890-goog
>


-- 
Luiz Augusto von Dentz