Re: [PATCH v1 1/1] iio: adc: tsc2046: add .read_raw support

From: Jonathan Cameron
Date: Sat Jan 15 2022 - 13:13:27 EST


On Tue, 11 Jan 2022 14:18:48 +0100
Oleksij Rempel <o.rempel@xxxxxxxxxxxxxx> wrote:

> Hi Jonathan,
>
> On Sun, Jan 09, 2022 at 04:00:09PM +0000, Jonathan Cameron wrote:
> > On Fri, 7 Jan 2022 10:35:27 +0100
> > Oleksij Rempel <o.rempel@xxxxxxxxxxxxxx> wrote:
> >
> > > Add read_raw() support to make use of iio_hwmon and other iio clients.
> > >
> > > Signed-off-by: Oleksij Rempel <o.rempel@xxxxxxxxxxxxxx>
> > Hi Oleksij
> >
> > Main questions in here are around settling time and the interface used for that.
> >
> > > ---
> > > drivers/iio/adc/ti-tsc2046.c | 114 ++++++++++++++++++++++++++++++++---
> > > 1 file changed, 106 insertions(+), 8 deletions(-)
> > >
> > > @@ -252,16 +266,47 @@ static u16 tsc2046_adc_get_value(struct tsc2046_adc_atom *buf)
> > > static int tsc2046_adc_read_one(struct tsc2046_adc_priv *priv, int ch_idx,
> > > u32 *effective_speed_hz)
> > > {
> > > + struct tsc2046_adc_ch_cfg *ch = &priv->ch_cfg[ch_idx];
> > > + struct tsc2046_adc_atom *rx_buf, *tx_buf;
> > > + unsigned int val, val_normalized = 0;
> > > + int ret, i, count_skip = 0, max_count;
> > > struct spi_transfer xfer;
> > > struct spi_message msg;
> > > - int ret;
> > > + u8 cmd;
> > > +
> > > + if (!effective_speed_hz) {
> > > + count_skip = tsc2046_adc_time_to_count(priv, ch->settling_time_us);
> > > + max_count = count_skip + ch->oversampling_ratio;
> > > + } else {
> > > + max_count = 1;
> > > + }
> > > +
> > > + tx_buf = kcalloc(max_count, sizeof(*tx_buf), GFP_KERNEL);
> > > + if (!tx_buf)
> > > + return -ENOMEM;
> > > +
> > > + rx_buf = kcalloc(max_count, sizeof(*rx_buf), GFP_KERNEL);
> > > + if (!rx_buf) {
> > > + ret = -ENOMEM;
> > > + goto free_tx;
> > > + }
> >
> > I guess these are fine to do everytime because you expect this to be used in
> > paths which aren't called at a particularly high frequency?
>
> Yes, this was my assumption as well. Instead of preallocating buffer of
> max size, I hope it is less ugly.
>
> > These buffers could get rather large so maybe you need a cap on settling time?
>
> What do you mean by "cap on settling"?

In theory the buffer needed could get very large, so perhap set a maximum reasonable
size (1 page perhaps) and report an error if the settling time is too large to fit
in that space.

>
> >
> > > +
> > > + /*
> > > + * Do not enable automatic power down on working samples. Otherwise the
> > > + * plates will never be completely charged.
> > > + */
> > > + cmd = tsc2046_adc_get_cmd(priv, ch_idx, true);
> > > +
> > > + for (i = 0; i < max_count - 1; i++)
> > > + tx_buf[i].cmd = cmd;
> > > +
> > > + /* automatically power down on last sample */
> > > + tx_buf[i].cmd = tsc2046_adc_get_cmd(priv, ch_idx, false);
> > >
> > > memset(&xfer, 0, sizeof(xfer));
> > > - priv->tx_one->cmd = tsc2046_adc_get_cmd(priv, ch_idx, false);
> > > - priv->tx_one->data = 0;
> > > - xfer.tx_buf = priv->tx_one;
> > > - xfer.rx_buf = priv->rx_one;
> >
> > Are these used for anything else? If not probably need to drop them and
> > their allocation.
>
> done
>
> > > - xfer.len = sizeof(*priv->tx_one);
> > > + xfer.tx_buf = tx_buf;
> > > + xfer.rx_buf = rx_buf;
> > > + xfer.len = sizeof(*tx_buf) * max_count;
> >
> > This could be very big and more than possible some spi controllers will fail
> > it (or does the SPI core handle splitting very large transfers?) Maybe a loop
> > is needed with smaller fixed size transfers?
>
> I can't exclude possible issue with some of SPI drivers. But SPI level
> optimizations should be done on SPI driver or framework level.

As above, I think you want to set a reasonable limit otherwise it will fail
on an awful lot of hardware if someone sets a silly value...

>
> > > spi_message_init_with_transfers(&msg, &xfer, 1);
> > >
> > > /*
> > > @@ -272,13 +317,25 @@ static int tsc2046_adc_read_one(struct tsc2046_adc_priv *priv, int ch_idx,
> > > if (ret) {
> > > dev_err_ratelimited(&priv->spi->dev, "SPI transfer failed %pe\n",
> > > ERR_PTR(ret));
> > > + *val2 = chan->scan_type.realbits;
> > > + return IIO_VAL_FRACTIONAL_LOG2;
> > > + case IIO_CHAN_INFO_OVERSAMPLING_RATIO:
> > > + *val = priv->ch_cfg[chan->channel].oversampling_ratio;
> > > + return IIO_VAL_INT;
> > > + case IIO_CHAN_INFO_DEBOUNCE_COUNT:
> >
> > These are unusual. I think they've only been used for the more literal bounce suppression
> > of a human step counting algorithm.
> >
> > I'd probably not expect to see the both even if we decide this is applicable.
>
> Ok, i do not need this information so far. I'll remove it
>
> > > + *val = tsc2046_adc_time_to_count(priv,
> > > + priv->ch_cfg[chan->channel].settling_time_us);
> >
> > Setting time is often about external circuitry so it's a bit unusual to expose
> > it to userspace rather than making it a device tree property and just making
> > sure the driver doesn't provide a reading until appropriate debounce has passed.
> > Here is coming from DT anyway, so what benefit do these two read only channel
> > properties provide?
>
> No benefit. Will remove it.
>
> Regards,
> Oleksij