Re: [RFC V1 07/16] Bluetooth: hci_ldisc: Add protocol check to hci_uart_tx_wakeup()

[Marcel Holtmann <marcel@holtmann.org>]

Hi Dean,

> Before attempting to schedule a work-item onto hu->write_work in
> hci_uart_tx_wakeup(), check that the Data Link protocol layer is
> still bound to the HCI UART driver.
> 
> Failure to perform this protocol check causes a race condition between
> the work queue hu->write_work running hci_uart_write_work() and the
> Data Link protocol layer being unbound (closed) in hci_uart_tty_close().
> 
> Note hci_uart_tty_close() does have a "cancel_work_sync(&hu->write_work)"
> but it is ineffective because it cannot prevent work-items being added
> to hu->write_work after cancel_work_sync() has run.
> 
> Therefore, add a check for HCI_UART_PROTO_READY into hci_uart_tx_wakeup()
> which prevents scheduling of the work queue when HCI_UART_PROTO_READY
> is in the clear state. However, note a small race condition remains
> because the hci_uart_tx_wakeup() thread can run in parallel with the
> hci_uart_tty_close() thread so it is possible that a schedule of
> hu->write_work can occur when HCI_UART_PROTO_READY is cleared. A complete
> solution needs locking of the threads which is implemented in a future
> commit.
> 
> Signed-off-by: Dean Jenkins <Dean_Jenkins@mentor.com>
> ---
> drivers/bluetooth/hci_ldisc.c | 6 +++++-
> 1 file changed, 5 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/bluetooth/hci_ldisc.c b/drivers/bluetooth/hci_ldisc.c
> index c4b801b..a1bb4d64b9 100644
> --- a/drivers/bluetooth/hci_ldisc.c
> +++ b/drivers/bluetooth/hci_ldisc.c
> @@ -125,15 +125,19 @@ static inline struct sk_buff *hci_uart_dequeue(struct hci_uart *hu)
> 
> int hci_uart_tx_wakeup(struct hci_uart *hu)
> {
> +	if (!test_bit(HCI_UART_PROTO_READY, &hu->flags))
> +		goto no_schedule;
> +
> 	if (test_and_set_bit(HCI_UART_SENDING, &hu->tx_state)) {
> 		set_bit(HCI_UART_TX_WAKEUP, &hu->tx_state);
> -		return 0;
> +		goto no_schedule;
> 	}
> 
> 	BT_DBG("");
> 
> 	schedule_work(&hu->write_work);
> 
> +no_schedule:
> 	return 0;
> }

so instead of calling no_schedule label, just keep calling return 0 directly.

Regards

Marcel