Hi Jonathan, Peter,
On Mon, Apr 22, 2019 at 01:01:38PM +0100, Jonathan Cameron wrote:
> On Mon, 22 Apr 2019 11:17:27 +0200 (CEST)
> Peter Meerwald-Stadler <pmeerw@xxxxxxxxxx> wrote:
>
> > On Sun, 21 Apr 2019, Ronald Tschalär wrote:
> >
> > > On 2016/2017 MacBook Pro's with a Touch Bar the ALS is attached to,
> > > and exposed via the iBridge device. This provides the driver for that
> > > sensor.
> >
> > some comments below inline
> I'll 'nest' on Peter's review to avoid repetition.
>
> A few additional comments inline.
Thank you both for your reviews. I've applied most of your
suggestions, so I'm limiting my responses below to just those places
where I need further clarification or can provide some answers.
[snip]
> > > +static int appleals_read_raw(struct iio_dev *iio_dev,
> > > + struct iio_chan_spec const *chan,
> > > + int *val, int *val2, long mask)
> > > +{
> > > + struct appleals_device *als_dev =
> > > + *(struct appleals_device **)iio_priv(iio_dev);
> > > + __s32 value;
> > > + int rc;
> > > +
> > > + *val = 0;
> > > + *val2 = 0;
> >
> > no need to set these here
> >
> > > +
> > > + switch (mask) {
> > > + case IIO_CHAN_INFO_RAW:
> > > + case IIO_CHAN_INFO_PROCESSED:
> > > + *val = appleals_get_field_value(als_dev, als_dev->illum_field);
>
> How can one read by both processed and raw?
>From my understanding, processed is a converted and normalized value
of raw? But since the raw value is already in Lux the processed value
is then the same. Furthermore, looking at userspace apps that read
these values (e.g. iio-sensor-proxy, lightsd) they seem to be all over
the map in terms of what sysfs entries they read:
in_illuminance_input, in_intensity_both_raw, etc. So I figured
providing both would provide maximal compatibility.
What then is currently the preferred approach here?
> > > + case IIO_CHAN_INFO_HYSTERESIS:
> > > + if (val == APPLEALS_DYN_SENS) {
> > > + if (als_dev->cur_hysteresis != APPLEALS_DYN_SENS) {
> > > + als_dev->cur_hysteresis = val;
> > > + illum = appleals_get_field_value(als_dev,
> > > + als_dev->illum_field);
> > > + appleals_update_dyn_sensitivity(als_dev, illum);
>
> There is some debate in another thread on whether dynamic sensitivity can be
> mapped to hysteresis or not...
Is that the "drivers: iio: Add more data field for iio driver
hysteresis parsing" thread?
In any case, I'm using the value 0 here to indicate that the
pseudo-percent-relative change sensitivity should be used. Better
suggestions certainly welcome.
[snip]
> > > +static const struct iio_chan_spec appleals_channels[] = {
> > > + {
> > > + .type = IIO_INTENSITY,
> > > + .modified = 1,
> > > + .channel2 = IIO_MOD_LIGHT_BOTH,
> > > + .info_mask_separate = BIT(IIO_CHAN_INFO_PROCESSED) |
> > > + BIT(IIO_CHAN_INFO_RAW),
> Why return both processed and raw? We don't generally allow that in IIO
> (there are a few historical exceptions due to us getting it wrong the
> first time).
Ok. But which should be used then, especially in view of the fact that
different user-space apps/daemons appear to use different values?
[snip]
> > > +static int appleals_config_iio(struct appleals_device *als_dev)
[snip]
> > > + iio_dev = iio_device_alloc(sizeof(als_dev));
> >
> > how about using the devm_ variants?
The problem is that there is no device with the proper lifetime here.
In particular we can't use hid_device (the only struct device
available here) because the instances of those can (and do) outlive
the module. This is a consequence of the hid-driver demuxing: e.g.
when this als module is unloaded, the hid_device is still in use by
the touchbar driver, and if this als module is then loaded again it
will get associated with that same hid_device again.
[snip]
> > > + rc = iio_device_register(iio_dev);
> > > + if (rc) {
> > > + dev_err(als_dev->log_dev, "failed to register iio device: %d\n",
> > > + rc);
> > > + goto unreg_iio_trig;
> > > + }
> > > +
> > > + als_dev->iio_dev = iio_dev;
>
> I really don't like nest of pointers going on in here. I haven't dug
> down to check if any of them can be remove, but they are definitely
> ugly to deal with.
I'm not sure how this can be avoided: *iio_dev is allocated by
iio_device_alloc(), and this is just storing that pointer in our data
structure for this als device.
[snip]
> > > +static int appleals_probe(struct hid_device *hdev,
> > > + const struct hid_device_id *id)
> > > +{
> > > + struct appleals_device *als_dev =
> > > + appleib_get_drvdata(hid_get_drvdata(hdev),
> > > + &appleals_hid_driver);
> > > + struct hid_field *state_field;
> > > + struct hid_field *illum_field;
> > > + int rc;
> > > +
> > > + /* find als fields and reports */
> > > + state_field = appleib_find_hid_field(hdev, HID_USAGE_SENSOR_ALS,
> > > + HID_USAGE_SENSOR_PROP_REPORT_STATE);
> > > + illum_field = appleib_find_hid_field(hdev, HID_USAGE_SENSOR_ALS,
> > > + HID_USAGE_SENSOR_LIGHT_ILLUM);
> > > + if (!state_field || !illum_field)
> > > + return -ENODEV;
> > > +
> > > + if (als_dev->hid_dev) {
> > > + dev_warn(als_dev->log_dev,
> > > + "Found duplicate ambient light sensor - ignoring\n");
>
> I'll bite. So how can this actually happen? What fundamentally breaks in
> your model if there really are two? Can you fix whatever that is so
> as to drop this handling?
This should indeed never happen - the check is just defensive
programming, in case either some new device in some new model appears,
or due to some bug somewhere.
To actually handle this I'd need to split up appleals_device into a
per iBridge-device part and a per ALS-device part, and allow multiple
ALS-device parts. This is certainly doable, but seemed a bit overkill
for something that is unlikely to ever be needed.
[snip]
> > > +static void appleals_remove(struct hid_device *hdev)
> > > +{
> > > + struct appleals_device *als_dev =
> > > + appleib_get_drvdata(hid_get_drvdata(hdev),
> > > + &appleals_hid_driver);
> > > +
> >
> > could be a lot less if devm_ were used?
Agreed, but see explanation above of why this can't be used.
> > > + if (als_dev->iio_dev) {
> > > + iio_device_unregister(als_dev->iio_dev);
> > > +
> > > + iio_trigger_unregister(als_dev->iio_trig);
> > > + iio_trigger_free(als_dev->iio_trig);
> > > + als_dev->iio_trig = NULL;
>
> I would be decidedly worried if you actually have any paths
> where setting these to NULL do anything useful. By the time
> we get here we should absolutely 'know' nothing will touch
> these pointers.
I guess I was being overly paranoid here. I've removed these now.
> It is these that are blocking Peter's suggestion of using
> devm to clean all of this up automatically for you.
No, that is for a different reason - see above.
[snip]
> > > +static int appleals_platform_probe(struct platform_device *pdev)
> > > +{
> > > + struct appleib_platform_data *pdata = pdev->dev.platform_data;
> > > + struct appleib_device *ib_dev = pdata->ib_dev;
> > > + struct appleals_device *als_dev;
> > > + int rc;
> > > +
> > > + als_dev = kzalloc(sizeof(*als_dev), GFP_KERNEL);
> > > + if (!als_dev)
> > > + return -ENOMEM;
> > > +
> > > + als_dev->ib_dev = ib_dev;
> > > + als_dev->log_dev = pdata->log_dev;
> > > +
> > > + rc = appleib_register_hid_driver(ib_dev, &appleals_hid_driver, als_dev);
>
> Hmm. This is all a bit odd. We register a platform device, to call it's driver
> so that it can then register another driver? I'll let Lee comment on that but
> it does seem overly complex.
"Normally" this call would be to hid_register_driver(). However, as I
tried to explain in the cover letter and module comments, at least one
of the hid_device's needs to be shared by both this als driver and the
touchbar driver. But the linux device/driver architecture does not
allow multiple drivers to be attached to a single device. Hence the
mfd driver implements demuxing hid_driver that allows multiple
hid_drivers to be registered for a single hid_device. And this is why
we register our hid_driver with this demuxing driver here instead of
directly via hid_register_driver().
Does this make sense?
Cheers,
Ronald
