Received-SPF: pass (google.com: domain of linux-bluetooth-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: <20220622082716.478486-1-lee.jones@linaro.org>
In-Reply-To: <20220622082716.478486-1-lee.jones@linaro.org>
From:   Eric Dumazet <edumazet@google.com>
Date:   Mon, 27 Jun 2022 16:41:23 +0200
Message-ID: <CANn89iK-uFP6Swgc0ZeEC38UsuywJ3wbybSNouH202Wa7X7Tzg@mail.gmail.com>
Subject: Re: [RESEND 1/1] Bluetooth: Use chan_list_lock to protect the whole
 put/destroy invokation
To:     Lee Jones <lee.jones@linaro.org>
Cc:     LKML <linux-kernel@vger.kernel.org>, stable@kernel.org,
        Marcel Holtmann <marcel@holtmann.org>,
        Johan Hedberg <johan.hedberg@gmail.com>,
        Luiz Augusto von Dentz <luiz.dentz@gmail.com>,
        "David S. Miller" <davem@davemloft.net>,
        Jakub Kicinski <kuba@kernel.org>,
        Paolo Abeni <pabeni@redhat.com>,
        linux-bluetooth@vger.kernel.org, netdev <netdev@vger.kernel.org>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Wed, Jun 22, 2022 at 10:27 AM Lee Jones <lee.jones@linaro.org> wrote:
>
> This change prevents a use-after-free caused by one of the worker
> threads starting up (see below) *after* the final channel reference
> has been put() during sock_close() but *before* the references to the
> channel have been destroyed.
>
>   refcount_t: increment on 0; use-after-free.
>   BUG: KASAN: use-after-free in refcount_dec_and_test+0x20/0xd0
>   Read of size 4 at addr ffffffc114f5bf18 by task kworker/u17:14/705
>
>   CPU: 4 PID: 705 Comm: kworker/u17:14 Tainted: G S      W       4.14.234-00003-g1fb6d0bd49a4-dirty #28
>   Hardware name: Qualcomm Technologies, Inc. SM8150 V2 PM8150 Google Inc. MSM sm8150 Flame DVT (DT)
>   Workqueue: hci0 hci_rx_work
>   Call trace:
>    dump_backtrace+0x0/0x378
>    show_stack+0x20/0x2c
>    dump_stack+0x124/0x148
>    print_address_description+0x80/0x2e8
>    __kasan_report+0x168/0x188
>    kasan_report+0x10/0x18
>    __asan_load4+0x84/0x8c
>    refcount_dec_and_test+0x20/0xd0
>    l2cap_chan_put+0x48/0x12c
>    l2cap_recv_frame+0x4770/0x6550
>    l2cap_recv_acldata+0x44c/0x7a4
>    hci_acldata_packet+0x100/0x188
>    hci_rx_work+0x178/0x23c
>    process_one_work+0x35c/0x95c
>    worker_thread+0x4cc/0x960
>    kthread+0x1a8/0x1c4
>    ret_from_fork+0x10/0x18
>
> Cc: stable@kernel.org

When was the bug added ? (Fixes: tag please)

> Cc: Marcel Holtmann <marcel@holtmann.org>
> Cc: Johan Hedberg <johan.hedberg@gmail.com>
> Cc: Luiz Augusto von Dentz <luiz.dentz@gmail.com>
> Cc: "David S. Miller" <davem@davemloft.net>
> Cc: Eric Dumazet <edumazet@google.com>
> Cc: Jakub Kicinski <kuba@kernel.org>
> Cc: Paolo Abeni <pabeni@redhat.com>
> Cc: linux-bluetooth@vger.kernel.org
> Cc: netdev@vger.kernel.org
> Signed-off-by: Lee Jones <lee.jones@linaro.org>
> ---
>  net/bluetooth/l2cap_core.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/net/bluetooth/l2cap_core.c b/net/bluetooth/l2cap_core.c
> index ae78490ecd3d4..82279c5919fd8 100644
> --- a/net/bluetooth/l2cap_core.c
> +++ b/net/bluetooth/l2cap_core.c
> @@ -483,9 +483,7 @@ static void l2cap_chan_destroy(struct kref *kref)
>
>         BT_DBG("chan %p", chan);
>
> -       write_lock(&chan_list_lock);
>         list_del(&chan->global_l);
> -       write_unlock(&chan_list_lock);
>
>         kfree(chan);
>  }
> @@ -501,7 +499,9 @@ void l2cap_chan_put(struct l2cap_chan *c)
>  {
>         BT_DBG("chan %p orig refcnt %u", c, kref_read(&c->kref));
>
> +       write_lock(&chan_list_lock);
>         kref_put(&c->kref, l2cap_chan_destroy);
> +       write_unlock(&chan_list_lock);
>  }
>  EXPORT_SYMBOL_GPL(l2cap_chan_put);
>
> --
> 2.36.1.255.ge46751e96f-goog
>

I do not think this patch is correct.

a kref does not need to be protected by a write lock.

This might shuffle things enough to work around a particular repro you have.

If the patch was correct why not protect kref_get() sides ?

Before the &hdev->rx_work is scheduled (queue_work(hdev->workqueue,
&hdev->rx_work),
a reference must be taken.

Then this reference must be released at the end of hci_rx_work() or
when hdev->workqueue
is canceled.

This refcount is not needed _if_ the workqueue is properly canceled at
device dismantle,
in a synchronous way.

I do not see this hdev->rx_work being canceled, maybe this is the real issue.

There is a call to drain_workqueue() but this is not enough I think,
because hci_recv_frame()
can re-arm
   queue_work(hdev->workqueue, &hdev->rx_work);