Re: [PATCH 2/3] Bluetooth: hci_qca: Skip serdev wait when no transfer is pending

[Matthias Kaehlcke]

Hi Bala,

On Sat, Jun 06, 2020 at 06:27:59PM +0530, bgodavar@xxxxxxxxxxxxxx wrote:
> Hi matthias,
> 
> On 2020-06-06 00:16, Matthias Kaehlcke wrote:
> > qca_suspend() calls serdev_device_wait_until_sent() regardless of
> > whether a transfer is pending. While it does no active harm since
> > the function should return immediately it makes the code more
> > confusing. Add a flag to track whether a transfer is pending and
> > only call serdev_device_wait_until_sent() is needed.
> > 
> > Signed-off-by: Matthias Kaehlcke <mka@xxxxxxxxxxxx>
> > ---
> > 
> >  drivers/bluetooth/hci_qca.c | 8 ++++++--
> >  1 file changed, 6 insertions(+), 2 deletions(-)
> > 
> > diff --git a/drivers/bluetooth/hci_qca.c b/drivers/bluetooth/hci_qca.c
> > index b1d82d32892e9..90ffd8ca1fb0d 100644
> > --- a/drivers/bluetooth/hci_qca.c
> > +++ b/drivers/bluetooth/hci_qca.c
> > @@ -2050,6 +2050,7 @@ static int __maybe_unused qca_suspend(struct
> > device *dev)
> >  	struct hci_uart *hu = &qcadev->serdev_hu;
> >  	struct qca_data *qca = hu->priv;
> >  	unsigned long flags;
> > +	bool tx_pending = false;
> >  	int ret = 0;
> >  	u8 cmd;
> > 
> > @@ -2083,6 +2084,7 @@ static int __maybe_unused qca_suspend(struct
> > device *dev)
> > 
> >  		qca->tx_ibs_state = HCI_IBS_TX_ASLEEP;
> >  		qca->ibs_sent_slps++;
> > +		tx_pending = true;
> >  		break;
> > 
> >  	case HCI_IBS_TX_ASLEEP:
> > @@ -2099,8 +2101,10 @@ static int __maybe_unused qca_suspend(struct
> > device *dev)
> >  	if (ret < 0)
> >  		goto error;
> > 
> > -	serdev_device_wait_until_sent(hu->serdev,
> > -				      msecs_to_jiffies(CMD_TRANS_TIMEOUT_MS));
> > +	if (tx_pending) {
> [Bala]: Good idea why don't we move this call to switch case under
> HCI_IBS_TX_AWAKE
> https://elixir.bootlin.com/linux/latest/source/drivers/bluetooth/hci_qca.c#L1994

I agree that this would make the code easier to follow, however the
reason this wasn't done in the first place is (probably) that the
IBS spinlock is held during the switch/case block.

In principle the unlock could be done before calling _wait_until_sent(),
but then the unlock also needs to happen in the other 'case' branches.
It's not ideal, but also not necessarily worse than introducing
'tx_pending', the repeated unlocks might be preferable in terms of
readability.

> i.e. i would recommend below sequence
> 
> 1. Send SLEEP BYTE
> 2. wait for some time to write SLEEP Byte on Tx line
> 3. call for Tx clock off qca_wq_serial_tx_clock_vote_off
> 
> > +		serdev_device_wait_until_sent(hu->serdev,
> > +					      msecs_to_jiffies(CMD_TRANS_TIMEOUT_MS));
> > +	}
> > 
> >  	/* Wait for HCI_IBS_SLEEP_IND sent by device to indicate its Tx is
> > going
> >  	 * to sleep, so that the packet does not wake the system later.