On 6/11/24 20:14, Jonathan Cameron wrote:
On Mon, 10 Jun 2024 08:58:44 +0300
Matti Vaittinen <mazziesaccount@xxxxxxxxx> wrote:
la 8. kesäk. 2024 klo 19.22 Jonathan Cameron (jic23@xxxxxxxxxx) kirjoitti:
On Thu, 6 Jun 2024 17:29:42 +0100
Mudit Sharma <muditsharma.info@xxxxxxxxx> wrote:
Add support for BH1745, which is an I2C colour sensor with red, green,
blue and clear channels. It has a programmable active low interrupt
pin. Interrupt occurs when the signal from the selected interrupt
source channel crosses set interrupt threshold high or low level.
This driver includes device attributes to configure the following:
- Interrupt pin latch: The interrupt pin can be configured to
be latched (until interrupt register (0x60) is read or initialized)
or update after each measurement.
- Interrupt source: The colour channel that will cause the interrupt
when channel will cross the set threshold high or low level.
This driver also includes device attributes to present valid
configuration options/values for:
- Integration time
- Interrupt colour source
- Hardware gain
+
+#define BH1745_CHANNEL(_colour, _si, _addr) \
+ { \
+ .type = IIO_INTENSITY, .modified = 1, \
+ .info_mask_separate = BIT(IIO_CHAN_INFO_RAW), \
+ .info_mask_shared_by_type = BIT(IIO_CHAN_INFO_HARDWAREGAIN) | \
Provide _SCALE instead of HARDWAREGAIN
As it's an intensity channel (and units are tricky for color sensors given
frequency dependence etc) all you need to do is ensure that if you halve
the _scale and measure the same light source, the computed
_RAW * _SCALE value remains constant.
...Which is likely to cause also the integration time setting to
impact the SCALE.
You may or may not want to see the GTS-helpers
(drivers/iio/industrialio-gts-helper.c) - which have their own tricky
corners. I think Jonathan once suggested to me to keep the
HARDWAREGAIN as a read-only attribute to ease seeing what is going on.
For the last couple of days I've been reworking the BU27034 driver to
work with the new sensor variant - and I can definitely see the value
of the read-only HARDWAREGAIN when we have per channel gain settings +
integration time setting which all contribute to the scale...
I'm wondering if that was good advice, but it's definitely better
than letting userspace control the gain and integration time separately
I woke up last night at 03.14 AM thinking of this :rolleyes:
as there is no sensible way to know how to control that beyond -
I agree and disagree :)
I agree that it is simpler to just change the scale when readings get
saturated - or when more accuracy is needed. Hence, implementing the
scale change as is done now makes very much sense.
However, I can imagine that sometimes the measurement time plays a role
- and people would like to have more fine grained control over things.
In that case, if driver only allows changing things via the scale, then
the driver is probably doing autonomous choices regarding the
integration time - which may not be optimal for all cases (*citation
needed).
As you may remember, I implemented the ROHM RGB and ALS sensors (the
BU270xx series) so that the integration time can be set as well as the
gain. These sensors (at least the BU27034, don't remember all the dirty
details of the RGB sensors) had per-channel gain and a global
integration time settings. Hence, the scale can be set separately for
each channel. I invented a restriction that setting the per-channel
scale tried to maintain the integration time and change the gain - but
if it was not possible, the scale change changes also the integration
time in order to yield the desired scale.
Problem was that the integration time was a global setting, and changing
it for one channel results scale change also on the other channel(s).
To mitigate such side-effects I implemented logic that the scale change
for other channels (caused by the integration time change) is
compensated by changing the gain for these unrelated channels. Eg, if
scale change for channel #1 required doubling the integration time -
which effectively doubled the "gain contributed by integration time"
also for the channel #2 and #3 - then the HARDWAREGAIN for the unrelated
channels #2 and #3 is halved in order to keep their scale unchanged. Great.
Except that this is not always possible. The HWGAIN for these unrelated
channels may have been already set to the other extreme, and further
reducing/increasing is not possible. Or, there may be unsupported
multipliers (gaps) in the gain range, so that setting the hardwaregain
to required value is not possible.
Here I just decided to return an error to caller and disallow such scale
change.
This is very much annoying solution but I ran out of good ideas. Adding
more logic to the driver to work around this felt like asking for a
nose-bleed. I was sure I ended up adding a bug or two, and resulting
code that was so hairy I could never look at it again :) We can call
that as an unmaintainable mess.
Still, what makes this even more annoying is that it might be possible
to support the requested scale by selecting yet another integration
time. Eg, imagine a situation where we have 2 channels. Both channels
support gains
1x, 2x, 8x, 16x, 32x. 4x is not supported.
Let's further say we have integration times 50mS 100mS, 200mS, 400mS -
causing "effective gains" 1x, 2x, 4x and, 8x
Now, say channel #1 is using gain 2x, channel #2 is using 8x.
Integration time is set to 400mS.
Assume the user would like to double the scale for channel #2. This
means the "total gain" should be halved. The HWGAIN can't be halved
because 8x => 4x is not supported, so driver decides to drop the
integration time from 400mS to 200mS instead. That'd do the trick.
Then the driver goes to check if the channel #1 can maintain the scale
with this integration time. Gain caused by integration time is now
halved so HWGAIN for channel #1 should be doubled to mitigate the
effect. Well, the new gain for channel #1 should now go from 2x => 4x -
which is not supported, and the driver returns error and rejects the change.
Still, the hardware could be set-up to use integration time 50mS
(dropping the gain for channels from 8x => 1x eg. 8 times smaller), and
channel #2 HWGAIN go from 8x => 2x (4 times smaller) thus doubling the
scale. The channel #1 wants to maintain scale, so HWGAIN for channel #1
should go 8 times greater, from 2x => 16x which is possible.
To make this even more annoying - the available_scales lists the 'halved
scale' for the channel #1 as supported because there is a way to achieve
it. So, the user can't really easily figure out what went wrong. Having
the read-only HARDWAREGAIN and knowing the gains sensor's channels
support would give a hint - but this is far from obvious. listing the
supported hardwaregains might make things a bit better - but using the
standard "available" entry in sysfs might make user to assume setting
the hardwaregain is possible.
We may invent a new entry to list the supported hardwaregains - and I
believe adding the logic to find supported gain-timing combinations is
then easier (and less error-prone) in user-land applications than it is
in driver - but I am wondering if it actually would be better just allow
setting both the hardwaregain and integration time individually for
those applications which may care... Well, I am probably just missing
some culprit supporting setting the hardwaregain causes.
I believe there are many use-cases where it would work if we just
allowed the channel #1 scale to change as a side-effect of changing
channel #1 scale. Still, I am very reluctant to do this as there could
be different type of data coming from these channels, and different
consumers for this data. Allowing another application to unintentionally
change the data for other app would in my opinion be very nasty.
The problem is not exactly urgent though and I am not working on an
application suffering from it. But it managed to interrupt my glymphatic
system while it was cleaning-up my brains last night. I will use it as
an excuse if you find any errors from this babbling :)
Yours
-- Matti
--
Matti Vaittinen
Linux kernel developer at ROHM Semiconductors
Oulu Finland
~~ When things go utterly wrong vim users can always type :help! ~~
