On Thu, 9 Jun 2022 11:31:59 +0300
Claudiu Beznea <claudiu.beznea@xxxxxxxxxxxxx> wrote:
> .read_raw()/.write_raw() could be called asynchronously from user space
> or other in kernel drivers. Without locking on st->lock these could be
> called asynchronously while there is a conversion in progress. Read will
> be harmless but changing registers while conversion is in progress may
> lead to inconsistent results. Thus, to avoid this lock st->lock.
The patch makes sense, but I'm not convinced all of the changes below
involve any changes to registers. E.g. at91_adc_adjust_val_osr()
is using the cached value of something in a register, but not the
register itself, so please update the description to mention cached state.
Other comments inline.
>
> Fixes: 27e177190891 ("iio:adc:at91_adc8xx: introduce new atmel adc driver")
> Fixes: 6794e23fa3fe ("iio: adc: at91-sama5d2_adc: add support for oversampling resolution")
> Signed-off-by: Claudiu Beznea <claudiu.beznea@xxxxxxxxxxxxx>
> ---
> drivers/iio/adc/at91-sama5d2_adc.c | 17 ++++++++++++++---
> 1 file changed, 14 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/iio/adc/at91-sama5d2_adc.c b/drivers/iio/adc/at91-sama5d2_adc.c
> index 32b6f157b803..a672a520cdc0 100644
> --- a/drivers/iio/adc/at91-sama5d2_adc.c
> +++ b/drivers/iio/adc/at91-sama5d2_adc.c
> @@ -1542,10 +1542,11 @@ static int at91_adc_read_info_raw(struct iio_dev *indio_dev,
> ret = at91_adc_read_position(st, chan->channel,
> &tmp_val);
> *val = tmp_val;
> + ret = at91_adc_adjust_val_osr(st, val);
> mutex_unlock(&st->lock);
> iio_device_release_direct_mode(indio_dev);
>
> - return at91_adc_adjust_val_osr(st, val);
> + return ret;
> }
> if (chan->type == IIO_PRESSURE) {
> ret = iio_device_claim_direct_mode(indio_dev);
> @@ -1556,10 +1557,11 @@ static int at91_adc_read_info_raw(struct iio_dev *indio_dev,
> ret = at91_adc_read_pressure(st, chan->channel,
> &tmp_val);
> *val = tmp_val;
> + ret = at91_adc_adjust_val_osr(st, val);
> mutex_unlock(&st->lock);
> iio_device_release_direct_mode(indio_dev);
>
> - return at91_adc_adjust_val_osr(st, val);
> + return ret;
> }
>
> /* in this case we have a voltage channel */
> @@ -1620,11 +1622,15 @@ static int at91_adc_read_raw(struct iio_dev *indio_dev,
> return IIO_VAL_FRACTIONAL_LOG2;
>
> case IIO_CHAN_INFO_SAMP_FREQ:
> + mutex_lock(&st->lock);
> *val = at91_adc_get_sample_freq(st);
So this is a straight read of a cached value. The only thing you 'might'
arguably be protecting against is read/write tearing due to it in theory
being possible to write part of the value whilst reading. I don't
see that being a concern for st->current_sample_rate
> + mutex_unlock(&st->lock);
> return IIO_VAL_INT;
>
> case IIO_CHAN_INFO_OVERSAMPLING_RATIO:
> + mutex_lock(&st->lock);
> *val = st->oversampling_ratio;
Likewise, what are you protecting against racing with this that can't
just occur before or after the lock?
> + mutex_unlock(&st->lock);
> return IIO_VAL_INT;
>
> default:
> @@ -1644,18 +1650,23 @@ static int at91_adc_write_raw(struct iio_dev *indio_dev,
> (val != AT91_OSR_16SAMPLES))
> return -EINVAL;
> /* if no change, optimize out */
> + mutex_lock(&st->lock);
> if (val == st->oversampling_ratio)
> - return 0;
It should be race free to check this outside the lock.
Definitely valid to lock around the cached value write and the config
write though.
> + goto unlock;
If you did want to have locking as now then flip the logic
if (val != st->oversampling_ratio) {
st->oversampling_ratio = val;
at91_adc_config_emr(st);
}
mutex_unlock()
..
Goto always have a cost in readability so if you can avoid them with
a simple flip of logic like this it is usually a good idea.
(exception is error code which should always be out of line as
that is more common so what we expect to see).
> st->oversampling_ratio = val;
> /* update ratio */
> at91_adc_config_emr(st);
> +unlock:
> + mutex_unlock(&st->lock);
> return 0;
> case IIO_CHAN_INFO_SAMP_FREQ:
> if (val < st->soc_info.min_sample_rate ||
> val > st->soc_info.max_sample_rate)
> return -EINVAL;
>
> + mutex_lock(&st->lock);
> at91_adc_setup_samp_freq(indio_dev, val);
> + mutex_unlock(&st->lock);
> return 0;
> default:
> return -EINVAL;
