Re: [PATCH v8 3/3] Bluetooth: hci_qca: Disable IBS state machine and flush Tx buffer

[Matthias Kaehlcke <mka@chromium.org>]

On Wed, Jan 16, 2019 at 05:16:03PM +0530, Balakrishna Godavarthi wrote:
> During hci down we observed IBS sleep commands are queued in the Tx
> buffer and hci_uart_write_work is sending data to the chip which is
> not required as the chip is powered off. This patch will disable IBS
> and flush the Tx buffer before we turn off the chip.
> 
> Signed-off-by: Balakrishna Godavarthi <bgodavar@codeaurora.org>
> ---
>  drivers/bluetooth/hci_qca.c | 8 ++++++++
>  1 file changed, 8 insertions(+)
> 
> diff --git a/drivers/bluetooth/hci_qca.c b/drivers/bluetooth/hci_qca.c
> index 7e4afcf40da2..7330ba71ada4 100644
> --- a/drivers/bluetooth/hci_qca.c
> +++ b/drivers/bluetooth/hci_qca.c
> @@ -1275,6 +1275,14 @@ static const struct qca_vreg_data qca_soc_data = {
>  
>  static void qca_power_shutdown(struct hci_uart *hu)
>  {
> +	struct qca_data *qca = hu->priv;
> +
> +	/* From this point we go into power off state. But serial port is
> +	 * still open, stop queueing the IBS data and flush all the buffered
> +	 * data in skb's.
> +	 */
> +	clear_bit(STATE_IN_BAND_SLEEP_ENABLED, &qca->flags);
> +	qca_flush(hu);
>  	host_set_baudrate(hu, 2400);
>  	qca_send_power_pulse(hu, QCA_WCN3990_POWEROFF_PULSE);
>  	qca_power_setup(hu, false);

Due to a race-condition there could be an IBS sleep command queued
even after clearing the bit and flushing the queue.

In qca_enqueue() we have this:

static int qca_enqueue(struct hci_uart *hu, struct sk_buff *skb)
{
        ...

        /* Don't go to sleep in middle of patch download or
         * Out-Of-Band(GPIOs control) sleep is selected.
         */
        if (!test_bit(STATE_IN_BAND_SLEEP_ENABLED, &qca->flags)) {
                skb_queue_tail(&qca->txq, skb);
                return 0;
        }

        spin_lock_irqsave(&qca->hci_ibs_lock, flags);
}

With process X executing qca_power_shutdown() and process Y running
qca_enqueue() this could happen:

[X] test_bit(STATE_IN_BAND_SLEEP_ENABLED)  => set
[Y] clear_bit(STATE_IN_BAND_SLEEP_ENABLED)
[Y] qca_flush(hu);
[X] skb_queue_tail(&qca->txq, skb);

The following should fix this race:

--- a/drivers/bluetooth/hci_qca.c
+++ b/drivers/bluetooth/hci_qca.c
@@ -770,16 +770,17 @@ static int qca_enqueue(struct hci_uart *hu, struct sk_buff *skb)
 	/* Prepend skb with frame type */
 	memcpy(skb_push(skb, 1), &hci_skb_pkt_type(skb), 1);
 
+	spin_lock_irqsave(&qca->hci_ibs_lock, flags);
+
 	/* Don't go to sleep in middle of patch download or
 	 * Out-Of-Band(GPIOs control) sleep is selected.
 	 */
 	if (!test_bit(STATE_IN_BAND_SLEEP_ENABLED, &qca->flags)) {
 		skb_queue_tail(&qca->txq, skb);
+		spin_unlock_irqrestore(&qca->hci_ibs_lock, flags);
 		return 0;
 	}
 
-	spin_lock_irqsave(&qca->hci_ibs_lock, flags);
-
 	/* Act according to current state */
 	switch (qca->tx_ibs_state) {
 	case HCI_IBS_TX_AWAKE:
@@ -1275,13 +1276,17 @@ static const struct qca_vreg_data qca_soc_data = {
 static void qca_power_shutdown(struct hci_uart *hu)
 {
 	struct qca_data *qca = hu->priv;
+	unsigned long flags;
 
 	/* From this point we go into power off state. But serial port is
 	 * still open, stop queueing the IBS data and flush all the buffered
 	 * data in skb's.
 	 */
+	spin_lock_irqsave(&qca->hci_ibs_lock, flags);
 	clear_bit(STATE_IN_BAND_SLEEP_ENABLED, &qca->flags);
 	qca_flush(hu);
+	spin_unlock_irqrestore(&qca->hci_ibs_lock, flags);
+
 	host_set_baudrate(hu, 2400);
 	qca_send_power_pulse(hu, QCA_WCN3990_POWEROFF_PULSE);
 	qca_power_setup(hu, false);

Cheers

Matthias