On Fri, 7 Jan 2022 08:40:17 +0100
Oleksij Rempel <o.rempel@xxxxxxxxxxxxxx> wrote:
> Initially this was designed to:
> | Fix sleeping in atomic context warning and a deadlock after iio_trigger_poll()
> | call
> |
> | If iio_trigger_poll() is called after IRQ was disabled, we will call
> | reenable_trigger() directly from hard IRQ or hrtimer context instead of
> | IRQ thread. In this case we will run in to multiple issue as sleeping in atomic
> | context and a deadlock.
> |
> | To avoid this issue, rework the trigger to use state machine. All state
> | changes are done over the hrtimer, so it allows us to drop fsleep() and
> | avoid the deadlock.
>
> This issue was fixed by: 9020ef659885 ("iio: trigger: Fix a scheduling
> whilst atomic issue seen on tsc2046").
>
> Even if the root cause of this issue probably will and can be fixed in the iio
> core, this patch can be seen as clean-up to provide better internal state
> machine.
Probably want to update this text?
A few comments below.
>
> Fixes: 9374e8f5a38d ("iio: adc: add ADC driver for the TI TSC2046 controller")
>From above isn't this now fixed? The cleanup here is just making things easier
to follow I think...
> Signed-off-by: Oleksij Rempel <o.rempel@xxxxxxxxxxxxxx>
> ---
> drivers/iio/adc/ti-tsc2046.c | 102 ++++++++++++++++++++---------------
> 1 file changed, 58 insertions(+), 44 deletions(-)
>
> diff --git a/drivers/iio/adc/ti-tsc2046.c b/drivers/iio/adc/ti-tsc2046.c
> index d84ae6b008c1..91f6bd5effe7 100644
> --- a/drivers/iio/adc/ti-tsc2046.c
> +++ b/drivers/iio/adc/ti-tsc2046.c
> @@ -123,14 +123,21 @@ struct tsc2046_adc_ch_cfg {
> unsigned int oversampling_ratio;
> };
>
> +enum tsc2046_state {
> + TSC2046_STATE_STANDBY,
> + TSC2046_STATE_ENABLE_IRQ_POLL,
> + TSC2046_STATE_POLL,
> + TSC2046_STATE_ENABLE_IRQ,
> +};
> +
> struct tsc2046_adc_priv {
> struct spi_device *spi;
> const struct tsc2046_adc_dcfg *dcfg;
>
> struct iio_trigger *trig;
> struct hrtimer trig_timer;
> - spinlock_t trig_lock;
> - unsigned int trig_more_count;
> + enum tsc2046_state state;
> + spinlock_t state_lock;
>
> struct spi_transfer xfer;
> struct spi_message msg;
> @@ -411,21 +418,47 @@ static const struct iio_info tsc2046_adc_info = {
> .update_scan_mode = tsc2046_adc_update_scan_mode,
> };
>
> -static enum hrtimer_restart tsc2046_adc_trig_more(struct hrtimer *hrtimer)
> +static enum hrtimer_restart tsc2046_adc_timer(struct hrtimer *hrtimer)
> {
> struct tsc2046_adc_priv *priv = container_of(hrtimer,
> struct tsc2046_adc_priv,
> trig_timer);
> unsigned long flags;
>
> - spin_lock_irqsave(&priv->trig_lock, flags);
> -
> - disable_irq_nosync(priv->spi->irq);
> -
> - priv->trig_more_count++;
> - iio_trigger_poll(priv->trig);
> -
> - spin_unlock_irqrestore(&priv->trig_lock, flags);
> + spin_lock_irqsave(&priv->state_lock, flags);
> + switch (priv->state) {
> + case TSC2046_STATE_ENABLE_IRQ_POLL:
> + /*
> + * IRQ handler called iio_trigger_poll() to sample ADC.
> + * Here we
> + * - re-enable IRQs
> + * - start hrtimer for timeout if no IRQ will occur
> + */
> + priv->state = TSC2046_STATE_POLL;
> + enable_irq(priv->spi->irq);
I comment on this below, but I'm not sure why you don't move the enable_irq()
here out of this timer function and then have the first entry of the timer
go directly to TSC2046_STATE_POLL after a longer initial wait.
It's been a long time since I looked at this, so perhaps I'm missing the
point. What you have here works as far as I can see, it just seems to push
more than necessary into the state machine.
> + hrtimer_start(&priv->trig_timer,
> + ns_to_ktime(priv->scan_interval_us *
> + NSEC_PER_USEC),
> + HRTIMER_MODE_REL_SOFT);
> + break;
> + case TSC2046_STATE_POLL:
> + disable_irq_nosync(priv->spi->irq);
> + priv->state = TSC2046_STATE_ENABLE_IRQ;
> + /* iio_trigger_poll() starts hrtimer */
> + iio_trigger_poll(priv->trig);
> + break;
> + case TSC2046_STATE_ENABLE_IRQ:
> + priv->state = TSC2046_STATE_STANDBY;
> + enable_irq(priv->spi->irq);
> + break;
> + case TSC2046_STATE_STANDBY:
> + fallthrough;
> + default:
> + dev_warn(&priv->spi->dev, "Got unexpected state: %i\n",
> + priv->state);
> + break;
> + }
> + spin_unlock_irqrestore(&priv->state_lock, flags);
>
> return HRTIMER_NORESTART;
> }
> @@ -434,16 +467,17 @@ static irqreturn_t tsc2046_adc_irq(int irq, void *dev_id)
> {
> struct iio_dev *indio_dev = dev_id;
> struct tsc2046_adc_priv *priv = iio_priv(indio_dev);
> -
> - spin_lock(&priv->trig_lock);
> + unsigned long flags;
>
> hrtimer_try_to_cancel(&priv->trig_timer);
>
> - priv->trig_more_count = 0;
> + spin_lock_irqsave(&priv->state_lock, flags);x`
> disable_irq_nosync(priv->spi->irq);
> - iio_trigger_poll(priv->trig);
> + priv->state = TSC2046_STATE_ENABLE_IRQ_POLL;
>
> - spin_unlock(&priv->trig_lock);
> + /* iio_trigger_poll() starts hrtimer */
> + iio_trigger_poll(priv->trig);
> + spin_unlock_irqrestore(&priv->state_lock, flags);
>
> return IRQ_HANDLED;
> }
> @@ -452,37 +486,16 @@ static void tsc2046_adc_reenable_trigger(struct iio_trigger *trig)
> {
> struct iio_dev *indio_dev = iio_trigger_get_drvdata(trig);
> struct tsc2046_adc_priv *priv = iio_priv(indio_dev);
> - unsigned long flags;
> - int delta;
> + ktime_t tim;
>
> /*
> * We can sample it as fast as we can, but usually we do not need so
> * many samples. Reduce the sample rate for default (touchscreen) use
> * case.
> - * Currently we do not need a highly precise sample rate. It is enough
> - * to have calculated numbers.
> - */
> - delta = priv->scan_interval_us - priv->time_per_scan_us;
> - if (delta > 0)
> - fsleep(delta);
> -
> - spin_lock_irqsave(&priv->trig_lock, flags);
> -
> - /*
> - * We need to trigger at least one extra sample to detect state
> - * difference on ADC side.
> */
> - if (!priv->trig_more_count) {
> - int timeout_ms = DIV_ROUND_UP(priv->scan_interval_us,
> - USEC_PER_MSEC);
> -
> - hrtimer_start(&priv->trig_timer, ms_to_ktime(timeout_ms),
> - HRTIMER_MODE_REL_SOFT);
> - }
> -
> - enable_irq(priv->spi->irq);
> -
> - spin_unlock_irqrestore(&priv->trig_lock, flags);
> + tim = ns_to_ktime((priv->scan_interval_us - priv->time_per_scan_us) *
> + NSEC_PER_USEC);
> + hrtimer_start(&priv->trig_timer, tim, HRTIMER_MODE_REL_SOFT);
This moves enabling the irq to the first instance of the timer - is that ever too late?
> }
>
> static int tsc2046_adc_set_trigger_state(struct iio_trigger *trig, bool enable)
> @@ -493,8 +506,8 @@ static int tsc2046_adc_set_trigger_state(struct iio_trigger *trig, bool enable)
> if (enable) {
> enable_irq(priv->spi->irq);
> } else {
> + hrtimer_cancel(&priv->trig_timer);
So this will wait for the callback to finish. However, is there a chance
of an interrupt just after this but before disable_irq that ends up
starting the timer again?
> disable_irq(priv->spi->irq);
> - hrtimer_try_to_cancel(&priv->trig_timer);
> }
>
> return 0;
> @@ -668,10 +681,11 @@ static int tsc2046_adc_probe(struct spi_device *spi)
> iio_trigger_set_drvdata(trig, indio_dev);
> trig->ops = &tsc2046_adc_trigger_ops;
>
> - spin_lock_init(&priv->trig_lock);
> + spin_lock_init(&priv->state_lock);
> + priv->state = TSC2046_STATE_STANDBY;
> hrtimer_init(&priv->trig_timer, CLOCK_MONOTONIC,
> HRTIMER_MODE_REL_SOFT);
> - priv->trig_timer.function = tsc2046_adc_trig_more;
> + priv->trig_timer.function = tsc2046_adc_timer;
>
> ret = devm_iio_trigger_register(dev, trig);
> if (ret) {
