Received-SPF: pass (google.com: domain of linux-bluetooth+bounces-1911-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:45e3:2400::1 as permitted sender) client-ip=2604:1380:45e3:2400::1;
Precedence: bulk
MIME-Version: 1.0
References: <20240214012328.BlueZ.1.I10eda6f375bc8dfedf4eef3a8cf6572c65803afc@changeid>
In-Reply-To: <20240214012328.BlueZ.1.I10eda6f375bc8dfedf4eef3a8cf6572c65803afc@changeid>
From: Luiz Augusto von Dentz <luiz.dentz@gmail.com>
Date: Thu, 15 Feb 2024 16:37:08 -0500
Message-ID: <CABBYNZ+MJRSAWHjgNTNwcWCdq5y0zBU5f8XJvr1tv3LCqvdKJw@mail.gmail.com>
Subject: Re: [BlueZ PATCH] Bluetooth: hci_core: Skip hci_cmd_work if
 hci_request is pending
To: Hsin-chen Chuang <chharry@chromium.org>
Cc: linux-bluetooth@vger.kernel.org, 
	chromeos-bluetooth-upstreaming@chromium.org, 
	Johan Hedberg <johan.hedberg@gmail.com>, Marcel Holtmann <marcel@holtmann.org>, 
	linux-kernel@vger.kernel.org
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

Hi Hsin-chen,

On Tue, Feb 13, 2024 at 12:24=E2=80=AFPM Hsin-chen Chuang <chharry@chromium=
.org> wrote:
>
> hci_cmd_work overwrites the hdev->sent_cmd which contains the required
> info for a hci_request to work. In the real world, it's observed that
> a request from hci_le_ext_create_conn_sync could be interrupted by
> the authentication (hci_conn_auth) caused by rfcomm_sock_connect. When
> it happends, hci_le_ext_create_conn_sync hangs until timeout; If the
> LE connection is triggered by MGMT, it freezes the whole MGMT interface.
>
> Signed-off-by: Hsin-chen Chuang <chharry@chromium.org>
> ---
>
>  net/bluetooth/hci_core.c | 7 +++++--
>  1 file changed, 5 insertions(+), 2 deletions(-)
>
> diff --git a/net/bluetooth/hci_core.c b/net/bluetooth/hci_core.c
> index 34c8dca2069f..e3706889976d 100644
> --- a/net/bluetooth/hci_core.c
> +++ b/net/bluetooth/hci_core.c
> @@ -4213,8 +4213,11 @@ static void hci_cmd_work(struct work_struct *work)
>         BT_DBG("%s cmd_cnt %d cmd queued %d", hdev->name,
>                atomic_read(&hdev->cmd_cnt), skb_queue_len(&hdev->cmd_q));
>
> -       /* Send queued commands */
> -       if (atomic_read(&hdev->cmd_cnt)) {
> +       /* Send queued commands. Don't send the command when there is a p=
ending
> +        * hci_request because the request callbacks would be overwritten=
.
> +        */
> +       if (atomic_read(&hdev->cmd_cnt) &&
> +           !hci_dev_test_flag(hdev, HCI_CMD_PENDING)) {
>                 skb =3D skb_dequeue(&hdev->cmd_q);
>                 if (!skb)
>                         return;
> --
> 2.43.0.687.g38aa6559b0-goog


This seems to be causing some mgmt-tester failures:

Pair Device - Sec Mode 3 Success 1                   Timed out   22.753 sec=
onds
Pair Device - Sec Mode 3 Reject 1                    Timed out   22.533 sec=
onds
Pair Device - Sec Mode 3 Reject 2                    Timed out   22.526 sec=
onds

I think this is because we need to respond to an event with a command like:

< HCI Command: Create Conn.. (0x01|0x0005) plen 13  #241 [hci0] 16:25:38.69=
9066
        Address: 00:AA:01:01:00:00 (Intel Corporation)
        Packet type: 0x0018
          DM1 may be used
          DH1 may be used
        Page scan repetition mode: R2 (0x02)
        Page scan mode: Mandatory (0x00)
        Clock offset: 0x0000
        Role switch: Allow peripheral (0x01)
> HCI Event: Command Status (0x0f) plen 4           #242 [hci0] 16:25:38.70=
1881
      Create Connection (0x01|0x0005) ncmd 1
        Status: Success (0x00)
> HCI Event: Link Key Request (0x17) plen 6         #243 [hci0] 16:25:38.70=
2375
        Address: 00:AA:01:01:00:00 (Intel Corporation)

But because Create Connection is pending we cannot respond to Link Key
Request, so it is actually a design problem if we cannot send commands
because something is pending so perhaps we need to redesign how we
store cmd_sent so we can have multiple outstanding commands rather
than just one.

--=20
Luiz Augusto von Dentz